Emit CAS loop for min/max atomics.

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 0fe6ffc..b3fb27d 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -12867,12 +12867,23 @@ static bool fpModeMatchesGlobalFPAtomicMode(const AtomicRMWInst *RMW) {
   return DenormMode == DenormalMode::getIEEE();
 }
 
+// The amdgpu-unsafe-fp-atomics attribute enables generation of unsafe
+// floating point atomic instructions. May generate more efficient code,
+// but may not respect rounding and denormal modes, and may give incorrect
+// results for certain memory destinations.
+bool unsafeFPAtomicsDisabled(Function *F) {
+  return F->getFnAttribute("amdgpu-unsafe-fp-atomics").getValueAsString() !=
+         "true";
+}
+
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   unsigned AS = RMW->getPointerAddressSpace();
   if (AS == AMDGPUAS::PRIVATE_ADDRESS)
     return AtomicExpansionKind::NotAtomic;
 
+  auto SSID = RMW->getSyncScopeID();
+
   auto ReportUnsafeHWInst = [&](TargetLowering::AtomicExpansionKind Kind) {
     OptimizationRemarkEmitter ORE(RMW->getFunction());
     LLVMContext &Ctx = RMW->getFunction()->getContext();
@@ -12891,6 +12902,10 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
     return Kind;
   };
 
+  bool HasSystemScope =
+      SSID == SyncScope::System ||
+      SSID == RMW->getContext().getOrInsertSyncScopeID("one-as");
+
   switch (RMW->getOperation()) {
   case AtomicRMWInst::FAdd: {
     Type *Ty = RMW->getType();
@@ -12901,21 +12916,13 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
     if (!Ty->isFloatTy() && (!Subtarget->hasGFX90AInsts() || !Ty->isDoubleTy()))
       return AtomicExpansionKind::CmpXChg;
 
-    if ((AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) &&
+    if (AMDGPU::isFlatGlobalAddrSpace(AS) &&
         Subtarget->hasAtomicFaddNoRtnInsts()) {
-      // The amdgpu-unsafe-fp-atomics attribute enables generation of unsafe
-      // floating point atomic instructions. May generate more efficient code,
-      // but may not respect rounding and denormal modes, and may give incorrect
-      // results for certain memory destinations.
-      if (RMW->getFunction()
-              ->getFnAttribute("amdgpu-unsafe-fp-atomics")
-              .getValueAsString() != "true")
+      if (unsafeFPAtomicsDisabled(RMW->getFunction()))
         return AtomicExpansionKind::CmpXChg;
 
       // Always expand system scope fp atomics.
-      auto SSID = RMW->getSyncScopeID();
-      if (SSID == SyncScope::System ||
-          SSID == RMW->getContext().getOrInsertSyncScopeID("one-as"))
+      if (HasSystemScope)
         return AtomicExpansionKind::CmpXChg;
 
       if (AS == AMDGPUAS::GLOBAL_ADDRESS && Ty->isFloatTy()) {
@@ -12971,6 +12978,23 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
 
     return AtomicExpansionKind::CmpXChg;
   }
+  case AtomicRMWInst::FMin:
+  case AtomicRMWInst::FMax:
+  case AtomicRMWInst::Min:
+  case AtomicRMWInst::Max:
+  case AtomicRMWInst::UMin:
+  case AtomicRMWInst::UMax: {
+    if (AMDGPU::isFlatGlobalAddrSpace(AS)) {
+      if (RMW->getType()->isFloatTy() &&
+          unsafeFPAtomicsDisabled(RMW->getFunction()))
+        return AtomicExpansionKind::CmpXChg;
+
+      // Always expand system scope min/max atomics.
+      if (HasSystemScope)
+        return AtomicExpansionKind::CmpXChg;
+    }
+    break;
+  }
   default:
     break;
   }

diff --git a/llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll b/llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll
index a427bc6..ee99093 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll
+++ b/llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll
@@ -151,8 +151,8 @@ define protected amdgpu_kernel void @nand(i32 addrspace(1)* %p, %S addrspace(1)*
   ret void
 }
 
-define protected amdgpu_kernel void @max(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
-; CHECK-LABEL: max:
+define protected amdgpu_kernel void @max_workgroup(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: max_workgroup:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 0
@@ -166,6 +166,41 @@ define protected amdgpu_kernel void @max(i32 addrspace(1)* %p, %S addrspace(1)*
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
+  %n32 = atomicrmw max i32 addrspace(1)* %p, i32 1 syncscope("workgroup") monotonic
+  %n64 = zext i32 %n32 to i64
+  %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
+  store float 1.0, float addrspace(1)* %p1
+  ret void
+}
+
+define protected amdgpu_kernel void @max(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: max:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB7_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_max_i32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB7_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
+; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
+; CHECK-NEXT:    global_store_dword v[0:1], v2, off
+; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw max i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
@@ -173,8 +208,8 @@ define protected amdgpu_kernel void @max(i32 addrspace(1)* %p, %S addrspace(1)*
   ret void
 }
 
-define protected amdgpu_kernel void @min(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
-; CHECK-LABEL: min:
+define protected amdgpu_kernel void @min_workgroup(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: min_workgroup:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 0
@@ -188,6 +223,41 @@ define protected amdgpu_kernel void @min(i32 addrspace(1)* %p, %S addrspace(1)*
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
+  %n32 = atomicrmw min i32 addrspace(1)* %p, i32 1 syncscope("workgroup") monotonic
+  %n64 = zext i32 %n32 to i64
+  %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
+  store float 1.0, float addrspace(1)* %p1
+  ret void
+}
+
+define protected amdgpu_kernel void @min(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: min:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB9_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_min_i32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB9_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
+; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
+; CHECK-NEXT:    global_store_dword v[0:1], v2, off
+; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw min i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
@@ -195,8 +265,8 @@ define protected amdgpu_kernel void @min(i32 addrspace(1)* %p, %S addrspace(1)*
   ret void
 }
 
-define protected amdgpu_kernel void @umax(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
-; CHECK-LABEL: umax:
+define protected amdgpu_kernel void @umax_workgroup(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: umax_workgroup:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 0
@@ -210,6 +280,41 @@ define protected amdgpu_kernel void @umax(i32 addrspace(1)* %p, %S addrspace(1)*
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
+  %n32 = atomicrmw umax i32 addrspace(1)* %p, i32 1 syncscope("workgroup") monotonic
+  %n64 = zext i32 %n32 to i64
+  %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
+  store float 1.0, float addrspace(1)* %p1
+  ret void
+}
+
+define protected amdgpu_kernel void @umax(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: umax:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB11_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_max_u32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB11_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
+; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
+; CHECK-NEXT:    global_store_dword v[0:1], v2, off
+; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw umax i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
@@ -217,8 +322,8 @@ define protected amdgpu_kernel void @umax(i32 addrspace(1)* %p, %S addrspace(1)*
   ret void
 }
 
-define protected amdgpu_kernel void @umin(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
-; CHECK-LABEL: umin:
+define protected amdgpu_kernel void @umin_workgroup(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: umin_workgroup:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 0
@@ -232,6 +337,41 @@ define protected amdgpu_kernel void @umin(i32 addrspace(1)* %p, %S addrspace(1)*
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
+  %n32 = atomicrmw umin i32 addrspace(1)* %p, i32 1 syncscope("workgroup") monotonic
+  %n64 = zext i32 %n32 to i64
+  %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
+  store float 1.0, float addrspace(1)* %p1
+  ret void
+}
+
+define protected amdgpu_kernel void @umin(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
+; CHECK-LABEL: umin:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB13_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_min_u32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB13_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
+; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
+; CHECK-NEXT:    global_store_dword v[0:1], v2, off
+; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw umin i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
@@ -337,7 +477,7 @@ define protected amdgpu_kernel void @fadd(float addrspace(1)* %p, %S addrspace(1
 ; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_mov_b32_e32 v0, s6
-; CHECK-NEXT:  .LBB14_1: ; %atomicrmw.start
+; CHECK-NEXT:  .LBB18_1: ; %atomicrmw.start
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    v_mov_b32_e32 v3, v0
 ; CHECK-NEXT:    v_add_f32_e32 v2, 1.0, v3
@@ -346,7 +486,7 @@ define protected amdgpu_kernel void @fadd(float addrspace(1)* %p, %S addrspace(1
 ; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
 ; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
 ; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_cbranch_execnz .LBB14_1
+; CHECK-NEXT:    s_cbranch_execnz .LBB18_1
 ; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; CHECK-NEXT:    v_cvt_u32_f32_e32 v2, v0
@@ -374,7 +514,7 @@ define protected amdgpu_kernel void @fsub(float addrspace(1)* %p, %S addrspace(1
 ; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_mov_b32_e32 v0, s6
-; CHECK-NEXT:  .LBB15_1: ; %atomicrmw.start
+; CHECK-NEXT:  .LBB19_1: ; %atomicrmw.start
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    v_mov_b32_e32 v3, v0
 ; CHECK-NEXT:    v_add_f32_e32 v2, -1.0, v3
@@ -383,7 +523,7 @@ define protected amdgpu_kernel void @fsub(float addrspace(1)* %p, %S addrspace(1
 ; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
 ; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
 ; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_cbranch_execnz .LBB15_1
+; CHECK-NEXT:    s_cbranch_execnz .LBB19_1
 ; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; CHECK-NEXT:    v_cvt_u32_f32_e32 v2, v0

diff --git a/llvm/test/CodeGen/AMDGPU/flat_atomics.ll b/llvm/test/CodeGen/AMDGPU/flat_atomics.ll
index b2b6532..b6337a3 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/flat_atomics.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat_atomics.ll
@@ -1587,10 +1587,9 @@ define amdgpu_kernel void @atomic_max_i32_offset(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s4
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_offset:
@@ -1603,10 +1602,9 @@ define amdgpu_kernel void @atomic_max_i32_offset(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s4
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_offset:
@@ -1617,14 +1615,13 @@ define amdgpu_kernel void @atomic_max_i32_offset(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1639,12 +1636,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1658,12 +1655,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -1675,17 +1672,17 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -1705,10 +1702,9 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_addr64_offset:
@@ -1725,10 +1721,9 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_addr64_offset:
@@ -1743,15 +1738,14 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1770,12 +1764,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1793,12 +1787,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -1814,18 +1808,18 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -1839,10 +1833,9 @@ define amdgpu_kernel void @atomic_max_i32(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32:
@@ -1853,10 +1846,9 @@ define amdgpu_kernel void @atomic_max_i32(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32:
@@ -1867,13 +1859,12 @@ define amdgpu_kernel void @atomic_max_i32(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1886,12 +1877,12 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1903,12 +1894,12 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -1920,16 +1911,16 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -1947,10 +1938,9 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_addr64:
@@ -1965,10 +1955,9 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_addr64:
@@ -1983,14 +1972,13 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2007,12 +1995,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2028,12 +2016,12 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2049,17 +2037,17 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile max i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2075,10 +2063,9 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s4
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_offset:
@@ -2091,10 +2078,9 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s4
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_offset:
@@ -2105,14 +2091,13 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2127,12 +2112,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2146,12 +2131,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2163,17 +2148,17 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2193,10 +2178,9 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_addr64_offset:
@@ -2213,10 +2197,9 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_addr64_offset:
@@ -2231,15 +2214,14 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2258,12 +2240,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2281,12 +2263,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2302,18 +2284,18 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2327,10 +2309,9 @@ define amdgpu_kernel void @atomic_umax_i32(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32:
@@ -2341,10 +2322,9 @@ define amdgpu_kernel void @atomic_umax_i32(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32:
@@ -2355,13 +2335,12 @@ define amdgpu_kernel void @atomic_umax_i32(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2374,12 +2353,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2391,12 +2370,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2408,16 +2387,16 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2435,10 +2414,9 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_addr64:
@@ -2453,10 +2431,9 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_addr64:
@@ -2471,14 +2448,13 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2495,12 +2471,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2516,12 +2492,12 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2537,17 +2513,17 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2563,10 +2539,9 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s4
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_offset:
@@ -2579,10 +2554,9 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s4
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_offset:
@@ -2593,14 +2567,13 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2615,12 +2588,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2634,12 +2607,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2651,17 +2624,17 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2681,10 +2654,9 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_addr64_offset:
@@ -2701,10 +2673,9 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_addr64_offset:
@@ -2719,15 +2690,14 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2746,12 +2716,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2769,12 +2739,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2790,18 +2760,18 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32* %out, i32* %out
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2815,10 +2785,9 @@ define amdgpu_kernel void @atomic_min_i32(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32:
@@ -2829,10 +2798,9 @@ define amdgpu_kernel void @atomic_min_i32(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32:
@@ -2843,13 +2811,12 @@ define amdgpu_kernel void @atomic_min_i32(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2862,12 +2829,12 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2879,12 +2846,12 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -2896,16 +2863,16 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -2923,10 +2890,9 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_addr64:
@@ -2941,10 +2907,9 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_addr64:
@@ -2959,14 +2924,13 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32* %out, i32 %in, i64 %index)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2983,12 +2947,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -3004,12 +2968,12 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -3025,17 +2989,17 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile min i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -3051,10 +3015,9 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s4
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_offset:
@@ -3067,10 +3030,9 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s4
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_offset:
@@ -3081,14 +3043,13 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3103,12 +3064,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -3122,12 +3083,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -3139,17 +3100,17 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
-  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -3169,10 +3130,9 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_addr64_offset:
@@ -3189,10 +3149,9 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_addr64_offset:
@@ -3207,15 +3166,14 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32* %out, i32 %in, i64
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3234,12 +3192,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -3257,12 +3215,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -3278,18 +3236,18 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32* %out, i32* %ou
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
-  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -3303,10 +3261,9 @@ define amdgpu_kernel void @atomic_umin_i32(i32* %out, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32:
@@ -3317,10 +3274,9 @@ define amdgpu_kernel void @atomic_umin_i32(i32* %out, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32:
@@ -3331,13 +3287,12 @@ define amdgpu_kernel void @atomic_umin_i32(i32* %out, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s4
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3350,12 +3305,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -3367,12 +3322,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -3384,16 +3339,16 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }
@@ -3411,10 +3366,9 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_addr64:
@@ -3429,10 +3383,9 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_addr64:
@@ -3447,14 +3400,13 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32* %out, i32 %in, i64 %index
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s6
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3471,12 +3423,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s8
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dword v[0:1], v2
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -3492,12 +3444,12 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s8
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dword v[0:1], v2
 ; GCN2-NEXT:    s_endpgm
 ;
@@ -3513,17 +3465,17 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32* %out, i32* %out2, i32
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v2, s8
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
-; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s7
+; GCN3-NEXT:    s_waitcnt vmcnt(0)
 ; GCN3-NEXT:    flat_store_dword v[0:1], v2
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
-  %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32* %out2
   ret void
 }

diff --git a/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
index 1c8f858..dc7b860 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
@@ -1057,10 +1057,9 @@ define amdgpu_kernel void @atomic_max_i64_offset(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_offset:
@@ -1073,14 +1072,13 @@ define amdgpu_kernel void @atomic_max_i64_offset(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1096,12 +1094,12 @@ define amdgpu_kernel void @atomic_max_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1116,17 +1114,17 @@ define amdgpu_kernel void @atomic_max_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1146,10 +1144,9 @@ define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_addr64_offset:
@@ -1166,15 +1163,14 @@ define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1192,12 +1188,12 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64* %out, i64* %out
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1214,18 +1210,18 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64* %out, i64* %out
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1239,10 +1235,9 @@ define amdgpu_kernel void @atomic_max_i64(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64:
@@ -1253,13 +1248,12 @@ define amdgpu_kernel void @atomic_max_i64(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1273,12 +1267,12 @@ define amdgpu_kernel void @atomic_max_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1291,16 +1285,16 @@ define amdgpu_kernel void @atomic_max_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1318,10 +1312,9 @@ define amdgpu_kernel void @atomic_max_i64_addr64(i64* %out, i64 %in, i64 %index)
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_addr64:
@@ -1336,14 +1329,13 @@ define amdgpu_kernel void @atomic_max_i64_addr64(i64* %out, i64 %in, i64 %index)
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1359,12 +1351,12 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1379,17 +1371,17 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1405,10 +1397,9 @@ define amdgpu_kernel void @atomic_umax_i64_offset(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_offset:
@@ -1421,14 +1412,13 @@ define amdgpu_kernel void @atomic_umax_i64_offset(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1444,12 +1434,12 @@ define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1464,17 +1454,17 @@ define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1494,10 +1484,9 @@ define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_addr64_offset:
@@ -1514,15 +1503,14 @@ define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1540,12 +1528,12 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64* %out, i64* %ou
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1562,18 +1550,18 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64* %out, i64* %ou
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1587,10 +1575,9 @@ define amdgpu_kernel void @atomic_umax_i64(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64:
@@ -1601,13 +1588,12 @@ define amdgpu_kernel void @atomic_umax_i64(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1621,12 +1607,12 @@ define amdgpu_kernel void @atomic_umax_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1639,16 +1625,16 @@ define amdgpu_kernel void @atomic_umax_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1666,10 +1652,9 @@ define amdgpu_kernel void @atomic_umax_i64_addr64(i64* %out, i64 %in, i64 %index
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_addr64:
@@ -1684,14 +1669,13 @@ define amdgpu_kernel void @atomic_umax_i64_addr64(i64* %out, i64 %in, i64 %index
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1707,12 +1691,12 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1727,17 +1711,17 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1753,10 +1737,9 @@ define amdgpu_kernel void @atomic_min_i64_offset(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_offset:
@@ -1769,14 +1752,13 @@ define amdgpu_kernel void @atomic_min_i64_offset(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1792,12 +1774,12 @@ define amdgpu_kernel void @atomic_min_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1812,17 +1794,17 @@ define amdgpu_kernel void @atomic_min_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1842,10 +1824,9 @@ define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_addr64_offset:
@@ -1862,15 +1843,14 @@ define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1888,12 +1868,12 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64* %out, i64* %out
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1910,18 +1890,18 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64* %out, i64* %out
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -1935,10 +1915,9 @@ define amdgpu_kernel void @atomic_min_i64(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64:
@@ -1949,13 +1928,12 @@ define amdgpu_kernel void @atomic_min_i64(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1969,12 +1947,12 @@ define amdgpu_kernel void @atomic_min_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -1987,16 +1965,16 @@ define amdgpu_kernel void @atomic_min_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -2014,10 +1992,9 @@ define amdgpu_kernel void @atomic_min_i64_addr64(i64* %out, i64 %in, i64 %index)
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_addr64:
@@ -2032,14 +2009,13 @@ define amdgpu_kernel void @atomic_min_i64_addr64(i64* %out, i64 %in, i64 %index)
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2055,12 +2031,12 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2075,17 +2051,17 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -2101,10 +2077,9 @@ define amdgpu_kernel void @atomic_umin_i64_offset(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_offset:
@@ -2117,14 +2092,13 @@ define amdgpu_kernel void @atomic_umin_i64_offset(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2140,12 +2114,12 @@ define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2160,17 +2134,17 @@ define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
-  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -2190,10 +2164,9 @@ define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_addr64_offset:
@@ -2210,15 +2183,14 @@ define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64* %out, i64 %in, i64
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2236,12 +2208,12 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64* %out, i64* %ou
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2258,18 +2230,18 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64* %out, i64* %ou
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -2283,10 +2255,9 @@ define amdgpu_kernel void @atomic_umin_i64(i64* %out, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64:
@@ -2297,13 +2268,12 @@ define amdgpu_kernel void @atomic_umin_i64(i64* %out, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2317,12 +2287,12 @@ define amdgpu_kernel void @atomic_umin_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2335,16 +2305,16 @@ define amdgpu_kernel void @atomic_umin_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[0:1], v[2:3] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
@@ -2362,10 +2332,9 @@ define amdgpu_kernel void @atomic_umin_i64_addr64(i64* %out, i64 %in, i64 %index
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_addr64:
@@ -2380,14 +2349,13 @@ define amdgpu_kernel void @atomic_umin_i64_addr64(i64* %out, i64 %in, i64 %index
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2403,12 +2371,12 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
@@ -2423,17 +2391,17 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64* %out, i64* %out2, i64
 ; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
-  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }

diff --git a/llvm/test/CodeGen/AMDGPU/flat_atomics_i64_min_max_system.ll b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64_min_max_system.ll
new file mode 100644 (file)
index 0000000..c823c56
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64_min_max_system.ll
@@ -0,0 +1,2515 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN1 %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN2 %s
+
+define amdgpu_kernel void @atomic_max_i64_offset(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_max_i64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB0_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB0_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_max_i64_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB1_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB1_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i64_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB2_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB2_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i64_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB3_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB3_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_max_i64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB4_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB4_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_max_i64_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB5_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB5_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_addr64(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i64_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB6_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB6_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i64_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB7_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i64_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB7_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_offset(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_umax_i64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB8_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB8_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_umax_i64_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB9_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB9_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i64_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB10_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB10_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i64_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB11_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB11_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_umax_i64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB12_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB12_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_umax_i64_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB13_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB13_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_addr64(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i64_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB14_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB14_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i64_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB15_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i64_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB15_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_offset(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_min_i64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB16_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB16_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_min_i64_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB17_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB17_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i64_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB18_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB18_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i64_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB19_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB19_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_min_i64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB20_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB20_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_min_i64_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB21_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB21_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_addr64(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i64_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB22_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB22_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i64_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB23_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i64_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB23_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_offset(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_umin_i64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_umin_i64_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64* %out, i64 4
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i64_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i64_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    s_add_u32 s0, s0, 32
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    s_add_u32 s0, s0, 32
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %gep = getelementptr i64, i64* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64(i64* %out, i64 %in) {
+; GCN1-LABEL: atomic_umin_i64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret(i64* %out, i64* %out2, i64 %in) {
+; GCN1-LABEL: atomic_umin_i64_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_addr64(i64* %out, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i64_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i64_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i64_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64* %out, i64 %index
+  %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64* %out2
+  ret void
+}

diff --git a/llvm/test/CodeGen/AMDGPU/flat_atomics_min_max_system.ll b/llvm/test/CodeGen/AMDGPU/flat_atomics_min_max_system.ll
new file mode 100644 (file)
index 0000000..3a8b4ea
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/flat_atomics_min_max_system.ll
@@ -0,0 +1,3252 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN1 %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN2 %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN3 %s
+
+define amdgpu_kernel void @atomic_max_i32_offset(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_max_i32_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s2, 16
+; GCN1-NEXT:    s_addc_u32 s1, s3, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB0_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s2, 16
+; GCN2-NEXT:    s_addc_u32 s1, s3, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB0_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB0_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_max_i32_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s4, 16
+; GCN1-NEXT:    s_addc_u32 s1, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB1_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s4, 16
+; GCN2-NEXT:    s_addc_u32 s1, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB1_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_ret_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB1_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i32_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB2_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB2_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB2_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i32_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB3_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB3_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_ret_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB3_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_max_i32:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB4_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB4_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB4_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_max_i32_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_max_i32_e32 v0, s2, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB5_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_max_i32_e32 v0, s2, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB5_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_ret:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB5_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_addr64(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i32_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB6_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB6_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB6_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_max_i32_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB7_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_max_i32_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB7_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_max_i32_ret_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB7_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_offset(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_umax_i32_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s2, 16
+; GCN1-NEXT:    s_addc_u32 s1, s3, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB8_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s2, 16
+; GCN2-NEXT:    s_addc_u32 s1, s3, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB8_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB8_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_umax_i32_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s4, 16
+; GCN1-NEXT:    s_addc_u32 s1, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB9_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s4, 16
+; GCN2-NEXT:    s_addc_u32 s1, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB9_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_ret_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB9_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i32_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB10_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB10_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB10_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i32_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB11_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB11_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_ret_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB11_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_umax_i32:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB12_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB12_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB12_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_umax_i32_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_max_u32_e32 v0, s2, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB13_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_max_u32_e32 v0, s2, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB13_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_ret:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB13_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_addr64(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i32_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB14_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB14_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB14_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umax_i32_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB15_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umax_i32_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB15_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umax_i32_ret_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB15_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_offset(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_min_i32_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s2, 16
+; GCN1-NEXT:    s_addc_u32 s1, s3, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB16_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s2, 16
+; GCN2-NEXT:    s_addc_u32 s1, s3, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB16_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB16_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_min_i32_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s4, 16
+; GCN1-NEXT:    s_addc_u32 s1, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB17_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s4, 16
+; GCN2-NEXT:    s_addc_u32 s1, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB17_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_ret_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB17_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i32_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB18_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB18_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB18_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i32_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB19_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB19_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_ret_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB19_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_min_i32:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB20_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB20_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB20_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_min_i32_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_min_i32_e32 v0, s2, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB21_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_min_i32_e32 v0, s2, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB21_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_ret:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB21_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_addr64(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i32_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB22_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB22_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB22_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_min_i32_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB23_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_min_i32_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB23_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_min_i32_ret_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB23_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_offset(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_umin_i32_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s2, 16
+; GCN1-NEXT:    s_addc_u32 s1, s3, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s2, 16
+; GCN2-NEXT:    s_addc_u32 s1, s3, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_umin_i32_ret_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_add_u32 s0, s4, 16
+; GCN1-NEXT:    s_addc_u32 s1, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_ret_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_add_u32 s0, s4, 16
+; GCN2-NEXT:    s_addc_u32 s1, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_ret_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i32_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i32_ret_addr64_offset:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s0, 16
+; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_ret_addr64_offset:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s0, 16
+; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_ret_addr64_offset:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %gep = getelementptr i32, i32* %ptr, i32 4
+  %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32(i32* %out, i32 %in) {
+; GCN1-LABEL: atomic_umin_i32:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s2
+; GCN3-NEXT:    v_mov_b32_e32 v1, s3
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret(i32* %out, i32* %out2, i32 %in) {
+; GCN1-LABEL: atomic_umin_i32_ret:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v0, s4
+; GCN1-NEXT:    v_mov_b32_e32 v1, s5
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_min_u32_e32 v0, s2, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_ret:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v0, s4
+; GCN2-NEXT:    v_mov_b32_e32 v1, s5
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_min_u32_e32 v0, s2, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_ret:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v0, s4
+; GCN3-NEXT:    v_mov_b32_e32 v1, s5
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_addr64(i32* %out, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i32_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
+; GCN1-LABEL: atomic_umin_i32_ret_addr64:
+; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
+; GCN1-NEXT:    s_endpgm
+;
+; GCN2-LABEL: atomic_umin_i32_ret_addr64:
+; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
+; GCN2-NEXT:    s_endpgm
+;
+; GCN3-LABEL: atomic_umin_i32_ret_addr64:
+; GCN3:       ; %bb.0: ; %entry
+; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GCN3-NEXT:    s_add_u32 s0, s4, s0
+; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    v_mov_b32_e32 v0, s0
+; GCN3-NEXT:    v_mov_b32_e32 v1, s1
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
+; GCN3-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32* %out, i64 %index
+  %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
+  store i32 %val, i32* %out2
+  ret void
+}

diff --git a/llvm/test/CodeGen/AMDGPU/global-saddr-atomics-min-max-system.ll b/llvm/test/CodeGen/AMDGPU/global-saddr-atomics-min-max-system.ll
new file mode 100644 (file)
index 0000000..040c8df
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-saddr-atomics-min-max-system.ll
@@ -0,0 +1,2928 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX10 %s
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX11 %s
+
+; Test using saddr addressing mode of global_* flat atomic instructions.
+
+; --------------------------------------------------------------------------------
+; atomicrmw max
+; --------------------------------------------------------------------------------
+
+define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_max_saddr_i32_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_max_saddr_i32_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_max_saddr_i32_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_max_saddr_i32_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_max_saddr_i32_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_max_saddr_i32_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps void @global_max_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_max_saddr_i32_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_max_saddr_i32_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_max_saddr_i32_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %unused = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_max_saddr_i32_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB3_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_max_saddr_i32_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB3_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_max_saddr_i32_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB3_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %unused = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_max_saddr_i64_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_max_saddr_i64_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_max_saddr_i64_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_max_saddr_i64_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB5_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_max_saddr_i64_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB5_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_max_saddr_i64_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB5_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps void @global_max_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_max_saddr_i64_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB6_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_max_saddr_i64_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB6_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_max_saddr_i64_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB6_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %unused = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_max_saddr_i64_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB7_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_max_saddr_i64_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB7_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_max_saddr_i64_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB7_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %unused = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  ret void
+}
+
+; --------------------------------------------------------------------------------
+; atomicrmw min
+; --------------------------------------------------------------------------------
+
+define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_min_saddr_i32_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB8_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_min_saddr_i32_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB8_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_min_saddr_i32_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB8_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_min_saddr_i32_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB9_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_min_saddr_i32_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB9_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_min_saddr_i32_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB9_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps void @global_min_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_min_saddr_i32_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB10_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_min_saddr_i32_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB10_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_min_saddr_i32_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB10_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %unused = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_min_saddr_i32_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB11_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_min_saddr_i32_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB11_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_min_saddr_i32_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB11_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %unused = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_min_saddr_i64_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB12_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_min_saddr_i64_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB12_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_min_saddr_i64_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB12_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_min_saddr_i64_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB13_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_min_saddr_i64_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB13_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_min_saddr_i64_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB13_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps void @global_min_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_min_saddr_i64_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB14_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_min_saddr_i64_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB14_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_min_saddr_i64_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB14_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %unused = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_min_saddr_i64_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB15_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_min_saddr_i64_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB15_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_min_saddr_i64_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB15_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %unused = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  ret void
+}
+
+; --------------------------------------------------------------------------------
+; atomicrmw umax
+; --------------------------------------------------------------------------------
+
+define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umax_saddr_i32_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB16_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umax_saddr_i32_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB16_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umax_saddr_i32_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB16_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umax_saddr_i32_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB17_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umax_saddr_i32_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB17_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umax_saddr_i32_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB17_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps void @global_umax_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umax_saddr_i32_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB18_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umax_saddr_i32_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB18_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umax_saddr_i32_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB18_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umax_saddr_i32_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB19_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umax_saddr_i32_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB19_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umax_saddr_i32_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB19_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umax_saddr_i64_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB20_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umax_saddr_i64_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB20_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umax_saddr_i64_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB20_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umax_saddr_i64_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB21_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umax_saddr_i64_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB21_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umax_saddr_i64_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB21_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps void @global_umax_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umax_saddr_i64_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB22_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umax_saddr_i64_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB22_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umax_saddr_i64_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB22_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umax_saddr_i64_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB23_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umax_saddr_i64_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB23_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umax_saddr_i64_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB23_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  ret void
+}
+
+; --------------------------------------------------------------------------------
+; atomicrmw umin
+; --------------------------------------------------------------------------------
+
+define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umin_saddr_i32_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umin_saddr_i32_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umin_saddr_i32_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umin_saddr_i32_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umin_saddr_i32_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umin_saddr_i32_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %cast.rtn = bitcast i32 %rtn to float
+  ret float %cast.rtn
+}
+
+define amdgpu_ps void @global_umin_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umin_saddr_i32_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umin_saddr_i32_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umin_saddr_i32_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
+  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
+; GFX9-LABEL: global_umin_saddr_i32_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umin_saddr_i32_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umin_saddr_i32_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
+  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umin_saddr_i64_rtn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umin_saddr_i64_rtn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umin_saddr_i64_rtn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umin_saddr_i64_rtn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX10-LABEL: global_umin_saddr_i64_rtn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: global_umin_saddr_i64_rtn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    ; return to shader part epilog
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %cast.rtn = bitcast i64 %rtn to <2 x float>
+  ret <2 x float> %cast.rtn
+}
+
+define amdgpu_ps void @global_umin_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umin_saddr_i64_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umin_saddr_i64_nortn:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umin_saddr_i64_nortn:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
+  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  ret void
+}
+
+define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
+; GFX9-LABEL: global_umin_saddr_i64_nortn_neg128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: global_umin_saddr_i64_nortn_neg128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    buffer_gl0_inv
+; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: global_umin_saddr_i64_nortn_neg128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    buffer_gl0_inv
+; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_endpgm
+  %zext.offset = zext i32 %voffset to i64
+  %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
+  %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
+  %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
+  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  ret void
+}
+
+attributes #0 = { argmemonly nounwind willreturn }

diff --git a/llvm/test/CodeGen/AMDGPU/global-saddr-atomics.ll b/llvm/test/CodeGen/AMDGPU/global-saddr-atomics.ll
index 1e51258..1a4bde7 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/global-saddr-atomics.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-saddr-atomics.ll
@@ -2069,10 +2069,9 @@ define amdgpu_ps void @global_xor_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i32_rtn:
@@ -2082,7 +2081,6 @@ define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i32_rtn:
@@ -2092,12 +2090,11 @@ define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2105,10 +2102,9 @@ define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i32_rtn_neg128:
@@ -2118,7 +2114,6 @@ define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 ; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i32_rtn_neg128:
@@ -2128,13 +2123,12 @@ define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 ; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %rtn = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2142,10 +2136,8 @@ define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 define amdgpu_ps void @global_max_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i32_nortn:
@@ -2155,7 +2147,6 @@ define amdgpu_ps void @global_max_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i32_nortn:
@@ -2165,22 +2156,19 @@ define amdgpu_ps void @global_max_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %unused = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %unused = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i32_nortn_neg128:
@@ -2190,7 +2178,6 @@ define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i32_nortn_neg128:
@@ -2200,23 +2187,21 @@ define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %unused = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %unused = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i64_rtn:
@@ -2226,7 +2211,6 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 ; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i64_rtn:
@@ -2236,12 +2220,11 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 ; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2249,10 +2232,9 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i64_rtn_neg128:
@@ -2262,7 +2244,6 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 ; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i64_rtn_neg128:
@@ -2272,13 +2253,12 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 ; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %rtn = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2286,10 +2266,8 @@ define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 define amdgpu_ps void @global_max_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i64_nortn:
@@ -2299,7 +2277,6 @@ define amdgpu_ps void @global_max_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i64_nortn:
@@ -2309,22 +2286,19 @@ define amdgpu_ps void @global_max_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %unused = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %unused = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i64_nortn_neg128:
@@ -2334,7 +2308,6 @@ define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i64_nortn_neg128:
@@ -2344,13 +2317,12 @@ define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %unused = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %unused = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2361,10 +2333,9 @@ define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i32_rtn:
@@ -2374,7 +2345,6 @@ define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i32_rtn:
@@ -2384,12 +2354,11 @@ define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2397,10 +2366,9 @@ define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i32_rtn_neg128:
@@ -2410,7 +2378,6 @@ define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 ; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i32_rtn_neg128:
@@ -2420,13 +2387,12 @@ define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 ; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %rtn = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2434,10 +2400,8 @@ define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %
 define amdgpu_ps void @global_min_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i32_nortn:
@@ -2447,7 +2411,6 @@ define amdgpu_ps void @global_min_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i32_nortn:
@@ -2457,22 +2420,19 @@ define amdgpu_ps void @global_min_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %unused = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %unused = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i32_nortn_neg128:
@@ -2482,7 +2442,6 @@ define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i32_nortn_neg128:
@@ -2492,23 +2451,21 @@ define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %unused = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %unused = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i64_rtn:
@@ -2518,7 +2475,6 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 ; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i64_rtn:
@@ -2528,12 +2484,11 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 ; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2541,10 +2496,9 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %s
 define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i64_rtn_neg128:
@@ -2554,7 +2508,6 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 ; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i64_rtn_neg128:
@@ -2564,13 +2517,12 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 ; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %rtn = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2578,10 +2530,8 @@ define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* i
 define amdgpu_ps void @global_min_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i64_nortn:
@@ -2591,7 +2541,6 @@ define amdgpu_ps void @global_min_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i64_nortn:
@@ -2601,22 +2550,19 @@ define amdgpu_ps void @global_min_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %unused = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %unused = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i64_nortn_neg128:
@@ -2626,7 +2572,6 @@ define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i64_nortn_neg128:
@@ -2636,13 +2581,12 @@ define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %unused = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %unused = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2653,10 +2597,9 @@ define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_rtn:
@@ -2666,7 +2609,6 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_rtn:
@@ -2676,12 +2618,11 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2689,10 +2630,9 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_rtn_neg128:
@@ -2702,7 +2642,6 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_rtn_neg128:
@@ -2712,13 +2651,12 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2726,10 +2664,8 @@ define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps void @global_umax_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_nortn:
@@ -2739,7 +2675,6 @@ define amdgpu_ps void @global_umax_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_nortn:
@@ -2749,22 +2684,19 @@ define amdgpu_ps void @global_umax_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_nortn_neg128:
@@ -2774,7 +2706,6 @@ define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_nortn_neg128:
@@ -2784,23 +2715,21 @@ define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %unused = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_rtn:
@@ -2810,7 +2739,6 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %
 ; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_rtn:
@@ -2820,12 +2748,11 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %
 ; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2833,10 +2760,9 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %
 define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_rtn_neg128:
@@ -2846,7 +2772,6 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)*
 ; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_rtn_neg128:
@@ -2856,13 +2781,12 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)*
 ; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -2870,10 +2794,8 @@ define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)*
 define amdgpu_ps void @global_umax_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_nortn:
@@ -2883,7 +2805,6 @@ define amdgpu_ps void @global_umax_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_nortn:
@@ -2893,22 +2814,19 @@ define amdgpu_ps void @global_umax_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_nortn_neg128:
@@ -2918,7 +2836,6 @@ define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_nortn_neg128:
@@ -2928,13 +2845,12 @@ define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %unused = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2945,10 +2861,9 @@ define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_rtn:
@@ -2958,7 +2873,6 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_rtn:
@@ -2968,12 +2882,11 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 ; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -2981,10 +2894,9 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase,
 define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_rtn_neg128:
@@ -2994,7 +2906,6 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_rtn_neg128:
@@ -3004,13 +2915,12 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
@@ -3018,10 +2928,8 @@ define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg
 define amdgpu_ps void @global_umin_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_nortn:
@@ -3031,7 +2939,6 @@ define amdgpu_ps void @global_umin_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_nortn:
@@ -3041,22 +2948,19 @@ define amdgpu_ps void @global_umin_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
-  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
+  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_nortn_neg128:
@@ -3066,7 +2970,6 @@ define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_nortn_neg128:
@@ -3076,23 +2979,21 @@ define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
-  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
+  %unused = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_rtn:
@@ -3102,7 +3003,6 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %
 ; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_rtn:
@@ -3112,12 +3012,11 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %
 ; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -3125,10 +3024,9 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %
 define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_rtn_neg128:
@@ -3138,7 +3036,6 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)*
 ; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_rtn_neg128:
@@ -3148,13 +3045,12 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)*
 ; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
@@ -3162,10 +3058,8 @@ define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)*
 define amdgpu_ps void @global_umin_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_nortn:
@@ -3175,7 +3069,6 @@ define amdgpu_ps void @global_umin_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_nortn:
@@ -3185,22 +3078,19 @@ define amdgpu_ps void @global_umin_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase
 ; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3]
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
-  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
+  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_nortn_neg128:
@@ -3210,7 +3100,6 @@ define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_gl0_inv
-; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_nortn_neg128:
@@ -3220,13 +3109,12 @@ define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg
 ; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3] offset:-128
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
-  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
+  %unused = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data syncscope("workgroup") seq_cst
   ret void
 }
 

diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics.ll b/llvm/test/CodeGen/AMDGPU/global_atomics.ll
index 04bbd1e..c388bfe 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/global_atomics.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics.ll
@@ -1588,43 +1588,86 @@ entry:
 define amdgpu_kernel void @atomic_max_i32_offset(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_max_i32_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB27_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB27_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:16
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
@@ -1645,10 +1688,9 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; SI-NEXT:    s_mov_b32 s6, s2
 ; SI-NEXT:    s_mov_b32 s7, s3
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -1664,10 +1706,9 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1678,15 +1719,14 @@ define amdgpu_kernel void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -1704,10 +1744,8 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out,
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_addr64_offset:
@@ -1724,10 +1762,8 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_addr64_offset:
@@ -1741,15 +1777,13 @@ define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1769,11 +1803,10 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -1791,14 +1824,13 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1813,16 +1845,15 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -1836,10 +1867,8 @@ define amdgpu_kernel void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32:
@@ -1850,10 +1879,8 @@ define amdgpu_kernel void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32:
@@ -1863,13 +1890,11 @@ define amdgpu_kernel void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1884,12 +1909,11 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -1903,12 +1927,11 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; VI-NEXT:    s_mov_b32 s0, s4
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1919,14 +1942,13 @@ define amdgpu_kernel void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -1944,10 +1966,8 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_addr64:
@@ -1962,10 +1982,8 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_addr64:
@@ -1979,14 +1997,12 @@ define amdgpu_kernel void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2006,11 +2022,10 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2026,14 +2041,13 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2048,15 +2062,14 @@ define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2070,10 +2083,8 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_offset:
@@ -2084,10 +2095,8 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_offset:
@@ -2097,14 +2106,12 @@ define amdgpu_kernel void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2121,10 +2128,9 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; SI-NEXT:    s_mov_b32 s6, s2
 ; SI-NEXT:    s_mov_b32 s7, s3
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2140,10 +2146,9 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2154,15 +2159,14 @@ define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2180,10 +2184,8 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out,
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_addr64_offset:
@@ -2200,10 +2202,8 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_addr64_offset:
@@ -2217,15 +2217,13 @@ define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2245,11 +2243,10 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2267,14 +2264,13 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2289,16 +2285,15 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2312,10 +2307,8 @@ define amdgpu_kernel void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32:
@@ -2326,10 +2319,8 @@ define amdgpu_kernel void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32:
@@ -2339,13 +2330,11 @@ define amdgpu_kernel void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2360,12 +2349,11 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2379,12 +2367,11 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; VI-NEXT:    s_mov_b32 s0, s4
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2395,14 +2382,13 @@ define amdgpu_kernel void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2420,10 +2406,8 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_addr64:
@@ -2438,10 +2422,8 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_addr64:
@@ -2455,14 +2437,12 @@ define amdgpu_kernel void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2482,11 +2462,10 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2502,14 +2481,13 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2524,15 +2502,14 @@ define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2546,10 +2523,8 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32 addrspace(1)* %out, i32 %in
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_offset:
@@ -2560,10 +2535,8 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32 addrspace(1)* %out, i32 %in
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_offset:
@@ -2573,14 +2546,12 @@ define amdgpu_kernel void @atomic_min_i32_offset(i32 addrspace(1)* %out, i32 %in
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2597,10 +2568,9 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; SI-NEXT:    s_mov_b32 s6, s2
 ; SI-NEXT:    s_mov_b32 s7, s3
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2616,10 +2586,9 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2630,15 +2599,14 @@ define amdgpu_kernel void @atomic_min_i32_ret_offset(i32 addrspace(1)* %out, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2656,10 +2624,8 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32 addrspace(1)* %out,
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_addr64_offset:
@@ -2676,10 +2642,8 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_addr64_offset:
@@ -2693,15 +2657,13 @@ define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2721,11 +2683,10 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2743,14 +2704,13 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2765,16 +2725,15 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32 addrspace(1)* %o
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2788,10 +2747,8 @@ define amdgpu_kernel void @atomic_min_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32:
@@ -2802,10 +2759,8 @@ define amdgpu_kernel void @atomic_min_i32(i32 addrspace(1)* %out, i32 %in) {
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32:
@@ -2815,13 +2770,11 @@ define amdgpu_kernel void @atomic_min_i32(i32 addrspace(1)* %out, i32 %in) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2836,12 +2789,11 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2855,12 +2807,11 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; VI-NEXT:    s_mov_b32 s0, s4
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2871,14 +2822,13 @@ define amdgpu_kernel void @atomic_min_i32_ret(i32 addrspace(1)* %out, i32 addrsp
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -2896,10 +2846,8 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_addr64:
@@ -2914,10 +2862,8 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_addr64:
@@ -2931,14 +2877,12 @@ define amdgpu_kernel void @atomic_min_i32_addr64(i32 addrspace(1)* %out, i32 %in
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2958,11 +2902,10 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -2978,14 +2921,13 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3000,15 +2942,14 @@ define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32 addrspace(1)* %out, i32
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -3022,10 +2963,8 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_offset:
@@ -3036,10 +2975,8 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_offset:
@@ -3049,14 +2986,12 @@ define amdgpu_kernel void @atomic_umin_i32_offset(i32 addrspace(1)* %out, i32 %i
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3073,10 +3008,9 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; SI-NEXT:    s_mov_b32 s6, s2
 ; SI-NEXT:    s_mov_b32 s7, s3
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -3092,10 +3026,9 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3106,15 +3039,14 @@ define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32 addrspace(1)* %out, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
-  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -3132,10 +3064,8 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32 addrspace(1)* %out,
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_addr64_offset:
@@ -3152,10 +3082,8 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_addr64_offset:
@@ -3169,15 +3097,13 @@ define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3197,11 +3123,10 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -3219,14 +3144,13 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3241,16 +3165,15 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32 addrspace(1)* %
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
-  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -3264,10 +3187,8 @@ define amdgpu_kernel void @atomic_umin_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32:
@@ -3278,10 +3199,8 @@ define amdgpu_kernel void @atomic_umin_i32(i32 addrspace(1)* %out, i32 %in) {
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32:
@@ -3291,13 +3210,11 @@ define amdgpu_kernel void @atomic_umin_i32(i32 addrspace(1)* %out, i32 %in) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3312,12 +3229,11 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -3331,12 +3247,11 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; VI-NEXT:    s_mov_b32 s0, s4
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3347,14 +3262,13 @@ define amdgpu_kernel void @atomic_umin_i32_ret(i32 addrspace(1)* %out, i32 addrs
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
@@ -3372,10 +3286,8 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; SI-NEXT:    v_mov_b32_e32 v2, s6
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_addr64:
@@ -3390,10 +3302,8 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_addr64:
@@ -3407,14 +3317,12 @@ define amdgpu_kernel void @atomic_umin_i32_addr64(i32 addrspace(1)* %out, i32 %i
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s6
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3434,11 +3342,10 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; SI-NEXT:    v_mov_b32_e32 v2, s2
 ; SI-NEXT:    v_mov_b32_e32 v0, s8
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
 ; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -3454,14 +3361,13 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s8
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3476,15 +3382,14 @@ define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32 addrspace(1)* %out, i3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s8
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
-  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
+  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in syncscope("workgroup") seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }

diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_i64.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_i64.ll
index b8e4b4a..3d019d2 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/global_atomics_i64.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_i64.ll
@@ -1452,10 +1452,8 @@ define amdgpu_kernel void @atomic_max_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_offset:
@@ -1466,10 +1464,8 @@ define amdgpu_kernel void @atomic_max_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_offset:
@@ -1479,14 +1475,12 @@ define amdgpu_kernel void @atomic_max_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1504,10 +1498,9 @@ define amdgpu_kernel void @atomic_max_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
 ; CI-NEXT:    s_mov_b32 s6, s2
 ; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -1524,10 +1517,9 @@ define amdgpu_kernel void @atomic_max_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1539,15 +1531,14 @@ define amdgpu_kernel void @atomic_max_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -1565,10 +1556,8 @@ define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64 addrspace(1)* %out,
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_addr64_offset:
@@ -1585,10 +1574,8 @@ define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_addr64_offset:
@@ -1602,15 +1589,13 @@ define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1630,10 +1615,9 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -1650,14 +1634,13 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1671,16 +1654,15 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -1696,10 +1678,8 @@ define amdgpu_kernel void @atomic_max_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-NEXT:    s_mov_b32 s5, s1
 ; CI-NEXT:    v_mov_b32_e32 v0, s2
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64:
@@ -1712,10 +1692,8 @@ define amdgpu_kernel void @atomic_max_i64(i64 addrspace(1)* %out, i64 %in) {
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s2
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64:
@@ -1725,13 +1703,11 @@ define amdgpu_kernel void @atomic_max_i64(i64 addrspace(1)* %out, i64 %in) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1747,12 +1723,11 @@ define amdgpu_kernel void @atomic_max_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; CI-NEXT:    s_mov_b32 s1, s5
 ; CI-NEXT:    v_mov_b32_e32 v0, s8
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -1767,12 +1742,11 @@ define amdgpu_kernel void @atomic_max_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1784,14 +1758,13 @@ define amdgpu_kernel void @atomic_max_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -1809,10 +1782,8 @@ define amdgpu_kernel void @atomic_max_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_addr64:
@@ -1827,10 +1798,8 @@ define amdgpu_kernel void @atomic_max_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_addr64:
@@ -1844,14 +1813,12 @@ define amdgpu_kernel void @atomic_max_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1871,10 +1838,9 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -1889,14 +1855,13 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; VI-NEXT:    s_addc_u32 s1, s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -1910,15 +1875,14 @@ define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -1932,10 +1896,8 @@ define amdgpu_kernel void @atomic_umax_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_offset:
@@ -1946,10 +1908,8 @@ define amdgpu_kernel void @atomic_umax_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_offset:
@@ -1959,14 +1919,12 @@ define amdgpu_kernel void @atomic_umax_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -1984,10 +1942,9 @@ define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
 ; CI-NEXT:    s_mov_b32 s6, s2
 ; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2004,10 +1961,9 @@ define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2019,15 +1975,14 @@ define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2045,10 +2000,8 @@ define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64 addrspace(1)* %out,
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_addr64_offset:
@@ -2065,10 +2018,8 @@ define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_addr64_offset:
@@ -2082,15 +2033,13 @@ define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2110,10 +2059,9 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2130,14 +2078,13 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2151,16 +2098,15 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2176,10 +2122,8 @@ define amdgpu_kernel void @atomic_umax_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-NEXT:    s_mov_b32 s5, s1
 ; CI-NEXT:    v_mov_b32_e32 v0, s2
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64:
@@ -2192,10 +2136,8 @@ define amdgpu_kernel void @atomic_umax_i64(i64 addrspace(1)* %out, i64 %in) {
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s2
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64:
@@ -2205,13 +2147,11 @@ define amdgpu_kernel void @atomic_umax_i64(i64 addrspace(1)* %out, i64 %in) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2227,12 +2167,11 @@ define amdgpu_kernel void @atomic_umax_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; CI-NEXT:    s_mov_b32 s1, s5
 ; CI-NEXT:    v_mov_b32_e32 v0, s8
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2247,12 +2186,11 @@ define amdgpu_kernel void @atomic_umax_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2264,14 +2202,13 @@ define amdgpu_kernel void @atomic_umax_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2289,10 +2226,8 @@ define amdgpu_kernel void @atomic_umax_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_addr64:
@@ -2307,10 +2242,8 @@ define amdgpu_kernel void @atomic_umax_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_addr64:
@@ -2324,14 +2257,12 @@ define amdgpu_kernel void @atomic_umax_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2351,10 +2282,9 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2369,14 +2299,13 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; VI-NEXT:    s_addc_u32 s1, s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2390,15 +2319,14 @@ define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2412,10 +2340,8 @@ define amdgpu_kernel void @atomic_min_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_offset:
@@ -2426,10 +2352,8 @@ define amdgpu_kernel void @atomic_min_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_offset:
@@ -2439,14 +2363,12 @@ define amdgpu_kernel void @atomic_min_i64_offset(i64 addrspace(1)* %out, i64 %in
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2464,10 +2386,9 @@ define amdgpu_kernel void @atomic_min_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
 ; CI-NEXT:    s_mov_b32 s6, s2
 ; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2484,10 +2405,9 @@ define amdgpu_kernel void @atomic_min_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2499,15 +2419,14 @@ define amdgpu_kernel void @atomic_min_i64_ret_offset(i64 addrspace(1)* %out, i64
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2525,10 +2444,8 @@ define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64 addrspace(1)* %out,
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_addr64_offset:
@@ -2545,10 +2462,8 @@ define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_addr64_offset:
@@ -2562,15 +2477,13 @@ define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2590,10 +2503,9 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2610,14 +2522,13 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2631,16 +2542,15 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64 addrspace(1)* %o
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2656,10 +2566,8 @@ define amdgpu_kernel void @atomic_min_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-NEXT:    s_mov_b32 s5, s1
 ; CI-NEXT:    v_mov_b32_e32 v0, s2
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64:
@@ -2672,10 +2580,8 @@ define amdgpu_kernel void @atomic_min_i64(i64 addrspace(1)* %out, i64 %in) {
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s2
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64:
@@ -2685,13 +2591,11 @@ define amdgpu_kernel void @atomic_min_i64(i64 addrspace(1)* %out, i64 %in) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2707,12 +2611,11 @@ define amdgpu_kernel void @atomic_min_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; CI-NEXT:    s_mov_b32 s1, s5
 ; CI-NEXT:    v_mov_b32_e32 v0, s8
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2727,12 +2630,11 @@ define amdgpu_kernel void @atomic_min_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2744,14 +2646,13 @@ define amdgpu_kernel void @atomic_min_i64_ret(i64 addrspace(1)* %out, i64 addrsp
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2769,10 +2670,8 @@ define amdgpu_kernel void @atomic_min_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_addr64:
@@ -2787,10 +2686,8 @@ define amdgpu_kernel void @atomic_min_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_addr64:
@@ -2804,14 +2701,12 @@ define amdgpu_kernel void @atomic_min_i64_addr64(i64 addrspace(1)* %out, i64 %in
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2831,10 +2726,9 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2849,14 +2743,13 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; VI-NEXT:    s_addc_u32 s1, s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2870,15 +2763,14 @@ define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64 addrspace(1)* %out, i64
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -2892,10 +2784,8 @@ define amdgpu_kernel void @atomic_umin_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_offset:
@@ -2906,10 +2796,8 @@ define amdgpu_kernel void @atomic_umin_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_offset:
@@ -2919,14 +2807,12 @@ define amdgpu_kernel void @atomic_umin_i64_offset(i64 addrspace(1)* %out, i64 %i
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -2944,10 +2830,9 @@ define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
 ; CI-NEXT:    s_mov_b32 s6, s2
 ; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -2964,10 +2849,9 @@ define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
 ; VI-NEXT:    s_mov_b32 s6, s2
 ; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -2979,15 +2863,14 @@ define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64 addrspace(1)* %out, i6
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -3005,10 +2888,8 @@ define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64 addrspace(1)* %out,
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_addr64_offset:
@@ -3025,10 +2906,8 @@ define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64 addrspace(1)* %out,
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_addr64_offset:
@@ -3042,15 +2921,13 @@ define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64 addrspace(1)* %out,
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1] offset:32
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3070,10 +2947,9 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -3090,14 +2966,13 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3111,16 +2986,15 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64 addrspace(1)* %
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -3136,10 +3010,8 @@ define amdgpu_kernel void @atomic_umin_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-NEXT:    s_mov_b32 s5, s1
 ; CI-NEXT:    v_mov_b32_e32 v0, s2
 ; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64:
@@ -3152,10 +3024,8 @@ define amdgpu_kernel void @atomic_umin_i64(i64 addrspace(1)* %out, i64 %in) {
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s2
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64:
@@ -3165,13 +3035,11 @@ define amdgpu_kernel void @atomic_umin_i64(i64 addrspace(1)* %out, i64 %in) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3187,12 +3055,11 @@ define amdgpu_kernel void @atomic_umin_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; CI-NEXT:    s_mov_b32 s1, s5
 ; CI-NEXT:    v_mov_b32_e32 v0, s8
 ; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -3207,12 +3074,11 @@ define amdgpu_kernel void @atomic_umin_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v0, s8
 ; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3224,14 +3090,13 @@ define amdgpu_kernel void @atomic_umin_i64_ret(i64 addrspace(1)* %out, i64 addrs
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
@@ -3249,10 +3114,8 @@ define amdgpu_kernel void @atomic_umin_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, 0
 ; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_addr64:
@@ -3267,10 +3130,8 @@ define amdgpu_kernel void @atomic_umin_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_addr64:
@@ -3284,14 +3145,12 @@ define amdgpu_kernel void @atomic_umin_i64_addr64(i64 addrspace(1)* %out, i64 %i
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   ret void
 }
 
@@ -3311,10 +3170,9 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_mov_b32 s3, s11
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
 ; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; CI-NEXT:    s_endpgm
 ;
@@ -3329,14 +3187,13 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; VI-NEXT:    s_addc_u32 s1, s1, s5
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -3350,15 +3207,14 @@ define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64 addrspace(1)* %out, i6
 ; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
 ; GFX9-NEXT:    s_add_u32 s0, s0, s4
 ; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
-  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in syncscope("workgroup") seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }

diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_i64_min_max_system.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_i64_min_max_system.ll
new file mode 100644 (file)
index 0000000..7187f5d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_i64_min_max_system.ll
@@ -0,0 +1,3912 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=bonaire -amdgpu-atomic-optimizations=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=CI %s
+; RUN: llc -march=amdgcn -mcpu=tonga -amdgpu-atomic-optimizations=false -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-atomic-optimizations=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s
+
+define amdgpu_kernel void @atomic_max_i64_offset(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_max_i64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB0_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB0_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB0_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB0_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_max_i64_ret_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB1_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB1_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB1_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB1_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_max_i64_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB2_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB2_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB2_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB2_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_max_i64_ret_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB3_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB3_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB3_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB3_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB3_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_max_i64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB4_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB4_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB4_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB4_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_max_i64_ret:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB5_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB5_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB5_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB5_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB5_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_max_i64_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB6_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB6_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB6_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB6_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB6_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_max_i64_ret_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB7_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB7_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i64_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB7_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB7_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i64_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB7_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_offset(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_umax_i64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB8_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB8_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB8_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB8_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB8_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_umax_i64_ret_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB9_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB9_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB9_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB9_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB9_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umax_i64_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB10_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB10_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB10_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB10_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB10_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umax_i64_ret_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB11_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB11_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB11_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB11_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB11_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_umax_i64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB12_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB12_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB12_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB12_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB12_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_umax_i64_ret:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB13_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB13_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB13_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB13_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB13_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umax_i64_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB14_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB14_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB14_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB14_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB14_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umax_i64_ret_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB15_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB15_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i64_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB15_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB15_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i64_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB15_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_offset(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_min_i64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB16_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB16_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB16_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB16_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB16_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_min_i64_ret_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB17_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB17_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB17_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB17_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB17_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_min_i64_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB18_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB18_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB18_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB18_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB18_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_min_i64_ret_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB19_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB19_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB19_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB19_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB19_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_min_i64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB20_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB20_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB20_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB20_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB20_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_min_i64_ret:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB21_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB21_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB21_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB21_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB21_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_min_i64_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB22_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB22_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB22_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB22_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB22_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_min_i64_ret_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB23_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB23_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i64_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB23_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB23_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i64_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB23_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_offset(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_umin_i64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB24_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB24_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_umin_i64_ret_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB25_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB25_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umin_i64_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB26_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB26_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umin_i64_ret_addr64_offset:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB27_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB27_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64(i64 addrspace(1)* %out, i64 %in) {
+; CI-LABEL: atomic_umin_i64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB28_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB28_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
+; CI-LABEL: atomic_umin_i64_ret:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB29_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_mov_b32 s0, s6
+; CI-NEXT:    s_mov_b32 s1, s7
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB29_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umin_i64_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB30_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB30_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
+; CI-LABEL: atomic_umin_i64_ret_addr64:
+; CI:       ; %bb.0: ; %entry
+; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; CI-NEXT:    s_mov_b32 s11, 0xf000
+; CI-NEXT:    s_mov_b32 s10, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB31_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; CI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i64_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB31_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b32 s4, s2
+; VI-NEXT:    s_mov_b32 s5, s3
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i64_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
+  %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
+  store i64 %tmp0, i64 addrspace(1)* %out2
+  ret void
+}

diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_min_max_system.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_min_max_system.ll
new file mode 100644 (file)
index 0000000..e61ac9a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_min_max_system.ll
@@ -0,0 +1,3312 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -amdgpu-atomic-optimizations=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI %s
+; RUN: llc -march=amdgcn -mcpu=tonga -amdgpu-atomic-optimizations=false -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-atomic-optimizations=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s
+
+define amdgpu_kernel void @atomic_max_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_max_i32_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB0_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB0_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB0_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB0_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB0_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_max_i32_ret_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB1_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB1_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB1_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB1_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_max_i32_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB2_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB2_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB2_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB2_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_max_i32_ret_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB3_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB3_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB3_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB3_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB3_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB3_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_max_i32:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB4_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB4_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB4_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB4_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_max_i32_ret:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB5_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB5_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB5_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB5_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB5_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB5_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_max_i32_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB6_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB6_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB6_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB6_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB6_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB6_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_max_i32_ret_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB7_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB7_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_max_i32_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB7_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB7_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_max_i32_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB7_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB7_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_umax_i32_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB8_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB8_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB8_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB8_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB8_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB8_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_umax_i32_ret_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB9_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB9_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB9_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB9_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB9_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB9_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umax_i32_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB10_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB10_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB10_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB10_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB10_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB10_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umax_i32_ret_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB11_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB11_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB11_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB11_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB11_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB11_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_umax_i32:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB12_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB12_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB12_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB12_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB12_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB12_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_umax_i32_ret:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB13_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB13_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB13_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB13_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB13_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB13_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umax_i32_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB14_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB14_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB14_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB14_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB14_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB14_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umax_i32_ret_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB15_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB15_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umax_i32_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB15_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB15_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umax_i32_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB15_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB15_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_min_i32_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB16_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB16_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB16_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB16_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB16_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB16_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_min_i32_ret_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB17_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB17_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB17_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB17_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB17_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB17_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_min_i32_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB18_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB18_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB18_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB18_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB18_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB18_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_min_i32_ret_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB19_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB19_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB19_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB19_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB19_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB19_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_min_i32:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB20_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB20_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB20_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB20_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB20_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB20_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_min_i32_ret:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB21_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB21_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB21_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB21_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB21_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB21_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_min_i32_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB22_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB22_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB22_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB22_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB22_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB22_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_min_i32_ret_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB23_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB23_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_min_i32_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB23_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB23_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_min_i32_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB23_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB23_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_umin_i32_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB24_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB24_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_umin_i32_ret_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB25_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_ret_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB25_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_ret_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umin_i32_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB26_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB26_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umin_i32_ret_addr64_offset:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB27_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_ret_addr64_offset:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
+; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_addc_u32 s1, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB27_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_ret_addr64_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
+  %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32(i32 addrspace(1)* %out, i32 %in) {
+; SI-LABEL: atomic_umin_i32:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB28_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB28_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+; SI-LABEL: atomic_umin_i32_ret:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB29_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_ret:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB29_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_ret:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umin_i32_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB30_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB30_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+; SI-LABEL: atomic_umin_i32_ret_addr64:
+; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB31_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s0, s6
+; SI-NEXT:    s_mov_b32 s1, s7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: atomic_umin_i32_ret_addr64:
+; VI:       ; %bb.0: ; %entry
+; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; VI-NEXT:    s_add_u32 s0, s4, s0
+; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB31_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: atomic_umin_i32_ret_addr64:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; GFX9-NEXT:    s_add_u32 s0, s4, s0
+; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s8, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX9-NEXT:    s_endpgm
+entry:
+  %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
+  %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out2
+  ret void
+}

diff --git a/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw-system.ll b/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw-system.ll
new file mode 100644 (file)
index 0000000..fcf100e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw-system.ll
@@ -0,0 +1,129 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; XUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,VI %s
+
+; FIXME: broken on VI because flat instructions need to be emitted
+; instead of addr64 equivalent of the _OFFSET variants.
+
+; Check that moving the pointer out of the resource descriptor to
+; vaddr works for atomics.
+
+declare i32 @llvm.amdgcn.workitem.id.x() #1
+
+define amdgpu_kernel void @atomic_max_i32(ptr addrspace(1) %out, ptr addrspace(1) %in, ptr addrspace(1) %x, i32 %y) #0 {
+; GCN-LABEL: atomic_max_i32:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s11, 0xf000
+; GCN-NEXT:    s_mov_b32 s10, 0
+; GCN-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GCN-NEXT:    buffer_load_dwordx2 v[1:2], v[1:2], s[8:11], 0 addr64 glc
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v0
+; GCN-NEXT:    s_and_saveexec_b64 s[2:3], vcc
+; GCN-NEXT:    s_cbranch_execz .LBB0_4
+; GCN-NEXT:  ; %bb.1: ; %atomic
+; GCN-NEXT:    s_mov_b32 s8, s10
+; GCN-NEXT:    s_mov_b32 s9, s10
+; GCN-NEXT:    buffer_load_dword v4, v[1:2], s[8:11], 0 addr64 offset:400
+; GCN-NEXT:    s_load_dword s2, s[0:1], 0xf
+; GCN-NEXT:    s_mov_b64 s[0:1], 0
+; GCN-NEXT:  .LBB0_2: ; %atomicrmw.start
+; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_max_i32_e32 v3, s2, v4
+; GCN-NEXT:    s_waitcnt expcnt(0)
+; GCN-NEXT:    v_mov_b32_e32 v6, v4
+; GCN-NEXT:    v_mov_b32_e32 v5, v3
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    buffer_atomic_cmpswap v[5:6], v[1:2], s[8:11], 0 addr64 offset:400 glc
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    buffer_wbinvl1
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v4
+; GCN-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v4, v5
+; GCN-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_cbranch_execnz .LBB0_2
+; GCN-NEXT:  ; %bb.3: ; %atomicrmw.end
+; GCN-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    buffer_store_dword v5, off, s[4:7], 0
+; GCN-NEXT:  .LBB0_4: ; %exit
+; GCN-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.gep = getelementptr ptr addrspace(1), ptr addrspace(1) %in, i32 %tid
+  %ptr = load volatile ptr addrspace(1), ptr addrspace(1) %tid.gep
+  %xor = xor i32 %tid, 1
+  %cmp = icmp ne i32 %xor, 0
+  br i1 %cmp, label %atomic, label %exit
+
+atomic:
+  %gep = getelementptr i32, ptr addrspace(1) %ptr, i32 100
+  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y seq_cst
+  store i32 %ret, ptr addrspace(1) %out
+  br label %exit
+
+exit:
+  ret void
+}
+
+define amdgpu_kernel void @atomic_max_i32_noret(ptr addrspace(1) %out, ptr addrspace(1) %in, ptr addrspace(1) %x, i32 %y) #0 {
+; GCN-LABEL: atomic_max_i32_noret:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xb
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    s_mov_b32 s6, 0
+; GCN-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    buffer_load_dwordx2 v[1:2], v[1:2], s[4:7], 0 addr64 glc
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v0
+; GCN-NEXT:    s_and_saveexec_b64 s[2:3], vcc
+; GCN-NEXT:    s_cbranch_execz .LBB1_3
+; GCN-NEXT:  ; %bb.1: ; %atomic
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s6
+; GCN-NEXT:    buffer_load_dword v4, v[1:2], s[4:7], 0 addr64 offset:400
+; GCN-NEXT:    s_load_dword s2, s[0:1], 0xf
+; GCN-NEXT:    s_mov_b64 s[0:1], 0
+; GCN-NEXT:  .LBB1_2: ; %atomicrmw.start
+; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_max_i32_e32 v3, s2, v4
+; GCN-NEXT:    s_waitcnt expcnt(0)
+; GCN-NEXT:    v_mov_b32_e32 v6, v4
+; GCN-NEXT:    v_mov_b32_e32 v5, v3
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    buffer_atomic_cmpswap v[5:6], v[1:2], s[4:7], 0 addr64 offset:400 glc
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    buffer_wbinvl1
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v4
+; GCN-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v4, v5
+; GCN-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_cbranch_execnz .LBB1_2
+; GCN-NEXT:  .LBB1_3: ; %exit
+; GCN-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.gep = getelementptr ptr addrspace(1), ptr addrspace(1) %in, i32 %tid
+  %ptr = load volatile ptr addrspace(1), ptr addrspace(1) %tid.gep
+  %xor = xor i32 %tid, 1
+  %cmp = icmp ne i32 %xor, 0
+  br i1 %cmp, label %atomic, label %exit
+
+atomic:
+  %gep = getelementptr i32, ptr addrspace(1) %ptr, i32 100
+  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y seq_cst
+  br label %exit
+
+exit:
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }

diff --git a/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll b/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll
index e04c6a9..a44ede5 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll
+++ b/llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll
@@ -33,10 +33,9 @@ define amdgpu_kernel void @atomic_max_i32(ptr addrspace(1) %out, ptr addrspace(1
 ; GCN-NEXT:    s_mov_b32 s7, s11
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v0, s0
-; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    buffer_atomic_smax v0, v[1:2], s[8:11], 0 addr64 offset:400 glc
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    buffer_wbinvl1
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:  .LBB0_2: ; %exit
 ; GCN-NEXT:    s_endpgm
@@ -49,7 +48,7 @@ define amdgpu_kernel void @atomic_max_i32(ptr addrspace(1) %out, ptr addrspace(1
 
 atomic:
   %gep = getelementptr i32, ptr addrspace(1) %ptr, i32 100
-  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y seq_cst
+  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y syncscope("workgroup") seq_cst
   store i32 %ret, ptr addrspace(1) %out
   br label %exit
 
@@ -77,10 +76,8 @@ define amdgpu_kernel void @atomic_max_i32_noret(ptr addrspace(1) %out, ptr addrs
 ; GCN-NEXT:    s_mov_b32 s5, s6
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v0, s0
-; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    buffer_atomic_smax v0, v[1:2], s[4:7], 0 addr64 offset:400
-; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    buffer_wbinvl1
 ; GCN-NEXT:  .LBB1_2: ; %exit
 ; GCN-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
@@ -92,7 +89,7 @@ define amdgpu_kernel void @atomic_max_i32_noret(ptr addrspace(1) %out, ptr addrs
 
 atomic:
   %gep = getelementptr i32, ptr addrspace(1) %ptr, i32 100
-  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y seq_cst
+  %ret = atomicrmw max ptr addrspace(1) %gep, i32 %y syncscope("workgroup") seq_cst
   br label %exit
 
 exit: