AMDGPU/GlobalISel: Select G_FENCE
authorMatt Arsenault <Matthew.Arsenault@amd.com>
Tue, 2 Jul 2019 14:17:38 +0000 (14:17 +0000)
committerMatt Arsenault <Matthew.Arsenault@amd.com>
Tue, 2 Jul 2019 14:17:38 +0000 (14:17 +0000)
Manually select to workaround tablegen emitter emitting checks for
G_CONSTANT.

llvm-svn: 364927

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
llvm/test/CodeGen/AMDGPU/GlobalISel/memory-legalizer-atomic-fence.ll [new file with mode: 0644]

index 868a742..c7e5f05 100644 (file)
@@ -1180,6 +1180,11 @@ bool AMDGPUInstructionSelector::select(MachineInstr &I,
     return selectG_BRCOND(I);
   case TargetOpcode::G_FRAME_INDEX:
     return selectG_FRAME_INDEX(I);
+  case TargetOpcode::G_FENCE:
+    // FIXME: Tablegen importer doesn't handle the imm operands correctly, and
+    // is checking for G_CONSTANT
+    I.setDesc(TII.get(AMDGPU::ATOMIC_FENCE));
+    return true;
   }
   return false;
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/memory-legalizer-atomic-fence.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/memory-legalizer-atomic-fence.ll
new file mode 100644 (file)
index 0000000..9d9f981
--- /dev/null
@@ -0,0 +1,719 @@
+; RUN: llc -global-isel -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX6,GFX68 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+code-object-v3 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10WGP %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+code-object-v3,+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10CU %s
+
+; FUNC-LABEL: {{^}}system_one_as_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0){{$}}
+; GFX6-NEXT:  buffer_wbinvl1{{$}}
+; GFX8:       s_waitcnt vmcnt(0){{$}}
+; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
+; GFX10:      s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_one_as_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_one_as_acquire() {
+entry:
+  fence syncscope("one-as") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_one_as_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_one_as_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_one_as_release() {
+entry:
+  fence syncscope("one-as") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_one_as_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_one_as_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_one_as_acq_rel() {
+entry:
+  fence syncscope("one-as") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_one_as_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_one_as_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_one_as_seq_cst() {
+entry:
+  fence syncscope("one-as") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_one_as_acquire() {
+entry:
+  fence syncscope("singlethread-one-as") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_one_as_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_one_as_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_one_as_release() {
+entry:
+  fence syncscope("singlethread-one-as") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_one_as_acq_rel() {
+entry:
+  fence syncscope("singlethread-one-as") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_one_as_seq_cst() {
+entry:
+  fence syncscope("singlethread-one-as") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_one_as_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0){{$}}
+; GFX6-NEXT:  buffer_wbinvl1{{$}}
+; GFX8:       s_waitcnt vmcnt(0){{$}}
+; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
+; GFX10:      s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_one_as_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_one_as_acquire() {
+entry:
+  fence syncscope("agent-one-as") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_one_as_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_one_as_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_one_as_release() {
+entry:
+  fence syncscope("agent-one-as") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_one_as_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_one_as_acq_rel() {
+entry:
+  fence syncscope("agent-one-as") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_waitcnt vmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_one_as_seq_cst() {
+entry:
+  fence syncscope("agent-one-as") seq_cst
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_one_as_acquire:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_one_as_acquire() {
+entry:
+  fence syncscope("workgroup-one-as") acquire
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_one_as_release:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NOT:     buffer_gl0_inv
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_one_as_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_one_as_release() {
+entry:
+  fence syncscope("workgroup-one-as") release
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_one_as_acq_rel:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_one_as_acq_rel() {
+entry:
+  fence syncscope("workgroup-one-as") acq_rel
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_one_as_seq_cst:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_one_as_seq_cst() {
+entry:
+  fence syncscope("workgroup-one-as") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_one_as_acquire() {
+entry:
+  fence syncscope("wavefront-one-as") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_one_as_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_one_as_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_one_as_release() {
+entry:
+  fence syncscope("wavefront-one-as") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_one_as_acq_rel() {
+entry:
+  fence syncscope("wavefront-one-as") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_one_as_seq_cst() {
+entry:
+  fence syncscope("wavefront-one-as") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX6-NEXT:  buffer_wbinvl1{{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_acquire() {
+entry:
+  fence acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_release() {
+entry:
+  fence release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_acq_rel() {
+entry:
+  fence acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}system_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel system_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @system_seq_cst() {
+entry:
+  fence seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_acquire() {
+entry:
+  fence syncscope("singlethread") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_release() {
+entry:
+  fence syncscope("singlethread") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_acq_rel() {
+entry:
+  fence syncscope("singlethread") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}singlethread_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel singlethread_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @singlethread_seq_cst() {
+entry:
+  fence syncscope("singlethread") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX6-NEXT:  buffer_wbinvl1{{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_acquire() {
+entry:
+  fence syncscope("agent") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_release() {
+entry:
+  fence syncscope("agent") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_acq_rel() {
+entry:
+  fence syncscope("agent") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}agent_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX6:       buffer_wbinvl1{{$}}
+; GFX8:       buffer_wbinvl1_vol{{$}}
+; GFX10-NEXT: buffer_gl0_inv{{$}}
+; GFX10-NEXT: buffer_gl1_inv{{$}}
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel agent_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @agent_seq_cst() {
+entry:
+  fence syncscope("agent") seq_cst
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_acquire:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_acquire() {
+entry:
+  fence syncscope("workgroup") acquire
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_release:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10-NOT:     buffer_gl0_inv
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_release() {
+entry:
+  fence syncscope("workgroup") release
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_acq_rel:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_acq_rel() {
+entry:
+  fence syncscope("workgroup") acq_rel
+  ret void
+}
+
+; FUNC-LABEL:    {{^}}workgroup_seq_cst:
+; GCN:           %bb.0
+; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
+; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
+; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
+; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
+; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
+; GFX10CU-NOT:   buffer_gl0_inv{{$}}
+; GCN-NOT:       ATOMIC_FENCE
+; GCN:           s_endpgm
+; GFX10:         .amdhsa_kernel workgroup_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @workgroup_seq_cst() {
+entry:
+  fence syncscope("workgroup") seq_cst
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_acquire:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_acquire
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_acquire() {
+entry:
+  fence syncscope("wavefront") acquire
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_release:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_release
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_release() {
+entry:
+  fence syncscope("wavefront") release
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_acq_rel:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_acq_rel
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_acq_rel() {
+entry:
+  fence syncscope("wavefront") acq_rel
+  ret void
+}
+
+; FUNC-LABEL: {{^}}wavefront_seq_cst:
+; GCN:        %bb.0
+; GCN-NOT:    ATOMIC_FENCE
+; GCN:        s_endpgm
+; GFX10:         .amdhsa_kernel wavefront_seq_cst
+; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
+; GFX10CU:       .amdhsa_workgroup_processor_mode 0
+; GFX10-NOT:     .amdhsa_memory_ordered 0
+define amdgpu_kernel void @wavefront_seq_cst() {
+entry:
+  fence syncscope("wavefront") seq_cst
+  ret void
+}