projects / platform / upstream / llvm.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 450073c)
GlobalISel: Implement lower for G_INTRINSIC_ROUND
authorMatt Arsenault <Matthew.Arsenault@amd.com>
Tue, 24 Dec 2019 19:49:31 +0000 (14:49 -0500)
committerMatt Arsenault <arsenm2@gmail.com>
Mon, 6 Jan 2020 23:26:42 +0000 (18:26 -0500)
Mostly copied from AMDGPU lowering implementation, except used
G_SITOFP instead of directly creating a select on -1.0, 0.0.

llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h patch | blob | history
llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h patch | blob | history
llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp patch | blob | history
llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp patch | blob | history
llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp patch | blob | history
llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir patch | blob | history
llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-intrinsic-round.mir [deleted file] patch | blob | history

diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
index a09edff..cc89a3b 100644 (file)
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
@@ -231,6 +231,7 @@ public:
   LegalizeResult lowerFCopySign(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
   LegalizeResult lowerFMinNumMaxNum(MachineInstr &MI);
   LegalizeResult lowerFMad(MachineInstr &MI);
+  LegalizeResult lowerIntrinsicRound(MachineInstr &MI);
   LegalizeResult lowerUnmergeValues(MachineInstr &MI);
   LegalizeResult lowerShuffleVector(MachineInstr &MI);
   LegalizeResult lowerDynStackAlloc(MachineInstr &MI);
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h b/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
index d3c17ae..072a141 100644 (file)
--- a/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
@@ -1413,6 +1413,12 @@ public:
     return buildInstr(TargetOpcode::G_FCANONICALIZE, {Dst}, {Src0}, Flags);
   }
 
+  /// Build and insert \p Dst = G_INTRINSIC_TRUNC \p Src0
+  MachineInstrBuilder buildIntrinsicTrunc(const DstOp &Dst, const SrcOp &Src0,
+                                         Optional<unsigned> Flags = None) {
+    return buildInstr(TargetOpcode::G_INTRINSIC_TRUNC, {Dst}, {Src0}, Flags);
+  }
+
   /// Build and insert \p Res = G_FCOPYSIGN \p Op0, \p Op1
   MachineInstrBuilder buildFCopysign(const DstOp &Dst, const SrcOp &Src0,
                                      const SrcOp &Src1) {
diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index bb3ece3..53bda81 100644 (file)
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -2053,6 +2053,8 @@ LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
   }
   case TargetOpcode::G_FMAD:
     return lowerFMad(MI);
+  case TargetOpcode::G_INTRINSIC_ROUND:
+    return lowerIntrinsicRound(MI);
   case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
     Register OldValRes = MI.getOperand(0).getReg();
     Register SuccessRes = MI.getOperand(1).getReg();
@@ -4130,6 +4132,33 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerFMad(MachineInstr &MI) {
 }
 
 LegalizerHelper::LegalizeResult
+LegalizerHelper::lowerIntrinsicRound(MachineInstr &MI) {
+  Register DstReg = MI.getOperand(0).getReg();
+  Register SrcReg = MI.getOperand(1).getReg();
+  unsigned Flags = MI.getFlags();
+  LLT Ty = MRI.getType(DstReg);
+  const LLT CondTy = Ty.changeElementSize(1);
+
+  // result = trunc(src);
+  // if (src < 0.0 && src != result)
+  //   result += -1.0.
+
+  auto Zero = MIRBuilder.buildFConstant(Ty, 0.0);
+  auto Trunc = MIRBuilder.buildIntrinsicTrunc(Ty, SrcReg, Flags);
+
+  auto Lt0 = MIRBuilder.buildFCmp(CmpInst::FCMP_OLT, CondTy,
+                                  SrcReg, Zero, Flags);
+  auto NeTrunc = MIRBuilder.buildFCmp(CmpInst::FCMP_ONE, CondTy,
+                                      SrcReg, Trunc, Flags);
+  auto And = MIRBuilder.buildAnd(CondTy, Lt0, NeTrunc);
+  auto AddVal = MIRBuilder.buildSITOFP(Ty, And);
+
+  MIRBuilder.buildFAdd(DstReg, Trunc, AddVal);
+  MI.eraseFromParent();
+  return Legalized;
+}
+
+LegalizerHelper::LegalizeResult
 LegalizerHelper::lowerUnmergeValues(MachineInstr &MI) {
   const unsigned NumDst = MI.getNumOperands() - 1;
   const Register SrcReg = MI.getOperand(NumDst).getReg();
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index 3498abe..64f058a 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -471,8 +471,8 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
        .scalarize(0);
 
   getActionDefinitionsBuilder(G_INTRINSIC_ROUND)
-    .legalFor({S32, S64})
-    .scalarize(0);
+    .scalarize(0)
+    .lower();
 
   if (ST.has16BitInsts()) {
     getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index 8916273..5c8afce 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -2377,7 +2377,6 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
   case AMDGPU::G_FMAXNUM_IEEE:
   case AMDGPU::G_FCANONICALIZE:
   case AMDGPU::G_INTRINSIC_TRUNC:
-  case AMDGPU::G_INTRINSIC_ROUND:
   case AMDGPU::G_AMDGPU_FFBH_U32:
     return getDefaultMappingVOP(MI);
   case AMDGPU::G_UMULH:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir
index 220ce91..5ea118f 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir
@@ -1,5 +1,7 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8 %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9 %s
 
 ---
 name: test_intrinsic_round_s32
@@ -7,12 +9,86 @@ body: |
   bb.0:
     liveins: $vgpr0
 
-    ; CHECK-LABEL: name: test_intrinsic_round_s32
-    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
-    ; CHECK: $vgpr0 = COPY [[COPY]](s32)
+    ; GFX6-LABEL: name: test_intrinsic_round_s32
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX6: $vgpr0 = COPY [[FADD]](s32)
+    ; GFX8-LABEL: name: test_intrinsic_round_s32
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: $vgpr0 = COPY [[FADD]](s32)
+    ; GFX9-LABEL: name: test_intrinsic_round_s32
+    ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: $vgpr0 = COPY [[FADD]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = G_INTRINSIC_ROUND %0
-    $vgpr0 = COPY %0
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_intrinsic_round_s32_flags
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; GFX6-LABEL: name: test_intrinsic_round_s32_flags
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX6: $vgpr0 = COPY [[FADD]](s32)
+    ; GFX8-LABEL: name: test_intrinsic_round_s32_flags
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: $vgpr0 = COPY [[FADD]](s32)
+    ; GFX9-LABEL: name: test_intrinsic_round_s32_flags
+    ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: $vgpr0 = COPY [[FADD]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = nsz G_INTRINSIC_ROUND %0
+    $vgpr0 = COPY %1
 ...
 
 ---
@@ -21,10 +97,59 @@ body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
-    ; CHECK-LABEL: name: test_intrinsic_round_s64
-    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
-    ; CHECK: [[INTRINSIC_ROUND:%[0-9]+]]:_(s64) = G_INTRINSIC_ROUND [[COPY]]
-    ; CHECK: $vgpr0_vgpr1 = COPY [[INTRINSIC_ROUND]](s64)
+    ; GFX6-LABEL: name: test_intrinsic_round_s64
+    ; GFX6: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX6: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX6: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](s64)
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
+    ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32)
+    ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C3]]
+    ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648
+    ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV1]], [[C4]]
+    ; GFX6: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495
+    ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND]](s32)
+    ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB]](s32)
+    ; GFX6: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1
+    ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C7]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[COPY]], [[XOR]]
+    ; GFX6: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 51
+    ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C6]]
+    ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C8]]
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[MV]], [[AND1]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP1]](s1), [[COPY]], [[SELECT]]
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]]
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND2]](s1), [[C9]], [[C]]
+    ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]]
+    ; GFX6: $vgpr0_vgpr1 = COPY [[FADD]](s64)
+    ; GFX8-LABEL: name: test_intrinsic_round_s64
+    ; GFX8: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: $vgpr0_vgpr1 = COPY [[FADD]](s64)
+    ; GFX9-LABEL: name: test_intrinsic_round_s64
+    ; GFX9: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: $vgpr0_vgpr1 = COPY [[FADD]](s64)
     %0:_(s64) = COPY $vgpr0_vgpr1
     %1:_(s64) = G_INTRINSIC_ROUND %0
     $vgpr0_vgpr1 = COPY %1
@@ -36,13 +161,63 @@ body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
-    ; CHECK-LABEL: name: test_intrinsic_round_v2s32
-    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
-    ; CHECK: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
-    ; CHECK: [[INTRINSIC_ROUND:%[0-9]+]]:_(s32) = G_INTRINSIC_ROUND [[UV]]
-    ; CHECK: [[INTRINSIC_ROUND1:%[0-9]+]]:_(s32) = G_INTRINSIC_ROUND [[UV1]]
-    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[INTRINSIC_ROUND]](s32), [[INTRINSIC_ROUND1]](s32)
-    ; CHECK: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX6-LABEL: name: test_intrinsic_round_v2s32
+    ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    ; GFX6: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]]
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]]
+    ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]]
+    ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32)
+    ; GFX6: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX8-LABEL: name: test_intrinsic_round_v2s32
+    ; GFX8: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    ; GFX8: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]]
+    ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]]
+    ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32)
+    ; GFX8: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-LABEL: name: test_intrinsic_round_v2s32
+    ; GFX9: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]]
+    ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]]
+    ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX9: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
     %0:_(<2 x s32>) = COPY $vgpr0_vgpr1
     %1:_(<2 x s32>) = G_INTRINSIC_ROUND %0
     $vgpr0_vgpr1 = COPY %1
@@ -54,14 +229,595 @@ body: |
   bb.0:
     liveins: $vgpr0_vgpr1_vgpr2_vgpr3
 
-    ; CHECK-LABEL: name: test_intrinsic_round_v2s64
-    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
-    ; CHECK: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
-    ; CHECK: [[INTRINSIC_ROUND:%[0-9]+]]:_(s64) = G_INTRINSIC_ROUND [[UV]]
-    ; CHECK: [[INTRINSIC_ROUND1:%[0-9]+]]:_(s64) = G_INTRINSIC_ROUND [[UV1]]
-    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[INTRINSIC_ROUND]](s64), [[INTRINSIC_ROUND1]](s64)
-    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX6-LABEL: name: test_intrinsic_round_v2s64
+    ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; GFX6: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
+    ; GFX6: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX6: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV]](s64)
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
+    ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32)
+    ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C3]]
+    ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648
+    ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV3]], [[C4]]
+    ; GFX6: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495
+    ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND]](s32)
+    ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB]](s32)
+    ; GFX6: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1
+    ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C7]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[UV]], [[XOR]]
+    ; GFX6: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 51
+    ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C6]]
+    ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C8]]
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[MV]], [[AND1]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP1]](s1), [[UV]], [[SELECT]]
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]]
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND2]](s1), [[C9]], [[C]]
+    ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]]
+    ; GFX6: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV1]](s64)
+    ; GFX6: [[INT1:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32)
+    ; GFX6: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[INT1]], [[C3]]
+    ; GFX6: [[AND3:%[0-9]+]]:_(s32) = G_AND [[UV5]], [[C4]]
+    ; GFX6: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND3]](s32)
+    ; GFX6: [[ASHR1:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB1]](s32)
+    ; GFX6: [[XOR1:%[0-9]+]]:_(s64) = G_XOR [[ASHR1]], [[C7]]
+    ; GFX6: [[AND4:%[0-9]+]]:_(s64) = G_AND [[UV1]], [[XOR1]]
+    ; GFX6: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB1]](s32), [[C6]]
+    ; GFX6: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB1]](s32), [[C8]]
+    ; GFX6: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP2]](s1), [[MV1]], [[AND4]]
+    ; GFX6: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[ICMP3]](s1), [[UV1]], [[SELECT2]]
+    ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]]
+    ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]]
+    ; GFX6: [[AND5:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX6: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[AND5]](s1), [[C9]], [[C]]
+    ; GFX6: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT3]]
+    ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64)
+    ; GFX6: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX8-LABEL: name: test_intrinsic_round_v2s64
+    ; GFX8: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; GFX8: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
+    ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]]
+    ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]]
+    ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND1]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64)
+    ; GFX8: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX9-LABEL: name: test_intrinsic_round_v2s64
+    ; GFX9: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; GFX9: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
+    ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]]
+    ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]]
+    ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]]
+    ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX9: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND1]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
     %0:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
     %1:_(<2 x s64>) = G_INTRINSIC_ROUND %0
     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
 ...
+
+---
+name: test_intrinsic_round_s16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; GFX6-LABEL: name: test_intrinsic_round_s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C]](s16)
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]]
+    ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16)
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]]
+    ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
+    ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC1]](s16)
+    ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX8-LABEL: name: test_intrinsic_round_s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
+    ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-LABEL: name: test_intrinsic_round_s16
+    ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
+    ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s16) = G_TRUNC %0
+    %2:_(s16) = G_INTRINSIC_ROUND %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_intrinsic_round_v2s16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; GFX6-LABEL: name: test_intrinsic_round_v2s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+    ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]]
+    ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16)
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]]
+    ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
+    ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]]
+    ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32)
+    ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]]
+    ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16)
+    ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]]
+    ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
+    ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FPTRUNC1]](s16), [[FPTRUNC3]](s16)
+    ; GFX6: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX8-LABEL: name: test_intrinsic_round_v2s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+    ; GFX8: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX8: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX9-LABEL: name: test_intrinsic_round_v2s16
+    ; GFX9: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+    ; GFX9: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX9: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    %0:_(<2 x s16>) = COPY $vgpr0
+    %1:_(<2 x s16>) = G_INTRINSIC_ROUND %0
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_intrinsic_round_v3s16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; GFX6-LABEL: name: test_intrinsic_round_v3s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX6: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[COPY]](<4 x s16>), 0
+    ; GFX6: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX6: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s16>), 0
+    ; GFX6: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT]](<4 x s16>)
+    ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]]
+    ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16)
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]]
+    ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
+    ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]]
+    ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32)
+    ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]]
+    ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16)
+    ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]]
+    ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
+    ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT14]]
+    ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32)
+    ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT15]](s32), [[FPEXT16]]
+    ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
+    ; GFX6: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT17]](s32), [[FPEXT18]]
+    ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
+    ; GFX6: [[FPEXT20:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT2]](s16)
+    ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]]
+    ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32)
+    ; GFX6: [[DEF1:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+    ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FPTRUNC1]](s16), [[FPTRUNC3]](s16)
+    ; GFX6: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FPTRUNC5]](s16), [[DEF1]](s16)
+    ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX6: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[CONCAT_VECTORS]](<4 x s16>), 0
+    ; GFX6: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX6: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; GFX6: $vgpr0_vgpr1 = COPY [[INSERT1]](<4 x s16>)
+    ; GFX8-LABEL: name: test_intrinsic_round_v3s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX8: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[COPY]](<4 x s16>), 0
+    ; GFX8: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX8: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s16>), 0
+    ; GFX8: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT]](<4 x s16>)
+    ; GFX8: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX8: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX8: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX8: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX8: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]]
+    ; GFX8: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]]
+    ; GFX8: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]]
+    ; GFX8: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]]
+    ; GFX8: [[DEF1:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+    ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX8: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD2]](s16), [[DEF1]](s16)
+    ; GFX8: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX8: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[CONCAT_VECTORS]](<4 x s16>), 0
+    ; GFX8: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX8: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; GFX8: $vgpr0_vgpr1 = COPY [[INSERT1]](<4 x s16>)
+    ; GFX9-LABEL: name: test_intrinsic_round_v3s16
+    ; GFX9: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX9: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[COPY]](<4 x s16>), 0
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s16>), 0
+    ; GFX9: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT]](<4 x s16>)
+    ; GFX9: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX9: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX9: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]]
+    ; GFX9: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]]
+    ; GFX9: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]]
+    ; GFX9: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]]
+    ; GFX9: [[DEF1:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD2]](s16), [[DEF1]](s16)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX9: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[CONCAT_VECTORS]](<4 x s16>), 0
+    ; GFX9: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; GFX9: $vgpr0_vgpr1 = COPY [[INSERT1]](<4 x s16>)
+    %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s16>) = G_EXTRACT %0, 0
+    %2:_(<3 x s16>) = G_INTRINSIC_ROUND %1
+    %3:_(<4 x s16>) = G_IMPLICIT_DEF
+    %4:_(<4 x s16>) = G_INSERT %3, %2, 0
+    $vgpr0_vgpr1 = COPY %4
+...
+
+---
+name: test_intrinsic_round_v4s16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; GFX6-LABEL: name: test_intrinsic_round_v4s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX6: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
+    ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
+    ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]]
+    ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]]
+    ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
+    ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16)
+    ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]]
+    ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
+    ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]]
+    ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32)
+    ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]]
+    ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
+    ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]]
+    ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
+    ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16)
+    ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]]
+    ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
+    ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT14]]
+    ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32)
+    ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT15]](s32), [[FPEXT16]]
+    ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
+    ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
+    ; GFX6: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT17]](s32), [[FPEXT18]]
+    ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
+    ; GFX6: [[FPEXT20:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT2]](s16)
+    ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]]
+    ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32)
+    ; GFX6: [[FPEXT21:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
+    ; GFX6: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT21]]
+    ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC3]](s32)
+    ; GFX6: [[FPEXT22:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
+    ; GFX6: [[FPEXT23:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16)
+    ; GFX6: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT22]](s32), [[FPEXT23]]
+    ; GFX6: [[FPEXT24:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
+    ; GFX6: [[FPEXT25:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16)
+    ; GFX6: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT24]](s32), [[FPEXT25]]
+    ; GFX6: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]]
+    ; GFX6: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]]
+    ; GFX6: [[FPEXT26:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16)
+    ; GFX6: [[FPEXT27:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT3]](s16)
+    ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT26]], [[FPEXT27]]
+    ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32)
+    ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FPTRUNC1]](s16), [[FPTRUNC3]](s16)
+    ; GFX6: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FPTRUNC5]](s16), [[FPTRUNC7]](s16)
+    ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX6: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; GFX8-LABEL: name: test_intrinsic_round_v4s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX8: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
+    ; GFX8: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX8: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX8: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX8: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX8: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX8: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]]
+    ; GFX8: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]]
+    ; GFX8: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]]
+    ; GFX8: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]]
+    ; GFX8: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC3]]
+    ; GFX8: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC3]](s16), [[C1]]
+    ; GFX8: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC3]](s16), [[INTRINSIC_TRUNC3]]
+    ; GFX8: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]]
+    ; GFX8: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]]
+    ; GFX8: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]]
+    ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX8: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD2]](s16), [[FADD3]](s16)
+    ; GFX8: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX8: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; GFX9-LABEL: name: test_intrinsic_round_v4s16
+    ; GFX9: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    ; GFX9: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
+    ; GFX9: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX9: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
+    ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000
+    ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]]
+    ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]]
+    ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]]
+    ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00
+    ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]]
+    ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]]
+    ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]]
+    ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]]
+    ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]]
+    ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]]
+    ; GFX9: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]]
+    ; GFX9: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]]
+    ; GFX9: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]]
+    ; GFX9: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]]
+    ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]]
+    ; GFX9: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC3]]
+    ; GFX9: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC3]](s16), [[C1]]
+    ; GFX9: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC3]](s16), [[INTRINSIC_TRUNC3]]
+    ; GFX9: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]]
+    ; GFX9: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]]
+    ; GFX9: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD]](s16), [[FADD1]](s16)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[FADD2]](s16), [[FADD3]](s16)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s16>) = G_INTRINSIC_ROUND %0
+    $vgpr0_vgpr1 = COPY %1
+...
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-intrinsic-round.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-intrinsic-round.mir
deleted file mode 100644 (file)
index 456baf4..0000000
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-intrinsic-round.mir
+++ /dev/null
@@ -1,31 +0,0 @@
-# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
-
----
-name: intrinsic_round_s
-legalized: true
-
-body: |
-  bb.0:
-    liveins: $sgpr0
-    ; CHECK-LABEL: name: intrinsic_round_s
-    ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
-    ; CHECK: [[INTRINSIC_ROUND:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_ROUND [[COPY]]
-    %0:_(s32) = COPY $sgpr0
-    %1:_(s32) = G_INTRINSIC_ROUND %0
-...
-
----
-name: intrinsic_round_v
-legalized: true
-
-body: |
-  bb.0:
-    liveins: $vgpr0
-    ; CHECK-LABEL: name: intrinsic_round_v
-    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
-    ; CHECK: [[INTRINSIC_ROUND:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_ROUND [[COPY]]
-    %0:_(s32) = COPY $vgpr0
-    %1:_(s32) = G_INTRINSIC_ROUND %0
-...
Domain: System / Toolchain;