AMDGPU: Support commuting with immediate in src0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@280970 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -497,9 +497,7 @@ void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   }
 }
 
-int SIInstrInfo::commuteOpcode(const MachineInstr &MI) const {
-  const unsigned Opcode = MI.getOpcode();
-
+int SIInstrInfo::commuteOpcode(unsigned Opcode) const {
   int NewOpc;
 
   // Try to map original to commuted opcode
@@ -908,91 +906,89 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
   return true;
 }
 
-/// Commutes the operands in the given instruction.
-/// The commutable operands are specified by their indices OpIdx0 and OpIdx1.
-///
-/// Do not call this method for a non-commutable instruction or for
-/// non-commutable pair of operand indices OpIdx0 and OpIdx1.
-/// Even though the instruction is commutable, the method may still
-/// fail to commute the operands, null pointer is returned in such cases.
-MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
-                                                  unsigned OpIdx0,
-                                                  unsigned OpIdx1) const {
-  int CommutedOpcode = commuteOpcode(MI);
-  if (CommutedOpcode == -1)
-    return nullptr;
+bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
+                                      MachineOperand &Src0,
+                                      unsigned Src0OpName,
+                                      MachineOperand &Src1,
+                                      unsigned Src1OpName) const {
+  MachineOperand *Src0Mods = getNamedOperand(MI, Src0OpName);
+  if (!Src0Mods)
+    return false;
 
-  int Src0Idx =
-      AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0);
-  MachineOperand &Src0 = MI.getOperand(Src0Idx);
-  if (!Src0.isReg())
+  MachineOperand *Src1Mods = getNamedOperand(MI, Src1OpName);
+  assert(Src1Mods &&
+         "All commutable instructions have both src0 and src1 modifiers");
+
+  int Src0ModsVal = Src0Mods->getImm();
+  int Src1ModsVal = Src1Mods->getImm();
+
+  Src1Mods->setImm(Src0ModsVal);
+  Src0Mods->setImm(Src1ModsVal);
+  return true;
+}
+
+static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
+                                             MachineOperand &RegOp,
+                                             MachineOperand &ImmOp) {
+  // TODO: Handle other immediate like types.
+  if (!ImmOp.isImm())
     return nullptr;
 
-  int Src1Idx =
-      AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src1);
+  int64_t ImmVal = ImmOp.getImm();
+  ImmOp.ChangeToRegister(RegOp.getReg(), false, false,
+                         RegOp.isKill(), RegOp.isDead(), RegOp.isUndef(),
+                         RegOp.isDebug());
+  ImmOp.setSubReg(RegOp.getSubReg());
+  RegOp.ChangeToImmediate(ImmVal);
+  return &MI;
+}
 
-  if ((OpIdx0 != static_cast<unsigned>(Src0Idx) ||
-       OpIdx1 != static_cast<unsigned>(Src1Idx)) &&
-      (OpIdx0 != static_cast<unsigned>(Src1Idx) ||
-       OpIdx1 != static_cast<unsigned>(Src0Idx)))
+MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
+                                                  unsigned Src0Idx,
+                                                  unsigned Src1Idx) const {
+  assert(!NewMI && "this should never be used");
+
+  unsigned Opc = MI.getOpcode();
+  int CommutedOpcode = commuteOpcode(Opc);
+  if (CommutedOpcode == -1)
     return nullptr;
 
-  MachineOperand &Src1 = MI.getOperand(Src1Idx);
+  assert(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) ==
+           static_cast<int>(Src0Idx) &&
+         AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1) ==
+           static_cast<int>(Src1Idx) &&
+         "inconsistency with findCommutedOpIndices");
 
-  if (isVOP2(MI) || isVOPC(MI)) {
-    const MCInstrDesc &InstrDesc = MI.getDesc();
-    // For VOP2 and VOPC instructions, any operand type is valid to use for
-    // src0.  Make sure we can use the src0 as src1.
-    //
-    // We could be stricter here and only allow commuting if there is a reason
-    // to do so. i.e. if both operands are VGPRs there is no real benefit,
-    // although MachineCSE attempts to find matches by commuting.
-    const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
-    if (!isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src0))
-      return nullptr;
-  }
+  MachineOperand &Src0 = MI.getOperand(Src0Idx);
+  MachineOperand &Src1 = MI.getOperand(Src1Idx);
 
-  MachineInstr *CommutedMI = &MI;
-  if (!Src1.isReg()) {
-    // Allow commuting instructions with Imm operands.
-    if (NewMI || !Src1.isImm() || (!isVOP2(MI) && !isVOP3(MI))) {
-      return nullptr;
-    }
-    // Be sure to copy the source modifiers to the right place.
-    if (MachineOperand *Src0Mods =
-            getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)) {
-      MachineOperand *Src1Mods =
-          getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
-
-      int Src0ModsVal = Src0Mods->getImm();
-      if (!Src1Mods && Src0ModsVal != 0)
-        return nullptr;
-
-      // XXX - This assert might be a lie. It might be useful to have a neg
-      // modifier with 0.0.
-      int Src1ModsVal = Src1Mods->getImm();
-      assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
-
-      Src1Mods->setImm(Src0ModsVal);
-      Src0Mods->setImm(Src1ModsVal);
+  MachineInstr *CommutedMI = nullptr;
+  if (Src0.isReg() && Src1.isReg()) {
+    if (isOperandLegal(MI, Src1Idx, &Src0)) {
+      // Be sure to copy the source modifiers to the right place.
+      CommutedMI
+        = TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
     }
 
-    unsigned Reg = Src0.getReg();
-    unsigned SubReg = Src0.getSubReg();
-    if (Src1.isImm())
-      Src0.ChangeToImmediate(Src1.getImm());
-    else
-      llvm_unreachable("Should only have immediates");
-
-    Src1.ChangeToRegister(Reg, false);
-    Src1.setSubReg(SubReg);
+  } else if (Src0.isReg() && !Src1.isReg()) {
+    // src0 should always be able to support any operand type, so no need to
+    // check operand legality.
+    CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
+  } else if (!Src0.isReg() && Src1.isReg()) {
+    if (isOperandLegal(MI, Src1Idx, &Src0))
+      CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
   } else {
-    CommutedMI =
-        TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx0, OpIdx1);
+    // FIXME: Found two non registers to commute. This does happen.
+    return nullptr;
   }
 
-  if (CommutedMI)
+
+  if (CommutedMI) {
+    swapSourceModifiers(MI, Src0, AMDGPU::OpName::src0_modifiers,
+                        Src1, AMDGPU::OpName::src1_modifiers);
+
     CommutedMI->setDesc(get(CommutedOpcode));
+  }
 
   return CommutedMI;
 }
@@ -1002,40 +998,18 @@ MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
 // TargetInstrInfo::commuteInstruction uses it.
 bool SIInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx0,
                                         unsigned &SrcOpIdx1) const {
-  const MCInstrDesc &MCID = MI.getDesc();
-  if (!MCID.isCommutable())
+  if (!MI.isCommutable())
     return false;
 
   unsigned Opc = MI.getOpcode();
   int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
   if (Src0Idx == -1)
     return false;
 
-  // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
-  // immediate. Also, immediate src0 operand is not handled in
-  // SIInstrInfo::commuteInstruction();
-  if (!MI.getOperand(Src0Idx).isReg())
-    return false;
-
   int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
   if (Src1Idx == -1)
     return false;
 
-  MachineOperand &Src1 = MI.getOperand(Src1Idx);
-  if (Src1.isImm()) {
-    // SIInstrInfo::commuteInstruction() does support commuting the immediate
-    // operand src1 in 2 and 3 operand instructions.
-    if (!isVOP2(MI.getOpcode()) && !isVOP3(MI.getOpcode()))
-      return false;
-  } else if (Src1.isReg()) {
-    // If any source modifiers are set, the generic instruction commuting won't
-    // understand how to copy the source modifiers.
-    if (hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) ||
-        hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers))
-      return false;
-  } else
-    return false;
-
   return fixCommutedOpIndices(SrcOpIdx0, SrcOpIdx1, Src0Idx, Src1Idx);
 }
 

lib/Target/AMDGPU/SIInstrInfo.h

@@ -86,6 +86,10 @@ class SIInstrInfo final : public AMDGPUInstrInfo {
   unsigned findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const;
 
 protected:
+  bool swapSourceModifiers(MachineInstr &MI,
+                           MachineOperand &Src0, unsigned Src0OpName,
+                           MachineOperand &Src1, unsigned Src1OpName) const;
+
   MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
                                        unsigned OpIdx0,
                                        unsigned OpIdx1) const override;
@@ -144,7 +148,12 @@ class SIInstrInfo final : public AMDGPUInstrInfo {
   unsigned getMovOpcode(const TargetRegisterClass *DstRC) const;
 
   LLVM_READONLY
-  int commuteOpcode(const MachineInstr &MI) const;
+  int commuteOpcode(unsigned Opc) const;
+
+  LLVM_READONLY
+  inline int commuteOpcode(const MachineInstr &MI) const {
+    return commuteOpcode(MI.getOpcode());
+  }
 
   bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
                              unsigned &SrcOpIdx2) const override;

test/CodeGen/AMDGPU/addrspacecast.ll

@@ -11,7 +11,7 @@
 ; HSA-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]]
 ; HSA-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
 
-; HSA-DAG: v_cmp_ne_i32_e64 vcc, -1, [[PTR]]
+; HSA-DAG: v_cmp_ne_i32_e64 vcc, [[PTR]], -1
 ; HSA-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]]
 ; HSA-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
 ; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
@@ -34,7 +34,7 @@ define void @use_group_to_flat_addrspacecast(i32 addrspace(3)* %ptr) #0 {
 ; HSA-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]]
 ; HSA-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
 
-; HSA-DAG: v_cmp_ne_i32_e64 vcc, -1, [[PTR]]
+; HSA-DAG: v_cmp_ne_i32_e64 vcc, [[PTR]], -1
 ; HSA-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]]
 ; HSA-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
 ; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
@@ -79,7 +79,7 @@ define void @use_constant_to_flat_addrspacecast(i32 addrspace(2)* %ptr) #0 {
 ; HSA: enable_sgpr_queue_ptr = 0
 
 ; HSA: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}
-; HSA-DAG: v_cmp_ne_i64_e64 vcc, 0, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}
+; HSA-DAG: v_cmp_ne_i64_e64 vcc, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0{{$}}
 ; HSA-DAG: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s[[PTR_LO]]
 ; HSA-DAG: v_cndmask_b32_e32 [[CASTPTR:v[0-9]+]], -1, v[[VPTR_LO]]
 ; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 0{{$}}
@@ -96,7 +96,7 @@ define void @use_flat_to_group_addrspacecast(i32 addrspace(4)* %ptr) #0 {
 ; HSA: enable_sgpr_queue_ptr = 0
 
 ; HSA: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}
-; HSA-DAG: v_cmp_ne_i64_e64 vcc, 0, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}
+; HSA-DAG: v_cmp_ne_i64_e64 vcc, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0{{$}}
 ; HSA-DAG: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s[[PTR_LO]]
 ; HSA-DAG: v_cndmask_b32_e32 [[CASTPTR:v[0-9]+]], -1, v[[VPTR_LO]]
 ; HSA-DAG: v_mov_b32_e32 v[[K:[0-9]+]], 0{{$}}

test/CodeGen/AMDGPU/cndmask-no-def-vcc.ll

@@ -5,7 +5,7 @@ declare i1 @llvm.amdgcn.class.f32(float, i32)
 ; Produces error after adding an implicit def to v_cndmask_b32
 
 ; GCN-LABEL: {{^}}vcc_shrink_vcc_def:
-; GCN: v_cmp_eq_i32_e64 vcc, 0, s{{[0-9]+}}
+; GCN: v_cmp_eq_i32_e64 vcc, s{{[0-9]+}}, 0{{$}}
 ; GCN: v_cndmask_b32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}, vcc
 ; GCN: v_cndmask_b32_e64 v1, 0, 1, s{{\[[0-9]+:[0-9]+\]}}
 define void @vcc_shrink_vcc_def(float %arg, i32 %arg1, float %arg2, i32 %arg3) {

test/CodeGen/AMDGPU/commute_modifiers.ll

@@ -6,7 +6,7 @@ declare float @llvm.fma.f32(float, float, float) nounwind readnone
 
 ; FUNC-LABEL: @commute_add_imm_fabs_f32
 ; SI: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; SI: v_add_f32_e64 [[REG:v[0-9]+]], 2.0, |[[X]]|
+; SI: v_add_f32_e64 [[REG:v[0-9]+]], |[[X]]|, 2.0
 ; SI: buffer_store_dword [[REG]]
 define void @commute_add_imm_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1
@@ -20,7 +20,7 @@ define void @commute_add_imm_fabs_f32(float addrspace(1)* %out, float addrspace(
 
 ; FUNC-LABEL: @commute_mul_imm_fneg_fabs_f32
 ; SI: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; SI: v_mul_f32_e64 [[REG:v[0-9]+]], -4.0, |[[X]]|
+; SI: v_mul_f32_e64 [[REG:v[0-9]+]], |[[X]]|, -4.0
 ; SI: buffer_store_dword [[REG]]
 define void @commute_mul_imm_fneg_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1
@@ -51,7 +51,7 @@ define void @commute_mul_imm_fneg_f32(float addrspace(1)* %out, float addrspace(
 ; FUNC-LABEL: @commute_add_lit_fabs_f32
 ; SI: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; SI: v_mov_b32_e32 [[K:v[0-9]+]], 0x44800000
-; SI: v_add_f32_e64 [[REG:v[0-9]+]], |[[X]]|, [[K]]
+; SI: v_add_f32_e64 [[REG:v[0-9]+]], [[K]], |[[X]]|
 ; SI: buffer_store_dword [[REG]]
 define void @commute_add_lit_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1

test/CodeGen/AMDGPU/ctlz.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
 
@@ -19,7 +19,7 @@ declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 ; FUNC-LABEL: {{^}}s_ctlz_i32:
 ; SI: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
 ; SI-DAG: s_flbit_i32_b32 [[CTLZ:s[0-9]+]], [[VAL]]
-; SI-DAG: v_cmp_eq_i32_e64 [[CMPZ:s\[[0-9]+:[0-9]+\]]], 0, [[VAL]]
+; SI-DAG: v_cmp_eq_i32_e64 [[CMPZ:s\[[0-9]+:[0-9]+\]]], [[VAL]], 0{{$}}
 ; SI-DAG: v_mov_b32_e32 [[VCTLZ:v[0-9]+]], [[CTLZ]]
 ; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], [[VCTLZ]], 32, [[CMPZ]]
 ; SI: buffer_store_dword [[RESULT]]
@@ -112,7 +112,7 @@ define void @v_ctlz_i8(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %
 
 ; FUNC-LABEL: {{^}}s_ctlz_i64:
 ; SI: s_load_dwordx2 s{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
-; SI-DAG: v_cmp_eq_i32_e64 vcc, 0, s[[HI]]
+; SI-DAG: v_cmp_eq_i32_e64 vcc, s[[HI]], 0{{$}}
 ; SI-DAG: s_flbit_i32_b32 [[FFBH_LO:s[0-9]+]], s[[LO]]
 ; SI-DAG: s_add_i32 [[ADD:s[0-9]+]], [[FFBH_LO]], 32
 ; SI-DAG: s_flbit_i32_b32 [[FFBH_HI:s[0-9]+]], s[[HI]]

test/CodeGen/AMDGPU/ctlz_zero_undef.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
 
@@ -92,7 +92,7 @@ define void @v_ctlz_zero_undef_i8(i8 addrspace(1)* noalias %out, i8 addrspace(1)
 
 ; FUNC-LABEL: {{^}}s_ctlz_zero_undef_i64:
 ; SI: s_load_dwordx2 s{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
-; SI-DAG: v_cmp_eq_i32_e64 vcc, 0, s[[HI]]
+; SI-DAG: v_cmp_eq_i32_e64 vcc, s[[HI]], 0{{$}}
 ; SI-DAG: s_flbit_i32_b32 [[FFBH_LO:s[0-9]+]], s[[LO]]
 ; SI-DAG: s_add_i32 [[ADD:s[0-9]+]], [[FFBH_LO]], 32
 ; SI-DAG: s_flbit_i32_b32 [[FFBH_HI:s[0-9]+]], s[[HI]]

test/CodeGen/AMDGPU/cvt_flr_i32_f32.ll

@@ -1,5 +1,5 @@
-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=SI -enable-no-nans-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-NONAN -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -enable-no-nans-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-NONAN -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 
 declare float @llvm.fabs.f32(float) #1
@@ -18,7 +18,7 @@ define void @cvt_flr_i32_f32_0(i32 addrspace(1)* %out, float %x) #0 {
 }
 
 ; FUNC-LABEL: {{^}}cvt_flr_i32_f32_1:
-; SI: v_add_f32_e64 [[TMP:v[0-9]+]], 1.0, s{{[0-9]+}}
+; SI: v_add_f32_e64 [[TMP:v[0-9]+]], s{{[0-9]+}}, 1.0
 ; SI-SAFE-NOT: v_cvt_flr_i32_f32
 ; SI-NONAN: v_cvt_flr_i32_f32_e32 v{{[0-9]+}}, [[TMP]]
 ; SI: s_endpgm

test/CodeGen/AMDGPU/fabs.f64.ll

@@ -55,7 +55,7 @@ define void @fabs_v4f64(<4 x double> addrspace(1)* %out, <4 x double> %in) {
 ; SI-LABEL: {{^}}fabs_fold_f64:
 ; SI: s_load_dwordx2 [[ABS_VALUE:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0xb
 ; SI-NOT: and
-; SI: v_mul_f64 {{v\[[0-9]+:[0-9]+\]}}, |[[ABS_VALUE]]|, {{v\[[0-9]+:[0-9]+\]}}
+; SI: v_mul_f64 {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}, |[[ABS_VALUE]]|
 ; SI: s_endpgm
 define void @fabs_fold_f64(double addrspace(1)* %out, double %in0, double %in1) {
   %fabs = call double @llvm.fabs.f64(double %in0)
@@ -67,7 +67,7 @@ define void @fabs_fold_f64(double addrspace(1)* %out, double %in0, double %in1)
 ; SI-LABEL: {{^}}fabs_fn_fold_f64:
 ; SI: s_load_dwordx2 [[ABS_VALUE:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0xb
 ; SI-NOT: and
-; SI: v_mul_f64 {{v\[[0-9]+:[0-9]+\]}}, |[[ABS_VALUE]]|, {{v\[[0-9]+:[0-9]+\]}}
+; SI: v_mul_f64 {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}, |[[ABS_VALUE]]|
 ; SI: s_endpgm
 define void @fabs_fn_fold_f64(double addrspace(1)* %out, double %in0, double %in1) {
   %fabs = call double @fabs(double %in0)