AMDGPU/SI: Use v_readfirstlane to legalize SMRD with VGPR base pointer

[• tstellarAMD]

Summary

Instead of trying to replace SMRD instructions with a VGPR base pointer
with an equivalent MUBUF instruction, we now copy the base pointer to
SGPRs using v_readfirstlane.

This is safe to do, because any load selected as an SMRD instruction
has been proven to have a uniform base pointer, so each thread in the
wave will have the same pointer value in VGPRs.

This will fix some errors on VI from trying to replace SMRD instructions
with addr64-enabled MUBUF instructions that don't exist.

[arsenm]

LGTM

[cfang]

A good approach. Can we use v_readfirstlane for some other cases in SIFixSGPRCopies? There might be optimization choice which
approaches are better.

[nhaehnle]

This does indeed fix the problem in Mesa that we talked about, thank you.

[arsenm]

There should probably be a comment about why this is OK because SMRD is only originally selected when the pointer is known uniform

lib/Target/AMDGPU/SIInstrInfo.h

public:
 
   /// Copy a value from a VGPR (\p SrcReg) to SGPR.  This function can only
   /// be used when it is know that the value in SrcReg is same across all
   /// threads in the wave.
   /// \returns The SGPR register that \p SrcReg was copied to.
   unsigned readlaneVGPRToSGPR(unsigned SrcReg, MachineInstr *UseMI,
                           MachineRegisterInfo &MRI) const;
 
+  void legalizeOperandsSMRD(MachineRegisterInfo &MRI, MachineInstr *MI) const;
+
   /// \brief Legalize all operands in this instruction.  This function may
   /// create new instruction and insert them before \p MI.
   void legalizeOperands(MachineInstr *MI) const;
 
-  /// \brief Split an SMRD instruction into two smaller loads of half the
-  //  size storing the results in \p Lo and \p Hi.
-  void splitSMRD(MachineInstr *MI, const TargetRegisterClass *HalfRC,
-                 unsigned HalfImmOp, unsigned HalfSGPROp,
-                 MachineInstr *&Lo, MachineInstr *&Hi) const;
-
-  void moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI,
-                      SmallVectorImpl<MachineInstr *> &Worklist) const;
-
   /// \brief Replace this instruction's opcode with the equivalent VALU
   /// opcode.  This function will also move the users of \p MI to the
   /// VALU if necessary.
   void moveToVALU(MachineInstr &MI) const;
 
   const TargetRegisterClass *getIndirectAddrRegClass() const override;
 
   void reserveIndirectRegisters(BitVector &Reserved,

lib/Target/AMDGPU/SIInstrInfo.cpp

unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
   case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
   case AMDGPU::S_CMP_EQ_U32: return AMDGPU::V_CMP_EQ_U32_e32;
   case AMDGPU::S_CMP_LG_U32: return AMDGPU::V_CMP_NE_U32_e32;
   case AMDGPU::S_CMP_GT_U32: return AMDGPU::V_CMP_GT_U32_e32;
   case AMDGPU::S_CMP_GE_U32: return AMDGPU::V_CMP_GE_U32_e32;
   case AMDGPU::S_CMP_LT_U32: return AMDGPU::V_CMP_LT_U32_e32;
   case AMDGPU::S_CMP_LE_U32: return AMDGPU::V_CMP_LE_U32_e32;
-  case AMDGPU::S_LOAD_DWORD_IMM:
-  case AMDGPU::S_LOAD_DWORD_SGPR:
-  case AMDGPU::S_LOAD_DWORD_IMM_ci:
-    return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
-  case AMDGPU::S_LOAD_DWORDX2_IMM:
-  case AMDGPU::S_LOAD_DWORDX2_SGPR:
-  case AMDGPU::S_LOAD_DWORDX2_IMM_ci:
-    return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
-  case AMDGPU::S_LOAD_DWORDX4_IMM:
-  case AMDGPU::S_LOAD_DWORDX4_SGPR:
-  case AMDGPU::S_LOAD_DWORDX4_IMM_ci:
-    return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
   case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
   case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
   case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
   case AMDGPU::S_FLBIT_I32: return AMDGPU::V_FFBH_I32_e64;
   case AMDGPU::S_CBRANCH_SCC0: return AMDGPU::S_CBRANCH_VCCZ;
   case AMDGPU::S_CBRANCH_SCC1: return AMDGPU::S_CBRANCH_VCCNZ;
   }
 }
MachineInstrBuilder MIB = BuildMI(*UseMI->getParent(), UseMI,
                                     get(AMDGPU::REG_SEQUENCE), DstReg);
   for (unsigned i = 0; i < SubRegs; ++i) {
     MIB.addReg(SRegs[i]);
     MIB.addImm(RI.getSubRegFromChannel(i));
   }
   return DstReg;
 }
 
+void SIInstrInfo::legalizeOperandsSMRD(MachineRegisterInfo &MRI,
+                                       MachineInstr *MI) const {
+  MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
+  if (SBase && !RI.isSGPRClass(MRI.getRegClass(SBase->getReg()))) {
+      unsigned SGPR = readlaneVGPRToSGPR(SBase->getReg(), MI, MRI);
+      SBase->setReg(SGPR);
+  }
+}
+
 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
   MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
 
   // Legalize VOP2
   if (isVOP2(*MI) || isVOPC(*MI)) {
     legalizeOperandsVOP2(MRI, MI);
     return;
   }
 
   // Legalize VOP3
   if (isVOP3(*MI)) {
     legalizeOperandsVOP3(MRI, MI);
     return;
   }
 
+  // Legalize SMRD
+  if (isSMRD(*MI)) {
+    legalizeOperandsSMRD(MRI, MI);
+    return;
+  }
+
   // Legalize REG_SEQUENCE and PHI
   // The register class of the operands much be the same type as the register
   // class of the output.
   if (MI->getOpcode() == AMDGPU::PHI) {
     const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
     for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
       if (!MI->getOperand(i).isReg() ||
           !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
if (SRsrcIdx != -1) {
 
     // Update the instruction to use NewVaddr
     VAddr->setReg(NewVAddr);
     // Update the instruction to use NewSRsrc
     SRsrc->setReg(NewSRsrc);
   }
 }
 
-void SIInstrInfo::splitSMRD(MachineInstr *MI,
-                            const TargetRegisterClass *HalfRC,
-                            unsigned HalfImmOp, unsigned HalfSGPROp,
-                            MachineInstr *&Lo, MachineInstr *&Hi) const {
-
-  DebugLoc DL = MI->getDebugLoc();
-  MachineBasicBlock *MBB = MI->getParent();
-  MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
-  unsigned RegLo = MRI.createVirtualRegister(HalfRC);
-  unsigned RegHi = MRI.createVirtualRegister(HalfRC);
-  unsigned HalfSize = HalfRC->getSize();
-  const MachineOperand *OffOp =
-      getNamedOperand(*MI, AMDGPU::OpName::offset);
-  const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
-
-  // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
-  // on VI.
-
-  bool IsKill = SBase->isKill();
-  if (OffOp) {
-    bool isVI =
-        MBB->getParent()->getSubtarget<AMDGPUSubtarget>().getGeneration() >=
-        AMDGPUSubtarget::VOLCANIC_ISLANDS;
-    unsigned OffScale = isVI ? 1 : 4;
-    // Handle the _IMM variant
-    unsigned LoOffset = OffOp->getImm() * OffScale;
-    unsigned HiOffset = LoOffset + HalfSize;
-    Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
-                  // Use addReg instead of addOperand
-                  // to make sure kill flag is cleared.
-                  .addReg(SBase->getReg(), 0, SBase->getSubReg())
-                  .addImm(LoOffset / OffScale);
-
-    if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
-      unsigned OffsetSGPR =
-          MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
-      BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
-              .addImm(HiOffset); // The offset in register is in bytes.
-      Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
-                    .addReg(SBase->getReg(), getKillRegState(IsKill),
-                            SBase->getSubReg())
-                    .addReg(OffsetSGPR);
-    } else {
-      Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
-                     .addReg(SBase->getReg(), getKillRegState(IsKill),
-                             SBase->getSubReg())
-                     .addImm(HiOffset / OffScale);
-    }
-  } else {
-    // Handle the _SGPR variant
-    MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
-    Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
-                  .addReg(SBase->getReg(), 0, SBase->getSubReg())
-                  .addOperand(*SOff);
-    unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
-    BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
-      .addReg(SOff->getReg(), 0, SOff->getSubReg())
-      .addImm(HalfSize);
-    Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
-                  .addReg(SBase->getReg(), getKillRegState(IsKill),
-                          SBase->getSubReg())
-                  .addReg(OffsetSGPR);
-  }
-
-  unsigned SubLo, SubHi;
-  const TargetRegisterClass *NewDstRC;
-  switch (HalfSize) {
-    case 4:
-      SubLo = AMDGPU::sub0;
-      SubHi = AMDGPU::sub1;
-      NewDstRC = &AMDGPU::VReg_64RegClass;
-      break;
-    case 8:
-      SubLo = AMDGPU::sub0_sub1;
-      SubHi = AMDGPU::sub2_sub3;
-      NewDstRC = &AMDGPU::VReg_128RegClass;
-      break;
-    case 16:
-      SubLo = AMDGPU::sub0_sub1_sub2_sub3;
-      SubHi = AMDGPU::sub4_sub5_sub6_sub7;
-      NewDstRC = &AMDGPU::VReg_256RegClass;
-      break;
-    case 32:
-      SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
-      SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
-      NewDstRC = &AMDGPU::VReg_512RegClass;
-      break;
-    default:
-      llvm_unreachable("Unhandled HalfSize");
-  }
-
-  unsigned OldDst = MI->getOperand(0).getReg();
-  unsigned NewDst = MRI.createVirtualRegister(NewDstRC);
-
-  MRI.replaceRegWith(OldDst, NewDst);
-
-  BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE), NewDst)
-    .addReg(RegLo)
-    .addImm(SubLo)
-    .addReg(RegHi)
-    .addImm(SubHi);
-}
-
-void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI,
-                                 MachineRegisterInfo &MRI,
-                                 SmallVectorImpl<MachineInstr *> &Worklist) const {
-  MachineBasicBlock *MBB = MI->getParent();
-  int DstIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
-  assert(DstIdx != -1);
-  unsigned DstRCID = get(MI->getOpcode()).OpInfo[DstIdx].RegClass;
-  switch(RI.getRegClass(DstRCID)->getSize()) {
-    case 4:
-    case 8:
-    case 16: {
-      unsigned NewOpcode = getVALUOp(*MI);
-      unsigned RegOffset;
-      unsigned ImmOffset;
-
-      if (MI->getOperand(2).isReg()) {
-        RegOffset = MI->getOperand(2).getReg();
-        ImmOffset = 0;
-      } else {
-        assert(MI->getOperand(2).isImm());
-        // SMRD instructions take a dword offsets on SI and byte offset on VI
-        // and MUBUF instructions always take a byte offset.
-        ImmOffset = MI->getOperand(2).getImm();
-        if (MBB->getParent()->getSubtarget<AMDGPUSubtarget>().getGeneration() <=
-            AMDGPUSubtarget::SEA_ISLANDS)
-          ImmOffset <<= 2;
-        RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
-
-        if (isUInt<12>(ImmOffset)) {
-          BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
-                  RegOffset)
-                  .addImm(0);
-        } else {
-          BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
-                  RegOffset)
-                  .addImm(ImmOffset);
-          ImmOffset = 0;
-        }
-      }
-
-      unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
-      unsigned DWord0 = RegOffset;
-      unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
-      unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
-      unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
-      uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
-
-      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
-              .addImm(0);
-      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
-              .addImm(RsrcDataFormat & 0xFFFFFFFF);
-      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
-              .addImm(RsrcDataFormat >> 32);
-      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
-        .addReg(DWord0)
-        .addImm(AMDGPU::sub0)
-        .addReg(DWord1)
-        .addImm(AMDGPU::sub1)
-        .addReg(DWord2)
-        .addImm(AMDGPU::sub2)
-        .addReg(DWord3)
-        .addImm(AMDGPU::sub3);
-
-      const MCInstrDesc &NewInstDesc = get(NewOpcode);
-      const TargetRegisterClass *NewDstRC
-        = RI.getRegClass(NewInstDesc.OpInfo[0].RegClass);
-      unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
-      unsigned DstReg = MI->getOperand(0).getReg();
-      MRI.replaceRegWith(DstReg, NewDstReg);
-
-      MachineInstr *NewInst =
-        BuildMI(*MBB, MI, MI->getDebugLoc(), NewInstDesc, NewDstReg)
-        .addOperand(MI->getOperand(1)) // sbase
-        .addReg(SRsrc)
-        .addImm(0)
-        .addImm(ImmOffset)
-        .addImm(0) // glc
-        .addImm(0) // slc
-        .addImm(0) // tfe
-        .setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
-      MI->eraseFromParent();
-
-      legalizeOperands(NewInst);
-      addUsersToMoveToVALUWorklist(NewDstReg, MRI, Worklist);
-      break;
-    }
-    case 32: {
-      MachineInstr *Lo, *Hi;
-      addUsersToMoveToVALUWorklist(MI->getOperand(0).getReg(), MRI, Worklist);
-      splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
-                AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
-      MI->eraseFromParent();
-      moveSMRDToVALU(Lo, MRI, Worklist);
-      moveSMRDToVALU(Hi, MRI, Worklist);
-      break;
-    }
-
-    case 64: {
-      MachineInstr *Lo, *Hi;
-      addUsersToMoveToVALUWorklist(MI->getOperand(0).getReg(), MRI, Worklist);
-      splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
-                AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
-      MI->eraseFromParent();
-      moveSMRDToVALU(Lo, MRI, Worklist);
-      moveSMRDToVALU(Hi, MRI, Worklist);
-      break;
-    }
-  }
-}
-
 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
   SmallVector<MachineInstr *, 128> Worklist;
   Worklist.push_back(&TopInst);
 
   while (!Worklist.empty()) {
     MachineInstr *Inst = Worklist.pop_back_val();
     MachineBasicBlock *MBB = Inst->getParent();
     MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
 
     unsigned Opcode = Inst->getOpcode();
     unsigned NewOpcode = getVALUOp(*Inst);
 
     // Handle some special cases
     switch (Opcode) {
     default:
-      if (isSMRD(*Inst)) {
-        moveSMRDToVALU(Inst, MRI, Worklist);
-        continue;
-      }
       break;
     case AMDGPU::S_AND_B64:
       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_AND_B32_e64);
       Inst->eraseFromParent();
       continue;
 
     case AMDGPU::S_OR_B64:
       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_OR_B32_e64);

test/CodeGen/AMDGPU/missing-store.ll

 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=FUNC -check-prefix=SI %s
 
 @ptr_load = addrspace(3) global i32 addrspace(2)* undef, align 8
 
 ; Make sure when the load from %ptr2 is folded the chain isn't lost,
 ; resulting in losing the store to gptr
 
 ; FUNC-LABEL: {{^}}missing_store_reduced:
 ; SI: ds_read_b64
 ; SI: buffer_store_dword
-; SI: buffer_load_dword
+; SI: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}}
+; SI: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}}
+; SI: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}
 ; SI: buffer_store_dword
 ; SI: s_endpgm
 define void @missing_store_reduced(i32 addrspace(1)* %out, i32 addrspace(1)* %gptr) #0 {
   %ptr0 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(3)* @ptr_load, align 8
   %ptr2 = getelementptr inbounds i32, i32 addrspace(2)* %ptr0, i64 2
 
   store i32 99, i32 addrspace(1)* %gptr, align 4
   %tmp2 = load i32, i32 addrspace(2)* %ptr2, align 4
 
   store i32 %tmp2, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 attributes #0 = { nounwind }
 

test/CodeGen/AMDGPU/salu-to-valu.ll

 
 done:                                             ; preds = %loop
   ret void
 }
 
 ; Test moving an SMRD instruction to the VALU
 
 ; GCN-LABEL: {{^}}smrd_valu:
-; FIXME: We should be using flat load for HSA.
-; GCN: buffer_load_dword [[OUT:v[0-9]+]]
-; GCN-NOHSA: buffer_store_dword [[OUT]]
-; GCN-HSA: flat_store_dword {{.*}}, [[OUT]]
+; SI: s_movk_i32 [[OFFSET:s[0-9]+]], 0x2ee0
+; GCN: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}}
+; GCN: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}}
+; SI: s_load_dword [[OUT:s[0-9]+]], s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, [[OFFSET]]
+; CI: s_load_dword [[OUT:s[0-9]+]], s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0xbb8
+; GCN: v_mov_b32_e32 [[V_OUT:v[0-9]+]], [[OUT]]
+; GCN-NOHSA: buffer_store_dword [[V_OUT]]
+; GCN-HSA: flat_store_dword {{.*}}, [[V_OUT]]
 define void @smrd_valu(i32 addrspace(2)* addrspace(1)* %in, i32 %a, i32 %b, i32 addrspace(1)* %out) #1 {
 entry:
   %tmp = icmp ne i32 %a, 0
   br i1 %tmp, label %if, label %else
 
 if:                                               ; preds = %entry
   %tmp1 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(1)* %in
   br label %endif

test/CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll

define void @no_reorder_barrier_local_load_global_store_local_load(i32 addrspace(1)* %out, i32 addrspace(1)* %gptr) #0 {
   %tmp2 = load i32, i32 addrspace(3)* %ptr2, align 4
 
   %add = add nsw i32 %tmp1, %tmp2
 
   store i32 %add, i32 addrspace(1)* %out, align 4
   ret void
 }
 
-; Technically we could reorder these, but just comparing the
-; instruction type of the load is insufficient.
-
-; FUNC-LABEL: @no_reorder_constant_load_global_store_constant_load
-; CI: buffer_load_dword
+; FUNC-LABEL: @reorder_constant_load_global_store_constant_load
 ; CI: buffer_store_dword
-; CI: buffer_load_dword
+; CI: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}}
+; CI: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}}
+; CI-DAG: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x1
+; CI-DAG: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x2
 ; CI: buffer_store_dword
-define void @no_reorder_constant_load_global_store_constant_load(i32 addrspace(1)* %out, i32 addrspace(1)* %gptr) #0 {
+define void @reorder_constant_load_global_store_constant_load(i32 addrspace(1)* %out, i32 addrspace(1)* %gptr) #0 {
   %ptr0 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(3)* @stored_constant_ptr, align 8
 
   %ptr1 = getelementptr inbounds i32, i32 addrspace(2)* %ptr0, i64 1
   %ptr2 = getelementptr inbounds i32, i32 addrspace(2)* %ptr0, i64 2
 
   %tmp1 = load i32, i32 addrspace(2)* %ptr1, align 4
   store i32 99, i32 addrspace(1)* %gptr, align 4
   %tmp2 = load i32, i32 addrspace(2)* %ptr2, align 4
 
   %add = add nsw i32 %tmp1, %tmp2
 
   store i32 %add, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; FUNC-LABEL: @reorder_constant_load_local_store_constant_load
-; CI: buffer_load_dword
-; CI: buffer_load_dword
+; CI: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}}
+; CI: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}}
+; CI-DAG: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x1
+; CI-DAG: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x2
 ; CI: ds_write_b32
 ; CI: buffer_store_dword
 define void @reorder_constant_load_local_store_constant_load(i32 addrspace(1)* %out, i32 addrspace(3)* %lptr) #0 {
   %ptr0 = load i32 addrspace(2)*, i32 addrspace(2)* addrspace(3)* @stored_constant_ptr, align 8
 
   %ptr1 = getelementptr inbounds i32, i32 addrspace(2)* %ptr0, i64 1
   %ptr2 = getelementptr inbounds i32, i32 addrspace(2)* %ptr0, i64 2