[GlobalISel][X86] Support G_STORE/G_LOAD operation

[igorb]

Summary

1. Support pointer type as function argumnet and return value
2. G_STORE/G_LOAD - set legal action for i8/i16/i32/i64/f32/f64/vec128
3. RegisterBank - support typeless operations like G_STORE/G_LOAD, for scalar use GPR bank.
4. Support instruction selection for G_LOAD/G_STORE

[qcolombet]

Hi Igor,

For G_LOAD, Fast and Greedy mode choose the same RegisterBank mapping (GprRegBank ) for the G_GLOAD + G_FADD , can't get rid of cross register bank copy GprRegBank->VecRegBank.

Yeah, we are aware of this problem. Greedy is not smart enough right now to do the right thing. I hope to fix that in the coming weeks.

I haven't looked at the patch, but is it something you worked around here?

Cheers,
-Quentin

[igorb]

In D30973#701667, @qcolombet wrote:

I haven't looked at the patch, but is it something you worked around here?

Cheers,
-Quentin

Hi Quentin,
No , It is straightforward implementation for G_LOAD/G_STORE operation.

[rovka]

Hi,

Could you commit the unrelated changes separately (e.g. the space changes and the include reordering)?
Other than that, LGTM.

lib/Target/X86/X86InstructionSelector.cpp

321 ↗

(On Diff #91846)

Why not else if, like the others?

test/CodeGen/X86/GlobalISel/x86_64-instructionselect.mir

659 ↗

(On Diff #91846)

Can you also add testcases for s32 and s64 living in the vecr?

[igorb]

Thanks for the review!

llvm/trunk/lib/Target/X86/X86CallLowering.cpp

void X86CallLowering::splitToValueTypes(const ArgInfo &OrigArg,
                                         SplitArgTy PerformArgSplit) const {
 
   const X86TargetLowering &TLI = *getTLI<X86TargetLowering>();
   LLVMContext &Context = OrigArg.Ty->getContext();
   EVT VT = TLI.getValueType(DL, OrigArg.Ty);
   unsigned NumParts = TLI.getNumRegisters(Context, VT);
 
   if (NumParts == 1) {
-    SplitArgs.push_back(OrigArg);
+    // replace the original type ( pointer -> GPR ).
+    SplitArgs.emplace_back(OrigArg.Reg, VT.getTypeForEVT(Context),
+                           OrigArg.Flags, OrigArg.IsFixed);
     return;
   }
 
   SmallVector<uint64_t, 4> BitOffsets;
   SmallVector<unsigned, 8> SplitRegs;
 
   EVT PartVT = TLI.getRegisterType(Context, VT);
   Type *PartTy = PartVT.getTypeForEVT(Context);

llvm/trunk/lib/Target/X86/X86GenRegisterBankInfo.def

 /// This file defines all the static objects used by X86RegisterBankInfo.
 /// \todo This should be generated by TableGen.
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_BUILD_GLOBAL_ISEL
 #error "You shouldn't build this"
 #endif
 
-namespace llvm {
+#ifdef GET_TARGET_REGBANK_INFO_IMPL
 RegisterBankInfo::PartialMapping X86GenRegisterBankInfo::PartMappings[]{
     /* StartIdx, Length, RegBank */
     // GPR value
     {0, 8, X86::GPRRegBank},   // :0
     {0, 16, X86::GPRRegBank},  // :1
     {0, 32, X86::GPRRegBank},  // :2
     {0, 64, X86::GPRRegBank},  // :3
     // FR32/64 , xmm registers
     {0, 32, X86::VECRRegBank},  // :4
     {0, 64, X86::VECRRegBank},  // :5
     // VR128/256/512
     {0, 128, X86::VECRRegBank}, // :6
     {0, 256, X86::VECRRegBank}, // :7
     {0, 512, X86::VECRRegBank}, // :8   
 };
+#endif // GET_TARGET_REGBANK_INFO_IMPL
 
+#ifdef GET_TARGET_REGBANK_INFO_CLASS
 enum PartialMappingIdx {
   PMI_None = -1,
   PMI_GPR8,
   PMI_GPR16,
   PMI_GPR32,
   PMI_GPR64,
   PMI_FP32,
   PMI_FP64,
   PMI_VEC128,
   PMI_VEC256,
   PMI_VEC512
 };
+#endif // GET_TARGET_REGBANK_INFO_CLASS
 
+#ifdef GET_TARGET_REGBANK_INFO_IMPL
 #define INSTR_3OP(INFO) INFO, INFO, INFO,
 #define BREAKDOWN(INDEX, NUM)                                                  \
   { &X86GenRegisterBankInfo::PartMappings[INDEX], NUM }
 // ValueMappings.
 RegisterBankInfo::ValueMapping X86GenRegisterBankInfo::ValMappings[]{
     /* BreakDown, NumBreakDowns */
     // 3-operands instructions (all binary operations should end up with one of
     // those mapping).
     INSTR_3OP(BREAKDOWN(PMI_GPR8, 1))  // 0: GPR_8
     INSTR_3OP(BREAKDOWN(PMI_GPR16, 1)) // 3: GPR_16
     INSTR_3OP(BREAKDOWN(PMI_GPR32, 1)) // 6: GPR_32
     INSTR_3OP(BREAKDOWN(PMI_GPR64, 1)) // 9: GPR_64    
     INSTR_3OP(BREAKDOWN(PMI_FP32, 1))   // 12: Fp32
     INSTR_3OP(BREAKDOWN(PMI_FP64, 1))   // 15: Fp64
     INSTR_3OP(BREAKDOWN(PMI_VEC128, 1)) // 18: Vec128
     INSTR_3OP(BREAKDOWN(PMI_VEC256, 1)) // 21: Vec256
     INSTR_3OP(BREAKDOWN(PMI_VEC512, 1)) // 24: Vec512    
 };
 #undef INSTR_3OP
 #undef BREAKDOWN
+#endif // GET_TARGET_REGBANK_INFO_IMPL
 
+#ifdef GET_TARGET_REGBANK_INFO_CLASS
 enum ValueMappingIdx {
   VMI_None = -1,
   VMI_3OpsGpr8Idx =  PMI_GPR8  * 3,
   VMI_3OpsGpr16Idx = PMI_GPR16 * 3,
   VMI_3OpsGpr32Idx = PMI_GPR32 * 3,
   VMI_3OpsGpr64Idx = PMI_GPR64 * 3,  
   VMI_3OpsFp32Idx = PMI_FP32 * 3,
   VMI_3OpsFp64Idx = PMI_FP64 * 3,
   VMI_3OpsVec128Idx = PMI_VEC128 * 3,
   VMI_3OpsVec256Idx = PMI_VEC256 * 3,
   VMI_3OpsVec512Idx = PMI_VEC512 * 3,
 };
+#undef GET_TARGET_REGBANK_INFO_CLASS
+#endif // GET_TARGET_REGBANK_INFO_CLASS
+
+#ifdef GET_TARGET_REGBANK_INFO_IMPL
+#undef GET_TARGET_REGBANK_INFO_IMPL
+const RegisterBankInfo::ValueMapping *
+X86GenRegisterBankInfo::getValueMapping(PartialMappingIdx Idx,
+                                        unsigned NumOperands) {
+  
+  // We can use VMI_3Ops Mapping for all the cases.
+  if (NumOperands <= 3 && (Idx >= PMI_GPR8 && Idx <= PMI_VEC512))
+    return &ValMappings[(unsigned)Idx * 3];
+  
+  llvm_unreachable("Unsupported PartialMappingIdx.");
+}
+
+#endif // GET_TARGET_REGBANK_INFO_IMPL
 
-} // End llvm namespace.

llvm/trunk/lib/Target/X86/X86InstructionSelector.h

 class X86InstrInfo;
 class X86RegisterBankInfo;
 class X86RegisterInfo;
 class X86Subtarget;
 class X86TargetMachine;
 class LLT;
 class RegisterBank;
 class MachineRegisterInfo;
+class MachineFunction;
 
 class X86InstructionSelector : public InstructionSelector {
 public:
   X86InstructionSelector(const X86Subtarget &STI,
                          const X86RegisterBankInfo &RBI);
 
   bool select(MachineInstr &I) const override;
 
 private:
   /// tblgen-erated 'select' implementation, used as the initial selector for
   /// the patterns that don't require complex C++.
   bool selectImpl(MachineInstr &I) const;
 
   // TODO: remove after selectImpl support pattern with a predicate.
   unsigned getFAddOp(LLT &Ty, const RegisterBank &RB) const;
   unsigned getFSubOp(LLT &Ty, const RegisterBank &RB) const;
   unsigned getAddOp(LLT &Ty, const RegisterBank &RB) const;
   unsigned getSubOp(LLT &Ty, const RegisterBank &RB) const;
-  bool selectBinaryOp(MachineInstr &I, MachineRegisterInfo &MRI) const;
+  unsigned getLoadStoreOp(LLT &Ty, const RegisterBank &RB, unsigned Opc,
+                          uint64_t Alignment) const;
+
+  bool selectBinaryOp(MachineInstr &I, MachineRegisterInfo &MRI,
+                      MachineFunction &MF) const;
+  bool selectLoadStoreOp(MachineInstr &I, MachineRegisterInfo &MRI,
+                         MachineFunction &MF) const;
 
   const X86Subtarget &STI;
   const X86InstrInfo &TII;
   const X86RegisterInfo &TRI;
   const X86RegisterBankInfo &RBI;
 
 #define GET_GLOBALISEL_TEMPORARIES_DECL
 #include "X86GenGlobalISel.inc"
 #undef GET_GLOBALISEL_TEMPORARIES_DECL
 };
 
 } // end namespace llvm
 
 #endif // LLVM_LIB_TARGET_X86_X86INSTRUCTIONSELECTOR_H
 #endif // LLVM_BUILD_GLOBAL_ISEL

llvm/trunk/lib/Target/X86/X86InstructionSelector.cpp

 //===- X86InstructionSelector.cpp ----------------------------*- C++ -*-==//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 /// \file
 /// This file implements the targeting of the InstructionSelector class for
 /// X86.
 /// \todo This should be generated by TableGen.
 //===----------------------------------------------------------------------===//
 
 #include "X86InstructionSelector.h"
+#include "X86InstrBuilder.h"
 #include "X86InstrInfo.h"
 #include "X86RegisterBankInfo.h"
 #include "X86RegisterInfo.h"
 #include "X86Subtarget.h"
 #include "X86TargetMachine.h"
 #include "llvm/CodeGen/MachineBasicBlock.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstr.h"
if (!isPreISelGenericOpcode(Opcode)) {
     return true;
   }
 
   assert(I.getNumOperands() == I.getNumExplicitOperands() &&
          "Generic instruction has unexpected implicit operands\n");
 
   // TODO: This should be implemented by tblgen, pattern with predicate not
   // supported yet.
-  if (selectBinaryOp(I, MRI))
+  if (selectBinaryOp(I, MRI, MF))
+    return true;
+  if (selectLoadStoreOp(I, MRI, MF))
     return true;
 
   return selectImpl(I);
 }
 
 unsigned X86InstructionSelector::getFAddOp(LLT &Ty,
                                            const RegisterBank &RB) const {
 
if (STI.hasAVX512() && STI.hasVLX()) {
       return X86::PSUBDrr;
     }
   }
 
   return TargetOpcode::G_SUB;
 }
 
 bool X86InstructionSelector::selectBinaryOp(MachineInstr &I,
-                                            MachineRegisterInfo &MRI) const {
+                                            MachineRegisterInfo &MRI,
+                                            MachineFunction &MF) const {
 
-  LLT Ty = MRI.getType(I.getOperand(0).getReg());
   const unsigned DefReg = I.getOperand(0).getReg();
+  LLT Ty = MRI.getType(DefReg);
   const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI);
 
   unsigned NewOpc = I.getOpcode();
 
-  switch (I.getOpcode()) {
+  switch (NewOpc) {
   case TargetOpcode::G_FADD:
     NewOpc = getFAddOp(Ty, RB);
     break;
   case TargetOpcode::G_FSUB:
     NewOpc = getFSubOp(Ty, RB);
     break;
   case TargetOpcode::G_ADD:
     NewOpc = getAddOp(Ty, RB);
     break;
   case TargetOpcode::G_SUB:
     NewOpc = getSubOp(Ty, RB);
     break;
   default:
     break;
   }
 
   if (NewOpc == I.getOpcode())
     return false;
 
   I.setDesc(TII.get(NewOpc));
 
   return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
 }
+
+unsigned X86InstructionSelector::getLoadStoreOp(LLT &Ty, const RegisterBank &RB,
+                                                unsigned Opc,
+                                                uint64_t Alignment) const {
+  bool Isload = (Opc == TargetOpcode::G_LOAD);
+  bool HasAVX = STI.hasAVX();
+  bool HasAVX512 = STI.hasAVX512();
+  bool HasVLX = STI.hasVLX();
+
+  if (Ty == LLT::scalar(8)) {
+    if (X86::GPRRegBankID == RB.getID())
+      return Isload ? X86::MOV8rm : X86::MOV8mr;
+  } else if (Ty == LLT::scalar(16)) {
+    if (X86::GPRRegBankID == RB.getID())
+      return Isload ? X86::MOV16rm : X86::MOV16mr;
+  } else if (Ty == LLT::scalar(32)) {
+    if (X86::GPRRegBankID == RB.getID())
+      return Isload ? X86::MOV32rm : X86::MOV32mr;
+    if (X86::VECRRegBankID == RB.getID())
+      return Isload ? (HasAVX512 ? X86::VMOVSSZrm
+                                 : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm)
+                    : (HasAVX512 ? X86::VMOVSSZmr
+                                 : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
+  } else if (Ty == LLT::scalar(64)) {
+    if (X86::GPRRegBankID == RB.getID())
+      return Isload ? X86::MOV64rm : X86::MOV64mr;
+    if (X86::VECRRegBankID == RB.getID())
+      return Isload ? (HasAVX512 ? X86::VMOVSDZrm
+                                 : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm)
+                    : (HasAVX512 ? X86::VMOVSDZmr
+                                 : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
+  } else if (Ty.isVector() && Ty.getSizeInBits() == 128) {
+    if (Alignment >= 16)
+      return Isload ? (HasVLX ? X86::VMOVAPSZ128rm
+                              : HasAVX512
+                                    ? X86::VMOVAPSZ128rm_NOVLX
+                                    : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm)
+                    : (HasVLX ? X86::VMOVAPSZ128mr
+                              : HasAVX512
+                                    ? X86::VMOVAPSZ128mr_NOVLX
+                                    : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
+    else
+      return Isload ? (HasVLX ? X86::VMOVUPSZ128rm
+                              : HasAVX512
+                                    ? X86::VMOVUPSZ128rm_NOVLX
+                                    : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm)
+                    : (HasVLX ? X86::VMOVUPSZ128mr
+                              : HasAVX512
+                                    ? X86::VMOVUPSZ128mr_NOVLX
+                                    : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
+  }
+  return Opc;
+}
+
+bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I,
+                                               MachineRegisterInfo &MRI,
+                                               MachineFunction &MF) const {
+
+  unsigned Opc = I.getOpcode();
+
+  if (Opc != TargetOpcode::G_STORE && Opc != TargetOpcode::G_LOAD)
+    return false;
+
+  const unsigned DefReg = I.getOperand(0).getReg();
+  LLT Ty = MRI.getType(DefReg);
+  const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI);
+
+  auto &MemOp = **I.memoperands_begin();
+  unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlignment());
+  if (NewOpc == Opc)
+    return false;
+
+  I.setDesc(TII.get(NewOpc));
+  MachineInstrBuilder MIB(MF, I);
+  if (Opc == TargetOpcode::G_LOAD)
+    addOffset(MIB, 0);
+  else {
+    // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL)
+    I.RemoveOperand(0);
+    addOffset(MIB, 0).addUse(DefReg);
+  }
+  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
+}
+

llvm/trunk/lib/Target/X86/X86LegalizerInfo.cpp

X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI) : Subtarget(STI) {
   setLegalizerInfoSSE1();
   setLegalizerInfoSSE2();
 
   computeTables();
 }
 
 void X86LegalizerInfo::setLegalizerInfo32bit() {
 
+  if (Subtarget.is64Bit())
+    return;
+
+  const LLT p0 = LLT::pointer(0, 32);
   const LLT s8 = LLT::scalar(8);
   const LLT s16 = LLT::scalar(16);
   const LLT s32 = LLT::scalar(32);
 
-  for (auto Ty : {s8, s16, s32}) {
-    setAction({G_ADD, Ty}, Legal);
-    setAction({G_SUB, Ty}, Legal);
+  for (unsigned BinOp : {G_ADD, G_SUB})
+    for (auto Ty : {s8, s16, s32})
+      setAction({BinOp, Ty}, Legal);
+
+  for (unsigned MemOp : {G_LOAD, G_STORE}) {
+    for (auto Ty : {s8, s16, s32, p0})
+      setAction({MemOp, Ty}, Legal);
+
+    // And everything's fine in addrspace 0.
+    setAction({MemOp, 1, p0}, Legal);
   }
 }
 
 void X86LegalizerInfo::setLegalizerInfo64bit() {
 
   if (!Subtarget.is64Bit())
     return;
 
+  const LLT p0 = LLT::pointer(0, 64);
+  const LLT s8 = LLT::scalar(8);
+  const LLT s16 = LLT::scalar(16);
+  const LLT s32 = LLT::scalar(32);
   const LLT s64 = LLT::scalar(64);
 
-  setAction({G_ADD, s64}, Legal);
-  setAction({G_SUB, s64}, Legal);
+  for (unsigned BinOp : {G_ADD, G_SUB})
+    for (auto Ty : {s8, s16, s32, s64})
+      setAction({BinOp, Ty}, Legal);
+
+  for (unsigned MemOp : {G_LOAD, G_STORE}) {
+    for (auto Ty : {s8, s16, s32, s64, p0})
+      setAction({MemOp, Ty}, Legal);
+
+    // And everything's fine in addrspace 0.
+    setAction({MemOp, 1, p0}, Legal);
+  }
 }
 
 void X86LegalizerInfo::setLegalizerInfoSSE1() {
   if (!Subtarget.hasSSE1())
     return;
 
   const LLT s32 = LLT::scalar(32);
   const LLT v4s32 = LLT::vector(4, 32);
+  const LLT v2s64 = LLT::vector(2, 64);
 
   for (unsigned BinOp : {G_FADD, G_FSUB, G_FMUL, G_FDIV})
     for (auto Ty : {s32, v4s32})
       setAction({BinOp, Ty}, Legal);
+
+  for (unsigned MemOp : {G_LOAD, G_STORE})
+    for (auto Ty : {v4s32, v2s64})
+      setAction({MemOp, Ty}, Legal);
 }
 
 void X86LegalizerInfo::setLegalizerInfoSSE2() {
   if (!Subtarget.hasSSE2())
     return;
 
   const LLT s64 = LLT::scalar(64);
   const LLT v4s32 = LLT::vector(4, 32);

llvm/trunk/lib/Target/X86/X86RegisterBankInfo.h

 
 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
 
 #define GET_REGBANK_DECLARATIONS
 #include "X86GenRegisterBank.inc"
 
 namespace llvm {
 
+class LLT;
+
 class X86GenRegisterBankInfo : public RegisterBankInfo {
 protected:
+#define GET_TARGET_REGBANK_CLASS
+#include "X86GenRegisterBank.inc"
+#define GET_TARGET_REGBANK_INFO_CLASS
+#include "X86GenRegisterBankInfo.def"
+
   static RegisterBankInfo::PartialMapping PartMappings[];
   static RegisterBankInfo::ValueMapping ValMappings[];
 
-#define GET_TARGET_REGBANK_CLASS
-#include "X86GenRegisterBank.inc"
+  static PartialMappingIdx getPartialMappingIdx(const LLT &Ty, bool isFP);
+  static const RegisterBankInfo::ValueMapping *
+  getValueMapping(PartialMappingIdx Idx, unsigned NumOperands);
 };
 
 class TargetRegisterInfo;
 
 /// This class provides the information for the target register banks.
 class X86RegisterBankInfo final : public X86GenRegisterBankInfo {
 private:
   /// Get an instruction mapping.
   /// \return An InstructionMappings with a statically allocated
   /// OperandsMapping.
-  static InstructionMapping getOperandsMapping(const MachineInstr &MI,
+  static InstructionMapping getSameOperandsMapping(const MachineInstr &MI,
                                                    bool isFP);
 
 public:
   X86RegisterBankInfo(const TargetRegisterInfo &TRI);
 
   const RegisterBank &
   getRegBankFromRegClass(const TargetRegisterClass &RC) const override;
 
   InstructionMapping getInstrMapping(const MachineInstr &MI) const override;
 };
 
 } // namespace llvm
 #endif

llvm/trunk/lib/Target/X86/X86RegisterBankInfo.cpp

 #include "llvm/CodeGen/GlobalISel/RegisterBank.h"
 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/Target/TargetRegisterInfo.h"
 
 #define GET_TARGET_REGBANK_IMPL
 #include "X86GenRegisterBank.inc"
 
+using namespace llvm;
 // This file will be TableGen'ed at some point.
+#define GET_TARGET_REGBANK_INFO_IMPL
 #include "X86GenRegisterBankInfo.def"
 
-using namespace llvm;
-
 #ifndef LLVM_BUILD_GLOBAL_ISEL
 #error "You shouldn't build this"
 #endif
 
 X86RegisterBankInfo::X86RegisterBankInfo(const TargetRegisterInfo &TRI)
     : X86GenRegisterBankInfo() {
 
   // validate RegBank initialization.
if (X86::FR32XRegClass.hasSubClassEq(&RC) ||
       X86::VR128XRegClass.hasSubClassEq(&RC) ||
       X86::VR256XRegClass.hasSubClassEq(&RC) ||
       X86::VR512RegClass.hasSubClassEq(&RC))
     return getRegBank(X86::VECRRegBankID);
 
   llvm_unreachable("Unsupported register kind yet.");
 }
 
-RegisterBankInfo::InstructionMapping
-X86RegisterBankInfo::getOperandsMapping(const MachineInstr &MI, bool isFP) {
-  const MachineFunction &MF = *MI.getParent()->getParent();
-  const MachineRegisterInfo &MRI = MF.getRegInfo();
-
-  unsigned NumOperands = MI.getNumOperands();
-  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
-
-  if (NumOperands != 3 || (Ty != MRI.getType(MI.getOperand(1).getReg())) ||
-      (Ty != MRI.getType(MI.getOperand(2).getReg())))
-    llvm_unreachable("Unsupported operand maping yet.");
-
-  ValueMappingIdx ValMapIdx = VMI_None;
-
-  if (Ty.isScalar()) {
-    if (!isFP) {
-      switch (Ty.getSizeInBits()) {
-      case 8:
-        ValMapIdx = VMI_3OpsGpr8Idx;
-        break;
-      case 16:
-        ValMapIdx = VMI_3OpsGpr16Idx;
-        break;
-      case 32:
-        ValMapIdx = VMI_3OpsGpr32Idx;
-        break;
-      case 64:
-        ValMapIdx = VMI_3OpsGpr64Idx;
-        break;
-      default:
-        llvm_unreachable("Unsupported register size.");
-      }
-    } else {
-      switch (Ty.getSizeInBits()) {
-      case 32:
-        ValMapIdx = VMI_3OpsFp32Idx;
-        break;
-      case 64:
-        ValMapIdx = VMI_3OpsFp64Idx;
-        break;
-      default:
-        llvm_unreachable("Unsupported register size.");
-      }
-    }
+X86GenRegisterBankInfo::PartialMappingIdx
+X86GenRegisterBankInfo::getPartialMappingIdx(const LLT &Ty, bool isFP) {
+  if ((Ty.isScalar() && !isFP) || Ty.isPointer()) {
+      return PMI_GPR8;
+      return PMI_GPR16;
+      return PMI_GPR32;
+      return PMI_GPR64;
+  } else if (Ty.isScalar()) {
+      return PMI_FP32;
+      return PMI_FP64;
   } else {
     switch (Ty.getSizeInBits()) {
     case 128:
-      ValMapIdx = VMI_3OpsVec128Idx;
-      break;
+      return PMI_VEC128;
     case 256:
-      ValMapIdx = VMI_3OpsVec256Idx;
-      break;
+      return PMI_VEC256;
     case 512:
-      ValMapIdx = VMI_3OpsVec512Idx;
-      break;
+      return PMI_VEC512;
     default:
       llvm_unreachable("Unsupported register size.");
     }
   }
 
-  return InstructionMapping{DefaultMappingID, 1, &ValMappings[ValMapIdx],
-                            NumOperands};
+  return PMI_None;
+}
+
+RegisterBankInfo::InstructionMapping
+X86RegisterBankInfo::getSameOperandsMapping(const MachineInstr &MI, bool isFP) {
+  const MachineFunction &MF = *MI.getParent()->getParent();
+  const MachineRegisterInfo &MRI = MF.getRegInfo();
+
+  unsigned NumOperands = MI.getNumOperands();
+  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
+
+  if (NumOperands != 3 || (Ty != MRI.getType(MI.getOperand(1).getReg())) ||
+      (Ty != MRI.getType(MI.getOperand(2).getReg())))
+    llvm_unreachable("Unsupported operand mapping yet.");
+
+  auto Mapping = getValueMapping(getPartialMappingIdx(Ty, isFP), 3);
+  return InstructionMapping{DefaultMappingID, 1, Mapping, NumOperands};
 }
 
 RegisterBankInfo::InstructionMapping
 X86RegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
+  const MachineFunction &MF = *MI.getParent()->getParent();
+  const MachineRegisterInfo &MRI = MF.getRegInfo();
   auto Opc = MI.getOpcode();
 
   // Try the default logic for non-generic instructions that are either copies
   // or already have some operands assigned to banks.
   if (!isPreISelGenericOpcode(Opc)) {
     InstructionMapping Mapping = getInstrMappingImpl(MI);
     if (Mapping.isValid())
       return Mapping;
   }
 
   switch (Opc) {
   case TargetOpcode::G_ADD:
   case TargetOpcode::G_SUB:
-    return getOperandsMapping(MI, false);
+    return getSameOperandsMapping(MI, false);
     break;
   case TargetOpcode::G_FADD:
   case TargetOpcode::G_FSUB:
   case TargetOpcode::G_FMUL:
   case TargetOpcode::G_FDIV:
-    return getOperandsMapping(MI, true);
+    return getSameOperandsMapping(MI, true);
     break;
   default:
-    return InstructionMapping{};
+    break;
+  }
+
+  unsigned NumOperands = MI.getNumOperands();
+  unsigned Cost = 1; // set dafault cost
+
+  // Track the bank of each register.
+  SmallVector<PartialMappingIdx, 4> OpRegBankIdx(NumOperands);
+  for (unsigned Idx = 0; Idx < NumOperands; ++Idx) {
+    auto &MO = MI.getOperand(Idx);
+    if (!MO.isReg())
+      continue;
+
+    // As a top-level guess, use NotFP mapping (all scalars in GPRs)
+    OpRegBankIdx[Idx] = getPartialMappingIdx(MRI.getType(MO.getReg()), false);
+  }
+
+  // Finally construct the computed mapping.
+  RegisterBankInfo::InstructionMapping Mapping =
+      InstructionMapping{DefaultMappingID, Cost, nullptr, NumOperands};
+  SmallVector<const ValueMapping *, 8> OpdsMapping(NumOperands);
+  for (unsigned Idx = 0; Idx < NumOperands; ++Idx) {
+    if (MI.getOperand(Idx).isReg()) {
+      auto Mapping = getValueMapping(OpRegBankIdx[Idx], 1);
+      if (!Mapping->isValid())
+        return InstructionMapping();
+
+      OpdsMapping[Idx] = Mapping;
+    }
   }
 
-  return InstructionMapping{};
+  Mapping.setOperandsMapping(getOperandsMapping(OpdsMapping));
+  return Mapping;
 }

llvm/trunk/test/CodeGen/X86/GlobalISel/X86-regbankselect.mir

define <4 x i32>  @test_add_v4i32(<4 x i32> %arg1, <4 x i32>  %arg2) {
     ret <4 x i32>  %ret
   }
 
   define <4 x float>  @test_add_v4f32(<4 x float> %arg1, <4 x float>  %arg2) {
     %ret = fadd <4 x float>  %arg1, %arg2
     ret <4 x float>  %ret
   }
 
+  define i8 @test_load_i8(i8* %p1) {
+    %r = load i8, i8* %p1
+    ret i8 %r
+  }
+
+  define i16 @test_load_i16(i16* %p1) {
+    %r = load i16, i16* %p1
+    ret i16 %r
+  }
+
+  define i32 @test_load_i32(i32* %p1) {
+    %r = load i32, i32* %p1
+    ret i32 %r
+  }
+
+  define i64 @test_load_i64(i64* %p1) {
+    %r = load i64, i64* %p1
+    ret i64 %r
+  }
+
+  define float @test_load_float(float* %p1) {
+    %r = load float, float* %p1
+    ret float %r
+  }
+
+  define double @test_load_double(double* %p1) {
+    %r = load double, double* %p1
+    ret double %r
+  }
+
+  define <4 x i32> @test_load_v4i32(<4 x i32>* %p1) {
+    %r = load <4 x i32>, <4 x i32>* %p1, align 16
+    ret <4 x i32> %r
+  }
+
+  define i32* @test_store_i32(i32 %val, i32* %p1) {
+    store i32 %val, i32* %p1
+    ret i32* %p1
+  }
+
+  define i64* @test_store_i64(i64 %val, i64* %p1) {
+    store i64 %val, i64* %p1
+    ret i64* %p1
+  }
+
+  define float* @test_store_float(float %val, float* %p1) {
+    store float %val, float* %p1
+    ret float* %p1
+  }
+
+  define double* @test_store_double(double %val, double* %p1) {
+    store double %val, double* %p1
+    ret double* %p1
+  }
+
 ...
 ---
 name:            test_add_i8
 alignment:       4
 legalized:       true
 regBankSelected: false
 selected:        false
 tracksRegLiveness: true
bb.1 (%ir-block.0):
 
     %0(<4 x s32>) = COPY %xmm0
     %1(<4 x s32>) = COPY %xmm1
     %2(<4 x s32>) = G_FADD %0, %1
     %xmm0 = COPY %2(<4 x s32>)
     RET 0, implicit %xmm0
 
 ...
+---
+name:            test_load_i8
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_i8
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s8) = G_LOAD %0(p0) :: (load 1 from %ir.p1)
+    %al = COPY %1(s8)
+    RET 0, implicit %al
+
+...
+---
+name:            test_load_i16
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_i16
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s16) = G_LOAD %0(p0) :: (load 2 from %ir.p1)
+    %ax = COPY %1(s16)
+    RET 0, implicit %ax
+
+...
+---
+name:            test_load_i32
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_i32
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s32) = G_LOAD %0(p0) :: (load 4 from %ir.p1)
+    %eax = COPY %1(s32)
+    RET 0, implicit %eax
+
+...
+---
+name:            test_load_i64
+alignment:       4
+exposesReturnsTwice: false
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_i64
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s64) = G_LOAD %0(p0) :: (load 8 from %ir.p1)
+    %rax = COPY %1(s64)
+    RET 0, implicit %rax
+
+...
+---
+name:            test_load_float
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_float
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s32) = G_LOAD %0(p0) :: (load 4 from %ir.p1)
+    %xmm0 = COPY %1(s32)
+    RET 0, implicit %xmm0
+
+...
+---
+name:            test_load_double
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_double
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s64) = G_LOAD %0(p0) :: (load 8 from %ir.p1)
+    %xmm0 = COPY %1(s64)
+    RET 0, implicit %xmm0
+
+...
+---
+name:            test_load_v4i32
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_load_v4i32
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: vecr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(<4 x s32>) = G_LOAD %0(p0) :: (load 16 from %ir.p1, align 1)
+    %xmm0 = COPY %1(<4 x s32>)
+    RET 0, implicit %xmm0
+
+...
+---
+name:            test_store_i32
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_store_i32
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %edi, %rsi
+
+    %0(s32) = COPY %edi
+    %1(p0) = COPY %rsi
+    G_STORE %0(s32), %1(p0) :: (store 4 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+name:            test_store_i64
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_store_i64
+# CHECK: registers:
+# CHECK:   - { id: 0, class: gpr }
+# CHECK:   - { id: 1, class: gpr }
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %rsi
+
+    %0(s64) = COPY %rdi
+    %1(p0) = COPY %rsi
+    G_STORE %0(s64), %1(p0) :: (store 8 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+name:            test_store_float
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_store_float
+# CHECK: registers:
+# CHECK:    - { id: 0, class: vecr }
+# CHECK:    - { id: 1, class: gpr }
+# CHECK:    - { id: 2, class: gpr }
+
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s32) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    ; CHECK:  %2(s32) = COPY %0(s32)
+    ; CHECK:  G_STORE %2(s32), %1(p0) :: (store 4 into %ir.p1)
+    G_STORE %0(s32), %1(p0) :: (store 4 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+name:            test_store_double
+alignment:       4
+legalized:       true
+regBankSelected: false
+selected:        false
+# CHECK-LABEL: name:            test_store_double
+# CHECK: registers:
+# CHECK:    - { id: 0, class: vecr }
+# CHECK:    - { id: 1, class: gpr }
+# CHECK:    - { id: 2, class: gpr }
+
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s64) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    ; CHECK:  %2(s64) = COPY %0(s64)
+    ; CHECK:  G_STORE %2(s64), %1(p0) :: (store 8 into %ir.p1)
+    G_STORE %0(s64), %1(p0) :: (store 8 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
 

llvm/trunk/test/CodeGen/X86/GlobalISel/binop-isel.ll

 ;
 ; ALL_AVX-LABEL: test_sub_v4f32:
 ; ALL_AVX:       # BB#0:
 ; ALL_AVX-NEXT:    vsubps %xmm1, %xmm0, %xmm0
 ; ALL_AVX-NEXT:    retq
   %ret = fsub <4 x float>  %arg1, %arg2
   ret <4 x float>  %ret
 }
+
+define i32  @test_copy_float(float %val) {
+; SSE-LABEL: test_copy_float:
+; SSE:       # BB#0:
+; SSE-NEXT:    movd %xmm0, %eax
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_copy_float:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovd %xmm0, %eax
+; ALL_AVX-NEXT:    retq
+  %r = bitcast float %val to i32
+  ret i32 %r
+}
+
+define float  @test_copy_i32(i32 %val) {
+; SSE-LABEL: test_copy_i32:
+; SSE:       # BB#0:
+; SSE-NEXT:    movd %edi, %xmm0
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_copy_i32:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovd %edi, %xmm0
+; ALL_AVX-NEXT:    retq
+  %r = bitcast i32 %val to float
+  ret float %r
+}
+

llvm/trunk/test/CodeGen/X86/GlobalISel/irtranslator-callingconv.ll

 ; X32-NEXT:  [[ARG1:%[0-9]+]](s64) = G_LOAD [[ARG1_ADDR:%[0-9]+]](p0) :: (invariant load 8 from %fixed-stack.[[STACK0]], align 0)
 ; X32-NEXT:  [[ARG2_ADDR:%[0-9]+]](p0) = G_FRAME_INDEX %fixed-stack.[[STACK4]]
 ; X32-NEXT:  [[ARG2:%[0-9]+]](s64) = G_LOAD [[ARG2_ADDR:%[0-9]+]](p0) :: (invariant load 8 from %fixed-stack.[[STACK4]], align 0)
 ; X32-NEXT:  %fp0 = COPY [[ARG2:%[0-9]+]](s64)
 ; X32-NEXT:  RET 0, implicit %fp0
 
   ret double %arg2
 }
+
+define i32 * @test_memop_i32(i32 * %p1) {
+; ALL-LABEL:name:            test_memop_i32
+;X64    liveins: %rdi
+;X64:       %0(p0) = COPY %rdi
+;X64-NEXT:  %rax = COPY %0(p0)
+;X64-NEXT:  RET 0, implicit %rax
+
+;X32: fixedStack:
+;X32:  id: [[STACK0:[0-9]+]], offset: 0, size: 4, alignment: 16, isImmutable: true, isAliased: false }
+;X32:         %1(p0) = G_FRAME_INDEX %fixed-stack.[[STACK0]]
+;X32-NEXT:    %0(p0) = G_LOAD %1(p0) :: (invariant load 4 from %fixed-stack.[[STACK0]], align 0)
+;X32-NEXT:    %eax = COPY %0(p0)
+;X32-NEXT:    RET 0, implicit %eax
+
+  ret i32 * %p1;
+}
 No newline at end of file

llvm/trunk/test/CodeGen/X86/GlobalISel/memop-isel.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=x86_64-linux-gnu                 -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=SSE
+; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx     -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=AVX
+; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=AVX512F
+; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=AVX512VL
+
+
+define i8 @test_load_i8(i8 * %p1) {
+; ALL-LABEL: test_load_i8:
+; ALL:       # BB#0:
+; ALL-NEXT:    movb (%rdi), %al
+; ALL-NEXT:    retq
+  %r = load i8, i8* %p1
+  ret i8 %r
+}
+
+define i16 @test_load_i16(i16 * %p1) {
+; ALL-LABEL: test_load_i16:
+; ALL:       # BB#0:
+; ALL-NEXT:    movzwl (%rdi), %eax
+; ALL-NEXT:    retq
+  %r = load i16, i16* %p1
+  ret i16 %r
+}
+
+define i32 @test_load_i32(i32 * %p1) {
+; ALL-LABEL: test_load_i32:
+; ALL:       # BB#0:
+; ALL-NEXT:    movl (%rdi), %eax
+; ALL-NEXT:    retq
+  %r = load i32, i32* %p1
+  ret i32 %r
+}
+
+define i64 @test_load_i64(i64 * %p1) {
+; ALL-LABEL: test_load_i64:
+; ALL:       # BB#0:
+; ALL-NEXT:    movq (%rdi), %rax
+; ALL-NEXT:    retq
+  %r = load i64, i64* %p1
+  ret i64 %r
+}
+
+define float @test_load_float(float * %p1) {
+; SSE-LABEL: test_load_float:
+; SSE:       # BB#0:
+; SSE-NEXT:    movl (%rdi), %eax
+; SSE-NEXT:    movd %eax, %xmm0
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_load_float:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    movl (%rdi), %eax
+; ALL_AVX-NEXT:    vmovd %eax, %xmm0
+; ALL_AVX-NEXT:    retq
+  %r = load float, float* %p1
+  ret float %r
+}
+
+define double @test_load_double(double * %p1) {
+; SSE-LABEL: test_load_double:
+; SSE:       # BB#0:
+; SSE-NEXT:    movq (%rdi), %rax
+; SSE-NEXT:    movd %rax, %xmm0
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_load_double:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    movq (%rdi), %rax
+; ALL_AVX-NEXT:    vmovq %rax, %xmm0
+; ALL_AVX-NEXT:    retq
+  %r = load double, double* %p1
+  ret double %r
+}
+
+define <4 x i32> @test_load_v4i32_noalign(<4 x i32> * %p1) {
+; SSE-LABEL: test_load_v4i32_noalign:
+; SSE:       # BB#0:
+; SSE-NEXT:    movups (%rdi), %xmm0
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_load_v4i32_noalign:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovups (%rdi), %xmm0
+; ALL_AVX-NEXT:    retq
+  %r = load <4 x i32>, <4 x i32>* %p1, align 1
+  ret <4 x i32> %r
+}
+
+define <4 x i32> @test_load_v4i32_align(<4 x i32> * %p1) {
+; SSE-LABEL: test_load_v4i32_align:
+; SSE:       # BB#0:
+; SSE-NEXT:    movaps (%rdi), %xmm0
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_load_v4i32_align:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovaps (%rdi), %xmm0
+; ALL_AVX-NEXT:    retq
+  %r = load <4 x i32>, <4 x i32>* %p1, align 16
+  ret <4 x i32> %r
+}
+
+define i32 * @test_store_i32(i32 %val, i32 * %p1) {
+; ALL-LABEL: test_store_i32:
+; ALL:       # BB#0:
+; ALL-NEXT:    movl %edi, (%rsi)
+; ALL-NEXT:    movq %rsi, %rax
+; ALL-NEXT:    retq
+  store i32 %val, i32* %p1
+  ret i32 * %p1;
+}
+
+define i64 * @test_store_i64(i64 %val, i64 * %p1) {
+; ALL-LABEL: test_store_i64:
+; ALL:       # BB#0:
+; ALL-NEXT:    movq %rdi, (%rsi)
+; ALL-NEXT:    movq %rsi, %rax
+; ALL-NEXT:    retq
+  store i64 %val, i64* %p1
+  ret i64 * %p1;
+}
+
+define float * @test_store_float(float %val, float * %p1) {
+; SSE-LABEL: test_store_float:
+; SSE:       # BB#0:
+; SSE-NEXT:    movd %xmm0, %eax
+; SSE-NEXT:    movl %eax, (%rdi)
+; SSE-NEXT:    movq %rdi, %rax
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_store_float:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovd %xmm0, %eax
+; ALL_AVX-NEXT:    movl %eax, (%rdi)
+; ALL_AVX-NEXT:    movq %rdi, %rax
+; ALL_AVX-NEXT:    retq
+  store float %val, float* %p1
+  ret float * %p1;
+}
+
+define double * @test_store_double(double %val, double * %p1) {
+; SSE-LABEL: test_store_double:
+; SSE:       # BB#0:
+; SSE-NEXT:    movd %xmm0, %rax
+; SSE-NEXT:    movq %rax, (%rdi)
+; SSE-NEXT:    movq %rdi, %rax
+; SSE-NEXT:    retq
+;
+; ALL_AVX-LABEL: test_store_double:
+; ALL_AVX:       # BB#0:
+; ALL_AVX-NEXT:    vmovq %xmm0, %rax
+; ALL_AVX-NEXT:    movq %rax, (%rdi)
+; ALL_AVX-NEXT:    movq %rdi, %rax
+; ALL_AVX-NEXT:    retq
+  store double %val, double* %p1
+  ret double * %p1;
+}
+

llvm/trunk/test/CodeGen/X86/GlobalISel/x86_64-instructionselect.mir

define <4 x float>  @test_add_v4f32(<4 x float> %arg1, <4 x float>  %arg2) {
     %ret = fadd <4 x float>  %arg1, %arg2
     ret <4 x float>  %ret
   }
 
   define <4 x float>  @test_sub_v4f32(<4 x float> %arg1, <4 x float>  %arg2) {
     %ret = fsub <4 x float>  %arg1, %arg2
     ret <4 x float>  %ret
   }
+  
+    define i8 @test_load_i8(i8* %p1) {
+    %r = load i8, i8* %p1
+    ret i8 %r
+  }
+
+  define i16 @test_load_i16(i16* %p1) {
+    %r = load i16, i16* %p1
+    ret i16 %r
+  }
+
+  define i32 @test_load_i32(i32* %p1) {
+    %r = load i32, i32* %p1
+    ret i32 %r
+  }
+
+  define i64 @test_load_i64(i64* %p1) {
+    %r = load i64, i64* %p1
+    ret i64 %r
+  }
+
+  define float @test_load_float(float* %p1) {
+    %r = load float, float* %p1
+    ret float %r
+  }
+  
+  define float @test_load_float_vecreg(float* %p1) {
+    %r = load float, float* %p1
+    ret float %r
+  }
+  
+
+  define double @test_load_double(double* %p1) {
+    %r = load double, double* %p1
+    ret double %r
+  }
+
+  define double @test_load_double_vecreg(double* %p1) {
+    %r = load double, double* %p1
+    ret double %r
+  }
+
+  define <4 x i32> @test_load_v4i32_noalign(<4 x i32>* %p1) {
+    %r = load <4 x i32>, <4 x i32>* %p1, align 1
+    ret <4 x i32> %r
+  }
+
+  define <4 x i32> @test_load_v4i32_align(<4 x i32>* %p1) {
+    %r = load <4 x i32>, <4 x i32>* %p1, align 16
+    ret <4 x i32> %r
+  }
+
+  define i32* @test_store_i32(i32 %val, i32* %p1) {
+    store i32 %val, i32* %p1
+    ret i32* %p1
+  }
+
+  define i64* @test_store_i64(i64 %val, i64* %p1) {
+    store i64 %val, i64* %p1
+    ret i64* %p1
+  }
+
+  define float* @test_store_float(float %val, float* %p1) {
+    store float %val, float* %p1
+    ret float* %p1
+  }
+
+  define float* @test_store_float_vec(float %val, float* %p1) {
+    store float %val, float* %p1
+    ret float* %p1
+  }
+
+  define double* @test_store_double(double %val, double* %p1) {
+    store double %val, double* %p1
+    ret double* %p1
+  }
+  
+  define double* @test_store_double_vec(double %val, double* %p1) {
+    store double %val, double* %p1
+    ret double* %p1
+  }
+
+  define <4 x i32>* @test_store_v4i32_align(<4 x i32> %val, <4 x i32>* %p1) {
+    store <4 x i32> %val, <4 x i32>* %p1, align 16
+    ret <4 x i32>* %p1
+  }
+
+  define <4 x i32>* @test_store_v4i32_noalign(<4 x i32> %val, <4 x i32>* %p1) {
+    store <4 x i32> %val, <4 x i32>* %p1, align 1
+    ret <4 x i32>* %p1
+  }
+
 ...
 
 ---
 name:            test_add_i64
 legalized:       true
 regBankSelected: true
 # ALL:      registers:
 # ALL-NEXT:  - { id: 0, class: gr64 }
bb.1 (%ir-block.0):
 
     %0(<4 x s32>) = COPY %xmm0
     %1(<4 x s32>) = COPY %xmm1
     %2(<4 x s32>) = G_FSUB %0, %1
     %xmm0 = COPY %2(<4 x s32>)
     RET 0, implicit %xmm0
 
 ...
+---
+# ALL-LABEL: name:            test_load_i8
+name:            test_load_i8
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr8 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV8rm %0, 1, _, 0, _ :: (load 1 from %ir.p1)
+# ALL:     %al = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s8) = G_LOAD %0(p0) :: (load 1 from %ir.p1)
+    %al = COPY %1(s8)
+    RET 0, implicit %al
+
+...
+---
+# ALL-LABEL: name:            test_load_i16
+name:            test_load_i16
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr16 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV16rm %0, 1, _, 0, _ :: (load 2 from %ir.p1)
+# ALL:     %ax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s16) = G_LOAD %0(p0) :: (load 2 from %ir.p1)
+    %ax = COPY %1(s16)
+    RET 0, implicit %ax
+
+...
+---
+# ALL-LABEL: name:            test_load_i32
+name:            test_load_i32
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr32 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV32rm %0, 1, _, 0, _ :: (load 4 from %ir.p1)
+# ALL:     %eax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s32) = G_LOAD %0(p0) :: (load 4 from %ir.p1)
+    %eax = COPY %1(s32)
+    RET 0, implicit %eax
+
+...
+---
+# ALL-LABEL: name:            test_load_i64
+name:            test_load_i64
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV64rm %0, 1, _, 0, _ :: (load 8 from %ir.p1)
+# ALL:     %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s64) = G_LOAD %0(p0) :: (load 8 from %ir.p1)
+    %rax = COPY %1(s64)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_load_float
+name:            test_load_float
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr32 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV32rm %0, 1, _, 0, _ :: (load 4 from %ir.p1)
+# ALL:     %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s32) = G_LOAD %0(p0) :: (load 4 from %ir.p1)
+    %xmm0 = COPY %1(s32)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_load_float_vecreg
+name:            test_load_float_vecreg
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:          - { id: 0, class: gr64 }
+# NO_AVX512F:   - { id: 1, class: fr32 }
+# AVX512ALL:    - { id: 1, class: fr32x }
+  - { id: 0, class: gpr }
+  - { id: 1, class: vecr }
+# ALL:       %0 = COPY %rdi
+# SSE:       %1 = MOVSSrm %0, 1, _, 0, _ :: (load 4 from %ir.p1)
+# AVX:       %1 = VMOVSSrm %0, 1, _, 0, _ :: (load 4 from %ir.p1)
+# AVX512ALL: %1 = VMOVSSZrm %0, 1, _, 0, _ :: (load 4 from %ir.p1)
+# ALL: %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s32) = G_LOAD %0(p0) :: (load 4 from %ir.p1)
+    %xmm0 = COPY %1(s32)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_load_double
+name:            test_load_double
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = MOV64rm %0, 1, _, 0, _ :: (load 8 from %ir.p1)
+# ALL:     %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s64) = G_LOAD %0(p0) :: (load 8 from %ir.p1)
+    %xmm0 = COPY %1(s64)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_load_double_vecreg
+name:            test_load_double_vecreg
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:          - { id: 0, class: gr64 }
+# NO_AVX512F:   - { id: 1, class: fr64 }
+# AVX512ALL:    - { id: 1, class: fr64x }
+  - { id: 0, class: gpr }
+  - { id: 1, class: vecr }
+# ALL:       %0 = COPY %rdi
+# SSE:       %1 = MOVSDrm %0, 1, _, 0, _ :: (load 8 from %ir.p1)
+# AVX:       %1 = VMOVSDrm %0, 1, _, 0, _ :: (load 8 from %ir.p1)
+# AVX512ALL: %1 = VMOVSDZrm %0, 1, _, 0, _ :: (load 8 from %ir.p1)
+# ALL: %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(s64) = G_LOAD %0(p0) :: (load 8 from %ir.p1)
+    %xmm0 = COPY %1(s64)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_load_v4i32_noalign
+name:            test_load_v4i32_noalign
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:          - { id: 0, class: gr64 }
+# NO_AVX512F:   - { id: 1, class: vr128 }
+# AVX512ALL:    - { id: 1, class: vr128x }
+  - { id: 0, class: gpr }
+  - { id: 1, class: vecr }
+# ALL:      %0 = COPY %rdi
+# SSE:      %1 = MOVUPSrm %0, 1, _, 0, _ :: (load 16 from %ir.p1, align 1)
+# AVX:      %1 = VMOVUPSrm %0, 1, _, 0, _ :: (load 16 from %ir.p1, align 1)
+# AVX512F:  %1 = VMOVUPSZ128rm_NOVLX %0, 1, _, 0, _ :: (load 16 from %ir.p1, align 1)
+# AVX512VL: %1 = VMOVUPSZ128rm %0, 1, _, 0, _ :: (load 16 from %ir.p1, align 1)
+# ALL: %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(<4 x s32>) = G_LOAD %0(p0) :: (load 16 from %ir.p1, align 1)
+    %xmm0 = COPY %1(<4 x s32>)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_load_v4i32_align
+name:            test_load_v4i32_align
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# NO_AVX512F:   - { id: 1, class: vr128 }
+# AVX512ALL:    - { id: 1, class: vr128x }
+  - { id: 0, class: gpr }
+  - { id: 1, class: vecr }
+# ALL:      %0 = COPY %rdi
+# SSE:      %1 = MOVAPSrm %0, 1, _, 0, _ :: (load 16 from %ir.p1)
+# AVX:      %1 = VMOVAPSrm %0, 1, _, 0, _ :: (load 16 from %ir.p1)
+# AVX512F:  %1 = VMOVAPSZ128rm_NOVLX %0, 1, _, 0, _ :: (load 16 from %ir.p1)
+# AVX512VL: %1 = VMOVAPSZ128rm %0, 1, _, 0, _ :: (load 16 from %ir.p1)
+# ALL: %xmm0 = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi
+
+    %0(p0) = COPY %rdi
+    %1(<4 x s32>) = G_LOAD %0(p0) :: (load 16 from %ir.p1)
+    %xmm0 = COPY %1(<4 x s32>)
+    RET 0, implicit %xmm0
+
+...
+---
+# ALL-LABEL: name:            test_store_i32
+name:            test_store_i32
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr32 }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %edi
+# ALL:     %1 = COPY %rsi
+# ALL:     MOV32mr %1, 1, _, 0, _, %0 :: (store 4 into %ir.p1)
+# ALL:     %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %edi, %rsi
+
+    %0(s32) = COPY %edi
+    %1(p0) = COPY %rsi
+    G_STORE %0(s32), %1(p0) :: (store 4 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_i64
+name:            test_store_i64
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: gr64 }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+# ALL:     %0 = COPY %rdi
+# ALL:     %1 = COPY %rsi
+# ALL:     MOV64mr %1, 1, _, 0, _, %0 :: (store 8 into %ir.p1)
+# ALL:     %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %rsi
+
+    %0(s64) = COPY %rdi
+    %1(p0) = COPY %rsi
+    G_STORE %0(s64), %1(p0) :: (store 8 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_float
+name:            test_store_float
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: fr32x }
+# ALL:   - { id: 1, class: gr64 }
+# ALL:   - { id: 2, class: gr32 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+# ALL:     %0 = COPY %xmm0
+# ALL:     %1 = COPY %rdi
+# ALL:     %2 = COPY %0
+# ALL:     MOV32mr %1, 1, _, 0, _, %2 :: (store 4 into %ir.p1)
+# ALL:     %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s32) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    %2(s32) = COPY %0(s32)
+    G_STORE %2(s32), %1(p0) :: (store 4 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_float_vec
+name:            test_store_float_vec
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# NO_AVX512F:   - { id: 0, class: fr32 }
+# AVX512ALL:    - { id: 0, class: fr32x }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+# ALL:       %0 = COPY %xmm0
+# ALL:       %1 = COPY %rdi
+# SSE:       MOVSSmr %1, 1, _, 0, _, %0 :: (store 4 into %ir.p1)
+# AVX:       VMOVSSmr %1, 1, _, 0, _, %0 :: (store 4 into %ir.p1)
+# AVX512ALL: VMOVSSZmr %1, 1, _, 0, _, %0 :: (store 4 into %ir.p1)
+# ALL:       %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s32) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    G_STORE %0(s32), %1(p0) :: (store 4 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_double
+name:            test_store_double
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# ALL:   - { id: 0, class: fr64x }
+# ALL:   - { id: 1, class: gr64 }
+# ALL:   - { id: 2, class: gr64 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+# ALL:     %0 = COPY %xmm0
+# ALL:     %1 = COPY %rdi
+# ALL:     %2 = COPY %0
+# ALL:     MOV64mr %1, 1, _, 0, _, %2 :: (store 8 into %ir.p1)
+# ALL:     %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s64) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    %2(s64) = COPY %0(s64)
+    G_STORE %2(s64), %1(p0) :: (store 8 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_double_vec
+name:            test_store_double_vec
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# NO_AVX512F:   - { id: 0, class: fr64 }
+# AVX512ALL:    - { id: 0, class: fr64x }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+# ALL:       %0 = COPY %xmm0
+# ALL:       %1 = COPY %rdi
+# SSE:       MOVSDmr %1, 1, _, 0, _, %0 :: (store 8 into %ir.p1)
+# AVX:       VMOVSDmr %1, 1, _, 0, _, %0 :: (store 8 into %ir.p1)
+# AVX512ALL: VMOVSDZmr %1, 1, _, 0, _, %0 :: (store 8 into %ir.p1)
+# ALL:       %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(s64) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    G_STORE %0(s64), %1(p0) :: (store 8 into %ir.p1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_v4i32_align
+name:            test_store_v4i32_align
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# NO_AVX512F:   - { id: 0, class: vr128 }
+# AVX512ALL:    - { id: 0, class: vr128x }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+# ALL:       %0 = COPY %xmm0
+# ALL:       %1 = COPY %rdi
+# SSE:       MOVAPSmr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1)
+# AVX:       VMOVAPSmr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1)
+# AVX512F:   VMOVAPSZ128mr_NOVLX %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1)
+# AVX512VL:  VMOVAPSZ128mr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1)
+# ALL:       %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(<4 x s32>) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    G_STORE %0(<4 x s32>), %1(p0) :: (store 16 into %ir.p1, align 16)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...
+---
+# ALL-LABEL: name:            test_store_v4i32_noalign
+name:            test_store_v4i32_noalign
+alignment:       4
+legalized:       true
+regBankSelected: true
+registers:
+# NO_AVX512F:   - { id: 0, class: vr128 }
+# AVX512ALL:    - { id: 0, class: vr128x }
+# ALL:   - { id: 1, class: gr64 }
+  - { id: 0, class: vecr }
+  - { id: 1, class: gpr }
+# ALL:       %0 = COPY %xmm0
+# ALL:       %1 = COPY %rdi
+# SSE:       MOVUPSmr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1, align 1)
+# AVX:       VMOVUPSmr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1, align 1)
+# AVX512F:   VMOVUPSZ128mr_NOVLX %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1, align 1)
+# AVX512VL:  VMOVUPSZ128mr %1, 1, _, 0, _, %0 :: (store 16 into %ir.p1, align 1)
+# ALL:       %rax = COPY %1
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: %rdi, %xmm0
+
+    %0(<4 x s32>) = COPY %xmm0
+    %1(p0) = COPY %rdi
+    G_STORE %0(<4 x s32>), %1(p0) :: (store 16 into %ir.p1, align 1)
+    %rax = COPY %1(p0)
+    RET 0, implicit %rax
+
+...