Google Git
Sign in
swiftshader / SwiftShader / 389854967d7803d5e3d909ca2b3995c56c080c01 / . / third_party / llvm-10.0 / llvm / lib / CodeGen / GlobalISel / LegalizerHelper.cpp
blob: fefce4b5f11a85e3436faa4a75bc35a7d563bf93 [file] [log] [blame]
//===-- llvm/CodeGen/GlobalISel/LegalizerHelper.cpp -----------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file This file implements the LegalizerHelper class to legalize
/// individual instructions and the LegalizeMachineIR wrapper pass for the
/// primary legalization.
//
//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
#include "llvm/CodeGen/GlobalISel/CallLowering.h"
#include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
#include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#define DEBUG_TYPE "legalizer"
using namespace llvm;
using namespace LegalizeActions;
/// Try to break down \p OrigTy into \p NarrowTy sized pieces.
///
/// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy,
/// with any leftover piece as type \p LeftoverTy
///
/// Returns -1 in the first element of the pair if the breakdown is not
/// satisfiable.
static std::pair<int, int>
getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) {
assert(!LeftoverTy.isValid() && "this is an out argument");
unsigned Size = OrigTy.getSizeInBits();
unsigned NarrowSize = NarrowTy.getSizeInBits();
unsigned NumParts = Size / NarrowSize;
unsigned LeftoverSize = Size - NumParts * NarrowSize;
assert(Size > NarrowSize);
if (LeftoverSize == 0)
return {NumParts, 0};
if (NarrowTy.isVector()) {
unsigned EltSize = OrigTy.getScalarSizeInBits();
if (LeftoverSize % EltSize != 0)
return {-1, -1};
LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize);
} else {
LeftoverTy = LLT::scalar(LeftoverSize);
}
int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits();
return std::make_pair(NumParts, NumLeftover);
}
LegalizerHelper::LegalizerHelper(MachineFunction &MF,
GISelChangeObserver &Observer,
MachineIRBuilder &Builder)
: MIRBuilder(Builder), MRI(MF.getRegInfo()),
LI(*MF.getSubtarget().getLegalizerInfo()), Observer(Observer) {
MIRBuilder.setMF(MF);
MIRBuilder.setChangeObserver(Observer);
}
LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI,
GISelChangeObserver &Observer,
MachineIRBuilder &B)
: MIRBuilder(B), MRI(MF.getRegInfo()), LI(LI), Observer(Observer) {
MIRBuilder.setMF(MF);
MIRBuilder.setChangeObserver(Observer);
}
LegalizerHelper::LegalizeResult
LegalizerHelper::legalizeInstrStep(MachineInstr &MI) {
LLVM_DEBUG(dbgs() << "Legalizing: "; MI.print(dbgs()));
if (MI.getOpcode() == TargetOpcode::G_INTRINSIC ||
MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS)
return LI.legalizeIntrinsic(MI, MRI, MIRBuilder) ? Legalized
: UnableToLegalize;
auto Step = LI.getAction(MI, MRI);
switch (Step.Action) {
case Legal:
LLVM_DEBUG(dbgs() << ".. Already legal\n");
return AlreadyLegal;
case Libcall:
LLVM_DEBUG(dbgs() << ".. Convert to libcall\n");
return libcall(MI);
case NarrowScalar:
LLVM_DEBUG(dbgs() << ".. Narrow scalar\n");
return narrowScalar(MI, Step.TypeIdx, Step.NewType);
case WidenScalar:
LLVM_DEBUG(dbgs() << ".. Widen scalar\n");
return widenScalar(MI, Step.TypeIdx, Step.NewType);
case Lower:
LLVM_DEBUG(dbgs() << ".. Lower\n");
return lower(MI, Step.TypeIdx, Step.NewType);
case FewerElements:
LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n");
return fewerElementsVector(MI, Step.TypeIdx, Step.NewType);
case MoreElements:
LLVM_DEBUG(dbgs() << ".. Increase number of elements\n");
return moreElementsVector(MI, Step.TypeIdx, Step.NewType);
case Custom:
LLVM_DEBUG(dbgs() << ".. Custom legalization\n");
return LI.legalizeCustom(MI, MRI, MIRBuilder, Observer) ? Legalized
: UnableToLegalize;
default:
LLVM_DEBUG(dbgs() << ".. Unable to legalize\n");
return UnableToLegalize;
}
}
void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts,
SmallVectorImpl<Register> &VRegs) {
for (int i = 0; i < NumParts; ++i)
VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
MIRBuilder.buildUnmerge(VRegs, Reg);
}
bool LegalizerHelper::extractParts(Register Reg, LLT RegTy,
LLT MainTy, LLT &LeftoverTy,
SmallVectorImpl<Register> &VRegs,
SmallVectorImpl<Register> &LeftoverRegs) {
assert(!LeftoverTy.isValid() && "this is an out argument");
unsigned RegSize = RegTy.getSizeInBits();
unsigned MainSize = MainTy.getSizeInBits();
unsigned NumParts = RegSize / MainSize;
unsigned LeftoverSize = RegSize - NumParts * MainSize;
// Use an unmerge when possible.
if (LeftoverSize == 0) {
for (unsigned I = 0; I < NumParts; ++I)
VRegs.push_back(MRI.createGenericVirtualRegister(MainTy));
MIRBuilder.buildUnmerge(VRegs, Reg);
return true;
}
if (MainTy.isVector()) {
unsigned EltSize = MainTy.getScalarSizeInBits();
if (LeftoverSize % EltSize != 0)
return false;
LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize);
} else {
LeftoverTy = LLT::scalar(LeftoverSize);
}
// For irregular sizes, extract the individual parts.
for (unsigned I = 0; I != NumParts; ++I) {
Register NewReg = MRI.createGenericVirtualRegister(MainTy);
VRegs.push_back(NewReg);
MIRBuilder.buildExtract(NewReg, Reg, MainSize * I);
}
for (unsigned Offset = MainSize * NumParts; Offset < RegSize;
Offset += LeftoverSize) {
Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy);
LeftoverRegs.push_back(NewReg);
MIRBuilder.buildExtract(NewReg, Reg, Offset);
}
return true;
}
static LLT getGCDType(LLT OrigTy, LLT TargetTy) {
if (OrigTy.isVector() && TargetTy.isVector()) {
assert(OrigTy.getElementType() == TargetTy.getElementType());
int GCD = greatestCommonDivisor(OrigTy.getNumElements(),
TargetTy.getNumElements());
return LLT::scalarOrVector(GCD, OrigTy.getElementType());
}
if (OrigTy.isVector() && !TargetTy.isVector()) {
assert(OrigTy.getElementType() == TargetTy);
return TargetTy;
}
assert(!OrigTy.isVector() && !TargetTy.isVector());
int GCD = greatestCommonDivisor(OrigTy.getSizeInBits(),
TargetTy.getSizeInBits());
return LLT::scalar(GCD);
}
void LegalizerHelper::insertParts(Register DstReg,
LLT ResultTy, LLT PartTy,
ArrayRef<Register> PartRegs,
LLT LeftoverTy,
ArrayRef<Register> LeftoverRegs) {
if (!LeftoverTy.isValid()) {
assert(LeftoverRegs.empty());
if (!ResultTy.isVector()) {
MIRBuilder.buildMerge(DstReg, PartRegs);
return;
}
if (PartTy.isVector())
MIRBuilder.buildConcatVectors(DstReg, PartRegs);
else
MIRBuilder.buildBuildVector(DstReg, PartRegs);
return;
}
unsigned PartSize = PartTy.getSizeInBits();
unsigned LeftoverPartSize = LeftoverTy.getSizeInBits();
Register CurResultReg = MRI.createGenericVirtualRegister(ResultTy);
MIRBuilder.buildUndef(CurResultReg);
unsigned Offset = 0;
for (Register PartReg : PartRegs) {
Register NewResultReg = MRI.createGenericVirtualRegister(ResultTy);
MIRBuilder.buildInsert(NewResultReg, CurResultReg, PartReg, Offset);
CurResultReg = NewResultReg;
Offset += PartSize;
}
for (unsigned I = 0, E = LeftoverRegs.size(); I != E; ++I) {
// Use the original output register for the final insert to avoid a copy.
Register NewResultReg = (I + 1 == E) ?
DstReg : MRI.createGenericVirtualRegister(ResultTy);
MIRBuilder.buildInsert(NewResultReg, CurResultReg, LeftoverRegs[I], Offset);
CurResultReg = NewResultReg;
Offset += LeftoverPartSize;
}
}
static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
switch (Opcode) {
case TargetOpcode::G_SDIV:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
switch (Size) {
case 32:
return RTLIB::SDIV_I32;
case 64:
return RTLIB::SDIV_I64;
case 128:
return RTLIB::SDIV_I128;
default:
llvm_unreachable("unexpected size");
}
case TargetOpcode::G_UDIV:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
switch (Size) {
case 32:
return RTLIB::UDIV_I32;
case 64:
return RTLIB::UDIV_I64;
case 128:
return RTLIB::UDIV_I128;
default:
llvm_unreachable("unexpected size");
}
case TargetOpcode::G_SREM:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::SREM_I64 : RTLIB::SREM_I32;
case TargetOpcode::G_UREM:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::UREM_I64 : RTLIB::UREM_I32;
case TargetOpcode::G_CTLZ_ZERO_UNDEF:
assert(Size == 32 && "Unsupported size");
return RTLIB::CTLZ_I32;
case TargetOpcode::G_FADD:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::ADD_F64 : RTLIB::ADD_F32;
case TargetOpcode::G_FSUB:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::SUB_F64 : RTLIB::SUB_F32;
case TargetOpcode::G_FMUL:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::MUL_F64 : RTLIB::MUL_F32;
case TargetOpcode::G_FDIV:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::DIV_F64 : RTLIB::DIV_F32;
case TargetOpcode::G_FEXP:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::EXP_F64 : RTLIB::EXP_F32;
case TargetOpcode::G_FEXP2:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::EXP2_F64 : RTLIB::EXP2_F32;
case TargetOpcode::G_FREM:
return Size == 64 ? RTLIB::REM_F64 : RTLIB::REM_F32;
case TargetOpcode::G_FPOW:
return Size == 64 ? RTLIB::POW_F64 : RTLIB::POW_F32;
case TargetOpcode::G_FMA:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::FMA_F64 : RTLIB::FMA_F32;
case TargetOpcode::G_FSIN:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
return Size == 128 ? RTLIB::SIN_F128
: Size == 64 ? RTLIB::SIN_F64 : RTLIB::SIN_F32;
case TargetOpcode::G_FCOS:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
return Size == 128 ? RTLIB::COS_F128
: Size == 64 ? RTLIB::COS_F64 : RTLIB::COS_F32;
case TargetOpcode::G_FLOG10:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
return Size == 128 ? RTLIB::LOG10_F128
: Size == 64 ? RTLIB::LOG10_F64 : RTLIB::LOG10_F32;
case TargetOpcode::G_FLOG:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
return Size == 128 ? RTLIB::LOG_F128
: Size == 64 ? RTLIB::LOG_F64 : RTLIB::LOG_F32;
case TargetOpcode::G_FLOG2:
assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
return Size == 128 ? RTLIB::LOG2_F128
: Size == 64 ? RTLIB::LOG2_F64 : RTLIB::LOG2_F32;
case TargetOpcode::G_FCEIL:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::CEIL_F64 : RTLIB::CEIL_F32;
case TargetOpcode::G_FFLOOR:
assert((Size == 32 || Size == 64) && "Unsupported size");
return Size == 64 ? RTLIB::FLOOR_F64 : RTLIB::FLOOR_F32;
}
llvm_unreachable("Unknown libcall function");
}
/// True if an instruction is in tail position in its caller. Intended for
/// legalizing libcalls as tail calls when possible.
static bool isLibCallInTailPosition(MachineInstr &MI) {
const Function &F = MI.getParent()->getParent()->getFunction();
// Conservatively require the attributes of the call to match those of
// the return. Ignore NoAlias and NonNull because they don't affect the
// call sequence.
AttributeList CallerAttrs = F.getAttributes();
if (AttrBuilder(CallerAttrs, AttributeList::ReturnIndex)
.removeAttribute(Attribute::NoAlias)
.removeAttribute(Attribute::NonNull)
.hasAttributes())
return false;
// It's not safe to eliminate the sign / zero extension of the return value.
if (CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt) ||
CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
return false;
// Only tail call if the following instruction is a standard return.
auto &TII = *MI.getMF()->getSubtarget().getInstrInfo();
MachineInstr *Next = MI.getNextNode();
if (!Next || TII.isTailCall(*Next) || !Next->isReturn())
return false;
return true;
}
LegalizerHelper::LegalizeResult
llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall,
const CallLowering::ArgInfo &Result,
ArrayRef<CallLowering::ArgInfo> Args) {
auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
const char *Name = TLI.getLibcallName(Libcall);
CallLowering::CallLoweringInfo Info;
Info.CallConv = TLI.getLibcallCallingConv(Libcall);
Info.Callee = MachineOperand::CreateES(Name);
Info.OrigRet = Result;
std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs));
if (!CLI.lowerCall(MIRBuilder, Info))
return LegalizerHelper::UnableToLegalize;
return LegalizerHelper::Legalized;
}
// Useful for libcalls where all operands have the same type.
static LegalizerHelper::LegalizeResult
simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size,
Type *OpType) {
auto Libcall = getRTLibDesc(MI.getOpcode(), Size);
SmallVector<CallLowering::ArgInfo, 3> Args;
for (unsigned i = 1; i < MI.getNumOperands(); i++)
Args.push_back({MI.getOperand(i).getReg(), OpType});
return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), OpType},
Args);
}
LegalizerHelper::LegalizeResult
llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
MachineInstr &MI) {
assert(MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
auto &Ctx = MIRBuilder.getMF().getFunction().getContext();
SmallVector<CallLowering::ArgInfo, 3> Args;
// Add all the args, except for the last which is an imm denoting 'tail'.
for (unsigned i = 1; i < MI.getNumOperands() - 1; i++) {
Register Reg = MI.getOperand(i).getReg();
// Need derive an IR type for call lowering.
LLT OpLLT = MRI.getType(Reg);
Type *OpTy = nullptr;
if (OpLLT.isPointer())
OpTy = Type::getInt8PtrTy(Ctx, OpLLT.getAddressSpace());
else
OpTy = IntegerType::get(Ctx, OpLLT.getSizeInBits());
Args.push_back({Reg, OpTy});
}
auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
Intrinsic::ID ID = MI.getOperand(0).getIntrinsicID();
RTLIB::Libcall RTLibcall;
switch (ID) {
case Intrinsic::memcpy:
RTLibcall = RTLIB::MEMCPY;
break;
case Intrinsic::memset:
RTLibcall = RTLIB::MEMSET;
break;
case Intrinsic::memmove:
RTLibcall = RTLIB::MEMMOVE;
break;
default:
return LegalizerHelper::UnableToLegalize;
}
const char *Name = TLI.getLibcallName(RTLibcall);
MIRBuilder.setInstr(MI);
CallLowering::CallLoweringInfo Info;
Info.CallConv = TLI.getLibcallCallingConv(RTLibcall);
Info.Callee = MachineOperand::CreateES(Name);
Info.OrigRet = CallLowering::ArgInfo({0}, Type::getVoidTy(Ctx));
Info.IsTailCall = MI.getOperand(MI.getNumOperands() - 1).getImm() == 1 &&
isLibCallInTailPosition(MI);
std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs));
if (!CLI.lowerCall(MIRBuilder, Info))
return LegalizerHelper::UnableToLegalize;
if (Info.LoweredTailCall) {
assert(Info.IsTailCall && "Lowered tail call when it wasn't a tail call?");
// We must have a return following the call to get past
// isLibCallInTailPosition.
assert(MI.getNextNode() && MI.getNextNode()->isReturn() &&
"Expected instr following MI to be a return?");
// We lowered a tail call, so the call is now the return from the block.
// Delete the old return.
MI.getNextNode()->eraseFromParent();
}
return LegalizerHelper::Legalized;
}
static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType,
Type *FromType) {
auto ToMVT = MVT::getVT(ToType);
auto FromMVT = MVT::getVT(FromType);
switch (Opcode) {
case TargetOpcode::G_FPEXT:
return RTLIB::getFPEXT(FromMVT, ToMVT);
case TargetOpcode::G_FPTRUNC:
return RTLIB::getFPROUND(FromMVT, ToMVT);
case TargetOpcode::G_FPTOSI:
return RTLIB::getFPTOSINT(FromMVT, ToMVT);
case TargetOpcode::G_FPTOUI:
return RTLIB::getFPTOUINT(FromMVT, ToMVT);
case TargetOpcode::G_SITOFP:
return RTLIB::getSINTTOFP(FromMVT, ToMVT);
case TargetOpcode::G_UITOFP:
return RTLIB::getUINTTOFP(FromMVT, ToMVT);
}
llvm_unreachable("Unsupported libcall function");
}
static LegalizerHelper::LegalizeResult
conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType,
Type *FromType) {
RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType);
return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), ToType},
{{MI.getOperand(1).getReg(), FromType}});
}
LegalizerHelper::LegalizeResult
LegalizerHelper::libcall(MachineInstr &MI) {
LLT LLTy = MRI.getType(MI.getOperand(0).getReg());
unsigned Size = LLTy.getSizeInBits();
auto &Ctx = MIRBuilder.getMF().getFunction().getContext();
MIRBuilder.setInstr(MI);
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_SDIV:
case TargetOpcode::G_UDIV:
case TargetOpcode::G_SREM:
case TargetOpcode::G_UREM:
case TargetOpcode::G_CTLZ_ZERO_UNDEF: {
Type *HLTy = IntegerType::get(Ctx, Size);
auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
if (Status != Legalized)
return Status;
break;
}
case TargetOpcode::G_FADD:
case TargetOpcode::G_FSUB:
case TargetOpcode::G_FMUL:
case TargetOpcode::G_FDIV:
case TargetOpcode::G_FMA:
case TargetOpcode::G_FPOW:
case TargetOpcode::G_FREM:
case TargetOpcode::G_FCOS:
case TargetOpcode::G_FSIN:
case TargetOpcode::G_FLOG10:
case TargetOpcode::G_FLOG:
case TargetOpcode::G_FLOG2:
case TargetOpcode::G_FEXP:
case TargetOpcode::G_FEXP2:
case TargetOpcode::G_FCEIL:
case TargetOpcode::G_FFLOOR: {
if (Size > 64) {
LLVM_DEBUG(dbgs() << "Size " << Size << " too large to legalize.\n");
return UnableToLegalize;
}
Type *HLTy = Size == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx);
auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
if (Status != Legalized)
return Status;
break;
}
case TargetOpcode::G_FPEXT: {
// FIXME: Support other floating point types (half, fp128 etc)
unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
if (ToSize != 64 || FromSize != 32)
return UnableToLegalize;
LegalizeResult Status = conversionLibcall(
MI, MIRBuilder, Type::getDoubleTy(Ctx), Type::getFloatTy(Ctx));
if (Status != Legalized)
return Status;
break;
}
case TargetOpcode::G_FPTRUNC: {
// FIXME: Support other floating point types (half, fp128 etc)
unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
if (ToSize != 32 || FromSize != 64)
return UnableToLegalize;
LegalizeResult Status = conversionLibcall(
MI, MIRBuilder, Type::getFloatTy(Ctx), Type::getDoubleTy(Ctx));
if (Status != Legalized)
return Status;
break;
}
case TargetOpcode::G_FPTOSI:
case TargetOpcode::G_FPTOUI: {
// FIXME: Support other types
unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64))
return UnableToLegalize;
LegalizeResult Status = conversionLibcall(
MI, MIRBuilder,
ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx),
FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx));
if (Status != Legalized)
return Status;
break;
}
case TargetOpcode::G_SITOFP:
case TargetOpcode::G_UITOFP: {
// FIXME: Support other types
unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64))
return UnableToLegalize;
LegalizeResult Status = conversionLibcall(
MI, MIRBuilder,
ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx),
FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx));
if (Status != Legalized)
return Status;
break;
}
}
MI.eraseFromParent();
return Legalized;
}
LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
unsigned TypeIdx,
LLT NarrowTy) {
MIRBuilder.setInstr(MI);
uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
uint64_t NarrowSize = NarrowTy.getSizeInBits();
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_IMPLICIT_DEF: {
// FIXME: add support for when SizeOp0 isn't an exact multiple of
// NarrowSize.
if (SizeOp0 % NarrowSize != 0)
return UnableToLegalize;
int NumParts = SizeOp0 / NarrowSize;
SmallVector<Register, 2> DstRegs;
for (int i = 0; i < NumParts; ++i)
DstRegs.push_back(
MIRBuilder.buildUndef(NarrowTy)->getOperand(0).getReg());
Register DstReg = MI.getOperand(0).getReg();
if(MRI.getType(DstReg).isVector())
MIRBuilder.buildBuildVector(DstReg, DstRegs);
else
MIRBuilder.buildMerge(DstReg, DstRegs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_CONSTANT: {
LLT Ty = MRI.getType(MI.getOperand(0).getReg());
const APInt &Val = MI.getOperand(1).getCImm()->getValue();
unsigned TotalSize = Ty.getSizeInBits();
unsigned NarrowSize = NarrowTy.getSizeInBits();
int NumParts = TotalSize / NarrowSize;
SmallVector<Register, 4> PartRegs;
for (int I = 0; I != NumParts; ++I) {
unsigned Offset = I * NarrowSize;
auto K = MIRBuilder.buildConstant(NarrowTy,
Val.lshr(Offset).trunc(NarrowSize));
PartRegs.push_back(K.getReg(0));
}
LLT LeftoverTy;
unsigned LeftoverBits = TotalSize - NumParts * NarrowSize;
SmallVector<Register, 1> LeftoverRegs;
if (LeftoverBits != 0) {
LeftoverTy = LLT::scalar(LeftoverBits);
auto K = MIRBuilder.buildConstant(
LeftoverTy,
Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits));
LeftoverRegs.push_back(K.getReg(0));
}
insertParts(MI.getOperand(0).getReg(),
Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_SEXT: {
if (TypeIdx != 0)
return UnableToLegalize;
Register SrcReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
// FIXME: support the general case where the requested NarrowTy may not be
// the same as the source type. E.g. s128 = sext(s32)
if ((SrcTy.getSizeInBits() != SizeOp0 / 2) ||
SrcTy.getSizeInBits() != NarrowTy.getSizeInBits()) {
LLVM_DEBUG(dbgs() << "Can't narrow sext to type " << NarrowTy << "\n");
return UnableToLegalize;
}
// Shift the sign bit of the low register through the high register.
auto ShiftAmt =
MIRBuilder.buildConstant(LLT::scalar(64), NarrowTy.getSizeInBits() - 1);
auto Shift = MIRBuilder.buildAShr(NarrowTy, SrcReg, ShiftAmt);
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {SrcReg, Shift.getReg(0)});
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_ZEXT: {
if (TypeIdx != 0)
return UnableToLegalize;
LLT SrcTy = MRI.getType(MI.getOperand(1).getReg());
uint64_t SizeOp1 = SrcTy.getSizeInBits();
if (SizeOp0 % SizeOp1 != 0)
return UnableToLegalize;
// Generate a merge where the bottom bits are taken from the source, and
// zero everything else.
Register ZeroReg = MIRBuilder.buildConstant(SrcTy, 0).getReg(0);
unsigned NumParts = SizeOp0 / SizeOp1;
SmallVector<Register, 4> Srcs = {MI.getOperand(1).getReg()};
for (unsigned Part = 1; Part < NumParts; ++Part)
Srcs.push_back(ZeroReg);
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), Srcs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_TRUNC: {
if (TypeIdx != 1)
return UnableToLegalize;
uint64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
if (NarrowTy.getSizeInBits() * 2 != SizeOp1) {
LLVM_DEBUG(dbgs() << "Can't narrow trunc to type " << NarrowTy << "\n");
return UnableToLegalize;
}
auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1).getReg());
MIRBuilder.buildCopy(MI.getOperand(0).getReg(), Unmerge.getReg(0));
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_ADD: {
// FIXME: add support for when SizeOp0 isn't an exact multiple of
// NarrowSize.
if (SizeOp0 % NarrowSize != 0)
return UnableToLegalize;
// Expand in terms of carry-setting/consuming G_ADDE instructions.
int NumParts = SizeOp0 / NarrowTy.getSizeInBits();
SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
Register CarryIn;
for (int i = 0; i < NumParts; ++i) {
Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
Register CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
if (i == 0)
MIRBuilder.buildUAddo(DstReg, CarryOut, Src1Regs[i], Src2Regs[i]);
else {
MIRBuilder.buildUAdde(DstReg, CarryOut, Src1Regs[i],
Src2Regs[i], CarryIn);
}
DstRegs.push_back(DstReg);
CarryIn = CarryOut;
}
Register DstReg = MI.getOperand(0).getReg();
if(MRI.getType(DstReg).isVector())
MIRBuilder.buildBuildVector(DstReg, DstRegs);
else
MIRBuilder.buildMerge(DstReg, DstRegs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_SUB: {
// FIXME: add support for when SizeOp0 isn't an exact multiple of
// NarrowSize.
if (SizeOp0 % NarrowSize != 0)
return UnableToLegalize;
int NumParts = SizeOp0 / NarrowTy.getSizeInBits();
SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
Register BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildInstr(TargetOpcode::G_USUBO, {DstReg, BorrowOut},
{Src1Regs[0], Src2Regs[0]});
DstRegs.push_back(DstReg);
Register BorrowIn = BorrowOut;
for (int i = 1; i < NumParts; ++i) {
DstReg = MRI.createGenericVirtualRegister(NarrowTy);
BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildInstr(TargetOpcode::G_USUBE, {DstReg, BorrowOut},
{Src1Regs[i], Src2Regs[i], BorrowIn});
DstRegs.push_back(DstReg);
BorrowIn = BorrowOut;
}
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_MUL:
case TargetOpcode::G_UMULH:
return narrowScalarMul(MI, NarrowTy);
case TargetOpcode::G_EXTRACT:
return narrowScalarExtract(MI, TypeIdx, NarrowTy);
case TargetOpcode::G_INSERT:
return narrowScalarInsert(MI, TypeIdx, NarrowTy);
case TargetOpcode::G_LOAD: {
const auto &MMO = **MI.memoperands_begin();
Register DstReg = MI.getOperand(0).getReg();
LLT DstTy = MRI.getType(DstReg);
if (DstTy.isVector())
return UnableToLegalize;
if (8 * MMO.getSize() != DstTy.getSizeInBits()) {
Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
auto &MMO = **MI.memoperands_begin();
MIRBuilder.buildLoad(TmpReg, MI.getOperand(1).getReg(), MMO);
MIRBuilder.buildAnyExt(DstReg, TmpReg);
MI.eraseFromParent();
return Legalized;
}
return reduceLoadStoreWidth(MI, TypeIdx, NarrowTy);
}
case TargetOpcode::G_ZEXTLOAD:
case TargetOpcode::G_SEXTLOAD: {
bool ZExt = MI.getOpcode() == TargetOpcode::G_ZEXTLOAD;
Register DstReg = MI.getOperand(0).getReg();
Register PtrReg = MI.getOperand(1).getReg();
Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
auto &MMO = **MI.memoperands_begin();
if (MMO.getSizeInBits() == NarrowSize) {
MIRBuilder.buildLoad(TmpReg, PtrReg, MMO);
} else {
unsigned ExtLoad = ZExt ? TargetOpcode::G_ZEXTLOAD
: TargetOpcode::G_SEXTLOAD;
MIRBuilder.buildInstr(ExtLoad)
.addDef(TmpReg)
.addUse(PtrReg)
.addMemOperand(&MMO);
}
if (ZExt)
MIRBuilder.buildZExt(DstReg, TmpReg);
else
MIRBuilder.buildSExt(DstReg, TmpReg);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_STORE: {
const auto &MMO = **MI.memoperands_begin();
Register SrcReg = MI.getOperand(0).getReg();
LLT SrcTy = MRI.getType(SrcReg);
if (SrcTy.isVector())
return UnableToLegalize;
int NumParts = SizeOp0 / NarrowSize;
unsigned HandledSize = NumParts * NarrowTy.getSizeInBits();
unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize;
if (SrcTy.isVector() && LeftoverBits != 0)
return UnableToLegalize;
if (8 * MMO.getSize() != SrcTy.getSizeInBits()) {
Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
auto &MMO = **MI.memoperands_begin();
MIRBuilder.buildTrunc(TmpReg, SrcReg);
MIRBuilder.buildStore(TmpReg, MI.getOperand(1).getReg(), MMO);
MI.eraseFromParent();
return Legalized;
}
return reduceLoadStoreWidth(MI, 0, NarrowTy);
}
case TargetOpcode::G_SELECT:
return narrowScalarSelect(MI, TypeIdx, NarrowTy);
case TargetOpcode::G_AND:
case TargetOpcode::G_OR:
case TargetOpcode::G_XOR: {
// Legalize bitwise operation:
// A = BinOp<Ty> B, C
// into:
// B1, ..., BN = G_UNMERGE_VALUES B
// C1, ..., CN = G_UNMERGE_VALUES C
// A1 = BinOp<Ty/N> B1, C2
// ...
// AN = BinOp<Ty/N> BN, CN
// A = G_MERGE_VALUES A1, ..., AN
return narrowScalarBasic(MI, TypeIdx, NarrowTy);
}
case TargetOpcode::G_SHL:
case TargetOpcode::G_LSHR:
case TargetOpcode::G_ASHR:
return narrowScalarShift(MI, TypeIdx, NarrowTy);
case TargetOpcode::G_CTLZ:
case TargetOpcode::G_CTLZ_ZERO_UNDEF:
case TargetOpcode::G_CTTZ:
case TargetOpcode::G_CTTZ_ZERO_UNDEF:
case TargetOpcode::G_CTPOP:
if (TypeIdx != 0)
return UnableToLegalize; // TODO
Observer.changingInstr(MI);
narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_INTTOPTR:
if (TypeIdx != 1)
return UnableToLegalize;
Observer.changingInstr(MI);
narrowScalarSrc(MI, NarrowTy, 1);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_PTRTOINT:
if (TypeIdx != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_PHI: {
unsigned NumParts = SizeOp0 / NarrowSize;
SmallVector<Register, 2> DstRegs;
SmallVector<SmallVector<Register, 2>, 2> SrcRegs;
DstRegs.resize(NumParts);
SrcRegs.resize(MI.getNumOperands() / 2);
Observer.changingInstr(MI);
for (unsigned i = 1; i < MI.getNumOperands(); i += 2) {
MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB();
MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts,
SrcRegs[i / 2]);
}
MachineBasicBlock &MBB = *MI.getParent();
MIRBuilder.setInsertPt(MBB, MI);
for (unsigned i = 0; i < NumParts; ++i) {
DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy);
MachineInstrBuilder MIB =
MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]);
for (unsigned j = 1; j < MI.getNumOperands(); j += 2)
MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1));
}
MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI());
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs);
Observer.changedInstr(MI);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_EXTRACT_VECTOR_ELT:
case TargetOpcode::G_INSERT_VECTOR_ELT: {
if (TypeIdx != 2)
return UnableToLegalize;
int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3;
Observer.changingInstr(MI);
narrowScalarSrc(MI, NarrowTy, OpIdx);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_ICMP: {
uint64_t SrcSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits();
if (NarrowSize * 2 != SrcSize)
return UnableToLegalize;
Observer.changingInstr(MI);
Register LHSL = MRI.createGenericVirtualRegister(NarrowTy);
Register LHSH = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildUnmerge({LHSL, LHSH}, MI.getOperand(2).getReg());
Register RHSL = MRI.createGenericVirtualRegister(NarrowTy);
Register RHSH = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildUnmerge({RHSL, RHSH}, MI.getOperand(3).getReg());
CmpInst::Predicate Pred =
static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
LLT ResTy = MRI.getType(MI.getOperand(0).getReg());
if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
MachineInstrBuilder XorL = MIRBuilder.buildXor(NarrowTy, LHSL, RHSL);
MachineInstrBuilder XorH = MIRBuilder.buildXor(NarrowTy, LHSH, RHSH);
MachineInstrBuilder Or = MIRBuilder.buildOr(NarrowTy, XorL, XorH);
MachineInstrBuilder Zero = MIRBuilder.buildConstant(NarrowTy, 0);
MIRBuilder.buildICmp(Pred, MI.getOperand(0).getReg(), Or, Zero);
} else {
MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, ResTy, LHSH, RHSH);
MachineInstrBuilder CmpHEQ =
MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, ResTy, LHSH, RHSH);
MachineInstrBuilder CmpLU = MIRBuilder.buildICmp(
ICmpInst::getUnsignedPredicate(Pred), ResTy, LHSL, RHSL);
MIRBuilder.buildSelect(MI.getOperand(0).getReg(), CmpHEQ, CmpLU, CmpH);
}
Observer.changedInstr(MI);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_SEXT_INREG: {
if (TypeIdx != 0)
return UnableToLegalize;
if (!MI.getOperand(2).isImm())
return UnableToLegalize;
int64_t SizeInBits = MI.getOperand(2).getImm();
// So long as the new type has more bits than the bits we're extending we
// don't need to break it apart.
if (NarrowTy.getScalarSizeInBits() >= SizeInBits) {
Observer.changingInstr(MI);
// We don't lose any non-extension bits by truncating the src and
// sign-extending the dst.
MachineOperand &MO1 = MI.getOperand(1);
auto TruncMIB = MIRBuilder.buildTrunc(NarrowTy, MO1.getReg());
MO1.setReg(TruncMIB->getOperand(0).getReg());
MachineOperand &MO2 = MI.getOperand(0);
Register DstExt = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
MIRBuilder.buildInstr(TargetOpcode::G_SEXT, {MO2.getReg()}, {DstExt});
MO2.setReg(DstExt);
Observer.changedInstr(MI);
return Legalized;
}
// Break it apart. Components below the extension point are unmodified. The
// component containing the extension point becomes a narrower SEXT_INREG.
// Components above it are ashr'd from the component containing the
// extension point.
if (SizeOp0 % NarrowSize != 0)
return UnableToLegalize;
int NumParts = SizeOp0 / NarrowSize;
// List the registers where the destination will be scattered.
SmallVector<Register, 2> DstRegs;
// List the registers where the source will be split.
SmallVector<Register, 2> SrcRegs;
// Create all the temporary registers.
for (int i = 0; i < NumParts; ++i) {
Register SrcReg = MRI.createGenericVirtualRegister(NarrowTy);
SrcRegs.push_back(SrcReg);
}
// Explode the big arguments into smaller chunks.
MIRBuilder.buildUnmerge(SrcRegs, MI.getOperand(1).getReg());
Register AshrCstReg =
MIRBuilder.buildConstant(NarrowTy, NarrowTy.getScalarSizeInBits() - 1)
->getOperand(0)
.getReg();
Register FullExtensionReg = 0;
Register PartialExtensionReg = 0;
// Do the operation on each small part.
for (int i = 0; i < NumParts; ++i) {
if ((i + 1) * NarrowTy.getScalarSizeInBits() < SizeInBits)
DstRegs.push_back(SrcRegs[i]);
else if (i * NarrowTy.getScalarSizeInBits() > SizeInBits) {
assert(PartialExtensionReg &&
"Expected to visit partial extension before full");
if (FullExtensionReg) {
DstRegs.push_back(FullExtensionReg);
continue;
}
DstRegs.push_back(MIRBuilder
.buildInstr(TargetOpcode::G_ASHR, {NarrowTy},
{PartialExtensionReg, AshrCstReg})
->getOperand(0)
.getReg());
FullExtensionReg = DstRegs.back();
} else {
DstRegs.push_back(
MIRBuilder
.buildInstr(
TargetOpcode::G_SEXT_INREG, {NarrowTy},
{SrcRegs[i], SizeInBits % NarrowTy.getScalarSizeInBits()})
->getOperand(0)
.getReg());
PartialExtensionReg = DstRegs.back();
}
}
// Gather the destination registers into the final destination.
Register DstReg = MI.getOperand(0).getReg();
MIRBuilder.buildMerge(DstReg, DstRegs);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_BSWAP:
case TargetOpcode::G_BITREVERSE: {
if (SizeOp0 % NarrowSize != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
SmallVector<Register, 2> SrcRegs, DstRegs;
unsigned NumParts = SizeOp0 / NarrowSize;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs);
for (unsigned i = 0; i < NumParts; ++i) {
auto DstPart = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy},
{SrcRegs[NumParts - 1 - i]});
DstRegs.push_back(DstPart.getReg(0));
}
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs);
Observer.changedInstr(MI);
MI.eraseFromParent();
return Legalized;
}
}
}
void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy,
unsigned OpIdx, unsigned ExtOpcode) {
MachineOperand &MO = MI.getOperand(OpIdx);
auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO.getReg()});
MO.setReg(ExtB->getOperand(0).getReg());
}
void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy,
unsigned OpIdx) {
MachineOperand &MO = MI.getOperand(OpIdx);
auto ExtB = MIRBuilder.buildInstr(TargetOpcode::G_TRUNC, {NarrowTy},
{MO.getReg()});
MO.setReg(ExtB->getOperand(0).getReg());
}
void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy,
unsigned OpIdx, unsigned TruncOpcode) {
MachineOperand &MO = MI.getOperand(OpIdx);
Register DstExt = MRI.createGenericVirtualRegister(WideTy);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
MIRBuilder.buildInstr(TruncOpcode, {MO.getReg()}, {DstExt});
MO.setReg(DstExt);
}
void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy,
unsigned OpIdx, unsigned ExtOpcode) {
MachineOperand &MO = MI.getOperand(OpIdx);
Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
MIRBuilder.buildInstr(ExtOpcode, {MO.getReg()}, {DstTrunc});
MO.setReg(DstTrunc);
}
void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy,
unsigned OpIdx) {
MachineOperand &MO = MI.getOperand(OpIdx);
Register DstExt = MRI.createGenericVirtualRegister(WideTy);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
MIRBuilder.buildExtract(MO.getReg(), DstExt, 0);
MO.setReg(DstExt);
}
void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy,
unsigned OpIdx) {
MachineOperand &MO = MI.getOperand(OpIdx);
LLT OldTy = MRI.getType(MO.getReg());
unsigned OldElts = OldTy.getNumElements();
unsigned NewElts = MoreTy.getNumElements();
unsigned NumParts = NewElts / OldElts;
// Use concat_vectors if the result is a multiple of the number of elements.
if (NumParts * OldElts == NewElts) {
SmallVector<Register, 8> Parts;
Parts.push_back(MO.getReg());
Register ImpDef = MIRBuilder.buildUndef(OldTy).getReg(0);
for (unsigned I = 1; I != NumParts; ++I)
Parts.push_back(ImpDef);
auto Concat = MIRBuilder.buildConcatVectors(MoreTy, Parts);
MO.setReg(Concat.getReg(0));
return;
}
Register MoreReg = MRI.createGenericVirtualRegister(MoreTy);
Register ImpDef = MIRBuilder.buildUndef(MoreTy).getReg(0);
MIRBuilder.buildInsert(MoreReg, ImpDef, MO.getReg(), 0);
MO.setReg(MoreReg);
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy) {
if (TypeIdx != 1)
return UnableToLegalize;
Register DstReg = MI.getOperand(0).getReg();
LLT DstTy = MRI.getType(DstReg);
if (DstTy.isVector())
return UnableToLegalize;
Register Src1 = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(Src1);
const int DstSize = DstTy.getSizeInBits();
const int SrcSize = SrcTy.getSizeInBits();
const int WideSize = WideTy.getSizeInBits();
const int NumMerge = (DstSize + WideSize - 1) / WideSize;
unsigned NumOps = MI.getNumOperands();
unsigned NumSrc = MI.getNumOperands() - 1;
unsigned PartSize = DstTy.getSizeInBits() / NumSrc;
if (WideSize >= DstSize) {
// Directly pack the bits in the target type.
Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1).getReg(0);
for (unsigned I = 2; I != NumOps; ++I) {
const unsigned Offset = (I - 1) * PartSize;
Register SrcReg = MI.getOperand(I).getReg();
assert(MRI.getType(SrcReg) == LLT::scalar(PartSize));
auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg);
Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg :
MRI.createGenericVirtualRegister(WideTy);
auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset);
auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt);
MIRBuilder.buildOr(NextResult, ResultReg, Shl);
ResultReg = NextResult;
}
if (WideSize > DstSize)
MIRBuilder.buildTrunc(DstReg, ResultReg);
else if (DstTy.isPointer())
MIRBuilder.buildIntToPtr(DstReg, ResultReg);
MI.eraseFromParent();
return Legalized;
}
// Unmerge the original values to the GCD type, and recombine to the next
// multiple greater than the original type.
//
// %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6
// %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0
// %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1
// %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2
// %10:_(s6) = G_MERGE_VALUES %4, %5, %6
// %11:_(s6) = G_MERGE_VALUES %7, %8, %9
// %12:_(s12) = G_MERGE_VALUES %10, %11
//
// Padding with undef if necessary:
//
// %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6
// %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0
// %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1
// %7:_(s2) = G_IMPLICIT_DEF
// %8:_(s6) = G_MERGE_VALUES %3, %4, %5
// %9:_(s6) = G_MERGE_VALUES %6, %7, %7
// %10:_(s12) = G_MERGE_VALUES %8, %9
const int GCD = greatestCommonDivisor(SrcSize, WideSize);
LLT GCDTy = LLT::scalar(GCD);
SmallVector<Register, 8> Parts;
SmallVector<Register, 8> NewMergeRegs;
SmallVector<Register, 8> Unmerges;
LLT WideDstTy = LLT::scalar(NumMerge * WideSize);
// Decompose the original operands if they don't evenly divide.
for (int I = 1, E = MI.getNumOperands(); I != E; ++I) {
Register SrcReg = MI.getOperand(I).getReg();
if (GCD == SrcSize) {
Unmerges.push_back(SrcReg);
} else {
auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg);
for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J)
Unmerges.push_back(Unmerge.getReg(J));
}
}
// Pad with undef to the next size that is a multiple of the requested size.
if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) {
Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0);
for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I)
Unmerges.push_back(UndefReg);
}
const int PartsPerGCD = WideSize / GCD;
// Build merges of each piece.
ArrayRef<Register> Slicer(Unmerges);
for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) {
auto Merge = MIRBuilder.buildMerge(WideTy, Slicer.take_front(PartsPerGCD));
NewMergeRegs.push_back(Merge.getReg(0));
}
// A truncate may be necessary if the requested type doesn't evenly divide the
// original result type.
if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) {
MIRBuilder.buildMerge(DstReg, NewMergeRegs);
} else {
auto FinalMerge = MIRBuilder.buildMerge(WideDstTy, NewMergeRegs);
MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0));
}
MI.eraseFromParent();
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy) {
if (TypeIdx != 0)
return UnableToLegalize;
unsigned NumDst = MI.getNumOperands() - 1;
Register SrcReg = MI.getOperand(NumDst).getReg();
LLT SrcTy = MRI.getType(SrcReg);
if (!SrcTy.isScalar())
return UnableToLegalize;
Register Dst0Reg = MI.getOperand(0).getReg();
LLT DstTy = MRI.getType(Dst0Reg);
if (!DstTy.isScalar())
return UnableToLegalize;
unsigned NewSrcSize = NumDst * WideTy.getSizeInBits();
LLT NewSrcTy = LLT::scalar(NewSrcSize);
unsigned SizeDiff = WideTy.getSizeInBits() - DstTy.getSizeInBits();
auto WideSrc = MIRBuilder.buildZExt(NewSrcTy, SrcReg);
for (unsigned I = 1; I != NumDst; ++I) {
auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, SizeDiff * I);
auto Shl = MIRBuilder.buildShl(NewSrcTy, WideSrc, ShiftAmt);
WideSrc = MIRBuilder.buildOr(NewSrcTy, WideSrc, Shl);
}
Observer.changingInstr(MI);
MI.getOperand(NumDst).setReg(WideSrc->getOperand(0).getReg());
for (unsigned I = 0; I != NumDst; ++I)
widenScalarDst(MI, WideTy, I);
Observer.changedInstr(MI);
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy) {
Register DstReg = MI.getOperand(0).getReg();
Register SrcReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
LLT DstTy = MRI.getType(DstReg);
unsigned Offset = MI.getOperand(2).getImm();
if (TypeIdx == 0) {
if (SrcTy.isVector() || DstTy.isVector())
return UnableToLegalize;
SrcOp Src(SrcReg);
if (SrcTy.isPointer()) {
// Extracts from pointers can be handled only if they are really just
// simple integers.
const DataLayout &DL = MIRBuilder.getDataLayout();
if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace()))
return UnableToLegalize;
LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits());
Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src);
SrcTy = SrcAsIntTy;
}
if (DstTy.isPointer())
return UnableToLegalize;
if (Offset == 0) {
// Avoid a shift in the degenerate case.
MIRBuilder.buildTrunc(DstReg,
MIRBuilder.buildAnyExtOrTrunc(WideTy, Src));
MI.eraseFromParent();
return Legalized;
}
// Do a shift in the source type.
LLT ShiftTy = SrcTy;
if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) {
Src = MIRBuilder.buildAnyExt(WideTy, Src);
ShiftTy = WideTy;
} else if (WideTy.getSizeInBits() > SrcTy.getSizeInBits())
return UnableToLegalize;
auto LShr = MIRBuilder.buildLShr(
ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset));
MIRBuilder.buildTrunc(DstReg, LShr);
MI.eraseFromParent();
return Legalized;
}
if (SrcTy.isScalar()) {
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
Observer.changedInstr(MI);
return Legalized;
}
if (!SrcTy.isVector())
return UnableToLegalize;
if (DstTy != SrcTy.getElementType())
return UnableToLegalize;
if (Offset % SrcTy.getScalarSizeInBits() != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) *
Offset);
widenScalarDst(MI, WideTy.getScalarType(), 0);
Observer.changedInstr(MI);
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy) {
if (TypeIdx != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
MIRBuilder.setInstr(MI);
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_EXTRACT:
return widenScalarExtract(MI, TypeIdx, WideTy);
case TargetOpcode::G_INSERT:
return widenScalarInsert(MI, TypeIdx, WideTy);
case TargetOpcode::G_MERGE_VALUES:
return widenScalarMergeValues(MI, TypeIdx, WideTy);
case TargetOpcode::G_UNMERGE_VALUES:
return widenScalarUnmergeValues(MI, TypeIdx, WideTy);
case TargetOpcode::G_UADDO:
case TargetOpcode::G_USUBO: {
if (TypeIdx == 1)
return UnableToLegalize; // TODO
auto LHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy},
{MI.getOperand(2).getReg()});
auto RHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy},
{MI.getOperand(3).getReg()});
unsigned Opcode = MI.getOpcode() == TargetOpcode::G_UADDO
? TargetOpcode::G_ADD
: TargetOpcode::G_SUB;
// Do the arithmetic in the larger type.
auto NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSZext, RHSZext});
LLT OrigTy = MRI.getType(MI.getOperand(0).getReg());
APInt Mask = APInt::getAllOnesValue(OrigTy.getSizeInBits());
auto AndOp = MIRBuilder.buildInstr(
TargetOpcode::G_AND, {WideTy},
{NewOp, MIRBuilder.buildConstant(WideTy, Mask.getZExtValue())});
// There is no overflow if the AndOp is the same as NewOp.
MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1).getReg(), NewOp,
AndOp);
// Now trunc the NewOp to the original result.
MIRBuilder.buildTrunc(MI.getOperand(0).getReg(), NewOp);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_CTTZ:
case TargetOpcode::G_CTTZ_ZERO_UNDEF:
case TargetOpcode::G_CTLZ:
case TargetOpcode::G_CTLZ_ZERO_UNDEF:
case TargetOpcode::G_CTPOP: {
if (TypeIdx == 0) {
Observer.changingInstr(MI);
widenScalarDst(MI, WideTy, 0);
Observer.changedInstr(MI);
return Legalized;
}
Register SrcReg = MI.getOperand(1).getReg();
// First ZEXT the input.
auto MIBSrc = MIRBuilder.buildZExt(WideTy, SrcReg);
LLT CurTy = MRI.getType(SrcReg);
if (MI.getOpcode() == TargetOpcode::G_CTTZ) {
// The count is the same in the larger type except if the original
// value was zero. This can be handled by setting the bit just off
// the top of the original type.
auto TopBit =
APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits());
MIBSrc = MIRBuilder.buildOr(
WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit));
}
// Perform the operation at the larger size.
auto MIBNewOp = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, {MIBSrc});
// This is already the correct result for CTPOP and CTTZs
if (MI.getOpcode() == TargetOpcode::G_CTLZ ||
MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) {
// The correct result is NewOp - (Difference in widety and current ty).
unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits();
MIBNewOp = MIRBuilder.buildInstr(
TargetOpcode::G_SUB, {WideTy},
{MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff)});
}
MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_BSWAP: {
Observer.changingInstr(MI);
Register DstReg = MI.getOperand(0).getReg();
Register ShrReg = MRI.createGenericVirtualRegister(WideTy);
Register DstExt = MRI.createGenericVirtualRegister(WideTy);
Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
MI.getOperand(0).setReg(DstExt);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
LLT Ty = MRI.getType(DstReg);
unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits();
MIRBuilder.buildConstant(ShiftAmtReg, DiffBits);
MIRBuilder.buildInstr(TargetOpcode::G_LSHR)
.addDef(ShrReg)
.addUse(DstExt)
.addUse(ShiftAmtReg);
MIRBuilder.buildTrunc(DstReg, ShrReg);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_BITREVERSE: {
Observer.changingInstr(MI);
Register DstReg = MI.getOperand(0).getReg();
LLT Ty = MRI.getType(DstReg);
unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits();
Register DstExt = MRI.createGenericVirtualRegister(WideTy);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
MI.getOperand(0).setReg(DstExt);
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
auto ShiftAmt = MIRBuilder.buildConstant(WideTy, DiffBits);
auto Shift = MIRBuilder.buildLShr(WideTy, DstExt, ShiftAmt);
MIRBuilder.buildTrunc(DstReg, Shift);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_ADD:
case TargetOpcode::G_AND:
case TargetOpcode::G_MUL:
case TargetOpcode::G_OR:
case TargetOpcode::G_XOR:
case TargetOpcode::G_SUB:
// Perform operation at larger width (any extension is fines here, high bits
// don't affect the result) and then truncate the result back to the
// original type.
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_SHL:
Observer.changingInstr(MI);
if (TypeIdx == 0) {
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy);
} else {
assert(TypeIdx == 1);
// The "number of bits to shift" operand must preserve its value as an
// unsigned integer:
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
}
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_SDIV:
case TargetOpcode::G_SREM:
case TargetOpcode::G_SMIN:
case TargetOpcode::G_SMAX:
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_ASHR:
case TargetOpcode::G_LSHR:
Observer.changingInstr(MI);
if (TypeIdx == 0) {
unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ?
TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
widenScalarSrc(MI, WideTy, 1, CvtOp);
widenScalarDst(MI, WideTy);
} else {
assert(TypeIdx == 1);
// The "number of bits to shift" operand must preserve its value as an
// unsigned integer:
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
}
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_UDIV:
case TargetOpcode::G_UREM:
case TargetOpcode::G_UMIN:
case TargetOpcode::G_UMAX:
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_SELECT:
Observer.changingInstr(MI);
if (TypeIdx == 0) {
// Perform operation at larger width (any extension is fine here, high
// bits don't affect the result) and then truncate the result back to the
// original type.
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy);
} else {
bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector();
// Explicit extension is required here since high bits affect the result.
widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false));
}
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_FPTOSI:
case TargetOpcode::G_FPTOUI:
Observer.changingInstr(MI);
if (TypeIdx == 0)
widenScalarDst(MI, WideTy);
else
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_SITOFP:
if (TypeIdx != 1)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_UITOFP:
if (TypeIdx != 1)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_LOAD:
case TargetOpcode::G_SEXTLOAD:
case TargetOpcode::G_ZEXTLOAD:
Observer.changingInstr(MI);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_STORE: {
if (TypeIdx != 0)
return UnableToLegalize;
LLT Ty = MRI.getType(MI.getOperand(0).getReg());
if (!isPowerOf2_32(Ty.getSizeInBits()))
return UnableToLegalize;
Observer.changingInstr(MI);
unsigned ExtType = Ty.getScalarSizeInBits() == 1 ?
TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT;
widenScalarSrc(MI, WideTy, 0, ExtType);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_CONSTANT: {
MachineOperand &SrcMO = MI.getOperand(1);
LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
unsigned ExtOpc = LI.getExtOpcodeForWideningConstant(
MRI.getType(MI.getOperand(0).getReg()));
assert((ExtOpc == TargetOpcode::G_ZEXT || ExtOpc == TargetOpcode::G_SEXT ||
ExtOpc == TargetOpcode::G_ANYEXT) &&
"Illegal Extend");
const APInt &SrcVal = SrcMO.getCImm()->getValue();
const APInt &Val = (ExtOpc == TargetOpcode::G_SEXT)
? SrcVal.sext(WideTy.getSizeInBits())
: SrcVal.zext(WideTy.getSizeInBits());
Observer.changingInstr(MI);
SrcMO.setCImm(ConstantInt::get(Ctx, Val));
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_FCONSTANT: {
MachineOperand &SrcMO = MI.getOperand(1);
LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
APFloat Val = SrcMO.getFPImm()->getValueAPF();
bool LosesInfo;
switch (WideTy.getSizeInBits()) {
case 32:
Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
&LosesInfo);
break;
case 64:
Val.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
&LosesInfo);
break;
default:
return UnableToLegalize;
}
assert(!LosesInfo && "extend should always be lossless");
Observer.changingInstr(MI);
SrcMO.setFPImm(ConstantFP::get(Ctx, Val));
widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_IMPLICIT_DEF: {
Observer.changingInstr(MI);
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_BRCOND:
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false));
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_FCMP:
Observer.changingInstr(MI);
if (TypeIdx == 0)
widenScalarDst(MI, WideTy);
else {
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT);
widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT);
}
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_ICMP:
Observer.changingInstr(MI);
if (TypeIdx == 0)
widenScalarDst(MI, WideTy);
else {
unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>(
MI.getOperand(1).getPredicate()))
? TargetOpcode::G_SEXT
: TargetOpcode::G_ZEXT;
widenScalarSrc(MI, WideTy, 2, ExtOpcode);
widenScalarSrc(MI, WideTy, 3, ExtOpcode);
}
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_PTR_ADD:
assert(TypeIdx == 1 && "unable to legalize pointer of G_PTR_ADD");
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_PHI: {
assert(TypeIdx == 0 && "Expecting only Idx 0");
Observer.changingInstr(MI);
for (unsigned I = 1; I < MI.getNumOperands(); I += 2) {
MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB();
MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT);
}
MachineBasicBlock &MBB = *MI.getParent();
MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI());
widenScalarDst(MI, WideTy);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_EXTRACT_VECTOR_ELT: {
if (TypeIdx == 0) {
Register VecReg = MI.getOperand(1).getReg();
LLT VecTy = MRI.getType(VecReg);
Observer.changingInstr(MI);
widenScalarSrc(MI, LLT::vector(VecTy.getNumElements(),
WideTy.getSizeInBits()),
1, TargetOpcode::G_SEXT);
widenScalarDst(MI, WideTy, 0);
Observer.changedInstr(MI);
return Legalized;
}
if (TypeIdx != 2)
return UnableToLegalize;
Observer.changingInstr(MI);
// TODO: Probably should be zext
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_INSERT_VECTOR_ELT: {
if (TypeIdx == 1) {
Observer.changingInstr(MI);
Register VecReg = MI.getOperand(1).getReg();
LLT VecTy = MRI.getType(VecReg);
LLT WideVecTy = LLT::vector(VecTy.getNumElements(), WideTy);
widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_ANYEXT);
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideVecTy, 0);
Observer.changedInstr(MI);
return Legalized;
}
if (TypeIdx == 2) {
Observer.changingInstr(MI);
// TODO: Probably should be zext
widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT);
Observer.changedInstr(MI);
}
return Legalized;
}
case TargetOpcode::G_FADD:
case TargetOpcode::G_FMUL:
case TargetOpcode::G_FSUB:
case TargetOpcode::G_FMA:
case TargetOpcode::G_FMAD:
case TargetOpcode::G_FNEG:
case TargetOpcode::G_FABS:
case TargetOpcode::G_FCANONICALIZE:
case TargetOpcode::G_FMINNUM:
case TargetOpcode::G_FMAXNUM:
case TargetOpcode::G_FMINNUM_IEEE:
case TargetOpcode::G_FMAXNUM_IEEE:
case TargetOpcode::G_FMINIMUM:
case TargetOpcode::G_FMAXIMUM:
case TargetOpcode::G_FDIV:
case TargetOpcode::G_FREM:
case TargetOpcode::G_FCEIL:
case TargetOpcode::G_FFLOOR:
case TargetOpcode::G_FCOS:
case TargetOpcode::G_FSIN:
case TargetOpcode::G_FLOG10:
case TargetOpcode::G_FLOG:
case TargetOpcode::G_FLOG2:
case TargetOpcode::G_FRINT:
case TargetOpcode::G_FNEARBYINT:
case TargetOpcode::G_FSQRT:
case TargetOpcode::G_FEXP:
case TargetOpcode::G_FEXP2:
case TargetOpcode::G_FPOW:
case TargetOpcode::G_INTRINSIC_TRUNC:
case TargetOpcode::G_INTRINSIC_ROUND:
assert(TypeIdx == 0);
Observer.changingInstr(MI);
for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I)
widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT);
widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_INTTOPTR:
if (TypeIdx != 1)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_PTRTOINT:
if (TypeIdx != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarDst(MI, WideTy, 0);
Observer.changedInstr(MI);
return Legalized;
case TargetOpcode::G_BUILD_VECTOR: {
Observer.changingInstr(MI);
const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType();
for (int I = 1, E = MI.getNumOperands(); I != E; ++I)
widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT);
// Avoid changing the result vector type if the source element type was
// requested.
if (TypeIdx == 1) {
auto &TII = *MI.getMF()->getSubtarget().getInstrInfo();
MI.setDesc(TII.get(TargetOpcode::G_BUILD_VECTOR_TRUNC));
} else {
widenScalarDst(MI, WideTy, 0);
}
Observer.changedInstr(MI);
return Legalized;
}
case TargetOpcode::G_SEXT_INREG:
if (TypeIdx != 0)
return UnableToLegalize;
Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy, 0, TargetOpcode::G_TRUNC);
Observer.changedInstr(MI);
return Legalized;
}
}
LegalizerHelper::LegalizeResult
LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
using namespace TargetOpcode;
MIRBuilder.setInstr(MI);
switch(MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_SREM:
case TargetOpcode::G_UREM: {
Register QuotReg = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV)
.addDef(QuotReg)
.addUse(MI.getOperand(1).getReg())
.addUse(MI.getOperand(2).getReg());
Register ProdReg = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildMul(ProdReg, QuotReg, MI.getOperand(2).getReg());
MIRBuilder.buildSub(MI.getOperand(0).getReg(), MI.getOperand(1).getReg(),
ProdReg);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_SADDO:
case TargetOpcode::G_SSUBO:
return lowerSADDO_SSUBO(MI);
case TargetOpcode::G_SMULO:
case TargetOpcode::G_UMULO: {
// Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the
// result.
Register Res = MI.getOperand(0).getReg();
Register Overflow = MI.getOperand(1).getReg();
Register LHS = MI.getOperand(2).getReg();
Register RHS = MI.getOperand(3).getReg();
MIRBuilder.buildMul(Res, LHS, RHS);
unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO
? TargetOpcode::G_SMULH
: TargetOpcode::G_UMULH;
Register HiPart = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildInstr(Opcode)
.addDef(HiPart)
.addUse(LHS)
.addUse(RHS);
Register Zero = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildConstant(Zero, 0);
// For *signed* multiply, overflow is detected by checking:
// (hi != (lo >> bitwidth-1))
if (Opcode == TargetOpcode::G_SMULH) {
Register Shifted = MRI.createGenericVirtualRegister(Ty);
Register ShiftAmt = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildConstant(ShiftAmt, Ty.getSizeInBits() - 1);
MIRBuilder.buildInstr(TargetOpcode::G_ASHR)
.addDef(Shifted)
.addUse(Res)
.addUse(ShiftAmt);
MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted);
} else {
MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero);
}
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_FNEG: {
// TODO: Handle vector types once we are able to
// represent them.
if (Ty.isVector())
return UnableToLegalize;
Register Res = MI.getOperand(0).getReg();
Type *ZeroTy;
LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
switch (Ty.getSizeInBits()) {
case 16:
ZeroTy = Type::getHalfTy(Ctx);
break;
case 32:
ZeroTy = Type::getFloatTy(Ctx);
break;
case 64:
ZeroTy = Type::getDoubleTy(Ctx);
break;
case 128:
ZeroTy = Type::getFP128Ty(Ctx);
break;
default:
llvm_unreachable("unexpected floating-point type");
}
ConstantFP &ZeroForNegation =
*cast<ConstantFP>(ConstantFP::getZeroValueForNegation(ZeroTy));
auto Zero = MIRBuilder.buildFConstant(Ty, ZeroForNegation);
Register SubByReg = MI.getOperand(1).getReg();
Register ZeroReg = Zero->getOperand(0).getReg();
MIRBuilder.buildInstr(TargetOpcode::G_FSUB, {Res}, {ZeroReg, SubByReg},
MI.getFlags());
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_FSUB: {
// Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)).
// First, check if G_FNEG is marked as Lower. If so, we may
// end up with an infinite loop as G_FSUB is used to legalize G_FNEG.
if (LI.getAction({G_FNEG, {Ty}}).Action == Lower)
return UnableToLegalize;
Register Res = MI.getOperand(0).getReg();
Register LHS = MI.getOperand(1).getReg();
Register RHS = MI.getOperand(2).getReg();
Register Neg = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildInstr(TargetOpcode::G_FNEG).addDef(Neg).addUse(RHS);
MIRBuilder.buildInstr(TargetOpcode::G_FADD, {Res}, {LHS, Neg}, MI.getFlags());
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_FMAD:
return lowerFMad(MI);
case TargetOpcode::G_INTRINSIC_ROUND:
return lowerIntrinsicRound(MI);
case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
Register OldValRes = MI.getOperand(0).getReg();
Register SuccessRes = MI.getOperand(1).getReg();
Register Addr = MI.getOperand(2).getReg();
Register CmpVal = MI.getOperand(3).getReg();
Register NewVal = MI.getOperand(4).getReg();
MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal,
**MI.memoperands_begin());
MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_LOAD:
case TargetOpcode::G_SEXTLOAD:
case TargetOpcode::G_ZEXTLOAD: {
// Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT
Register DstReg = MI.getOperand(0).getReg();
Register PtrReg = MI.getOperand(1).getReg();
LLT DstTy = MRI.getType(DstReg);
auto &MMO = **MI.memoperands_begin();
if (DstTy.getSizeInBits() == MMO.getSizeInBits()) {
if (MI.getOpcode() == TargetOpcode::G_LOAD) {
// This load needs splitting into power of 2 sized loads.
if (DstTy.isVector())
return UnableToLegalize;
if (isPowerOf2_32(DstTy.getSizeInBits()))
return UnableToLegalize; // Don't know what we're being asked to do.
// Our strategy here is to generate anyextending loads for the smaller
// types up to next power-2 result type, and then combine the two larger
// result values together, before truncating back down to the non-pow-2
// type.
// E.g. v1 = i24 load =>
// v2 = i32 load (2 byte)
// v3 = i32 load (1 byte)
// v4 = i32 shl v3, 16
// v5 = i32 or v4, v2
// v1 = i24 trunc v5
// By doing this we generate the correct truncate which should get
// combined away as an artifact with a matching extend.
uint64_t LargeSplitSize = PowerOf2Floor(DstTy.getSizeInBits());
uint64_t SmallSplitSize = DstTy.getSizeInBits() - LargeSplitSize;
MachineFunction &MF = MIRBuilder.getMF();
MachineMemOperand *LargeMMO =
MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
MachineMemOperand *SmallMMO = MF.getMachineMemOperand(
&MMO, LargeSplitSize / 8, SmallSplitSize / 8);
LLT PtrTy = MRI.getType(PtrReg);
unsigned AnyExtSize = NextPowerOf2(DstTy.getSizeInBits());
LLT AnyExtTy = LLT::scalar(AnyExtSize);
Register LargeLdReg = MRI.createGenericVirtualRegister(AnyExtTy);
Register SmallLdReg = MRI.createGenericVirtualRegister(AnyExtTy);
auto LargeLoad =
MIRBuilder.buildLoad(LargeLdReg, PtrReg, *LargeMMO);
auto OffsetCst =
MIRBuilder.buildConstant(LLT::scalar(64), LargeSplitSize / 8);
Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy);
auto SmallPtr =
MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst.getReg(0));
auto SmallLoad = MIRBuilder.buildLoad(SmallLdReg, SmallPtr.getReg(0),
*SmallMMO);
auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize);
auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt);
auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad);
MIRBuilder.buildTrunc(DstReg, {Or.getReg(0)});
MI.eraseFromParent();
return Legalized;
}
MIRBuilder.buildLoad(DstReg, PtrReg, MMO);
MI.eraseFromParent();
return Legalized;
}
if (DstTy.isScalar()) {
Register TmpReg =
MRI.createGenericVirtualRegister(LLT::scalar(MMO.getSizeInBits()));
MIRBuilder.buildLoad(TmpReg, PtrReg, MMO);
switch (MI.getOpcode()) {
default:
llvm_unreachable("Unexpected opcode");
case TargetOpcode::G_LOAD:
MIRBuilder.buildExtOrTrunc(TargetOpcode::G_ANYEXT, DstReg, TmpReg);
break;
case TargetOpcode::G_SEXTLOAD:
MIRBuilder.buildSExt(DstReg, TmpReg);
break;
case TargetOpcode::G_ZEXTLOAD:
MIRBuilder.buildZExt(DstReg, TmpReg);
break;
}
MI.eraseFromParent();
return Legalized;
}
return UnableToLegalize;
}
case TargetOpcode::G_STORE: {
// Lower a non-power of 2 store into multiple pow-2 stores.
// E.g. split an i24 store into an i16 store + i8 store.
// We do this by first extending the stored value to the next largest power
// of 2 type, and then using truncating stores to store the components.
// By doing this, likewise with G_LOAD, generate an extend that can be
// artifact-combined away instead of leaving behind extracts.
Register SrcReg = MI.getOperand(0).getReg();
Register PtrReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
MachineMemOperand &MMO = **MI.memoperands_begin();
if (SrcTy.getSizeInBits() != MMO.getSizeInBits())
return UnableToLegalize;
if (SrcTy.isVector())
return UnableToLegalize;
if (isPowerOf2_32(SrcTy.getSizeInBits()))
return UnableToLegalize; // Don't know what we're being asked to do.
// Extend to the next pow-2.
const LLT ExtendTy = LLT::scalar(NextPowerOf2(SrcTy.getSizeInBits()));
auto ExtVal = MIRBuilder.buildAnyExt(ExtendTy, SrcReg);
// Obtain the smaller value by shifting away the larger value.
uint64_t LargeSplitSize = PowerOf2Floor(SrcTy.getSizeInBits());
uint64_t SmallSplitSize = SrcTy.getSizeInBits() - LargeSplitSize;
auto ShiftAmt = MIRBuilder.buildConstant(ExtendTy, LargeSplitSize);
auto SmallVal = MIRBuilder.buildLShr(ExtendTy, ExtVal, ShiftAmt);
// Generate the PtrAdd and truncating stores.
LLT PtrTy = MRI.getType(PtrReg);
auto OffsetCst =
MIRBuilder.buildConstant(LLT::scalar(64), LargeSplitSize / 8);
Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy);
auto SmallPtr =
MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst.getReg(0));
MachineFunction &MF = MIRBuilder.getMF();
MachineMemOperand *LargeMMO =
MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
MachineMemOperand *SmallMMO =
MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8);
MIRBuilder.buildStore(ExtVal.getReg(0), PtrReg, *LargeMMO);
MIRBuilder.buildStore(SmallVal.getReg(0), SmallPtr.getReg(0), *SmallMMO);
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_CTLZ_ZERO_UNDEF:
case TargetOpcode::G_CTTZ_ZERO_UNDEF:
case TargetOpcode::G_CTLZ:
case TargetOpcode::G_CTTZ:
case TargetOpcode::G_CTPOP:
return lowerBitCount(MI, TypeIdx, Ty);
case G_UADDO: {
Register Res = MI.getOperand(0).getReg();
Register CarryOut = MI.getOperand(1).getReg();
Register LHS = MI.getOperand(2).getReg();
Register RHS = MI.getOperand(3).getReg();
MIRBuilder.buildAdd(Res, LHS, RHS);
MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS);
MI.eraseFromParent();
return Legalized;
}
case G_UADDE: {
Register Res = MI.getOperand(0).getReg();
Register CarryOut = MI.getOperand(1).getReg();
Register LHS = MI.getOperand(2).getReg();
Register RHS = MI.getOperand(3).getReg();
Register CarryIn = MI.getOperand(4).getReg();
Register TmpRes = MRI.createGenericVirtualRegister(Ty);
Register ZExtCarryIn = MRI.createGenericVirtualRegister(Ty);
MIRBuilder.buildAdd(TmpRes, LHS, RHS);
MIRBuilder.buildZExt(ZExtCarryIn, CarryIn);
MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn);
MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, LHS);
MI.eraseFromParent();
return Legalized;
}
case G_USUBO: {
Register Res = MI.getOperand(0).getReg();
Register BorrowOut = MI.getOperand(1).getReg();
Register LHS = MI.getOperand(2).getReg();
Register RHS = MI.getOperand(3).getReg();
MIRBuilder.buildSub(Res, LHS, RHS);
MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS);
MI.eraseFromParent();
return Legalized;
}
case G_USUBE: {
Register Res = MI.getOperand(0).getReg();
Register BorrowOut = MI.getOperand(1).getReg();
Register LHS = MI.getOperand(2).getReg();
Register RHS = MI.getOperand(3).getReg();
Register BorrowIn = MI.getOperand(4).getReg();
Register TmpRes = MRI.createGenericVirtualRegister(Ty);
Register ZExtBorrowIn = MRI.createGenericVirtualRegister(Ty);
Register LHS_EQ_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1));
Register LHS_ULT_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildSub(TmpRes, LHS, RHS);
MIRBuilder.buildZExt(ZExtBorrowIn, BorrowIn);
MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn);
MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LHS_EQ_RHS, LHS, RHS);
MIRBuilder.buildICmp(CmpInst::ICMP_ULT, LHS_ULT_RHS, LHS, RHS);
MIRBuilder.buildSelect(BorrowOut, LHS_EQ_RHS, BorrowIn, LHS_ULT_RHS);
MI.eraseFromParent();
return Legalized;
}
case G_UITOFP:
return lowerUITOFP(MI, TypeIdx, Ty);
case G_SITOFP:
return lowerSITOFP(MI, TypeIdx, Ty);
case G_FPTOUI:
return lowerFPTOUI(MI, TypeIdx, Ty);
case G_SMIN:
case G_SMAX:
case G_UMIN:
case G_UMAX:
return lowerMinMax(MI, TypeIdx, Ty);
case G_FCOPYSIGN:
return lowerFCopySign(MI, TypeIdx, Ty);
case G_FMINNUM:
case G_FMAXNUM:
return lowerFMinNumMaxNum(MI);
case G_UNMERGE_VALUES:
return lowerUnmergeValues(MI);
case TargetOpcode::G_SEXT_INREG: {
assert(MI.getOperand(2).isImm() && "Expected immediate");
int64_t SizeInBits = MI.getOperand(2).getImm();
Register DstReg = MI.getOperand(0).getReg();
Register SrcReg = MI.getOperand(1).getReg();
LLT DstTy = MRI.getType(DstReg);
Register TmpRes = MRI.createGenericVirtualRegister(DstTy);
auto MIBSz = MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - SizeInBits);
MIRBuilder.buildInstr(TargetOpcode::G_SHL, {TmpRes}, {SrcReg, MIBSz->getOperand(0).getReg()});
MIRBuilder.buildInstr(TargetOpcode::G_ASHR, {DstReg}, {TmpRes, MIBSz->getOperand(0).getReg()});
MI.eraseFromParent();
return Legalized;
}
case G_SHUFFLE_VECTOR:
return lowerShuffleVector(MI);
case G_DYN_STACKALLOC:
return lowerDynStackAlloc(MI);
case G_EXTRACT:
return lowerExtract(MI);
case G_INSERT:
return lowerInsert(MI);
case G_BSWAP:
return lowerBswap(MI);
case G_BITREVERSE:
return lowerBitreverse(MI);
case G_READ_REGISTER:
return lowerReadRegister(MI);
}
}
LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorImplicitDef(
MachineInstr &MI, unsigned TypeIdx, LLT NarrowTy) {
SmallVector<Register, 2> DstRegs;
unsigned NarrowSize = NarrowTy.getSizeInBits();
Register DstReg = MI.getOperand(0).getReg();
unsigned Size = MRI.getType(DstReg).getSizeInBits();
int NumParts = Size / NarrowSize;
// FIXME: Don't know how to handle the situation where the small vectors
// aren't all the same size yet.
if (Size % NarrowSize != 0)
return UnableToLegalize;
for (int i = 0; i < NumParts; ++i) {
Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildUndef(TmpReg);
DstRegs.push_back(TmpReg);
}
if (NarrowTy.isVector())
MIRBuilder.buildConcatVectors(DstReg, DstRegs);
else
MIRBuilder.buildBuildVector(DstReg, DstRegs);
MI.eraseFromParent();
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorBasic(MachineInstr &MI, unsigned TypeIdx,
LLT NarrowTy) {
const unsigned Opc = MI.getOpcode();
const unsigned NumOps = MI.getNumOperands() - 1;
const unsigned NarrowSize = NarrowTy.getSizeInBits();
const Register DstReg = MI.getOperand(0).getReg();
const unsigned Flags = MI.getFlags();
const LLT DstTy = MRI.getType(DstReg);
const unsigned Size = DstTy.getSizeInBits();
const int NumParts = Size / NarrowSize;
const LLT EltTy = DstTy.getElementType();
const unsigned EltSize = EltTy.getSizeInBits();
const unsigned BitsForNumParts = NarrowSize * NumParts;
// Check if we have any leftovers. If we do, then only handle the case where
// the leftover is one element.
if (BitsForNumParts != Size && BitsForNumParts + EltSize != Size)
return UnableToLegalize;
if (BitsForNumParts != Size) {
Register AccumDstReg = MRI.createGenericVirtualRegister(DstTy);
MIRBuilder.buildUndef(AccumDstReg);
// Handle the pieces which evenly divide into the requested type with
// extract/op/insert sequence.
for (unsigned Offset = 0; Offset < BitsForNumParts; Offset += NarrowSize) {
SmallVector<SrcOp, 4> SrcOps;
for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) {
Register PartOpReg = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(), Offset);
SrcOps.push_back(PartOpReg);
}
Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags);
Register PartInsertReg = MRI.createGenericVirtualRegister(DstTy);
MIRBuilder.buildInsert(PartInsertReg, AccumDstReg, PartDstReg, Offset);
AccumDstReg = PartInsertReg;
}
// Handle the remaining element sized leftover piece.
SmallVector<SrcOp, 4> SrcOps;
for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) {
Register PartOpReg = MRI.createGenericVirtualRegister(EltTy);
MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(),
BitsForNumParts);
SrcOps.push_back(PartOpReg);
}
Register PartDstReg = MRI.createGenericVirtualRegister(EltTy);
MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags);
MIRBuilder.buildInsert(DstReg, AccumDstReg, PartDstReg, BitsForNumParts);
MI.eraseFromParent();
return Legalized;
}
SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src0Regs);
if (NumOps >= 2)
extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src1Regs);
if (NumOps >= 3)
extractParts(MI.getOperand(3).getReg(), NarrowTy, NumParts, Src2Regs);
for (int i = 0; i < NumParts; ++i) {
Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
if (NumOps == 1)
MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i]}, Flags);
else if (NumOps == 2) {
MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i], Src1Regs[i]}, Flags);
} else if (NumOps == 3) {
MIRBuilder.buildInstr(Opc, {DstReg},
{Src0Regs[i], Src1Regs[i], Src2Regs[i]}, Flags);
}
DstRegs.push_back(DstReg);
}
if (NarrowTy.isVector())
MIRBuilder.buildConcatVectors(DstReg, DstRegs);
else
MIRBuilder.buildBuildVector(DstReg, DstRegs);
MI.eraseFromParent();
return Legalized;
}
// Handle splitting vector operations which need to have the same number of
// elements in each type index, but each type index may have a different element
// type.
//
// e.g. <4 x s64> = G_SHL <4 x s64>, <4 x s32> ->
// <2 x s64> = G_SHL <2 x s64>, <2 x s32>
// <2 x s64> = G_SHL <2 x s64>, <2 x s32>
//
// Also handles some irregular breakdown cases, e.g.
// e.g. <3 x s64> = G_SHL <3 x s64>, <3 x s32> ->