blob: 5ebec6b8fa137ee16ddd8c75334dcf38914a8a96 [file] [log] [blame]
//===- SPIRVInstructionSelector.cpp ------------------------------*- C++ -*-==//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file implements the targeting of the InstructionSelector class for
// SPIRV.
// TODO: This should be generated by TableGen.
//
//===----------------------------------------------------------------------===//
#include "SPIRV.h"
#include "SPIRVGlobalRegistry.h"
#include "SPIRVInstrInfo.h"
#include "SPIRVRegisterBankInfo.h"
#include "SPIRVRegisterInfo.h"
#include "SPIRVTargetMachine.h"
#include "SPIRVUtils.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/IntrinsicsSPIRV.h"
#include "llvm/Support/Debug.h"
#define DEBUG_TYPE "spirv-isel"
using namespace llvm;
namespace CL = SPIRV::OpenCLExtInst;
namespace GL = SPIRV::GLSLExtInst;
using ExtInstList =
std::vector<std::pair<SPIRV::InstructionSet::InstructionSet, uint32_t>>;
namespace {
#define GET_GLOBALISEL_PREDICATE_BITSET
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATE_BITSET
class SPIRVInstructionSelector : public InstructionSelector {
const SPIRVSubtarget &STI;
const SPIRVInstrInfo &TII;
const SPIRVRegisterInfo &TRI;
const RegisterBankInfo &RBI;
SPIRVGlobalRegistry &GR;
MachineRegisterInfo *MRI;
public:
SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
const SPIRVSubtarget &ST,
const RegisterBankInfo &RBI);
void setupMF(MachineFunction &MF, GISelKnownBits *KB,
CodeGenCoverage &CoverageInfo, ProfileSummaryInfo *PSI,
BlockFrequencyInfo *BFI) override;
// Common selection code. Instruction-specific selection occurs in spvSelect.
bool select(MachineInstr &I) override;
static const char *getName() { return DEBUG_TYPE; }
#define GET_GLOBALISEL_PREDICATES_DECL
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_DECL
#define GET_GLOBALISEL_TEMPORARIES_DECL
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_DECL
private:
// tblgen-erated 'select' implementation, used as the initial selector for
// the patterns that don't require complex C++.
bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
// All instruction-specific selection that didn't happen in "select()".
// Is basically a large Switch/Case delegating to all other select method.
bool spvSelect(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectGlobalValue(Register ResVReg, MachineInstr &I,
const MachineInstr *Init = nullptr) const;
bool selectUnOpWithSrc(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I, Register SrcReg,
unsigned Opcode) const;
bool selectUnOp(Register ResVReg, const SPIRVType *ResType, MachineInstr &I,
unsigned Opcode) const;
bool selectLoad(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectStore(MachineInstr &I) const;
bool selectMemOperation(Register ResVReg, MachineInstr &I) const;
bool selectAtomicRMW(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I, unsigned NewOpcode) const;
bool selectAtomicCmpXchg(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectFence(MachineInstr &I) const;
bool selectAddrSpaceCast(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectBitreverse(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectConstVector(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectCmp(Register ResVReg, const SPIRVType *ResType,
unsigned comparisonOpcode, MachineInstr &I) const;
bool selectICmp(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectFCmp(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
void renderImm32(MachineInstrBuilder &MIB, const MachineInstr &I,
int OpIdx) const;
void renderFImm32(MachineInstrBuilder &MIB, const MachineInstr &I,
int OpIdx) const;
bool selectConst(Register ResVReg, const SPIRVType *ResType, const APInt &Imm,
MachineInstr &I) const;
bool selectSelect(Register ResVReg, const SPIRVType *ResType, MachineInstr &I,
bool IsSigned) const;
bool selectIToF(Register ResVReg, const SPIRVType *ResType, MachineInstr &I,
bool IsSigned, unsigned Opcode) const;
bool selectExt(Register ResVReg, const SPIRVType *ResType, MachineInstr &I,
bool IsSigned) const;
bool selectTrunc(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectIntToBool(Register IntReg, Register ResVReg, MachineInstr &I,
const SPIRVType *intTy, const SPIRVType *boolTy) const;
bool selectOpUndef(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectIntrinsic(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectExtractVal(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectInsertVal(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectExtractElt(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectInsertElt(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectGEP(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectFrameIndex(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectBranch(MachineInstr &I) const;
bool selectBranchCond(MachineInstr &I) const;
bool selectPhi(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I) const;
bool selectExtInst(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I, CL::OpenCLExtInst CLInst) const;
bool selectExtInst(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I, CL::OpenCLExtInst CLInst,
GL::GLSLExtInst GLInst) const;
bool selectExtInst(Register ResVReg, const SPIRVType *ResType,
MachineInstr &I, const ExtInstList &ExtInsts) const;
Register buildI32Constant(uint32_t Val, MachineInstr &I,
const SPIRVType *ResType = nullptr) const;
Register buildZerosVal(const SPIRVType *ResType, MachineInstr &I) const;
Register buildOnesVal(bool AllOnes, const SPIRVType *ResType,
MachineInstr &I) const;
};
} // end anonymous namespace
#define GET_GLOBALISEL_IMPL
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_IMPL
SPIRVInstructionSelector::SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
const SPIRVSubtarget &ST,
const RegisterBankInfo &RBI)
: InstructionSelector(), STI(ST), TII(*ST.getInstrInfo()),
TRI(*ST.getRegisterInfo()), RBI(RBI), GR(*ST.getSPIRVGlobalRegistry()),
#define GET_GLOBALISEL_PREDICATES_INIT
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
#include "SPIRVGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_INIT
{
}
void SPIRVInstructionSelector::setupMF(MachineFunction &MF, GISelKnownBits *KB,
CodeGenCoverage &CoverageInfo,
ProfileSummaryInfo *PSI,
BlockFrequencyInfo *BFI) {
MRI = &MF.getRegInfo();
GR.setCurrentFunc(MF);
InstructionSelector::setupMF(MF, KB, CoverageInfo, PSI, BFI);
}
static bool isImm(const MachineOperand &MO, MachineRegisterInfo *MRI);
// Defined in SPIRVLegalizerInfo.cpp.
extern bool isTypeFoldingSupported(unsigned Opcode);
bool SPIRVInstructionSelector::select(MachineInstr &I) {
assert(I.getParent() && "Instruction should be in a basic block!");
assert(I.getParent()->getParent() && "Instruction should be in a function!");
Register Opcode = I.getOpcode();
// If it's not a GMIR instruction, we've selected it already.
if (!isPreISelGenericOpcode(Opcode)) {
if (Opcode == SPIRV::ASSIGN_TYPE) { // These pseudos aren't needed any more.
auto *Def = MRI->getVRegDef(I.getOperand(1).getReg());
if (isTypeFoldingSupported(Def->getOpcode())) {
auto Res = selectImpl(I, *CoverageInfo);
assert(Res || Def->getOpcode() == TargetOpcode::G_CONSTANT);
if (Res)
return Res;
}
MRI->replaceRegWith(I.getOperand(1).getReg(), I.getOperand(0).getReg());
I.removeFromParent();
return true;
} else if (I.getNumDefs() == 1) {
// Make all vregs 32 bits (for SPIR-V IDs).
MRI->setType(I.getOperand(0).getReg(), LLT::scalar(32));
}
return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}
if (I.getNumOperands() != I.getNumExplicitOperands()) {
LLVM_DEBUG(errs() << "Generic instr has unexpected implicit operands\n");
return false;
}
// Common code for getting return reg+type, and removing selected instr
// from parent occurs here. Instr-specific selection happens in spvSelect().
bool HasDefs = I.getNumDefs() > 0;
Register ResVReg = HasDefs ? I.getOperand(0).getReg() : Register(0);
SPIRVType *ResType = HasDefs ? GR.getSPIRVTypeForVReg(ResVReg) : nullptr;
assert(!HasDefs || ResType || I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE);
if (spvSelect(ResVReg, ResType, I)) {
if (HasDefs) // Make all vregs 32 bits (for SPIR-V IDs).
MRI->setType(ResVReg, LLT::scalar(32));
I.removeFromParent();
return true;
}
return false;
}
bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
assert(!isTypeFoldingSupported(I.getOpcode()) ||
I.getOpcode() == TargetOpcode::G_CONSTANT);
const unsigned Opcode = I.getOpcode();
switch (Opcode) {
case TargetOpcode::G_CONSTANT:
return selectConst(ResVReg, ResType, I.getOperand(1).getCImm()->getValue(),
I);
case TargetOpcode::G_GLOBAL_VALUE:
return selectGlobalValue(ResVReg, I);
case TargetOpcode::G_IMPLICIT_DEF:
return selectOpUndef(ResVReg, ResType, I);
case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
return selectIntrinsic(ResVReg, ResType, I);
case TargetOpcode::G_BITREVERSE:
return selectBitreverse(ResVReg, ResType, I);
case TargetOpcode::G_BUILD_VECTOR:
return selectConstVector(ResVReg, ResType, I);
case TargetOpcode::G_SHUFFLE_VECTOR: {
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(1).getReg())
.addUse(I.getOperand(2).getReg());
for (auto V : I.getOperand(3).getShuffleMask())
MIB.addImm(V);
return MIB.constrainAllUses(TII, TRI, RBI);
}
case TargetOpcode::G_MEMMOVE:
case TargetOpcode::G_MEMCPY:
case TargetOpcode::G_MEMSET:
return selectMemOperation(ResVReg, I);
case TargetOpcode::G_ICMP:
return selectICmp(ResVReg, ResType, I);
case TargetOpcode::G_FCMP:
return selectFCmp(ResVReg, ResType, I);
case TargetOpcode::G_FRAME_INDEX:
return selectFrameIndex(ResVReg, ResType, I);
case TargetOpcode::G_LOAD:
return selectLoad(ResVReg, ResType, I);
case TargetOpcode::G_STORE:
return selectStore(I);
case TargetOpcode::G_BR:
return selectBranch(I);
case TargetOpcode::G_BRCOND:
return selectBranchCond(I);
case TargetOpcode::G_PHI:
return selectPhi(ResVReg, ResType, I);
case TargetOpcode::G_FPTOSI:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS);
case TargetOpcode::G_FPTOUI:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU);
case TargetOpcode::G_SITOFP:
return selectIToF(ResVReg, ResType, I, true, SPIRV::OpConvertSToF);
case TargetOpcode::G_UITOFP:
return selectIToF(ResVReg, ResType, I, false, SPIRV::OpConvertUToF);
case TargetOpcode::G_CTPOP:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitCount);
case TargetOpcode::G_SMIN:
return selectExtInst(ResVReg, ResType, I, CL::s_min, GL::SMin);
case TargetOpcode::G_UMIN:
return selectExtInst(ResVReg, ResType, I, CL::u_min, GL::UMin);
case TargetOpcode::G_SMAX:
return selectExtInst(ResVReg, ResType, I, CL::s_max, GL::SMax);
case TargetOpcode::G_UMAX:
return selectExtInst(ResVReg, ResType, I, CL::u_max, GL::UMax);
case TargetOpcode::G_FMA:
return selectExtInst(ResVReg, ResType, I, CL::fma, GL::Fma);
case TargetOpcode::G_FPOW:
return selectExtInst(ResVReg, ResType, I, CL::pow, GL::Pow);
case TargetOpcode::G_FPOWI:
return selectExtInst(ResVReg, ResType, I, CL::pown);
case TargetOpcode::G_FEXP:
return selectExtInst(ResVReg, ResType, I, CL::exp, GL::Exp);
case TargetOpcode::G_FEXP2:
return selectExtInst(ResVReg, ResType, I, CL::exp2, GL::Exp2);
case TargetOpcode::G_FLOG:
return selectExtInst(ResVReg, ResType, I, CL::log, GL::Log);
case TargetOpcode::G_FLOG2:
return selectExtInst(ResVReg, ResType, I, CL::log2, GL::Log2);
case TargetOpcode::G_FLOG10:
return selectExtInst(ResVReg, ResType, I, CL::log10);
case TargetOpcode::G_FABS:
return selectExtInst(ResVReg, ResType, I, CL::fabs, GL::FAbs);
case TargetOpcode::G_ABS:
return selectExtInst(ResVReg, ResType, I, CL::s_abs, GL::SAbs);
case TargetOpcode::G_FMINNUM:
case TargetOpcode::G_FMINIMUM:
return selectExtInst(ResVReg, ResType, I, CL::fmin, GL::FMin);
case TargetOpcode::G_FMAXNUM:
case TargetOpcode::G_FMAXIMUM:
return selectExtInst(ResVReg, ResType, I, CL::fmax, GL::FMax);
case TargetOpcode::G_FCOPYSIGN:
return selectExtInst(ResVReg, ResType, I, CL::copysign);
case TargetOpcode::G_FCEIL:
return selectExtInst(ResVReg, ResType, I, CL::ceil, GL::Ceil);
case TargetOpcode::G_FFLOOR:
return selectExtInst(ResVReg, ResType, I, CL::floor, GL::Floor);
case TargetOpcode::G_FCOS:
return selectExtInst(ResVReg, ResType, I, CL::cos, GL::Cos);
case TargetOpcode::G_FSIN:
return selectExtInst(ResVReg, ResType, I, CL::sin, GL::Sin);
case TargetOpcode::G_FSQRT:
return selectExtInst(ResVReg, ResType, I, CL::sqrt, GL::Sqrt);
case TargetOpcode::G_CTTZ:
case TargetOpcode::G_CTTZ_ZERO_UNDEF:
return selectExtInst(ResVReg, ResType, I, CL::ctz);
case TargetOpcode::G_CTLZ:
case TargetOpcode::G_CTLZ_ZERO_UNDEF:
return selectExtInst(ResVReg, ResType, I, CL::clz);
case TargetOpcode::G_INTRINSIC_ROUND:
return selectExtInst(ResVReg, ResType, I, CL::round, GL::Round);
case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
case TargetOpcode::G_INTRINSIC_TRUNC:
return selectExtInst(ResVReg, ResType, I, CL::trunc, GL::Trunc);
case TargetOpcode::G_FRINT:
case TargetOpcode::G_FNEARBYINT:
return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
case TargetOpcode::G_SMULH:
return selectExtInst(ResVReg, ResType, I, CL::s_mul_hi);
case TargetOpcode::G_UMULH:
return selectExtInst(ResVReg, ResType, I, CL::u_mul_hi);
case TargetOpcode::G_SEXT:
return selectExt(ResVReg, ResType, I, true);
case TargetOpcode::G_ANYEXT:
case TargetOpcode::G_ZEXT:
return selectExt(ResVReg, ResType, I, false);
case TargetOpcode::G_TRUNC:
return selectTrunc(ResVReg, ResType, I);
case TargetOpcode::G_FPTRUNC:
case TargetOpcode::G_FPEXT:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpFConvert);
case TargetOpcode::G_PTRTOINT:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertPtrToU);
case TargetOpcode::G_INTTOPTR:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertUToPtr);
case TargetOpcode::G_BITCAST:
return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast);
case TargetOpcode::G_ADDRSPACE_CAST:
return selectAddrSpaceCast(ResVReg, ResType, I);
case TargetOpcode::G_PTR_ADD: {
// Currently, we get G_PTR_ADD only as a result of translating
// global variables, initialized with constant expressions like GV + Const
// (see test opencl/basic/progvar_prog_scope_init.ll).
// TODO: extend the handler once we have other cases.
assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
Register GV = I.getOperand(1).getReg();
MachineRegisterInfo::def_instr_iterator II = MRI->def_instr_begin(GV);
assert(((*II).getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
(*II).getOpcode() == TargetOpcode::COPY ||
(*II).getOpcode() == SPIRV::OpVariable) &&
isImm(I.getOperand(2), MRI));
Register Idx = buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I);
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addImm(static_cast<uint32_t>(
SPIRV::Opcode::InBoundsPtrAccessChain))
.addUse(GV)
.addUse(Idx)
.addUse(I.getOperand(2).getReg());
return MIB.constrainAllUses(TII, TRI, RBI);
}
case TargetOpcode::G_ATOMICRMW_OR:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicOr);
case TargetOpcode::G_ATOMICRMW_ADD:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicIAdd);
case TargetOpcode::G_ATOMICRMW_AND:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicAnd);
case TargetOpcode::G_ATOMICRMW_MAX:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMax);
case TargetOpcode::G_ATOMICRMW_MIN:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMin);
case TargetOpcode::G_ATOMICRMW_SUB:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicISub);
case TargetOpcode::G_ATOMICRMW_XOR:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicXor);
case TargetOpcode::G_ATOMICRMW_UMAX:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMax);
case TargetOpcode::G_ATOMICRMW_UMIN:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMin);
case TargetOpcode::G_ATOMICRMW_XCHG:
return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicExchange);
case TargetOpcode::G_ATOMIC_CMPXCHG:
return selectAtomicCmpXchg(ResVReg, ResType, I);
case TargetOpcode::G_FENCE:
return selectFence(I);
default:
return false;
}
}
bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
CL::OpenCLExtInst CLInst) const {
return selectExtInst(ResVReg, ResType, I,
{{SPIRV::InstructionSet::OpenCL_std, CLInst}});
}
bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
CL::OpenCLExtInst CLInst,
GL::GLSLExtInst GLInst) const {
ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CLInst},
{SPIRV::InstructionSet::GLSL_std_450, GLInst}};
return selectExtInst(ResVReg, ResType, I, ExtInsts);
}
bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
const ExtInstList &Insts) const {
for (const auto &Ex : Insts) {
SPIRV::InstructionSet::InstructionSet Set = Ex.first;
uint32_t Opcode = Ex.second;
if (STI.canUseExtInstSet(Set)) {
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addImm(static_cast<uint32_t>(Set))
.addImm(Opcode);
const unsigned NumOps = I.getNumOperands();
for (unsigned i = 1; i < NumOps; ++i)
MIB.add(I.getOperand(i));
return MIB.constrainAllUses(TII, TRI, RBI);
}
}
return false;
}
bool SPIRVInstructionSelector::selectUnOpWithSrc(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
Register SrcReg,
unsigned Opcode) const {
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(SrcReg)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectUnOp(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
unsigned Opcode) const {
return selectUnOpWithSrc(ResVReg, ResType, I, I.getOperand(1).getReg(),
Opcode);
}
static SPIRV::Scope::Scope getScope(SyncScope::ID Ord) {
switch (Ord) {
case SyncScope::SingleThread:
return SPIRV::Scope::Invocation;
case SyncScope::System:
return SPIRV::Scope::Device;
default:
llvm_unreachable("Unsupported synchronization Scope ID.");
}
}
static void addMemoryOperands(MachineMemOperand *MemOp,
MachineInstrBuilder &MIB) {
uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
if (MemOp->isVolatile())
SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
if (MemOp->isNonTemporal())
SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
if (MemOp->getAlign().value())
SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned);
if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None)) {
MIB.addImm(SpvMemOp);
if (SpvMemOp & static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned))
MIB.addImm(MemOp->getAlign().value());
}
}
static void addMemoryOperands(uint64_t Flags, MachineInstrBuilder &MIB) {
uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
if (Flags & MachineMemOperand::Flags::MOVolatile)
SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
if (Flags & MachineMemOperand::Flags::MONonTemporal)
SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None))
MIB.addImm(SpvMemOp);
}
bool SPIRVInstructionSelector::selectLoad(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
unsigned OpOffset =
I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ? 1 : 0;
Register Ptr = I.getOperand(1 + OpOffset).getReg();
auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(Ptr);
if (!I.getNumMemOperands()) {
assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
} else {
addMemoryOperands(*I.memoperands_begin(), MIB);
}
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectStore(MachineInstr &I) const {
unsigned OpOffset =
I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ? 1 : 0;
Register StoreVal = I.getOperand(0 + OpOffset).getReg();
Register Ptr = I.getOperand(1 + OpOffset).getReg();
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpStore))
.addUse(Ptr)
.addUse(StoreVal);
if (!I.getNumMemOperands()) {
assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
} else {
addMemoryOperands(*I.memoperands_begin(), MIB);
}
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectMemOperation(Register ResVReg,
MachineInstr &I) const {
MachineBasicBlock &BB = *I.getParent();
Register SrcReg = I.getOperand(1).getReg();
if (I.getOpcode() == TargetOpcode::G_MEMSET) {
assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
unsigned Val = getIConstVal(I.getOperand(1).getReg(), MRI);
unsigned Num = getIConstVal(I.getOperand(2).getReg(), MRI);
SPIRVType *ValTy = GR.getOrCreateSPIRVIntegerType(8, I, TII);
SPIRVType *ArrTy = GR.getOrCreateSPIRVArrayType(ValTy, Num, I, TII);
Register Const = GR.getOrCreateConsIntArray(Val, I, ArrTy, TII);
SPIRVType *VarTy = GR.getOrCreateSPIRVPointerType(
ArrTy, I, TII, SPIRV::StorageClass::UniformConstant);
// TODO: check if we have such GV, add init, use buildGlobalVariable.
Type *LLVMArrTy = ArrayType::get(
IntegerType::get(GR.CurMF->getFunction().getContext(), 8), Num);
GlobalVariable *GV =
new GlobalVariable(LLVMArrTy, true, GlobalValue::InternalLinkage);
Register VarReg = MRI->createGenericVirtualRegister(LLT::scalar(32));
GR.add(GV, GR.CurMF, VarReg);
buildOpDecorate(VarReg, I, TII, SPIRV::Decoration::Constant, {});
BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpVariable))
.addDef(VarReg)
.addUse(GR.getSPIRVTypeID(VarTy))
.addImm(SPIRV::StorageClass::UniformConstant)
.addUse(Const)
.constrainAllUses(TII, TRI, RBI);
SPIRVType *SourceTy = GR.getOrCreateSPIRVPointerType(
ValTy, I, TII, SPIRV::StorageClass::UniformConstant);
SrcReg = MRI->createGenericVirtualRegister(LLT::scalar(32));
selectUnOpWithSrc(SrcReg, SourceTy, I, VarReg, SPIRV::OpBitcast);
}
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemorySized))
.addUse(I.getOperand(0).getReg())
.addUse(SrcReg)
.addUse(I.getOperand(2).getReg());
if (I.getNumMemOperands())
addMemoryOperands(*I.memoperands_begin(), MIB);
bool Result = MIB.constrainAllUses(TII, TRI, RBI);
if (ResVReg.isValid() && ResVReg != MIB->getOperand(0).getReg())
BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), ResVReg)
.addUse(MIB->getOperand(0).getReg());
return Result;
}
bool SPIRVInstructionSelector::selectAtomicRMW(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
unsigned NewOpcode) const {
assert(I.hasOneMemOperand());
const MachineMemOperand *MemOp = *I.memoperands_begin();
uint32_t Scope = static_cast<uint32_t>(getScope(MemOp->getSyncScopeID()));
Register ScopeReg = buildI32Constant(Scope, I);
Register Ptr = I.getOperand(1).getReg();
// TODO: Changed as it's implemented in the translator. See test/atomicrmw.ll
// auto ScSem =
// getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr));
AtomicOrdering AO = MemOp->getSuccessOrdering();
uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
Register MemSemReg = buildI32Constant(MemSem /*| ScSem*/, I);
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(NewOpcode))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(Ptr)
.addUse(ScopeReg)
.addUse(MemSemReg)
.addUse(I.getOperand(2).getReg())
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectFence(MachineInstr &I) const {
AtomicOrdering AO = AtomicOrdering(I.getOperand(0).getImm());
uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
Register MemSemReg = buildI32Constant(MemSem, I);
SyncScope::ID Ord = SyncScope::ID(I.getOperand(1).getImm());
uint32_t Scope = static_cast<uint32_t>(getScope(Ord));
Register ScopeReg = buildI32Constant(Scope, I);
MachineBasicBlock &BB = *I.getParent();
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMemoryBarrier))
.addUse(ScopeReg)
.addUse(MemSemReg)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectAtomicCmpXchg(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
Register ScopeReg;
Register MemSemEqReg;
Register MemSemNeqReg;
Register Ptr = I.getOperand(2).getReg();
if (I.getOpcode() != TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS) {
assert(I.hasOneMemOperand());
const MachineMemOperand *MemOp = *I.memoperands_begin();
unsigned Scope = static_cast<uint32_t>(getScope(MemOp->getSyncScopeID()));
ScopeReg = buildI32Constant(Scope, I);
unsigned ScSem = static_cast<uint32_t>(
getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr)));
AtomicOrdering AO = MemOp->getSuccessOrdering();
unsigned MemSemEq = static_cast<uint32_t>(getMemSemantics(AO)) | ScSem;
MemSemEqReg = buildI32Constant(MemSemEq, I);
AtomicOrdering FO = MemOp->getFailureOrdering();
unsigned MemSemNeq = static_cast<uint32_t>(getMemSemantics(FO)) | ScSem;
MemSemNeqReg =
MemSemEq == MemSemNeq ? MemSemEqReg : buildI32Constant(MemSemNeq, I);
} else {
ScopeReg = I.getOperand(5).getReg();
MemSemEqReg = I.getOperand(6).getReg();
MemSemNeqReg = I.getOperand(7).getReg();
}
Register Cmp = I.getOperand(3).getReg();
Register Val = I.getOperand(4).getReg();
SPIRVType *SpvValTy = GR.getSPIRVTypeForVReg(Val);
Register ACmpRes = MRI->createVirtualRegister(&SPIRV::IDRegClass);
const DebugLoc &DL = I.getDebugLoc();
bool Result =
BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpAtomicCompareExchange))
.addDef(ACmpRes)
.addUse(GR.getSPIRVTypeID(SpvValTy))
.addUse(Ptr)
.addUse(ScopeReg)
.addUse(MemSemEqReg)
.addUse(MemSemNeqReg)
.addUse(Val)
.addUse(Cmp)
.constrainAllUses(TII, TRI, RBI);
Register CmpSuccReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
SPIRVType *BoolTy = GR.getOrCreateSPIRVBoolType(I, TII);
Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpIEqual))
.addDef(CmpSuccReg)
.addUse(GR.getSPIRVTypeID(BoolTy))
.addUse(ACmpRes)
.addUse(Cmp)
.constrainAllUses(TII, TRI, RBI);
Register TmpReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
.addDef(TmpReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(ACmpRes)
.addUse(GR.getOrCreateUndef(I, ResType, TII))
.addImm(0)
.constrainAllUses(TII, TRI, RBI);
Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(CmpSuccReg)
.addUse(TmpReg)
.addImm(1)
.constrainAllUses(TII, TRI, RBI);
return Result;
}
static bool isGenericCastablePtr(SPIRV::StorageClass::StorageClass SC) {
switch (SC) {
case SPIRV::StorageClass::Workgroup:
case SPIRV::StorageClass::CrossWorkgroup:
case SPIRV::StorageClass::Function:
return true;
default:
return false;
}
}
// In SPIR-V address space casting can only happen to and from the Generic
// storage class. We can also only case Workgroup, CrossWorkgroup, or Function
// pointers to and from Generic pointers. As such, we can convert e.g. from
// Workgroup to Function by going via a Generic pointer as an intermediary. All
// other combinations can only be done by a bitcast, and are probably not safe.
bool SPIRVInstructionSelector::selectAddrSpaceCast(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
// If the AddrSpaceCast user is single and in OpConstantComposite or
// OpVariable, we should select OpSpecConstantOp.
auto UIs = MRI->use_instructions(ResVReg);
if (!UIs.empty() && ++UIs.begin() == UIs.end() &&
(UIs.begin()->getOpcode() == SPIRV::OpConstantComposite ||
UIs.begin()->getOpcode() == SPIRV::OpVariable ||
isSpvIntrinsic(*UIs.begin(), Intrinsic::spv_init_global))) {
Register NewReg = I.getOperand(1).getReg();
MachineBasicBlock &BB = *I.getParent();
SPIRVType *SpvBaseTy = GR.getOrCreateSPIRVIntegerType(8, I, TII);
ResType = GR.getOrCreateSPIRVPointerType(SpvBaseTy, I, TII,
SPIRV::StorageClass::Generic);
bool Result =
BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addImm(static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric))
.addUse(NewReg)
.constrainAllUses(TII, TRI, RBI);
return Result;
}
Register SrcPtr = I.getOperand(1).getReg();
SPIRVType *SrcPtrTy = GR.getSPIRVTypeForVReg(SrcPtr);
SPIRV::StorageClass::StorageClass SrcSC = GR.getPointerStorageClass(SrcPtr);
SPIRV::StorageClass::StorageClass DstSC = GR.getPointerStorageClass(ResVReg);
// Casting from an eligable pointer to Generic.
if (DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC))
return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric);
// Casting from Generic to an eligable pointer.
if (SrcSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(DstSC))
return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr);
// Casting between 2 eligable pointers using Generic as an intermediary.
if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) {
Register Tmp = MRI->createVirtualRegister(&SPIRV::IDRegClass);
SPIRVType *GenericPtrTy = GR.getOrCreateSPIRVPointerType(
SrcPtrTy, I, TII, SPIRV::StorageClass::Generic);
MachineBasicBlock &BB = *I.getParent();
const DebugLoc &DL = I.getDebugLoc();
bool Success = BuildMI(BB, I, DL, TII.get(SPIRV::OpPtrCastToGeneric))
.addDef(Tmp)
.addUse(GR.getSPIRVTypeID(GenericPtrTy))
.addUse(SrcPtr)
.constrainAllUses(TII, TRI, RBI);
return Success && BuildMI(BB, I, DL, TII.get(SPIRV::OpGenericCastToPtr))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(Tmp)
.constrainAllUses(TII, TRI, RBI);
}
// TODO Should this case just be disallowed completely?
// We're casting 2 other arbitrary address spaces, so have to bitcast.
return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast);
}
static unsigned getFCmpOpcode(unsigned PredNum) {
auto Pred = static_cast<CmpInst::Predicate>(PredNum);
switch (Pred) {
case CmpInst::FCMP_OEQ:
return SPIRV::OpFOrdEqual;
case CmpInst::FCMP_OGE:
return SPIRV::OpFOrdGreaterThanEqual;
case CmpInst::FCMP_OGT:
return SPIRV::OpFOrdGreaterThan;
case CmpInst::FCMP_OLE:
return SPIRV::OpFOrdLessThanEqual;
case CmpInst::FCMP_OLT:
return SPIRV::OpFOrdLessThan;
case CmpInst::FCMP_ONE:
return SPIRV::OpFOrdNotEqual;
case CmpInst::FCMP_ORD:
return SPIRV::OpOrdered;
case CmpInst::FCMP_UEQ:
return SPIRV::OpFUnordEqual;
case CmpInst::FCMP_UGE:
return SPIRV::OpFUnordGreaterThanEqual;
case CmpInst::FCMP_UGT:
return SPIRV::OpFUnordGreaterThan;
case CmpInst::FCMP_ULE:
return SPIRV::OpFUnordLessThanEqual;
case CmpInst::FCMP_ULT:
return SPIRV::OpFUnordLessThan;
case CmpInst::FCMP_UNE:
return SPIRV::OpFUnordNotEqual;
case CmpInst::FCMP_UNO:
return SPIRV::OpUnordered;
default:
llvm_unreachable("Unknown predicate type for FCmp");
}
}
static unsigned getICmpOpcode(unsigned PredNum) {
auto Pred = static_cast<CmpInst::Predicate>(PredNum);
switch (Pred) {
case CmpInst::ICMP_EQ:
return SPIRV::OpIEqual;
case CmpInst::ICMP_NE:
return SPIRV::OpINotEqual;
case CmpInst::ICMP_SGE:
return SPIRV::OpSGreaterThanEqual;
case CmpInst::ICMP_SGT:
return SPIRV::OpSGreaterThan;
case CmpInst::ICMP_SLE:
return SPIRV::OpSLessThanEqual;
case CmpInst::ICMP_SLT:
return SPIRV::OpSLessThan;
case CmpInst::ICMP_UGE:
return SPIRV::OpUGreaterThanEqual;
case CmpInst::ICMP_UGT:
return SPIRV::OpUGreaterThan;
case CmpInst::ICMP_ULE:
return SPIRV::OpULessThanEqual;
case CmpInst::ICMP_ULT:
return SPIRV::OpULessThan;
default:
llvm_unreachable("Unknown predicate type for ICmp");
}
}
static unsigned getPtrCmpOpcode(unsigned Pred) {
switch (static_cast<CmpInst::Predicate>(Pred)) {
case CmpInst::ICMP_EQ:
return SPIRV::OpPtrEqual;
case CmpInst::ICMP_NE:
return SPIRV::OpPtrNotEqual;
default:
llvm_unreachable("Unknown predicate type for pointer comparison");
}
}
// Return the logical operation, or abort if none exists.
static unsigned getBoolCmpOpcode(unsigned PredNum) {
auto Pred = static_cast<CmpInst::Predicate>(PredNum);
switch (Pred) {
case CmpInst::ICMP_EQ:
return SPIRV::OpLogicalEqual;
case CmpInst::ICMP_NE:
return SPIRV::OpLogicalNotEqual;
default:
llvm_unreachable("Unknown predicate type for Bool comparison");
}
}
bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
MachineBasicBlock &BB = *I.getParent();
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitReverse))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(1).getReg())
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectConstVector(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
// TODO: only const case is supported for now.
assert(std::all_of(
I.operands_begin(), I.operands_end(), [this](const MachineOperand &MO) {
if (MO.isDef())
return true;
if (!MO.isReg())
return false;
SPIRVType *ConstTy = this->MRI->getVRegDef(MO.getReg());
assert(ConstTy && ConstTy->getOpcode() == SPIRV::ASSIGN_TYPE &&
ConstTy->getOperand(1).isReg());
Register ConstReg = ConstTy->getOperand(1).getReg();
const MachineInstr *Const = this->MRI->getVRegDef(ConstReg);
assert(Const);
return (Const->getOpcode() == TargetOpcode::G_CONSTANT ||
Const->getOpcode() == TargetOpcode::G_FCONSTANT);
}));
auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
TII.get(SPIRV::OpConstantComposite))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType));
for (unsigned i = I.getNumExplicitDefs(); i < I.getNumExplicitOperands(); ++i)
MIB.addUse(I.getOperand(i).getReg());
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectCmp(Register ResVReg,
const SPIRVType *ResType,
unsigned CmpOpc,
MachineInstr &I) const {
Register Cmp0 = I.getOperand(2).getReg();
Register Cmp1 = I.getOperand(3).getReg();
assert(GR.getSPIRVTypeForVReg(Cmp0)->getOpcode() ==
GR.getSPIRVTypeForVReg(Cmp1)->getOpcode() &&
"CMP operands should have the same type");
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(CmpOpc))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(Cmp0)
.addUse(Cmp1)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectICmp(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
auto Pred = I.getOperand(1).getPredicate();
unsigned CmpOpc;
Register CmpOperand = I.getOperand(2).getReg();
if (GR.isScalarOfType(CmpOperand, SPIRV::OpTypePointer))
CmpOpc = getPtrCmpOpcode(Pred);
else if (GR.isScalarOrVectorOfType(CmpOperand, SPIRV::OpTypeBool))
CmpOpc = getBoolCmpOpcode(Pred);
else
CmpOpc = getICmpOpcode(Pred);
return selectCmp(ResVReg, ResType, CmpOpc, I);
}
void SPIRVInstructionSelector::renderFImm32(MachineInstrBuilder &MIB,
const MachineInstr &I,
int OpIdx) const {
assert(I.getOpcode() == TargetOpcode::G_FCONSTANT && OpIdx == -1 &&
"Expected G_FCONSTANT");
const ConstantFP *FPImm = I.getOperand(1).getFPImm();
addNumImm(FPImm->getValueAPF().bitcastToAPInt(), MIB);
}
void SPIRVInstructionSelector::renderImm32(MachineInstrBuilder &MIB,
const MachineInstr &I,
int OpIdx) const {
assert(I.getOpcode() == TargetOpcode::G_CONSTANT && OpIdx == -1 &&
"Expected G_CONSTANT");
addNumImm(I.getOperand(1).getCImm()->getValue(), MIB);
}
Register
SPIRVInstructionSelector::buildI32Constant(uint32_t Val, MachineInstr &I,
const SPIRVType *ResType) const {
Type *LLVMTy = IntegerType::get(GR.CurMF->getFunction().getContext(), 32);
const SPIRVType *SpvI32Ty =
ResType ? ResType : GR.getOrCreateSPIRVIntegerType(32, I, TII);
// Find a constant in DT or build a new one.
auto ConstInt = ConstantInt::get(LLVMTy, Val);
Register NewReg = GR.find(ConstInt, GR.CurMF);
if (!NewReg.isValid()) {
NewReg = MRI->createGenericVirtualRegister(LLT::scalar(32));
GR.add(ConstInt, GR.CurMF, NewReg);
MachineInstr *MI;
MachineBasicBlock &BB = *I.getParent();
if (Val == 0) {
MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
.addDef(NewReg)
.addUse(GR.getSPIRVTypeID(SpvI32Ty));
} else {
MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI))
.addDef(NewReg)
.addUse(GR.getSPIRVTypeID(SpvI32Ty))
.addImm(APInt(32, Val).getZExtValue());
}
constrainSelectedInstRegOperands(*MI, TII, TRI, RBI);
}
return NewReg;
}
bool SPIRVInstructionSelector::selectFCmp(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
unsigned CmpOp = getFCmpOpcode(I.getOperand(1).getPredicate());
return selectCmp(ResVReg, ResType, CmpOp, I);
}
Register SPIRVInstructionSelector::buildZerosVal(const SPIRVType *ResType,
MachineInstr &I) const {
if (ResType->getOpcode() == SPIRV::OpTypeVector)
return GR.getOrCreateConsIntVector(0, I, ResType, TII);
return GR.getOrCreateConstInt(0, I, ResType, TII);
}
Register SPIRVInstructionSelector::buildOnesVal(bool AllOnes,
const SPIRVType *ResType,
MachineInstr &I) const {
unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
APInt One = AllOnes ? APInt::getAllOnesValue(BitWidth)
: APInt::getOneBitSet(BitWidth, 0);
if (ResType->getOpcode() == SPIRV::OpTypeVector)
return GR.getOrCreateConsIntVector(One.getZExtValue(), I, ResType, TII);
return GR.getOrCreateConstInt(One.getZExtValue(), I, ResType, TII);
}
bool SPIRVInstructionSelector::selectSelect(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I,
bool IsSigned) const {
// To extend a bool, we need to use OpSelect between constants.
Register ZeroReg = buildZerosVal(ResType, I);
Register OneReg = buildOnesVal(IsSigned, ResType, I);
bool IsScalarBool =
GR.isScalarOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool);
unsigned Opcode =
IsScalarBool ? SPIRV::OpSelectSISCond : SPIRV::OpSelectSIVCond;
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(1).getReg())
.addUse(OneReg)
.addUse(ZeroReg)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectIToF(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I, bool IsSigned,
unsigned Opcode) const {
Register SrcReg = I.getOperand(1).getReg();
// We can convert bool value directly to float type without OpConvert*ToF,
// however the translator generates OpSelect+OpConvert*ToF, so we do the same.
if (GR.isScalarOrVectorOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool)) {
unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
SPIRVType *TmpType = GR.getOrCreateSPIRVIntegerType(BitWidth, I, TII);
if (ResType->getOpcode() == SPIRV::OpTypeVector) {
const unsigned NumElts = ResType->getOperand(2).getImm();
TmpType = GR.getOrCreateSPIRVVectorType(TmpType, NumElts, I, TII);
}
SrcReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
selectSelect(SrcReg, TmpType, I, false);
}
return selectUnOpWithSrc(ResVReg, ResType, I, SrcReg, Opcode);
}
bool SPIRVInstructionSelector::selectExt(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I, bool IsSigned) const {
if (GR.isScalarOrVectorOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool))
return selectSelect(ResVReg, ResType, I, IsSigned);
unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
return selectUnOp(ResVReg, ResType, I, Opcode);
}
bool SPIRVInstructionSelector::selectIntToBool(Register IntReg,
Register ResVReg,
MachineInstr &I,
const SPIRVType *IntTy,
const SPIRVType *BoolTy) const {
// To truncate to a bool, we use OpBitwiseAnd 1 and OpINotEqual to zero.
Register BitIntReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
bool IsVectorTy = IntTy->getOpcode() == SPIRV::OpTypeVector;
unsigned Opcode = IsVectorTy ? SPIRV::OpBitwiseAndV : SPIRV::OpBitwiseAndS;
Register Zero = buildZerosVal(IntTy, I);
Register One = buildOnesVal(false, IntTy, I);
MachineBasicBlock &BB = *I.getParent();
BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
.addDef(BitIntReg)
.addUse(GR.getSPIRVTypeID(IntTy))
.addUse(IntReg)
.addUse(One)
.constrainAllUses(TII, TRI, RBI);
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(BoolTy))
.addUse(BitIntReg)
.addUse(Zero)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectTrunc(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
if (GR.isScalarOrVectorOfType(ResVReg, SPIRV::OpTypeBool)) {
Register IntReg = I.getOperand(1).getReg();
const SPIRVType *ArgType = GR.getSPIRVTypeForVReg(IntReg);
return selectIntToBool(IntReg, ResVReg, I, ArgType, ResType);
}
bool IsSigned = GR.isScalarOrVectorSigned(ResType);
unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
return selectUnOp(ResVReg, ResType, I, Opcode);
}
bool SPIRVInstructionSelector::selectConst(Register ResVReg,
const SPIRVType *ResType,
const APInt &Imm,
MachineInstr &I) const {
unsigned TyOpcode = ResType->getOpcode();
assert(TyOpcode != SPIRV::OpTypePointer || Imm.isNullValue());
MachineBasicBlock &BB = *I.getParent();
if ((TyOpcode == SPIRV::OpTypePointer || TyOpcode == SPIRV::OpTypeEvent) &&
Imm.isNullValue())
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.constrainAllUses(TII, TRI, RBI);
if (TyOpcode == SPIRV::OpTypeInt) {
Register Reg = GR.getOrCreateConstInt(Imm.getZExtValue(), I, ResType, TII);
if (Reg == ResVReg)
return true;
return BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY))
.addDef(ResVReg)
.addUse(Reg)
.constrainAllUses(TII, TRI, RBI);
}
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType));
// <=32-bit integers should be caught by the sdag pattern.
assert(Imm.getBitWidth() > 32);
addNumImm(Imm, MIB);
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectOpUndef(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.constrainAllUses(TII, TRI, RBI);
}
static bool isImm(const MachineOperand &MO, MachineRegisterInfo *MRI) {
assert(MO.isReg());
const SPIRVType *TypeInst = MRI->getVRegDef(MO.getReg());
if (TypeInst->getOpcode() != SPIRV::ASSIGN_TYPE)
return false;
assert(TypeInst->getOperand(1).isReg());
MachineInstr *ImmInst = MRI->getVRegDef(TypeInst->getOperand(1).getReg());
return ImmInst->getOpcode() == TargetOpcode::G_CONSTANT;
}
static int64_t foldImm(const MachineOperand &MO, MachineRegisterInfo *MRI) {
const SPIRVType *TypeInst = MRI->getVRegDef(MO.getReg());
MachineInstr *ImmInst = MRI->getVRegDef(TypeInst->getOperand(1).getReg());
assert(ImmInst->getOpcode() == TargetOpcode::G_CONSTANT);
return ImmInst->getOperand(1).getCImm()->getZExtValue();
}
bool SPIRVInstructionSelector::selectInsertVal(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeInsert))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
// object to insert
.addUse(I.getOperand(3).getReg())
// composite to insert into
.addUse(I.getOperand(2).getReg());
for (unsigned i = 4; i < I.getNumOperands(); i++)
MIB.addImm(foldImm(I.getOperand(i), MRI));
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectExtractVal(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
MachineBasicBlock &BB = *I.getParent();
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(2).getReg());
for (unsigned i = 3; i < I.getNumOperands(); i++)
MIB.addImm(foldImm(I.getOperand(i), MRI));
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectInsertElt(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
if (isImm(I.getOperand(4), MRI))
return selectInsertVal(ResVReg, ResType, I);
MachineBasicBlock &BB = *I.getParent();
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorInsertDynamic))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(2).getReg())
.addUse(I.getOperand(3).getReg())
.addUse(I.getOperand(4).getReg())
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectExtractElt(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
if (isImm(I.getOperand(3), MRI))
return selectExtractVal(ResVReg, ResType, I);
MachineBasicBlock &BB = *I.getParent();
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorExtractDynamic))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addUse(I.getOperand(2).getReg())
.addUse(I.getOperand(3).getReg())
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectGEP(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
// In general we should also support OpAccessChain instrs here (i.e. not
// PtrAccessChain) but SPIRV-LLVM Translator doesn't emit them at all and so
// do we to stay compliant with its test and more importantly consumers.
unsigned Opcode = I.getOperand(2).getImm() ? SPIRV::OpInBoundsPtrAccessChain
: SPIRV::OpPtrAccessChain;
auto Res = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
// Object to get a pointer to.
.addUse(I.getOperand(3).getReg());
// Adding indices.
for (unsigned i = 4; i < I.getNumExplicitOperands(); ++i)
Res.addUse(I.getOperand(i).getReg());
return Res.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
MachineBasicBlock &BB = *I.getParent();
switch (I.getIntrinsicID()) {
case Intrinsic::spv_load:
return selectLoad(ResVReg, ResType, I);
break;
case Intrinsic::spv_store:
return selectStore(I);
break;
case Intrinsic::spv_extractv:
return selectExtractVal(ResVReg, ResType, I);
break;
case Intrinsic::spv_insertv:
return selectInsertVal(ResVReg, ResType, I);
break;
case Intrinsic::spv_extractelt:
return selectExtractElt(ResVReg, ResType, I);
break;
case Intrinsic::spv_insertelt:
return selectInsertElt(ResVReg, ResType, I);
break;
case Intrinsic::spv_gep:
return selectGEP(ResVReg, ResType, I);
break;
case Intrinsic::spv_unref_global:
case Intrinsic::spv_init_global: {
MachineInstr *MI = MRI->getVRegDef(I.getOperand(1).getReg());
MachineInstr *Init = I.getNumExplicitOperands() > 2
? MRI->getVRegDef(I.getOperand(2).getReg())
: nullptr;
assert(MI);
return selectGlobalValue(MI->getOperand(0).getReg(), *MI, Init);
} break;
case Intrinsic::spv_const_composite: {
// If no values are attached, the composite is null constant.
bool IsNull = I.getNumExplicitDefs() + 1 == I.getNumExplicitOperands();
unsigned Opcode =
IsNull ? SPIRV::OpConstantNull : SPIRV::OpConstantComposite;
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType));
// skip type MD node we already used when generated assign.type for this
if (!IsNull) {
for (unsigned i = I.getNumExplicitDefs() + 1;
i < I.getNumExplicitOperands(); ++i) {
MIB.addUse(I.getOperand(i).getReg());
}
}
return MIB.constrainAllUses(TII, TRI, RBI);
} break;
case Intrinsic::spv_assign_name: {
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpName));
MIB.addUse(I.getOperand(I.getNumExplicitDefs() + 1).getReg());
for (unsigned i = I.getNumExplicitDefs() + 2;
i < I.getNumExplicitOperands(); ++i) {
MIB.addImm(I.getOperand(i).getImm());
}
return MIB.constrainAllUses(TII, TRI, RBI);
} break;
case Intrinsic::spv_switch: {
auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSwitch));
for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
if (I.getOperand(i).isReg())
MIB.addReg(I.getOperand(i).getReg());
else if (I.getOperand(i).isCImm())
addNumImm(I.getOperand(i).getCImm()->getValue(), MIB);
else if (I.getOperand(i).isMBB())
MIB.addMBB(I.getOperand(i).getMBB());
else
llvm_unreachable("Unexpected OpSwitch operand");
}
return MIB.constrainAllUses(TII, TRI, RBI);
} break;
case Intrinsic::spv_cmpxchg:
return selectAtomicCmpXchg(ResVReg, ResType, I);
break;
case Intrinsic::spv_unreachable:
BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUnreachable));
break;
case Intrinsic::spv_alloca:
return selectFrameIndex(ResVReg, ResType, I);
break;
default:
llvm_unreachable("Intrinsic selection not implemented");
}
return true;
}
bool SPIRVInstructionSelector::selectFrameIndex(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpVariable))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType))
.addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function))
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectBranch(MachineInstr &I) const {
// InstructionSelector walks backwards through the instructions. We can use
// both a G_BR and a G_BRCOND to create an OpBranchConditional. We hit G_BR
// first, so can generate an OpBranchConditional here. If there is no
// G_BRCOND, we just use OpBranch for a regular unconditional branch.
const MachineInstr *PrevI = I.getPrevNode();
MachineBasicBlock &MBB = *I.getParent();
if (PrevI != nullptr && PrevI->getOpcode() == TargetOpcode::G_BRCOND) {
return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
.addUse(PrevI->getOperand(0).getReg())
.addMBB(PrevI->getOperand(1).getMBB())
.addMBB(I.getOperand(0).getMBB())
.constrainAllUses(TII, TRI, RBI);
}
return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranch))
.addMBB(I.getOperand(0).getMBB())
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectBranchCond(MachineInstr &I) const {
// InstructionSelector walks backwards through the instructions. For an
// explicit conditional branch with no fallthrough, we use both a G_BR and a
// G_BRCOND to create an OpBranchConditional. We should hit G_BR first, and
// generate the OpBranchConditional in selectBranch above.
//
// If an OpBranchConditional has been generated, we simply return, as the work
// is alread done. If there is no OpBranchConditional, LLVM must be relying on
// implicit fallthrough to the next basic block, so we need to create an
// OpBranchConditional with an explicit "false" argument pointing to the next
// basic block that LLVM would fall through to.
const MachineInstr *NextI = I.getNextNode();
// Check if this has already been successfully selected.
if (NextI != nullptr && NextI->getOpcode() == SPIRV::OpBranchConditional)
return true;
// Must be relying on implicit block fallthrough, so generate an
// OpBranchConditional with the "next" basic block as the "false" target.
MachineBasicBlock &MBB = *I.getParent();
unsigned NextMBBNum = MBB.getNextNode()->getNumber();
MachineBasicBlock *NextMBB = I.getMF()->getBlockNumbered(NextMBBNum);
return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
.addUse(I.getOperand(0).getReg())
.addMBB(I.getOperand(1).getMBB())
.addMBB(NextMBB)
.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectPhi(Register ResVReg,
const SPIRVType *ResType,
MachineInstr &I) const {
auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpPhi))
.addDef(ResVReg)
.addUse(GR.getSPIRVTypeID(ResType));
const unsigned NumOps = I.getNumOperands();
for (unsigned i = 1; i < NumOps; i += 2) {
MIB.addUse(I.getOperand(i + 0).getReg());
MIB.addMBB(I.getOperand(i + 1).getMBB());
}
return MIB.constrainAllUses(TII, TRI, RBI);
}
bool SPIRVInstructionSelector::selectGlobalValue(
Register ResVReg, MachineInstr &I, const MachineInstr *Init) const {
// FIXME: don't use MachineIRBuilder here, replace it with BuildMI.
MachineIRBuilder MIRBuilder(I);
const GlobalValue *GV = I.getOperand(1).getGlobal();
SPIRVType *ResType = GR.getOrCreateSPIRVType(
GV->getType(), MIRBuilder, SPIRV::AccessQualifier::ReadWrite, false);
std::string GlobalIdent = GV->getGlobalIdentifier();
// We have functions as operands in tests with blocks of instruction e.g. in
// transcoding/global_block.ll. These operands are not used and should be
// substituted by zero constants. Their type is expected to be always
// OpTypePointer Function %uchar.
if (isa<Function>(GV)) {
const Constant *ConstVal = GV;
MachineBasicBlock &BB = *I.getParent();
Register NewReg = GR.find(ConstVal, GR.CurMF);
if (!NewReg.isValid()) {
SPIRVType *SpvBaseTy = GR.getOrCreateSPIRVIntegerType(8, I, TII);
ResType = GR.getOrCreateSPIRVPointerType(SpvBaseTy, I, TII);
Register NewReg = ResVReg;
GR.add(ConstVal, GR.CurMF, NewReg);
return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
.addDef(NewReg)
.addUse(GR.getSPIRVTypeID(ResType))
.constrainAllUses(TII, TRI, RBI);
}
assert(NewReg != ResVReg);
return BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY))
.addDef(ResVReg)
.addUse(NewReg)
.constrainAllUses(TII, TRI, RBI);
}
auto GlobalVar = cast<GlobalVariable>(GV);
assert(GlobalVar->getName() != "llvm.global.annotations");
bool HasInit = GlobalVar->hasInitializer() &&
!isa<UndefValue>(GlobalVar->getInitializer());
// Skip empty declaration for GVs with initilaizers till we get the decl with
// passed initializer.
if (HasInit && !Init)
return true;
unsigned AddrSpace = GV->getAddressSpace();
SPIRV::StorageClass::StorageClass Storage =
addressSpaceToStorageClass(AddrSpace);
bool HasLnkTy = GV->getLinkage() != GlobalValue::InternalLinkage &&
Storage != SPIRV::StorageClass::Function;
SPIRV::LinkageType::LinkageType LnkType =
(GV->isDeclaration() || GV->hasAvailableExternallyLinkage())
? SPIRV::LinkageType::Import
: SPIRV::LinkageType::Export;
Register Reg = GR.buildGlobalVariable(ResVReg, ResType, GlobalIdent, GV,
Storage, Init, GlobalVar->isConstant(),
HasLnkTy, LnkType, MIRBuilder, true);
return Reg.isValid();
}
namespace llvm {
InstructionSelector *
createSPIRVInstructionSelector(const SPIRVTargetMachine &TM,
const SPIRVSubtarget &Subtarget,
const RegisterBankInfo &RBI) {
return new SPIRVInstructionSelector(TM, Subtarget, RBI);
}
} // namespace llvm