blob: 94bf0c0ca1229c33ee4af0a3852836d2d2df609d [file] [log] [blame]
//
// The Subzero Code Generator
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
///
/// \file
/// \brief Implements the TargetLoweringMIPS32 class, which consists almost
/// entirely of the lowering sequence for each high-level instruction.
///
//===----------------------------------------------------------------------===//
#include "IceTargetLoweringMIPS32.h"
#include "IceCfg.h"
#include "IceCfgNode.h"
#include "IceClFlags.h"
#include "IceDefs.h"
#include "IceELFObjectWriter.h"
#include "IceGlobalInits.h"
#include "IceInstMIPS32.h"
#include "IceInstVarIter.h"
#include "IceLiveness.h"
#include "IceOperand.h"
#include "IcePhiLoweringImpl.h"
#include "IceRegistersMIPS32.h"
#include "IceTargetLoweringMIPS32.def"
#include "IceUtils.h"
#include "llvm/Support/MathExtras.h"
namespace MIPS32 {
std::unique_ptr<::Ice::TargetLowering> createTargetLowering(::Ice::Cfg *Func) {
return ::Ice::MIPS32::TargetMIPS32::create(Func);
}
std::unique_ptr<::Ice::TargetDataLowering>
createTargetDataLowering(::Ice::GlobalContext *Ctx) {
return ::Ice::MIPS32::TargetDataMIPS32::create(Ctx);
}
std::unique_ptr<::Ice::TargetHeaderLowering>
createTargetHeaderLowering(::Ice::GlobalContext *Ctx) {
return ::Ice::MIPS32::TargetHeaderMIPS32::create(Ctx);
}
void staticInit(::Ice::GlobalContext *Ctx) {
::Ice::MIPS32::TargetMIPS32::staticInit(Ctx);
}
bool shouldBePooled(const ::Ice::Constant *C) {
return ::Ice::MIPS32::TargetMIPS32::shouldBePooled(C);
}
::Ice::Type getPointerType() {
return ::Ice::MIPS32::TargetMIPS32::getPointerType();
}
} // end of namespace MIPS32
namespace Ice {
namespace MIPS32 {
using llvm::isInt;
namespace {
// The maximum number of arguments to pass in GPR registers.
constexpr uint32_t MIPS32_MAX_GPR_ARG = 4;
std::array<RegNumT, MIPS32_MAX_GPR_ARG> GPRArgInitializer;
std::array<RegNumT, MIPS32_MAX_GPR_ARG / 2> I64ArgInitializer;
constexpr uint32_t MIPS32_MAX_FP_ARG = 2;
std::array<RegNumT, MIPS32_MAX_FP_ARG> FP32ArgInitializer;
std::array<RegNumT, MIPS32_MAX_FP_ARG> FP64ArgInitializer;
const char *getRegClassName(RegClass C) {
auto ClassNum = static_cast<RegClassMIPS32>(C);
assert(ClassNum < RCMIPS32_NUM);
switch (ClassNum) {
default:
assert(C < RC_Target);
return regClassString(C);
// Add handling of new register classes below.
}
}
// Stack alignment
constexpr uint32_t MIPS32_STACK_ALIGNMENT_BYTES = 16;
// Value is in bytes. Return Value adjusted to the next highest multiple of the
// stack alignment required for the given type.
uint32_t applyStackAlignmentTy(uint32_t Value, Type Ty) {
size_t typeAlignInBytes = typeWidthInBytes(Ty);
// Vectors are stored on stack with the same alignment as that of int type
if (isVectorType(Ty))
typeAlignInBytes = typeWidthInBytes(IceType_i64);
return Utils::applyAlignment(Value, typeAlignInBytes);
}
// Value is in bytes. Return Value adjusted to the next highest multiple of the
// stack alignment.
uint32_t applyStackAlignment(uint32_t Value) {
return Utils::applyAlignment(Value, MIPS32_STACK_ALIGNMENT_BYTES);
}
} // end of anonymous namespace
TargetMIPS32::TargetMIPS32(Cfg *Func) : TargetLowering(Func) {}
void TargetMIPS32::assignVarStackSlots(VarList &SortedSpilledVariables,
size_t SpillAreaPaddingBytes,
size_t SpillAreaSizeBytes,
size_t GlobalsAndSubsequentPaddingSize) {
const VariablesMetadata *VMetadata = Func->getVMetadata();
size_t GlobalsSpaceUsed = SpillAreaPaddingBytes;
size_t NextStackOffset = SpillAreaPaddingBytes;
CfgVector<size_t> LocalsSize(Func->getNumNodes());
const bool SimpleCoalescing = !callsReturnsTwice();
for (Variable *Var : SortedSpilledVariables) {
size_t Increment = typeWidthInBytesOnStack(Var->getType());
if (SimpleCoalescing && VMetadata->isTracked(Var)) {
if (VMetadata->isMultiBlock(Var)) {
GlobalsSpaceUsed += Increment;
NextStackOffset = GlobalsSpaceUsed;
} else {
SizeT NodeIndex = VMetadata->getLocalUseNode(Var)->getIndex();
LocalsSize[NodeIndex] += Increment;
NextStackOffset = SpillAreaPaddingBytes +
GlobalsAndSubsequentPaddingSize +
LocalsSize[NodeIndex];
}
} else {
NextStackOffset += Increment;
}
Var->setStackOffset(SpillAreaSizeBytes - NextStackOffset);
}
}
void TargetMIPS32::staticInit(GlobalContext *Ctx) {
(void)Ctx;
RegNumT::setLimit(RegMIPS32::Reg_NUM);
SmallBitVector IntegerRegisters(RegMIPS32::Reg_NUM);
SmallBitVector I64PairRegisters(RegMIPS32::Reg_NUM);
SmallBitVector Float32Registers(RegMIPS32::Reg_NUM);
SmallBitVector Float64Registers(RegMIPS32::Reg_NUM);
SmallBitVector VectorRegisters(RegMIPS32::Reg_NUM);
SmallBitVector InvalidRegisters(RegMIPS32::Reg_NUM);
#define X(val, encode, name, scratch, preserved, stackptr, frameptr, isInt, \
isI64Pair, isFP32, isFP64, isVec128, alias_init) \
IntegerRegisters[RegMIPS32::val] = isInt; \
I64PairRegisters[RegMIPS32::val] = isI64Pair; \
Float32Registers[RegMIPS32::val] = isFP32; \
Float64Registers[RegMIPS32::val] = isFP64; \
VectorRegisters[RegMIPS32::val] = isVec128; \
RegisterAliases[RegMIPS32::val].resize(RegMIPS32::Reg_NUM); \
for (SizeT RegAlias : alias_init) { \
assert(!RegisterAliases[RegMIPS32::val][RegAlias] && \
"Duplicate alias for " #val); \
RegisterAliases[RegMIPS32::val].set(RegAlias); \
} \
RegisterAliases[RegMIPS32::val].resize(RegMIPS32::Reg_NUM); \
assert(RegisterAliases[RegMIPS32::val][RegMIPS32::val]);
REGMIPS32_TABLE;
#undef X
// TODO(mohit.bhakkad): Change these inits once we provide argument related
// field in register tables
for (size_t i = 0; i < MIPS32_MAX_GPR_ARG; i++)
GPRArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_A0 + i);
for (size_t i = 0; i < MIPS32_MAX_GPR_ARG / 2; i++)
I64ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_A0A1 + i);
for (size_t i = 0; i < MIPS32_MAX_FP_ARG; i++) {
FP32ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_F12 + i * 2);
FP64ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_F12F13 + i);
}
TypeToRegisterSet[IceType_void] = InvalidRegisters;
TypeToRegisterSet[IceType_i1] = IntegerRegisters;
TypeToRegisterSet[IceType_i8] = IntegerRegisters;
TypeToRegisterSet[IceType_i16] = IntegerRegisters;
TypeToRegisterSet[IceType_i32] = IntegerRegisters;
TypeToRegisterSet[IceType_i64] = IntegerRegisters;
TypeToRegisterSet[IceType_f32] = Float32Registers;
TypeToRegisterSet[IceType_f64] = Float64Registers;
TypeToRegisterSet[IceType_v4i1] = VectorRegisters;
TypeToRegisterSet[IceType_v8i1] = VectorRegisters;
TypeToRegisterSet[IceType_v16i1] = VectorRegisters;
TypeToRegisterSet[IceType_v16i8] = VectorRegisters;
TypeToRegisterSet[IceType_v8i16] = VectorRegisters;
TypeToRegisterSet[IceType_v4i32] = VectorRegisters;
TypeToRegisterSet[IceType_v4f32] = VectorRegisters;
for (size_t i = 0; i < llvm::array_lengthof(TypeToRegisterSet); ++i)
TypeToRegisterSetUnfiltered[i] = TypeToRegisterSet[i];
filterTypeToRegisterSet(Ctx, RegMIPS32::Reg_NUM, TypeToRegisterSet,
llvm::array_lengthof(TypeToRegisterSet),
RegMIPS32::getRegName, getRegClassName);
}
void TargetMIPS32::unsetIfNonLeafFunc() {
for (CfgNode *Node : Func->getNodes()) {
for (Inst &Instr : Node->getInsts()) {
if (llvm::isa<InstCall>(&Instr)) {
// Unset MaybeLeafFunc if call instruction exists.
MaybeLeafFunc = false;
return;
}
}
}
}
uint32_t TargetMIPS32::getStackAlignment() const {
return MIPS32_STACK_ALIGNMENT_BYTES;
}
uint32_t TargetMIPS32::getCallStackArgumentsSizeBytes(const InstCall *Call) {
TargetMIPS32::CallingConv CC;
size_t OutArgsSizeBytes = 0;
Variable *Dest = Call->getDest();
bool PartialOnStack = false;
if (Dest != nullptr && isVectorFloatingType(Dest->getType())) {
CC.discardReg(RegMIPS32::Reg_A0);
// Next vector is partially on stack
PartialOnStack = true;
}
for (SizeT i = 0, NumArgs = Call->getNumArgs(); i < NumArgs; ++i) {
Operand *Arg = legalizeUndef(Call->getArg(i));
const Type Ty = Arg->getType();
RegNumT RegNum;
if (CC.argInReg(Ty, i, &RegNum)) {
// If PartialOnStack is true and if this is a vector type then last two
// elements are on stack
if (PartialOnStack && isVectorType(Ty)) {
OutArgsSizeBytes = applyStackAlignmentTy(OutArgsSizeBytes, IceType_i64);
OutArgsSizeBytes += typeWidthInBytesOnStack(IceType_i32) * 2;
}
continue;
}
OutArgsSizeBytes = applyStackAlignmentTy(OutArgsSizeBytes, Ty);
OutArgsSizeBytes += typeWidthInBytesOnStack(Ty);
}
// Add size of argument save area
constexpr int BytesPerStackArg = 4;
OutArgsSizeBytes += MIPS32_MAX_GPR_ARG * BytesPerStackArg;
return applyStackAlignment(OutArgsSizeBytes);
}
namespace {
inline uint64_t getConstantMemoryOrder(Operand *Opnd) {
if (auto *Integer = llvm::dyn_cast<ConstantInteger32>(Opnd))
return Integer->getValue();
return Intrinsics::MemoryOrderInvalid;
}
} // namespace
void TargetMIPS32::genTargetHelperCallFor(Inst *Instr) {
constexpr bool NoTailCall = false;
constexpr bool IsTargetHelperCall = true;
Variable *Dest = Instr->getDest();
const Type DestTy = Dest ? Dest->getType() : IceType_void;
switch (Instr->getKind()) {
default:
return;
case Inst::Select: {
if (isVectorType(DestTy)) {
Operand *SrcT = llvm::cast<InstSelect>(Instr)->getTrueOperand();
Operand *SrcF = llvm::cast<InstSelect>(Instr)->getFalseOperand();
Operand *Cond = llvm::cast<InstSelect>(Instr)->getCondition();
Variable *T = Func->makeVariable(DestTy);
auto *Undef = ConstantUndef::create(Ctx, DestTy);
Context.insert<InstAssign>(T, Undef);
auto *VarVecOn32 = llvm::cast<VariableVecOn32>(T);
VarVecOn32->initVecElement(Func);
for (SizeT I = 0; I < typeNumElements(DestTy); ++I) {
auto *Index = Ctx->getConstantInt32(I);
auto *OpC = Func->makeVariable(typeElementType(Cond->getType()));
Context.insert<InstExtractElement>(OpC, Cond, Index);
auto *OpT = Func->makeVariable(typeElementType(DestTy));
Context.insert<InstExtractElement>(OpT, SrcT, Index);
auto *OpF = Func->makeVariable(typeElementType(DestTy));
Context.insert<InstExtractElement>(OpF, SrcF, Index);
auto *Dst = Func->makeVariable(typeElementType(DestTy));
Variable *DestT = Func->makeVariable(DestTy);
Context.insert<InstSelect>(Dst, OpC, OpT, OpF);
Context.insert<InstInsertElement>(DestT, T, Dst, Index);
T = DestT;
}
Context.insert<InstAssign>(Dest, T);
Instr->setDeleted();
}
return;
}
case Inst::Fcmp: {
if (isVectorType(DestTy)) {
InstFcmp::FCond Cond = llvm::cast<InstFcmp>(Instr)->getCondition();
Operand *Src0 = Instr->getSrc(0);
Operand *Src1 = Instr->getSrc(1);
Variable *T = Func->makeVariable(IceType_v4f32);
auto *Undef = ConstantUndef::create(Ctx, IceType_v4f32);
Context.insert<InstAssign>(T, Undef);
auto *VarVecOn32 = llvm::cast<VariableVecOn32>(T);
VarVecOn32->initVecElement(Func);
for (SizeT I = 0; I < typeNumElements(IceType_v4f32); ++I) {
auto *Index = Ctx->getConstantInt32(I);
auto *Op0 = Func->makeVariable(IceType_f32);
Context.insert<InstExtractElement>(Op0, Src0, Index);
auto *Op1 = Func->makeVariable(IceType_f32);
Context.insert<InstExtractElement>(Op1, Src1, Index);
auto *Dst = Func->makeVariable(IceType_f32);
Variable *DestT = Func->makeVariable(IceType_v4f32);
Context.insert<InstFcmp>(Cond, Dst, Op0, Op1);
Context.insert<InstInsertElement>(DestT, T, Dst, Index);
T = DestT;
}
Context.insert<InstAssign>(Dest, T);
Instr->setDeleted();
}
return;
}
case Inst::Icmp: {
if (isVectorType(DestTy)) {
InstIcmp::ICond Cond = llvm::cast<InstIcmp>(Instr)->getCondition();
Operand *Src0 = Instr->getSrc(0);
Operand *Src1 = Instr->getSrc(1);
const Type SrcType = Src0->getType();
Variable *T = Func->makeVariable(DestTy);
auto *Undef = ConstantUndef::create(Ctx, DestTy);
Context.insert<InstAssign>(T, Undef);
auto *VarVecOn32 = llvm::cast<VariableVecOn32>(T);
VarVecOn32->initVecElement(Func);
for (SizeT I = 0; I < typeNumElements(SrcType); ++I) {
auto *Index = Ctx->getConstantInt32(I);
auto *Op0 = Func->makeVariable(typeElementType(SrcType));
Context.insert<InstExtractElement>(Op0, Src0, Index);
auto *Op1 = Func->makeVariable(typeElementType(SrcType));
Context.insert<InstExtractElement>(Op1, Src1, Index);
auto *Dst = Func->makeVariable(typeElementType(DestTy));
Variable *DestT = Func->makeVariable(DestTy);
Context.insert<InstIcmp>(Cond, Dst, Op0, Op1);
Context.insert<InstInsertElement>(DestT, T, Dst, Index);
T = DestT;
}
Context.insert<InstAssign>(Dest, T);
Instr->setDeleted();
}
return;
}
case Inst::Arithmetic: {
const InstArithmetic::OpKind Op =
llvm::cast<InstArithmetic>(Instr)->getOp();
if (isVectorType(DestTy)) {
scalarizeArithmetic(Op, Dest, Instr->getSrc(0), Instr->getSrc(1));
Instr->setDeleted();
return;
}
switch (DestTy) {
default:
return;
case IceType_i64: {
RuntimeHelper HelperID = RuntimeHelper::H_Num;
switch (Op) {
default:
return;
case InstArithmetic::Udiv:
HelperID = RuntimeHelper::H_udiv_i64;
break;
case InstArithmetic::Sdiv:
HelperID = RuntimeHelper::H_sdiv_i64;
break;
case InstArithmetic::Urem:
HelperID = RuntimeHelper::H_urem_i64;
break;
case InstArithmetic::Srem:
HelperID = RuntimeHelper::H_srem_i64;
break;
}
if (HelperID == RuntimeHelper::H_Num) {
return;
}
Operand *TargetHelper = Ctx->getRuntimeHelperFunc(HelperID);
constexpr SizeT MaxArgs = 2;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Instr->getSrc(0));
Call->addArg(Instr->getSrc(1));
Instr->setDeleted();
return;
}
case IceType_f32:
case IceType_f64: {
if (Op != InstArithmetic::Frem) {
return;
}
constexpr SizeT MaxArgs = 2;
Operand *TargetHelper = Ctx->getRuntimeHelperFunc(
DestTy == IceType_f32 ? RuntimeHelper::H_frem_f32
: RuntimeHelper::H_frem_f64);
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Instr->getSrc(0));
Call->addArg(Instr->getSrc(1));
Instr->setDeleted();
return;
}
}
llvm::report_fatal_error("Control flow should never have reached here.");
}
case Inst::Cast: {
Operand *Src0 = Instr->getSrc(0);
const Type SrcTy = Src0->getType();
auto *CastInstr = llvm::cast<InstCast>(Instr);
const InstCast::OpKind CastKind = CastInstr->getCastKind();
if (isVectorType(DestTy)) {
Variable *T = Func->makeVariable(DestTy);
auto *VarVecOn32 = llvm::cast<VariableVecOn32>(T);
VarVecOn32->initVecElement(Func);
auto *Undef = ConstantUndef::create(Ctx, DestTy);
Context.insert<InstAssign>(T, Undef);
for (SizeT I = 0; I < typeNumElements(DestTy); ++I) {
auto *Index = Ctx->getConstantInt32(I);
auto *Op = Func->makeVariable(typeElementType(SrcTy));
Context.insert<InstExtractElement>(Op, Src0, Index);
auto *Dst = Func->makeVariable(typeElementType(DestTy));
Variable *DestT = Func->makeVariable(DestTy);
Context.insert<InstCast>(CastKind, Dst, Op);
Context.insert<InstInsertElement>(DestT, T, Dst, Index);
T = DestT;
}
Context.insert<InstAssign>(Dest, T);
Instr->setDeleted();
return;
}
switch (CastKind) {
default:
return;
case InstCast::Fptosi:
case InstCast::Fptoui: {
if ((DestTy != IceType_i32) && (DestTy != IceType_i64)) {
return;
}
const bool DestIs32 = DestTy == IceType_i32;
const bool DestIsSigned = CastKind == InstCast::Fptosi;
const bool Src0IsF32 = isFloat32Asserting32Or64(SrcTy);
RuntimeHelper RTHFunc = RuntimeHelper::H_Num;
if (DestIsSigned) {
if (DestIs32) {
return;
}
RTHFunc = Src0IsF32 ? RuntimeHelper::H_fptosi_f32_i64
: RuntimeHelper::H_fptosi_f64_i64;
} else {
RTHFunc = Src0IsF32 ? (DestIs32 ? RuntimeHelper::H_fptoui_f32_i32
: RuntimeHelper::H_fptoui_f32_i64)
: (DestIs32 ? RuntimeHelper::H_fptoui_f64_i32
: RuntimeHelper::H_fptoui_f64_i64);
}
Operand *TargetHelper = Ctx->getRuntimeHelperFunc(RTHFunc);
static constexpr SizeT MaxArgs = 1;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Src0);
Instr->setDeleted();
return;
}
case InstCast::Sitofp:
case InstCast::Uitofp: {
if ((SrcTy != IceType_i32) && (SrcTy != IceType_i64)) {
return;
}
const bool SourceIs32 = SrcTy == IceType_i32;
const bool SourceIsSigned = CastKind == InstCast::Sitofp;
const bool DestIsF32 = isFloat32Asserting32Or64(DestTy);
RuntimeHelper RTHFunc = RuntimeHelper::H_Num;
if (SourceIsSigned) {
if (SourceIs32) {
return;
}
RTHFunc = DestIsF32 ? RuntimeHelper::H_sitofp_i64_f32
: RuntimeHelper::H_sitofp_i64_f64;
} else {
RTHFunc = DestIsF32 ? (SourceIs32 ? RuntimeHelper::H_uitofp_i32_f32
: RuntimeHelper::H_uitofp_i64_f32)
: (SourceIs32 ? RuntimeHelper::H_uitofp_i32_f64
: RuntimeHelper::H_uitofp_i64_f64);
}
Operand *TargetHelper = Ctx->getRuntimeHelperFunc(RTHFunc);
static constexpr SizeT MaxArgs = 1;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Src0);
Instr->setDeleted();
return;
}
case InstCast::Bitcast: {
if (DestTy == SrcTy) {
return;
}
Variable *CallDest = Dest;
RuntimeHelper HelperID = RuntimeHelper::H_Num;
switch (DestTy) {
default:
return;
case IceType_i8:
assert(SrcTy == IceType_v8i1);
HelperID = RuntimeHelper::H_bitcast_8xi1_i8;
CallDest = Func->makeVariable(IceType_i32);
break;
case IceType_i16:
assert(SrcTy == IceType_v16i1);
HelperID = RuntimeHelper::H_bitcast_16xi1_i16;
CallDest = Func->makeVariable(IceType_i32);
break;
case IceType_v8i1: {
assert(SrcTy == IceType_i8);
HelperID = RuntimeHelper::H_bitcast_i8_8xi1;
Variable *Src0AsI32 = Func->makeVariable(stackSlotType());
// Arguments to functions are required to be at least 32 bits wide.
Context.insert<InstCast>(InstCast::Zext, Src0AsI32, Src0);
Src0 = Src0AsI32;
} break;
case IceType_v16i1: {
assert(SrcTy == IceType_i16);
HelperID = RuntimeHelper::H_bitcast_i16_16xi1;
Variable *Src0AsI32 = Func->makeVariable(stackSlotType());
// Arguments to functions are required to be at least 32 bits wide.
Context.insert<InstCast>(InstCast::Zext, Src0AsI32, Src0);
Src0 = Src0AsI32;
} break;
}
constexpr SizeT MaxSrcs = 1;
InstCall *Call = makeHelperCall(HelperID, CallDest, MaxSrcs);
Call->addArg(Src0);
Context.insert(Call);
// The PNaCl ABI disallows i8/i16 return types, so truncate the helper
// call result to the appropriate type as necessary.
if (CallDest->getType() != DestTy)
Context.insert<InstCast>(InstCast::Trunc, Dest, CallDest);
Instr->setDeleted();
return;
}
case InstCast::Trunc: {
if (DestTy == SrcTy) {
return;
}
if (!isVectorType(SrcTy)) {
return;
}
assert(typeNumElements(DestTy) == typeNumElements(SrcTy));
assert(typeElementType(DestTy) == IceType_i1);
assert(isVectorIntegerType(SrcTy));
return;
}
case InstCast::Sext:
case InstCast::Zext: {
if (DestTy == SrcTy) {
return;
}
if (!isVectorType(DestTy)) {
return;
}
assert(typeNumElements(DestTy) == typeNumElements(SrcTy));
assert(typeElementType(SrcTy) == IceType_i1);
assert(isVectorIntegerType(DestTy));
return;
}
}
llvm::report_fatal_error("Control flow should never have reached here.");
}
case Inst::Intrinsic: {
auto *Intrinsic = llvm::cast<InstIntrinsic>(Instr);
Intrinsics::IntrinsicID ID = Intrinsic->getIntrinsicID();
if (isVectorType(DestTy) && ID == Intrinsics::Fabs) {
Operand *Src0 = Intrinsic->getArg(0);
Intrinsics::IntrinsicInfo Info = Intrinsic->getIntrinsicInfo();
Variable *T = Func->makeVariable(IceType_v4f32);
auto *Undef = ConstantUndef::create(Ctx, IceType_v4f32);
Context.insert<InstAssign>(T, Undef);
auto *VarVecOn32 = llvm::cast<VariableVecOn32>(T);
VarVecOn32->initVecElement(Func);
for (SizeT i = 0; i < typeNumElements(IceType_v4f32); ++i) {
auto *Index = Ctx->getConstantInt32(i);
auto *Op = Func->makeVariable(IceType_f32);
Context.insert<InstExtractElement>(Op, Src0, Index);
auto *Res = Func->makeVariable(IceType_f32);
Variable *DestT = Func->makeVariable(IceType_v4f32);
auto *Intrinsic = Context.insert<InstIntrinsic>(1, Res, Info);
Intrinsic->addArg(Op);
Context.insert<InstInsertElement>(DestT, T, Res, Index);
T = DestT;
}
Context.insert<InstAssign>(Dest, T);
Instr->setDeleted();
return;
}
switch (ID) {
default:
return;
case Intrinsics::AtomicLoad: {
if (DestTy != IceType_i64)
return;
if (!Intrinsics::isMemoryOrderValid(
ID, getConstantMemoryOrder(Intrinsic->getArg(1)))) {
Func->setError("Unexpected memory ordering for AtomicLoad");
return;
}
Operand *Addr = Intrinsic->getArg(0);
Operand *TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_val_compare_and_swap_8"));
static constexpr SizeT MaxArgs = 3;
auto *_0 = Ctx->getConstantZero(IceType_i64);
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Addr);
Call->addArg(_0);
Call->addArg(_0);
Context.insert<InstMIPS32Sync>();
Instr->setDeleted();
return;
}
case Intrinsics::AtomicStore: {
Operand *Val = Intrinsic->getArg(0);
if (Val->getType() != IceType_i64)
return;
if (!Intrinsics::isMemoryOrderValid(
ID, getConstantMemoryOrder(Intrinsic->getArg(2)))) {
Func->setError("Unexpected memory ordering for AtomicStore");
return;
}
Operand *Addr = Intrinsic->getArg(1);
Variable *NoDest = nullptr;
Operand *TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_lock_test_and_set_8"));
Context.insert<InstMIPS32Sync>();
static constexpr SizeT MaxArgs = 2;
auto *Call = Context.insert<InstCall>(MaxArgs, NoDest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Addr);
Call->addArg(Val);
Context.insert<InstMIPS32Sync>();
Instr->setDeleted();
return;
}
case Intrinsics::AtomicCmpxchg: {
if (DestTy != IceType_i64)
return;
if (!Intrinsics::isMemoryOrderValid(
ID, getConstantMemoryOrder(Intrinsic->getArg(3)),
getConstantMemoryOrder(Intrinsic->getArg(4)))) {
Func->setError("Unexpected memory ordering for AtomicCmpxchg");
return;
}
Operand *Addr = Intrinsic->getArg(0);
Operand *Oldval = Intrinsic->getArg(1);
Operand *Newval = Intrinsic->getArg(2);
Operand *TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_val_compare_and_swap_8"));
Context.insert<InstMIPS32Sync>();
static constexpr SizeT MaxArgs = 3;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Addr);
Call->addArg(Oldval);
Call->addArg(Newval);
Context.insert<InstMIPS32Sync>();
Instr->setDeleted();
return;
}
case Intrinsics::AtomicRMW: {
if (DestTy != IceType_i64)
return;
if (!Intrinsics::isMemoryOrderValid(
ID, getConstantMemoryOrder(Intrinsic->getArg(3)))) {
Func->setError("Unexpected memory ordering for AtomicRMW");
return;
}
auto Operation = static_cast<Intrinsics::AtomicRMWOperation>(
llvm::cast<ConstantInteger32>(Intrinsic->getArg(0))->getValue());
auto *Addr = Intrinsic->getArg(1);
auto *Newval = Intrinsic->getArg(2);
Operand *TargetHelper;
switch (Operation) {
case Intrinsics::AtomicAdd:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_fetch_and_add_8"));
break;
case Intrinsics::AtomicSub:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_fetch_and_sub_8"));
break;
case Intrinsics::AtomicOr:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_fetch_and_or_8"));
break;
case Intrinsics::AtomicAnd:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_fetch_and_and_8"));
break;
case Intrinsics::AtomicXor:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_fetch_and_xor_8"));
break;
case Intrinsics::AtomicExchange:
TargetHelper = Ctx->getConstantExternSym(
Ctx->getGlobalString("__sync_lock_test_and_set_8"));
break;
default:
llvm::report_fatal_error("Unknown AtomicRMW operation");
return;
}
Context.insert<InstMIPS32Sync>();
static constexpr SizeT MaxArgs = 2;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Addr);
Call->addArg(Newval);
Context.insert<InstMIPS32Sync>();
Instr->setDeleted();
return;
}
case Intrinsics::Ctpop: {
Operand *Src0 = Intrinsic->getArg(0);
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(isInt32Asserting32Or64(Src0->getType())
? RuntimeHelper::H_call_ctpop_i32
: RuntimeHelper::H_call_ctpop_i64);
static constexpr SizeT MaxArgs = 1;
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Src0);
Instr->setDeleted();
return;
}
case Intrinsics::Longjmp: {
static constexpr SizeT MaxArgs = 2;
static constexpr Variable *NoDest = nullptr;
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(RuntimeHelper::H_call_longjmp);
auto *Call = Context.insert<InstCall>(MaxArgs, NoDest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Intrinsic->getArg(0));
Call->addArg(Intrinsic->getArg(1));
Instr->setDeleted();
return;
}
case Intrinsics::Memcpy: {
static constexpr SizeT MaxArgs = 3;
static constexpr Variable *NoDest = nullptr;
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(RuntimeHelper::H_call_memcpy);
auto *Call = Context.insert<InstCall>(MaxArgs, NoDest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Intrinsic->getArg(0));
Call->addArg(Intrinsic->getArg(1));
Call->addArg(Intrinsic->getArg(2));
Instr->setDeleted();
return;
}
case Intrinsics::Memmove: {
static constexpr SizeT MaxArgs = 3;
static constexpr Variable *NoDest = nullptr;
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(RuntimeHelper::H_call_memmove);
auto *Call = Context.insert<InstCall>(MaxArgs, NoDest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Intrinsic->getArg(0));
Call->addArg(Intrinsic->getArg(1));
Call->addArg(Intrinsic->getArg(2));
Instr->setDeleted();
return;
}
case Intrinsics::Memset: {
Operand *ValOp = Intrinsic->getArg(1);
assert(ValOp->getType() == IceType_i8);
Variable *ValExt = Func->makeVariable(stackSlotType());
Context.insert<InstCast>(InstCast::Zext, ValExt, ValOp);
static constexpr SizeT MaxArgs = 3;
static constexpr Variable *NoDest = nullptr;
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(RuntimeHelper::H_call_memset);
auto *Call = Context.insert<InstCall>(MaxArgs, NoDest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Intrinsic->getArg(0));
Call->addArg(ValExt);
Call->addArg(Intrinsic->getArg(2));
Instr->setDeleted();
return;
}
case Intrinsics::Setjmp: {
static constexpr SizeT MaxArgs = 1;
Operand *TargetHelper =
Ctx->getRuntimeHelperFunc(RuntimeHelper::H_call_setjmp);
auto *Call = Context.insert<InstCall>(MaxArgs, Dest, TargetHelper,
NoTailCall, IsTargetHelperCall);
Call->addArg(Intrinsic->getArg(0));
Instr->setDeleted();
return;
}
}
llvm::report_fatal_error("Control flow should never have reached here.");
}
}
}
void TargetMIPS32::findMaxStackOutArgsSize() {
// MinNeededOutArgsBytes should be updated if the Target ever creates a
// high-level InstCall that requires more stack bytes.
size_t MinNeededOutArgsBytes = 0;
if (!MaybeLeafFunc)
MinNeededOutArgsBytes = MIPS32_MAX_GPR_ARG * 4;
MaxOutArgsSizeBytes = MinNeededOutArgsBytes;
for (CfgNode *Node : Func->getNodes()) {
Context.init(Node);
while (!Context.atEnd()) {
PostIncrLoweringContext PostIncrement(Context);
Inst *CurInstr = iteratorToInst(Context.getCur());
if (auto *Call = llvm::dyn_cast<InstCall>(CurInstr)) {
SizeT OutArgsSizeBytes = getCallStackArgumentsSizeBytes(Call);
MaxOutArgsSizeBytes = std::max(MaxOutArgsSizeBytes, OutArgsSizeBytes);
}
}
}
CurrentAllocaOffset = MaxOutArgsSizeBytes;
}
void TargetMIPS32::translateO2() {
TimerMarker T(TimerStack::TT_O2, Func);
// TODO(stichnot): share passes with X86?
// https://code.google.com/p/nativeclient/issues/detail?id=4094
genTargetHelperCalls();
unsetIfNonLeafFunc();
findMaxStackOutArgsSize();
// Merge Alloca instructions, and lay out the stack.
static constexpr bool SortAndCombineAllocas = true;
Func->processAllocas(SortAndCombineAllocas);
Func->dump("After Alloca processing");
if (!getFlags().getEnablePhiEdgeSplit()) {
// Lower Phi instructions.
Func->placePhiLoads();
if (Func->hasError())
return;
Func->placePhiStores();
if (Func->hasError())
return;
Func->deletePhis();
if (Func->hasError())
return;
Func->dump("After Phi lowering");
}
// Address mode optimization.
Func->getVMetadata()->init(VMK_SingleDefs);
Func->doAddressOpt();
// Argument lowering
Func->doArgLowering();
// Target lowering. This requires liveness analysis for some parts of the
// lowering decisions, such as compare/branch fusing. If non-lightweight
// liveness analysis is used, the instructions need to be renumbered first.
// TODO: This renumbering should only be necessary if we're actually
// calculating live intervals, which we only do for register allocation.
Func->renumberInstructions();
if (Func->hasError())
return;
// TODO: It should be sufficient to use the fastest liveness calculation,
// i.e. livenessLightweight(). However, for some reason that slows down the
// rest of the translation. Investigate.
Func->liveness(Liveness_Basic);
if (Func->hasError())
return;
Func->dump("After MIPS32 address mode opt");
Func->genCode();
if (Func->hasError())
return;
Func->dump("After MIPS32 codegen");
// Register allocation. This requires instruction renumbering and full
// liveness analysis.
Func->renumberInstructions();
if (Func->hasError())
return;
Func->liveness(Liveness_Intervals);
if (Func->hasError())
return;
// The post-codegen dump is done here, after liveness analysis and associated
// cleanup, to make the dump cleaner and more useful.
Func->dump("After initial MIPS32 codegen");
// Validate the live range computations. The expensive validation call is
// deliberately only made when assertions are enabled.
assert(Func->validateLiveness());
Func->getVMetadata()->init(VMK_All);
regAlloc(RAK_Global);
if (Func->hasError())
return;
Func->dump("After linear scan regalloc");
if (getFlags().getEnablePhiEdgeSplit()) {
Func->advancedPhiLowering();
Func->dump("After advanced Phi lowering");
}
// Stack frame mapping.
Func->genFrame();
if (Func->hasError())
return;
Func->dump("After stack frame mapping");
postLowerLegalization();
if (Func->hasError())
return;
Func->dump("After postLowerLegalization");
Func->contractEmptyNodes();
Func->reorderNodes();
// Branch optimization. This needs to be done just before code emission. In
// particular, no transformations that insert or reorder CfgNodes should be
// done after branch optimization. We go ahead and do it before nop insertion
// to reduce the amount of work needed for searching for opportunities.
Func->doBranchOpt();
Func->dump("After branch optimization");
}
void TargetMIPS32::translateOm1() {
TimerMarker T(TimerStack::TT_Om1, Func);
// TODO: share passes with X86?
genTargetHelperCalls();
unsetIfNonLeafFunc();
findMaxStackOutArgsSize();
// Do not merge Alloca instructions, and lay out the stack.
static constexpr bool SortAndCombineAllocas = false;
Func->processAllocas(SortAndCombineAllocas);
Func->dump("After Alloca processing");
Func->placePhiLoads();
if (Func->hasError())
return;
Func->placePhiStores();
if (Func->hasError())
return;
Func->deletePhis();
if (Func->hasError())
return;
Func->dump("After Phi lowering");
Func->doArgLowering();
Func->genCode();
if (Func->hasError())
return;
Func->dump("After initial MIPS32 codegen");
regAlloc(RAK_InfOnly);
if (Func->hasError())
return;
Func->dump("After regalloc of infinite-weight variables");
Func->genFrame();
if (Func->hasError())
return;
Func->dump("After stack frame mapping");
postLowerLegalization();
if (Func->hasError())
return;
Func->dump("After postLowerLegalization");
}
bool TargetMIPS32::doBranchOpt(Inst *Instr, const CfgNode *NextNode) {
if (auto *Br = llvm::dyn_cast<InstMIPS32Br>(Instr)) {
return Br->optimizeBranch(NextNode);
}
return false;
}
namespace {
const char *RegNames[RegMIPS32::Reg_NUM] = {
#define X(val, encode, name, scratch, preserved, stackptr, frameptr, isInt, \
isI64Pair, isFP32, isFP64, isVec128, alias_init) \
name,
REGMIPS32_TABLE
#undef X
};
} // end of anonymous namespace
const char *RegMIPS32::getRegName(RegNumT RegNum) {
RegNum.assertIsValid();
return RegNames[RegNum];
}
const char *TargetMIPS32::getRegName(RegNumT RegNum, Type Ty) const {
(void)Ty;
return RegMIPS32::getRegName(RegNum);
}
Variable *TargetMIPS32::getPhysicalRegister(RegNumT RegNum, Type Ty) {
if (Ty == IceType_void)
Ty = IceType_i32;
if (PhysicalRegisters[Ty].empty())
PhysicalRegisters[Ty].resize(RegMIPS32::Reg_NUM);
RegNum.assertIsValid();
Variable *Reg = PhysicalRegisters[Ty][RegNum];
if (Reg == nullptr) {
Reg = Func->makeVariable(Ty);
Reg->setRegNum(RegNum);
PhysicalRegisters[Ty][RegNum] = Reg;
// Specially mark a named physical register as an "argument" so that it is
// considered live upon function entry. Otherwise it's possible to get
// liveness validation errors for saving callee-save registers.
Func->addImplicitArg(Reg);
// Don't bother tracking the live range of a named physical register.
Reg->setIgnoreLiveness();
}
return Reg;
}
void TargetMIPS32::emitJumpTable(const Cfg *Func,
const InstJumpTable *JumpTable) const {
(void)Func;
(void)JumpTable;
UnimplementedError(getFlags());
}
/// Provide a trivial wrapper to legalize() for this common usage.
Variable *TargetMIPS32::legalizeToReg(Operand *From, RegNumT RegNum) {
return llvm::cast<Variable>(legalize(From, Legal_Reg, RegNum));
}
/// Legalize undef values to concrete values.
Operand *TargetMIPS32::legalizeUndef(Operand *From, RegNumT RegNum) {
(void)RegNum;
Type Ty = From->getType();
if (llvm::isa<ConstantUndef>(From)) {
// Lower undefs to zero. Another option is to lower undefs to an
// uninitialized register; however, using an uninitialized register
// results in less predictable code.
//
// If in the future the implementation is changed to lower undef
// values to uninitialized registers, a FakeDef will be needed:
// Context.insert(InstFakeDef::create(Func, Reg));
// This is in order to ensure that the live range of Reg is not
// overestimated. If the constant being lowered is a 64 bit value,
// then the result should be split and the lo and hi components will
// need to go in uninitialized registers.
if (isVectorType(Ty)) {
Variable *Var = makeReg(Ty, RegNum);
auto *Reg = llvm::cast<VariableVecOn32>(Var);
Reg->initVecElement(Func);
auto *Zero = getZero();
for (Variable *Var : Reg->getContainers()) {
_mov(Var, Zero);
}
return Reg;
}
return Ctx->getConstantZero(Ty);
}
return From;
}
Variable *TargetMIPS32::makeReg(Type Type, RegNumT RegNum) {
// There aren't any 64-bit integer registers for Mips32.
assert(Type != IceType_i64);
Variable *Reg = Func->makeVariable(Type);
if (RegNum.hasValue())
Reg->setRegNum(RegNum);
else
Reg->setMustHaveReg();
return Reg;
}
OperandMIPS32Mem *TargetMIPS32::formMemoryOperand(Operand *Operand, Type Ty) {
// It may be the case that address mode optimization already creates an
// OperandMIPS32Mem, so in that case it wouldn't need another level of
// transformation.
if (auto *Mem = llvm::dyn_cast<OperandMIPS32Mem>(Operand)) {
return llvm::cast<OperandMIPS32Mem>(legalize(Mem));
}
// If we didn't do address mode optimization, then we only have a base/offset
// to work with. MIPS always requires a base register, so just use that to
// hold the operand.
auto *Base = llvm::cast<Variable>(
legalize(Operand, Legal_Reg | Legal_Rematerializable));
const int32_t Offset = Base->hasStackOffset() ? Base->getStackOffset() : 0;
return OperandMIPS32Mem::create(
Func, Ty, Base,
llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(Offset)));
}
void TargetMIPS32::emitVariable(const Variable *Var) const {
if (!BuildDefs::dump())
return;
Ostream &Str = Ctx->getStrEmit();
const Type FrameSPTy = IceType_i32;
if (Var->hasReg()) {
Str << '$' << getRegName(Var->getRegNum(), Var->getType());
return;
}
if (Var->mustHaveReg()) {
llvm::report_fatal_error("Infinite-weight Variable (" + Var->getName() +
") has no register assigned - function " +
Func->getFunctionName());
}
const int32_t Offset = Var->getStackOffset();
Str << Offset;
Str << "($" << getRegName(getFrameOrStackReg(), FrameSPTy);
Str << ")";
}
TargetMIPS32::CallingConv::CallingConv()
: GPRegsUsed(RegMIPS32::Reg_NUM),
GPRArgs(GPRArgInitializer.rbegin(), GPRArgInitializer.rend()),
I64Args(I64ArgInitializer.rbegin(), I64ArgInitializer.rend()),
VFPRegsUsed(RegMIPS32::Reg_NUM),
FP32Args(FP32ArgInitializer.rbegin(), FP32ArgInitializer.rend()),
FP64Args(FP64ArgInitializer.rbegin(), FP64ArgInitializer.rend()) {}
// In MIPS O32 abi FP argument registers can be used only if first argument is
// of type float/double. UseFPRegs flag is used to care of that. Also FP arg
// registers can be used only for first 2 arguments, so we require argument
// number to make register allocation decisions.
bool TargetMIPS32::CallingConv::argInReg(Type Ty, uint32_t ArgNo,
RegNumT *Reg) {
if (isScalarIntegerType(Ty) || isVectorType(Ty))
return argInGPR(Ty, Reg);
if (isScalarFloatingType(Ty)) {
if (ArgNo == 0) {
UseFPRegs = true;
return argInVFP(Ty, Reg);
}
if (UseFPRegs && ArgNo == 1) {
UseFPRegs = false;
return argInVFP(Ty, Reg);
}
return argInGPR(Ty, Reg);
}
llvm::report_fatal_error("argInReg: Invalid type.");
return false;
}
bool TargetMIPS32::CallingConv::argInGPR(Type Ty, RegNumT *Reg) {
CfgVector<RegNumT> *Source;
switch (Ty) {
default: {
llvm::report_fatal_error("argInGPR: Invalid type.");
return false;
} break;
case IceType_v4i1:
case IceType_v8i1:
case IceType_v16i1:
case IceType_v16i8:
case IceType_v8i16:
case IceType_v4i32:
case IceType_v4f32:
case IceType_i32:
case IceType_f32: {
Source = &GPRArgs;
} break;
case IceType_i64:
case IceType_f64: {
Source = &I64Args;
} break;
}
discardUnavailableGPRsAndTheirAliases(Source);
// If $4 is used for any scalar type (or returining v4f32) then the next
// vector type if passed in $6:$7:stack:stack
if (isVectorType(Ty)) {
alignGPR(Source);
}
if (Source->empty()) {
GPRegsUsed.set();
return false;
}
*Reg = Source->back();
// Note that we don't Source->pop_back() here. This is intentional. Notice how
// we mark all of Reg's aliases as Used. So, for the next argument,
// Source->back() is marked as unavailable, and it is thus implicitly popped
// from the stack.
GPRegsUsed |= RegisterAliases[*Reg];
// All vector arguments irrespective of their base type are passed in GP
// registers. First vector argument is passed in $4:$5:$6:$7 and 2nd
// is passed in $6:$7:stack:stack. If it is 1st argument then discard
// $4:$5:$6:$7 otherwise discard $6:$7 only.
if (isVectorType(Ty)) {
if (((unsigned)*Reg) == RegMIPS32::Reg_A0) {
GPRegsUsed |= RegisterAliases[RegMIPS32::Reg_A1];
GPRegsUsed |= RegisterAliases[RegMIPS32::Reg_A2];
GPRegsUsed |= RegisterAliases[RegMIPS32::Reg_A3];
} else {
GPRegsUsed |= RegisterAliases[RegMIPS32::Reg_A3];
}
}
return true;
}
inline void TargetMIPS32::CallingConv::discardNextGPRAndItsAliases(
CfgVector<RegNumT> *Regs) {
GPRegsUsed |= RegisterAliases[Regs->back()];
Regs->pop_back();
}
inline void TargetMIPS32::CallingConv::alignGPR(CfgVector<RegNumT> *Regs) {
if (Regs->back() == RegMIPS32::Reg_A1 || Regs->back() == RegMIPS32::Reg_A3)
discardNextGPRAndItsAliases(Regs);
}
// GPR are not packed when passing parameters. Thus, a function foo(i32, i64,
// i32) will have the first argument in a0, the second in a2-a3, and the third
// on the stack. To model this behavior, whenever we pop a register from Regs,
// we remove all of its aliases from the pool of available GPRs. This has the
// effect of computing the "closure" on the GPR registers.
void TargetMIPS32::CallingConv::discardUnavailableGPRsAndTheirAliases(
CfgVector<RegNumT> *Regs) {
while (!Regs->empty() && GPRegsUsed[Regs->back()]) {
discardNextGPRAndItsAliases(Regs);
}
}
bool TargetMIPS32::CallingConv::argInVFP(Type Ty, RegNumT *Reg) {
CfgVector<RegNumT> *Source;
switch (Ty) {
default: {
llvm::report_fatal_error("argInVFP: Invalid type.");
return false;
} break;
case IceType_f32: {
Source = &FP32Args;
} break;
case IceType_f64: {
Source = &FP64Args;
} break;
}
discardUnavailableVFPRegsAndTheirAliases(Source);
if (Source->empty()) {
VFPRegsUsed.set();
return false;
}
*Reg = Source->back();
VFPRegsUsed |= RegisterAliases[*Reg];
// In MIPS O32 abi if fun arguments are (f32, i32) then one can not use reg_a0
// for second argument even though it's free. f32 arg goes in reg_f12, i32 arg
// goes in reg_a1. Similarly if arguments are (f64, i32) second argument goes
// in reg_a3 and a0, a1 are not used.
Source = &GPRArgs;
// Discard one GPR reg for f32(4 bytes), two for f64(4 + 4 bytes)
if (Ty == IceType_f64) {
// In MIPS o32 abi, when we use GPR argument pairs to store F64 values, pair
// must be aligned at even register. Similarly when we discard GPR registers
// when some arguments from starting 16 bytes goes in FPR, we must take care
// of alignment. For example if fun args are (f32, f64, f32), for first f32
// we discard a0, now for f64 argument, which will go in F14F15, we must
// first align GPR vector to even register by discarding a1, then discard
// two GPRs a2 and a3. Now last f32 argument will go on stack.
alignGPR(Source);
discardNextGPRAndItsAliases(Source);
}
discardNextGPRAndItsAliases(Source);
return true;
}
void TargetMIPS32::CallingConv::discardUnavailableVFPRegsAndTheirAliases(
CfgVector<RegNumT> *Regs) {
while (!Regs->empty() && VFPRegsUsed[Regs->back()]) {
Regs->pop_back();
}
}
void TargetMIPS32::lowerArguments() {
VarList &Args = Func->getArgs();
TargetMIPS32::CallingConv CC;
// For each register argument, replace Arg in the argument list with the home
// register. Then generate an instruction in the prolog to copy the home
// register to the assigned location of Arg.
Context.init(Func->getEntryNode());
Context.setInsertPoint(Context.getCur());
// v4f32 is returned through stack. $4 is setup by the caller and passed as
// first argument implicitly. Callee then copies the return vector at $4.
Variable *ImplicitRetVec = nullptr;
if (isVectorFloatingType(Func->getReturnType())) {
ImplicitRetVec = Func->makeVariable(IceType_i32);
ImplicitRetVec->setName(Func, "ImplicitRet_v4f32");
ImplicitRetVec->setIsArg();
Args.insert(Args.begin(), ImplicitRetVec);
setImplicitRet(ImplicitRetVec);
}
for (SizeT i = 0, E = Args.size(); i < E; ++i) {
Variable *Arg = Args[i];
Type Ty = Arg->getType();
RegNumT RegNum;
if (!CC.argInReg(Ty, i, &RegNum)) {
continue;
}
Variable *RegisterArg = Func->makeVariable(Ty);
if (BuildDefs::dump()) {
RegisterArg->setName(Func, "home_reg:" + Arg->getName());
}
RegisterArg->setIsArg();
Arg->setIsArg(false);
Args[i] = RegisterArg;
if (isVectorType(Ty)) {
auto *RegisterArgVec = llvm::cast<VariableVecOn32>(RegisterArg);
RegisterArgVec->initVecElement(Func);
RegisterArgVec->getContainers()[0]->setRegNum(
RegNumT::fixme((unsigned)RegNum + 0));
RegisterArgVec->getContainers()[1]->setRegNum(
RegNumT::fixme((unsigned)RegNum + 1));
// First two elements of second vector argument are passed
// in $6:$7 and remaining two on stack. Do not assign register
// to this is second vector argument.
if (i == 0) {
RegisterArgVec->getContainers()[2]->setRegNum(
RegNumT::fixme((unsigned)RegNum + 2));
RegisterArgVec->getContainers()[3]->setRegNum(
RegNumT::fixme((unsigned)RegNum + 3));
} else {
RegisterArgVec->getContainers()[2]->setRegNum(
RegNumT::fixme(RegNumT()));
RegisterArgVec->getContainers()[3]->setRegNum(
RegNumT::fixme(RegNumT()));
}
} else {
switch (Ty) {
default: {
RegisterArg->setRegNum(RegNum);
} break;
case IceType_i64: {
auto *RegisterArg64 = llvm::cast<Variable64On32>(RegisterArg);
RegisterArg64->initHiLo(Func);
RegisterArg64->getLo()->setRegNum(
RegNumT::fixme(RegMIPS32::get64PairFirstRegNum(RegNum)));
RegisterArg64->getHi()->setRegNum(
RegNumT::fixme(RegMIPS32::get64PairSecondRegNum(RegNum)));
} break;
}
}
Context.insert<InstAssign>(Arg, RegisterArg);
}
// Insert fake use of ImplicitRet_v4f32 to keep it live
if (ImplicitRetVec) {
for (CfgNode *Node : Func->getNodes()) {
for (Inst &Instr : Node->getInsts()) {
if (llvm::isa<InstRet>(&Instr)) {
Context.setInsertPoint(instToIterator(&Instr));
Context.insert<InstFakeUse>(ImplicitRetVec);
break;
}
}
}
}
}
Type TargetMIPS32::stackSlotType() { return IceType_i32; }
// Helper function for addProlog().
//
// This assumes Arg is an argument passed on the stack. This sets the frame
// offset for Arg and updates InArgsSizeBytes according to Arg's width. For an
// I64 arg that has been split into Lo and Hi components, it calls itself
// recursively on the components, taking care to handle Lo first because of the
// little-endian architecture. Lastly, this function generates an instruction
// to copy Arg into its assigned register if applicable.
void TargetMIPS32::finishArgumentLowering(Variable *Arg, bool PartialOnStack,
Variable *FramePtr,
size_t BasicFrameOffset,
size_t *InArgsSizeBytes) {
const Type Ty = Arg->getType();
*InArgsSizeBytes = applyStackAlignmentTy(*InArgsSizeBytes, Ty);
// If $4 is used for any scalar type (or returining v4f32) then the next
// vector type if passed in $6:$7:stack:stack. Load 3nd and 4th element
// from agument stack.
if (auto *ArgVecOn32 = llvm::dyn_cast<VariableVecOn32>(Arg)) {
if (PartialOnStack == false) {
auto *Elem0 = ArgVecOn32->getContainers()[0];
auto *Elem1 = ArgVecOn32->getContainers()[1];
finishArgumentLowering(Elem0, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
finishArgumentLowering(Elem1, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
}
auto *Elem2 = ArgVecOn32->getContainers()[2];
auto *Elem3 = ArgVecOn32->getContainers()[3];
finishArgumentLowering(Elem2, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
finishArgumentLowering(Elem3, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
return;
}
if (auto *Arg64On32 = llvm::dyn_cast<Variable64On32>(Arg)) {
Variable *const Lo = Arg64On32->getLo();
Variable *const Hi = Arg64On32->getHi();
finishArgumentLowering(Lo, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
finishArgumentLowering(Hi, PartialOnStack, FramePtr, BasicFrameOffset,
InArgsSizeBytes);
return;
}
assert(Ty != IceType_i64);
assert(!isVectorType(Ty));
const int32_t ArgStackOffset = BasicFrameOffset + *InArgsSizeBytes;
*InArgsSizeBytes += typeWidthInBytesOnStack(Ty);
if (!Arg->hasReg()) {
Arg->setStackOffset(ArgStackOffset);
return;
}
// If the argument variable has been assigned a register, we need to copy the
// value from the stack slot.
Variable *Parameter = Func->makeVariable(Ty);
Parameter->setMustNotHaveReg();
Parameter->setStackOffset(ArgStackOffset);
_mov(Arg, Parameter);
}
void TargetMIPS32::addProlog(CfgNode *Node) {
// Stack frame layout:
//
// +------------------------+
// | 1. preserved registers |
// +------------------------+
// | 2. padding |
// +------------------------+
// | 3. global spill area |
// +------------------------+
// | 4. padding |
// +------------------------+
// | 5. local spill area |
// +------------------------+
// | 6. padding |
// +------------------------+
// | 7. allocas |
// +------------------------+
// | 8. padding |
// +------------------------+
// | 9. out args |
// +------------------------+ <--- StackPointer
//
// The following variables record the size in bytes of the given areas:
// * PreservedRegsSizeBytes: area 1
// * SpillAreaPaddingBytes: area 2
// * GlobalsSize: area 3
// * GlobalsAndSubsequentPaddingSize: areas 3 - 4
// * LocalsSpillAreaSize: area 5
// * SpillAreaSizeBytes: areas 2 - 9
// * maxOutArgsSizeBytes(): area 9
Context.init(Node);
Context.setInsertPoint(Context.getCur());
SmallBitVector CalleeSaves = getRegisterSet(RegSet_CalleeSave, RegSet_None);
RegsUsed = SmallBitVector(CalleeSaves.size());
VarList SortedSpilledVariables;
size_t GlobalsSize = 0;
// If there is a separate locals area, this represents that area. Otherwise
// it counts any variable not counted by GlobalsSize.
SpillAreaSizeBytes = 0;
// If there is a separate locals area, this specifies the alignment for it.
uint32_t LocalsSlotsAlignmentBytes = 0;
// The entire spill locations area gets aligned to largest natural alignment
// of the variables that have a spill slot.
uint32_t SpillAreaAlignmentBytes = 0;
// For now, we don't have target-specific variables that need special
// treatment (no stack-slot-linked SpillVariable type).
std::function<bool(Variable *)> TargetVarHook = [](Variable *Var) {
static constexpr bool AssignStackSlot = false;
static constexpr bool DontAssignStackSlot = !AssignStackSlot;
if (llvm::isa<Variable64On32>(Var)) {
return DontAssignStackSlot;
}
return AssignStackSlot;
};
// Compute the list of spilled variables and bounds for GlobalsSize, etc.
getVarStackSlotParams(SortedSpilledVariables, RegsUsed, &GlobalsSize,
&SpillAreaSizeBytes, &SpillAreaAlignmentBytes,
&LocalsSlotsAlignmentBytes, TargetVarHook);
uint32_t LocalsSpillAreaSize = SpillAreaSizeBytes;
SpillAreaSizeBytes += GlobalsSize;
PreservedGPRs.reserve(CalleeSaves.size());
// Consider FP and RA as callee-save / used as needed.
if (UsesFramePointer) {
if (RegsUsed[RegMIPS32::Reg_FP]) {
llvm::report_fatal_error("Frame pointer has been used.");
}
CalleeSaves[RegMIPS32::Reg_FP] = true;
RegsUsed[RegMIPS32::Reg_FP] = true;
}
if (!MaybeLeafFunc) {
CalleeSaves[RegMIPS32::Reg_RA] = true;
RegsUsed[RegMIPS32::Reg_RA] = true;
}
// Make two passes over the used registers. The first pass records all the
// used registers -- and their aliases. Then, we figure out which GPR
// registers should be saved.
SmallBitVector ToPreserve(RegMIPS32::Reg_NUM);
for (SizeT i = 0; i < CalleeSaves.size(); ++i) {
if (CalleeSaves[i] && RegsUsed[i]) {
ToPreserve |= RegisterAliases[i];
}
}
uint32_t NumCallee = 0;
// RegClasses is a tuple of
//
// <First Register in Class, Last Register in Class, Vector of Save Registers>
//
// We use this tuple to figure out which register we should save/restore
// during
// prolog/epilog.
using RegClassType = std::tuple<uint32_t, uint32_t, VarList *>;
const RegClassType RegClass = RegClassType(
RegMIPS32::Reg_GPR_First, RegMIPS32::Reg_FPR_Last, &PreservedGPRs);
const uint32_t FirstRegInClass = std::get<0>(RegClass);
const uint32_t LastRegInClass = std::get<1>(RegClass);
VarList *const PreservedRegsInClass = std::get<2>(RegClass);
for (uint32_t Reg = LastRegInClass; Reg > FirstRegInClass; Reg--) {
if (!ToPreserve[Reg]) {
continue;
}
++NumCallee;
Variable *PhysicalRegister = getPhysicalRegister(RegNumT::fromInt(Reg));
PreservedRegsSizeBytes +=
typeWidthInBytesOnStack(PhysicalRegister->getType());
PreservedRegsInClass->push_back(PhysicalRegister);
}
Ctx->statsUpdateRegistersSaved(NumCallee);
// Align the variables area. SpillAreaPaddingBytes is the size of the region
// after the preserved registers and before the spill areas.
// LocalsSlotsPaddingBytes is the amount of padding between the globals and
// locals area if they are separate.
assert(SpillAreaAlignmentBytes <= MIPS32_STACK_ALIGNMENT_BYTES);
(void)MIPS32_STACK_ALIGNMENT_BYTES;
assert(LocalsSlotsAlignmentBytes <= SpillAreaAlignmentBytes);
uint32_t SpillAreaPaddingBytes = 0;
uint32_t LocalsSlotsPaddingBytes = 0;
alignStackSpillAreas(PreservedRegsSizeBytes, SpillAreaAlignmentBytes,
GlobalsSize, LocalsSlotsAlignmentBytes,
&SpillAreaPaddingBytes, &LocalsSlotsPaddingBytes);
SpillAreaSizeBytes += SpillAreaPaddingBytes + LocalsSlotsPaddingBytes;
uint32_t GlobalsAndSubsequentPaddingSize =
GlobalsSize + LocalsSlotsPaddingBytes;
// Adds the out args space to the stack, and align SP if necessary.
if (!NeedsStackAlignment) {
SpillAreaSizeBytes += MaxOutArgsSizeBytes * (VariableAllocaUsed ? 0 : 1);
} else {
SpillAreaSizeBytes = applyStackAlignment(
SpillAreaSizeBytes +
(VariableAllocaUsed ? VariableAllocaAlignBytes : MaxOutArgsSizeBytes));
}
// Combine fixed alloca with SpillAreaSize.
SpillAreaSizeBytes += FixedAllocaSizeBytes;
TotalStackSizeBytes =
applyStackAlignment(PreservedRegsSizeBytes + SpillAreaSizeBytes);
// Generate "addiu sp, sp, -TotalStackSizeBytes"
if (TotalStackSizeBytes) {
// Use the scratch register if needed to legalize the immediate.
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
_addiu(SP, SP, -TotalStackSizeBytes);
}
Ctx->statsUpdateFrameBytes(TotalStackSizeBytes);
if (!PreservedGPRs.empty()) {
uint32_t StackOffset = TotalStackSizeBytes;
for (Variable *Var : *PreservedRegsInClass) {
Type RegType;
if (RegMIPS32::isFPRReg(Var->getRegNum()))
RegType = IceType_f32;
else
RegType = IceType_i32;
auto *PhysicalRegister = makeReg(RegType, Var->getRegNum());
StackOffset -= typeWidthInBytesOnStack(RegType);
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
OperandMIPS32Mem *MemoryLocation = OperandMIPS32Mem::create(
Func, RegType, SP,
llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(StackOffset)));
_sw(PhysicalRegister, MemoryLocation);
}
}
Variable *FP = getPhysicalRegister(RegMIPS32::Reg_FP);
// Generate "mov FP, SP" if needed.
if (UsesFramePointer) {
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
_mov(FP, SP);
// Keep FP live for late-stage liveness analysis (e.g. asm-verbose mode).
Context.insert<InstFakeUse>(FP);
}
// Fill in stack offsets for stack args, and copy args into registers for
// those that were register-allocated. Args are pushed right to left, so
// Arg[0] is closest to the stack/frame pointer.
const VarList &Args = Func->getArgs();
size_t InArgsSizeBytes = MIPS32_MAX_GPR_ARG * 4;
TargetMIPS32::CallingConv CC;
uint32_t ArgNo = 0;
for (Variable *Arg : Args) {
RegNumT DummyReg;
const Type Ty = Arg->getType();
bool PartialOnStack;
// Skip arguments passed in registers.
if (CC.argInReg(Ty, ArgNo, &DummyReg)) {
// Load argument from stack:
// 1. If this is first vector argument and return type is v4f32.
// In this case $4 is used to pass stack address implicitly.
// 3rd and 4th element of vector argument is passed through stack.
// 2. If this is second vector argument.
if (ArgNo != 0 && isVectorType(Ty)) {
PartialOnStack = true;
finishArgumentLowering(Arg, PartialOnStack, FP, TotalStackSizeBytes,
&InArgsSizeBytes);
}
} else {
PartialOnStack = false;
finishArgumentLowering(Arg, PartialOnStack, FP, TotalStackSizeBytes,
&InArgsSizeBytes);
}
++ArgNo;
}
// Fill in stack offsets for locals.
assignVarStackSlots(SortedSpilledVariables, SpillAreaPaddingBytes,
SpillAreaSizeBytes, GlobalsAndSubsequentPaddingSize);
this->HasComputedFrame = true;
if (BuildDefs::dump() && Func->isVerbose(IceV_Frame)) {
OstreamLocker _(Func->getContext());
Ostream &Str = Func->getContext()->getStrDump();
Str << "Stack layout:\n";
uint32_t SPAdjustmentPaddingSize =
SpillAreaSizeBytes - LocalsSpillAreaSize -
GlobalsAndSubsequentPaddingSize - SpillAreaPaddingBytes -
MaxOutArgsSizeBytes;
Str << " in-args = " << InArgsSizeBytes << " bytes\n"
<< " preserved registers = " << PreservedRegsSizeBytes << " bytes\n"
<< " spill area padding = " << SpillAreaPaddingBytes << " bytes\n"
<< " globals spill area = " << GlobalsSize << " bytes\n"
<< " globals-locals spill areas intermediate padding = "
<< GlobalsAndSubsequentPaddingSize - GlobalsSize << " bytes\n"
<< " locals spill area = " << LocalsSpillAreaSize << " bytes\n"
<< " SP alignment padding = " << SPAdjustmentPaddingSize << " bytes\n";
Str << "Stack details:\n"
<< " SP adjustment = " << SpillAreaSizeBytes << " bytes\n"
<< " spill area alignment = " << SpillAreaAlignmentBytes << " bytes\n"
<< " outgoing args size = " << MaxOutArgsSizeBytes << " bytes\n"
<< " locals spill area alignment = " << LocalsSlotsAlignmentBytes
<< " bytes\n"
<< " is FP based = " << 1 << "\n";
}
return;
}
void TargetMIPS32::addEpilog(CfgNode *Node) {
InstList &Insts = Node->getInsts();
InstList::reverse_iterator RI, E;
for (RI = Insts.rbegin(), E = Insts.rend(); RI != E; ++RI) {
if (llvm::isa<InstMIPS32Ret>(*RI))
break;
}
if (RI == E)
return;
// Convert the reverse_iterator position into its corresponding (forward)
// iterator position.
InstList::iterator InsertPoint = reverseToForwardIterator(RI);
--InsertPoint;
Context.init(Node);
Context.setInsertPoint(InsertPoint);
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
if (UsesFramePointer) {
Variable *FP = getPhysicalRegister(RegMIPS32::Reg_FP);
// For late-stage liveness analysis (e.g. asm-verbose mode), adding a fake
// use of SP before the assignment of SP=FP keeps previous SP adjustments
// from being dead-code eliminated.
Context.insert<InstFakeUse>(SP);
_mov(SP, FP);
}
VarList::reverse_iterator RIter, END;
if (!PreservedGPRs.empty()) {
uint32_t StackOffset = TotalStackSizeBytes - PreservedRegsSizeBytes;
for (RIter = PreservedGPRs.rbegin(), END = PreservedGPRs.rend();
RIter != END; ++RIter) {
Type RegType;
if (RegMIPS32::isFPRReg((*RIter)->getRegNum()))
RegType = IceType_f32;
else
RegType = IceType_i32;
auto *PhysicalRegister = makeReg(RegType, (*RIter)->getRegNum());
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
OperandMIPS32Mem *MemoryLocation = OperandMIPS32Mem::create(
Func, RegType, SP,
llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(StackOffset)));
_lw(PhysicalRegister, MemoryLocation);
StackOffset += typeWidthInBytesOnStack(PhysicalRegister->getType());
}
}
if (TotalStackSizeBytes) {
_addiu(SP, SP, TotalStackSizeBytes);
}
}
Variable *TargetMIPS32::PostLoweringLegalizer::newBaseRegister(
Variable *Base, int32_t Offset, RegNumT ScratchRegNum) {
// Legalize will likely need a lui/ori combination, but if the top bits are
// all 0 from negating the offset and subtracting, we could use that instead.
const bool ShouldSub = Offset != 0 && (-Offset & 0xFFFF0000) == 0;
Variable *ScratchReg = Target->makeReg(IceType_i32, ScratchRegNum);
if (ShouldSub) {
Target->_addi(ScratchReg, Base, -Offset);
} else {
constexpr bool SignExt = true;
if (!OperandMIPS32Mem::canHoldOffset(Base->getType(), SignExt, Offset)) {
const uint32_t UpperBits = (Offset >> 16) & 0xFFFF;
const uint32_t LowerBits = Offset & 0xFFFF;
Target->_lui(ScratchReg, Target->Ctx->getConstantInt32(UpperBits));
if (LowerBits)
Target->_ori(ScratchReg, ScratchReg, LowerBits);
Target->_addu(ScratchReg, ScratchReg, Base);
} else {
Target->_addiu(ScratchReg, Base, Offset);
}
}
return ScratchReg;
}
void TargetMIPS32::PostLoweringLegalizer::legalizeMovFp(
InstMIPS32MovFP64ToI64 *MovInstr) {
Variable *Dest = MovInstr->getDest();
Operand *Src = MovInstr->getSrc(0);
const Type SrcTy = Src->getType();
if (Dest != nullptr && SrcTy == IceType_f64) {
int32_t Offset = Dest->getStackOffset();
auto *Base = Target->getPhysicalRegister(Target->getFrameOrStackReg());
OperandMIPS32Mem *TAddr = OperandMIPS32Mem::create(
Target->Func, IceType_f32, Base,
llvm::cast<ConstantInteger32>(Target->Ctx->getConstantInt32(Offset)));
OperandMIPS32Mem *Addr = legalizeMemOperand(TAddr);
auto *SrcV = llvm::cast<Variable>(Src);
Variable *SrcR;
if (MovInstr->getInt64Part() == Int64_Lo) {
SrcR = Target->makeReg(
IceType_f32, RegMIPS32::get64PairFirstRegNum(SrcV->getRegNum()));
} else {
SrcR = Target->makeReg(
IceType_f32, RegMIPS32::get64PairSecondRegNum(SrcV->getRegNum()));
}
Target->_sw(SrcR, Addr);
if (MovInstr->isDestRedefined()) {
Target->_set_dest_redefined();
}
MovInstr->setDeleted();
return;
}
llvm::report_fatal_error("legalizeMovFp: Invalid operands");
}
void TargetMIPS32::PostLoweringLegalizer::legalizeMov(InstMIPS32Mov *MovInstr) {
Variable *Dest = MovInstr->getDest();
assert(Dest != nullptr);
const Type DestTy = Dest->getType();
assert(DestTy != IceType_i64);
Operand *Src = MovInstr->getSrc(0);
const Type SrcTy = Src->getType();
(void)SrcTy;
assert(SrcTy != IceType_i64);
bool Legalized = false;
auto *SrcR = llvm::cast<Variable>(Src);
if (Dest->hasReg() && SrcR->hasReg()) {
// This might be a GP to/from FP move generated due to argument passing.
// Use mtc1/mfc1 instead of mov.[s/d] if src and dst registers are of
// different types.
const bool IsDstGPR = RegMIPS32::isGPRReg(Dest->getRegNum());
const bool IsSrcGPR = RegMIPS32::isGPRReg(SrcR->getRegNum());
const RegNumT SRegNum = SrcR->getRegNum();
const RegNumT DRegNum = Dest->getRegNum();
if (IsDstGPR != IsSrcGPR) {
if (IsDstGPR) {
// Dest is GPR and SrcR is FPR. Use mfc1.
int32_t TypeWidth = typeWidthInBytes(DestTy);
if (MovInstr->getDestHi() != nullptr)
TypeWidth += typeWidthInBytes(MovInstr->getDestHi()->getType());
if (TypeWidth == 8) {
// Split it into two mfc1 instructions
Variable *SrcGPRHi = Target->makeReg(
IceType_f32, RegMIPS32::get64PairFirstRegNum(SRegNum));
Variable *SrcGPRLo = Target->makeReg(
IceType_f32, RegMIPS32::get64PairSecondRegNum(SRegNum));
Variable *DstFPRHi, *DstFPRLo;
if (MovInstr->getDestHi() != nullptr && Dest != nullptr) {
DstFPRHi = Target->makeReg(IceType_i32,
MovInstr->getDestHi()->getRegNum());
DstFPRLo = Target->makeReg(IceType_i32, Dest->getRegNum());
} else {
DstFPRHi = Target->makeReg(
IceType_i32, RegMIPS32::get64PairFirstRegNum(DRegNum));
DstFPRLo = Target->makeReg(
IceType_i32, RegMIPS32::get64PairSecondRegNum(DRegNum));
}
Target->_mov(DstFPRHi, SrcGPRHi);
Target->_mov(DstFPRLo, SrcGPRLo);
Legalized = true;
} else {
Variable *SrcGPR = Target->makeReg(IceType_f32, SRegNum);
Variable *DstFPR = Target->makeReg(IceType_i32, DRegNum);
Target->_mov(DstFPR, SrcGPR);
Legalized = true;
}
} else {
// Dest is FPR and SrcR is GPR. Use mtc1.
if (typeWidthInBytes(Dest->getType()) == 8) {
Variable *SrcGPRHi, *SrcGPRLo;
// SrcR could be $zero which is i32
if (SRegNum == RegMIPS32::Reg_ZERO) {
SrcGPRHi = Target->makeReg(IceType_i32, SRegNum);
SrcGPRLo = SrcGPRHi;
} else {
// Split it into two mtc1 instructions
if (MovInstr->getSrcSize() == 2) {
const auto FirstReg =
(llvm::cast<Variable>(MovInstr->getSrc(0)))->getRegNum();
const auto SecondReg =
(llvm::cast<Variable>(MovInstr->getSrc(1)))->getRegNum();
SrcGPRHi = Target->makeReg(IceType_i32, FirstReg);
SrcGPRLo = Target->makeReg(IceType_i32, SecondReg);
} else {
SrcGPRLo = Target->makeReg(
IceType_i32, RegMIPS32::get64PairFirstRegNum(SRegNum));
SrcGPRHi = Target->makeReg(
IceType_i32, RegMIPS32::get64PairSecondRegNum(SRegNum));
}
}
Variable *DstFPRHi = Target->makeReg(
IceType_f32, RegMIPS32::get64PairFirstRegNum(DRegNum));
Variable *DstFPRLo = Target->makeReg(
IceType_f32, RegMIPS32::get64PairSecondRegNum(DRegNum));
Target->_mov(DstFPRHi, SrcGPRLo);
Target->_mov(DstFPRLo, SrcGPRHi);
Legalized = true;
} else {
Variable *SrcGPR = Target->makeReg(IceType_i32, SRegNum);
Variable *DstFPR = Target->makeReg(IceType_f32, DRegNum);
Target->_mov(DstFPR, SrcGPR);
Legalized = true;
}
}
}
if (Legalized) {
if (MovInstr->isDestRedefined()) {
Target->_set_dest_redefined();
}
MovInstr->setDeleted();
return;
}
}
if (!Dest->hasReg()) {
auto *SrcR = llvm::cast<Variable>(Src);
assert(SrcR->hasReg());
assert(!SrcR->isRematerializable());
int32_t Offset = Dest->getStackOffset();
// This is a _mov(Mem(), Variable), i.e., a store.
auto *Base = Target->getPhysicalRegister(Target->getFrameOrStackReg());
OperandMIPS32Mem *TAddr = OperandMIPS32Mem::create(
Target->Func, DestTy, Base,
llvm::cast<ConstantInteger32>(Target->Ctx->getConstantInt32(Offset)));
OperandMIPS32Mem *TAddrHi = OperandMIPS32Mem::create(
Target->Func, DestTy, Base,
llvm::cast<ConstantInteger32>(
Target->Ctx->getConstantInt32(Offset + 4)));
OperandMIPS32Mem *Addr = legalizeMemOperand(TAddr);
// FP arguments are passed in GP reg if first argument is in GP. In this
// case type of the SrcR is still FP thus we need to explicitly generate sw
// instead of swc1.
const RegNumT RegNum = SrcR->getRegNum();
const bool IsSrcGPReg = RegMIPS32::isGPRReg(SrcR->getRegNum());
if (SrcTy == IceType_f32 && IsSrcGPReg) {
Variable *SrcGPR = Target->makeReg(IceType_i32, RegNum);
Target->_sw(SrcGPR, Addr);
} else if (SrcTy == IceType_f64 && IsSrcGPReg) {
Variable *SrcGPRHi =
Target->makeReg(IceType_i32, RegMIPS32::get64PairFirstRegNum(RegNum));
Variable *SrcGPRLo = Target->makeReg(
IceType_i32, RegMIPS32::get64PairSecondRegNum(RegNum));
Target->_sw(SrcGPRHi, Addr);
OperandMIPS32Mem *AddrHi = legalizeMemOperand(TAddrHi);
Target->_sw(SrcGPRLo, AddrHi);
} else if (DestTy == IceType_f64 && IsSrcGPReg) {
const auto FirstReg =
(llvm::cast<Variable>(MovInstr->getSrc(0)))->getRegNum();
const auto SecondReg =
(llvm::cast<Variable>(MovInstr->getSrc(1)))->getRegNum();
Variable *SrcGPRHi = Target->makeReg(IceType_i32, FirstReg);
Variable *SrcGPRLo = Target->makeReg(IceType_i32, SecondReg);
Target->_sw(SrcGPRLo, Addr);
OperandMIPS32Mem *AddrHi = legalizeMemOperand(TAddrHi);
Target->_sw(SrcGPRHi, AddrHi);
} else {
Target->_sw(SrcR, Addr);
}
Target->Context.insert<InstFakeDef>(Dest);
Legalized = true;
} else if (auto *Var = llvm::dyn_cast<Variable>(Src)) {
if (Var->isRematerializable()) {
// This is equivalent to an x86 _lea(RematOffset(%esp/%ebp), Variable).
// ExtraOffset is only needed for stack-pointer based frames as we have
// to account for spill storage.
const int32_t ExtraOffset =
(Var->getRegNum() == Target->getFrameOrStackReg())
? Target->getFrameFixedAllocaOffset()
: 0;
const int32_t Offset = Var->getStackOffset() + ExtraOffset;
Variable *Base = Target->getPhysicalRegister(Var->getRegNum());
Variable *T = newBaseRegister(Base, Offset, Dest->getRegNum());
Target->_mov(Dest, T);
Legalized = true;
} else {
if (!Var->hasReg()) {
// This is a _mov(Variable, Mem()), i.e., a load.
const int32_t Offset = Var->getStackOffset();
auto *Base = Target->getPhysicalRegister(Target->getFrameOrStackReg());
const RegNumT RegNum = Dest->getRegNum();
const bool IsDstGPReg = RegMIPS32::isGPRReg(Dest->getRegNum());
// If we are moving i64 to a double using stack then the address may
// not be aligned to 8-byte boundary as we split i64 into Hi-Lo parts
// and store them individually with 4-byte alignment. Load the Hi-Lo
// parts in TmpReg and move them to the dest using mtc1.
if (DestTy == IceType_f64 && !Utils::IsAligned(Offset, 8) &&
!IsDstGPReg) {
auto *Reg = Target->makeReg(IceType_i32, Target->getReservedTmpReg());
const RegNumT RegNum = Dest->getRegNum();
Variable *DestLo = Target->makeReg(
IceType_f32, RegMIPS32::get64PairFirstRegNum(RegNum));
Variable *DestHi = Target->makeReg(
IceType_f32, RegMIPS32::get64PairSecondRegNum(RegNum));
OperandMIPS32Mem *AddrLo = OperandMIPS32Mem::create(
Target->Func, IceType_i32, Base,
llvm::cast<ConstantInteger32>(
Target->Ctx->getConstantInt32(Offset)));
OperandMIPS32Mem *AddrHi = OperandMIPS32Mem::create(
Target->Func, IceType_i32, Base,
llvm::cast<ConstantInteger32>(
Target->Ctx->getConstantInt32(Offset + 4)));
Target->_lw(Reg, AddrLo);
Target->_mov(DestLo, Reg);
Target->_lw(Reg, AddrHi);
Target->_mov(DestHi, Reg);
} else {
OperandMIPS32Mem *TAddr = OperandMIPS32Mem::create(
Target->Func, DestTy, Base,
llvm::cast<ConstantInteger32>(
Target->Ctx->getConstantInt32(Offset)));
OperandMIPS32Mem *Addr = legalizeMemOperand(TAddr);
OperandMIPS32Mem *TAddrHi = OperandMIPS32Mem::create(
Target->Func, DestTy, Base,
llvm::cast<ConstantInteger32>(
Target->Ctx->getConstantInt32(Offset + 4)));
// FP arguments are passed in GP reg if first argument is in GP.
// In this case type of the Dest is still FP thus we need to
// explicitly generate lw instead of lwc1.
if (DestTy == IceType_f32 && IsDstGPReg) {
Variable *DstGPR = Target->makeReg(IceType_i32, RegNum);
Target->_lw(DstGPR, Addr);
} else if (DestTy == IceType_f64 && IsDstGPReg) {
Variable *DstGPRHi = Target->makeReg(
IceType_i32, RegMIPS32::get64PairFirstRegNum(RegNum));
Variable *DstGPRLo = Target->makeReg(
IceType_i32, RegMIPS32::get64PairSecondRegNum(RegNum));
Target->_lw(DstGPRHi, Addr);
OperandMIPS32Mem *AddrHi = legalizeMemOperand(TAddrHi);
Target->_lw(DstGPRLo, AddrHi);
} else if (DestTy == IceType_f64 && IsDstGPReg) {
const auto FirstReg =
(llvm::cast<Variable>(MovInstr->getSrc(0)))->getRegNum();
const auto SecondReg =
(llvm::cast<Variable>(MovInstr->getSrc(1)))->getRegNum();
Variable *DstGPRHi = Target->makeReg(IceType_i32, FirstReg);
Variable *DstGPRLo = Target->makeReg(IceType_i32, SecondReg);
Target->_lw(DstGPRLo, Addr);
OperandMIPS32Mem *AddrHi = legalizeMemOperand(TAddrHi);
Target->_lw(DstGPRHi, AddrHi);
} else {
Target->_lw(Dest, Addr);
}
}
Legalized = true;
}
}
}
if (Legalized) {
if (MovInstr->isDestRedefined()) {
Target->_set_dest_redefined();
}
MovInstr->setDeleted();
}
}
OperandMIPS32Mem *
TargetMIPS32::PostLoweringLegalizer::legalizeMemOperand(OperandMIPS32Mem *Mem) {
if (llvm::isa<ConstantRelocatable>(Mem->getOffset())) {
return nullptr;
}
Variable *Base = Mem->getBase();
auto *Ci32 = llvm::cast<ConstantInteger32>(Mem->getOffset());
int32_t Offset = Ci32->getValue();
if (Base->isRematerializable()) {
const int32_t ExtraOffset =
(Base->getRegNum() == Target->getFrameOrStackReg())
? Target->getFrameFixedAllocaOffset()
: 0;
Offset += Base->getStackOffset() + ExtraOffset;
Base = Target->getPhysicalRegister(Base->getRegNum());
}
constexpr bool SignExt = true;
if (!OperandMIPS32Mem::canHoldOffset(Mem->getType(), SignExt, Offset)) {
Base = newBaseRegister(Base, Offset, Target->getReservedTmpReg());
Offset = 0;
}
return OperandMIPS32Mem::create(
Target->Func, Mem->getType(), Base,
llvm::cast<ConstantInteger32>(Target->Ctx->getConstantInt32(Offset)));
}
Variable *TargetMIPS32::PostLoweringLegalizer::legalizeImmediate(int32_t Imm) {
Variable *Reg = nullptr;
if (!((std::numeric_limits<int16_t>::min() <= Imm) &&
(Imm <= std::numeric_limits<int16_t>::max()))) {
const uint32_t UpperBits = (Imm >> 16) & 0xFFFF;
const uint32_t LowerBits = Imm & 0xFFFF;
Variable *TReg = Target->makeReg(IceType_i32, Target->getReservedTmpReg());
Reg = Target->makeReg(IceType_i32, Target->getReservedTmpReg());
if (LowerBits) {
Target->_lui(TReg, Target->Ctx->getConstantInt32(UpperBits));
Target->_ori(Reg, TReg, LowerBits);
} else {
Target->_lui(Reg, Target->Ctx->getConstantInt32(UpperBits));
}
}
return Reg;
}
void TargetMIPS32::postLowerLegalization() {
Func->dump("Before postLowerLegalization");
assert(hasComputedFrame());
for (CfgNode *Node : Func->getNodes()) {
Context.init(Node);
PostLoweringLegalizer Legalizer(this);
while (!Context.atEnd()) {
PostIncrLoweringContext PostIncrement(Context);
Inst *CurInstr = iteratorToInst(Context.getCur());
const SizeT NumSrcs = CurInstr->getSrcSize();
Operand *Src0 = NumSrcs < 1 ? nullptr : CurInstr->getSrc(0);
Operand *Src1 = NumSrcs < 2 ? nullptr : CurInstr->getSrc(1);
auto *Src0V = llvm::dyn_cast_or_null<Variable>(Src0);
auto *Src0M = llvm::dyn_cast_or_null<OperandMIPS32Mem>(Src0);
auto *Src1M = llvm::dyn_cast_or_null<OperandMIPS32Mem>(Src1);
Variable *Dst = CurInstr->getDest();
if (auto *MovInstr = llvm::dyn_cast<InstMIPS32Mov>(CurInstr)) {
Legalizer.legalizeMov(MovInstr);
continue;
}
if (auto *MovInstr = llvm::dyn_cast<InstMIPS32MovFP64ToI64>(CurInstr)) {
Legalizer.legalizeMovFp(MovInstr);
continue;
}
if (llvm::isa<InstMIPS32Sw>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src1M)) {
_sw(Src0V, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (llvm::isa<InstMIPS32Swc1>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src1M)) {
_swc1(Src0V, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (llvm::isa<InstMIPS32Sdc1>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src1M)) {
_sdc1(Src0V, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (llvm::isa<InstMIPS32Lw>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src0M)) {
_lw(Dst, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (llvm::isa<InstMIPS32Lwc1>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src0M)) {
_lwc1(Dst, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (llvm::isa<InstMIPS32Ldc1>(CurInstr)) {
if (auto *LegalMem = Legalizer.legalizeMemOperand(Src0M)) {
_ldc1(Dst, LegalMem);
CurInstr->setDeleted();
}
continue;
}
if (auto *AddiuInstr = llvm::dyn_cast<InstMIPS32Addiu>(CurInstr)) {
if (auto *LegalImm = Legalizer.legalizeImmediate(
static_cast<int32_t>(AddiuInstr->getImmediateValue()))) {
_addu(Dst, Src0V, LegalImm);
CurInstr->setDeleted();
}
continue;
}
}
}
}
Operand *TargetMIPS32::loOperand(Operand *Operand) {
assert(Operand->getType() == IceType_i64);
if (auto *Var64On32 = llvm::dyn_cast<Variable64On32>(Operand))
return Var64On32->getLo();
if (auto *Const = llvm::dyn_cast<ConstantInteger64>(Operand)) {
return Ctx->getConstantInt32(static_cast<uint32_t>(Const->getValue()));
}
if (auto *Mem = llvm::dyn_cast<OperandMIPS32Mem>(Operand)) {
// Conservatively disallow memory operands with side-effects (pre/post
// increment) in case of duplication.
assert(Mem->getAddrMode() == OperandMIPS32Mem::Offset);
return OperandMIPS32Mem::create(Func, IceType_i32, Mem->getBase(),
Mem->getOffset(), Mem->getAddrMode());
}
llvm_unreachable("Unsupported operand type");
return nullptr;
}
Operand *TargetMIPS32::getOperandAtIndex(Operand *Operand, Type BaseType,
uint32_t Index) {
if (!isVectorType(Operand->getType())) {
llvm::report_fatal_error("getOperandAtIndex: Operand is not vector");
return nullptr;
}
if (auto *Mem = llvm::dyn_cast<OperandMIPS32Mem>(Operand)) {
assert(Mem->getAddrMode() == OperandMIPS32Mem::Offset);
Variable *Base = Mem->getBase();
auto *Offset = llvm::cast<ConstantInteger32>(Mem->getOffset());
assert(!Utils::WouldOverflowAdd(Offset->getValue(), 4));
int32_t NextOffsetVal =
Offset->getValue() + (Index * typeWidthInBytes(BaseType));
constexpr bool NoSignExt = false;
if (!OperandMIPS32Mem::canHoldOffset(BaseType, NoSignExt, NextOffsetVal)) {
Constant *_4 = Ctx->getConstantInt32(4);
Variable *NewBase = Func->makeVariable(Base->getType());
lowerArithmetic(
InstArithmetic::create(Func, InstArithmetic::Add, NewBase, Base, _4));
Base = NewBase;
} else {
Offset =
llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(NextOffsetVal));
}
return OperandMIPS32Mem::create(Func, BaseType, Base, Offset,
Mem->getAddrMode());
}
if (auto *VarVecOn32 = llvm::dyn_cast<VariableVecOn32>(Operand))
return VarVecOn32->getContainers()[Index];
llvm_unreachable("Unsupported operand type");
return nullptr;
}
Operand *TargetMIPS32::hiOperand(Operand *Operand) {
assert(Operand->getType() == IceType_i64);
if (Operand->getType() != IceType_i64)
return Operand;
if (auto *Var64On32 = llvm::dyn_cast<Variable64On32>(Operand))
return Var64On32->getHi();
if (auto *Const = llvm::dyn_cast<ConstantInteger64>(Operand)) {
return Ctx->getConstantInt32(
static_cast<uint32_t>(Const->getValue() >> 32));
}
if (auto *Mem = llvm::dyn_cast<OperandMIPS32Mem>(Operand)) {
// Conservatively disallow memory operands with side-effects
// in case of duplication.
assert(Mem->getAddrMode() == OperandMIPS32Mem::Offset);
const Type SplitType = IceType_i32;
Variable *Base = Mem->getBase();
auto *Offset = llvm::cast<ConstantInteger32>(Mem->getOffset());
assert(!Utils::WouldOverflowAdd(Offset->getValue(), 4));
int32_t NextOffsetVal = Offset->getValue() + 4;
constexpr bool SignExt = false;
if (!OperandMIPS32Mem::canHoldOffset(SplitType, SignExt, NextOffsetVal)) {
// We have to make a temp variable and add 4 to either Base or Offset.
// If we add 4 to Offset, this will convert a non-RegReg addressing
// mode into a RegReg addressing mode. Since NaCl sandboxing disallows
// RegReg addressing modes, prefer adding to base and replacing instead.
// Thus we leave the old offset alone.
Constant *Four = Ctx->getConstantInt32(4);
Variable *NewBase = Func->makeVariable(Base->getType());
lowerArithmetic(InstArithmetic::create(Func, InstArithmetic::Add, NewBase,
Base, Four));
Base = NewBase;
} else {
Offset =
llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(NextOffsetVal));
}
return OperandMIPS32Mem::create(Func, SplitType, Base, Offset,
Mem->getAddrMode());
}
llvm_unreachable("Unsupported operand type");
return nullptr;
}
SmallBitVector TargetMIPS32::getRegisterSet(RegSetMask Include,
RegSetMask Exclude) const {
SmallBitVector Registers(RegMIPS32::Reg_NUM);
#define X(val, encode, name, scratch, preserved, stackptr, frameptr, isInt, \
isI64Pair, isFP32, isFP64, isVec128, alias_init) \
if (scratch && (Include & RegSet_CallerSave)) \
Registers[RegMIPS32::val] = true; \
if (preserved && (Include & RegSet_CalleeSave)) \
Registers[RegMIPS32::val] = true; \
if (stackptr && (Include & RegSet_StackPointer)) \
Registers[RegMIPS32::val] = true; \
if (frameptr && (Include & RegSet_FramePointer)) \
Registers[RegMIPS32::val] = true; \
if (scratch && (Exclude & RegSet_CallerSave)) \
Registers[RegMIPS32::val] = false; \
if (preserved && (Exclude & RegSet_CalleeSave)) \
Registers[RegMIPS32::val] = false; \
if (stackptr && (Exclude & RegSet_StackPointer)) \
Registers[RegMIPS32::val] = false; \
if (frameptr && (Exclude & RegSet_FramePointer)) \
Registers[RegMIPS32::val] = false;
REGMIPS32_TABLE
#undef X
return Registers;
}
void TargetMIPS32::lowerAlloca(const InstAlloca *Instr) {
// Conservatively require the stack to be aligned. Some stack adjustment
// operations implemented below assume that the stack is aligned before the
// alloca. All the alloca code ensures that the stack alignment is preserved
// after the alloca. The stack alignment restriction can be relaxed in some
// cases.
NeedsStackAlignment = true;
// For default align=0, set it to the real value 1, to avoid any
// bit-manipulation problems below.
const uint32_t AlignmentParam = std::max(1u, Instr->getAlignInBytes());
// LLVM enforces power of 2 alignment.
assert(llvm::isPowerOf2_32(AlignmentParam));
assert(llvm::isPowerOf2_32(MIPS32_STACK_ALIGNMENT_BYTES));
const uint32_t Alignment =
std::max(AlignmentParam, MIPS32_STACK_ALIGNMENT_BYTES);
const bool OverAligned = Alignment > MIPS32_STACK_ALIGNMENT_BYTES;
const bool OptM1 = Func->getOptLevel() == Opt_m1;
const bool AllocaWithKnownOffset = Instr->getKnownFrameOffset();
const bool UseFramePointer =
hasFramePointer() || OverAligned || !AllocaWithKnownOffset || OptM1;
if (UseFramePointer)
setHasFramePointer();
Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
Variable *Dest = Instr->getDest();
Operand *TotalSize = Instr->getSizeInBytes();
if (const auto *ConstantTotalSize =
llvm::dyn_cast<ConstantInteger32>(TotalSize)) {
const uint32_t Value =
Utils::applyAlignment(ConstantTotalSize->getValue(), Alignment);
FixedAllocaSizeBytes += Value;
// Constant size alloca.
if (!UseFramePointer) {
// If we don't need a Frame Pointer, this alloca has a known offset to the
// stack pointer. We don't need adjust the stack pointer, nor assign any
// value to Dest, as Dest is rematerializable.
assert(Dest->isRematerializable());
Context.insert<InstFakeDef>(Dest);
return;
}
if (Alignment > MIPS32_STACK_ALIGNMENT_BYTES) {
CurrentAllocaOffset =
Utils::applyAlignment(CurrentAllocaOffset, Alignment);
}
auto *T = I32Reg();
_addiu(T, SP, CurrentAllocaOffset);
_mov(Dest, T);
CurrentAllocaOffset += Value;
return;
} else {
// Non-constant sizes need to be adjusted to the next highest multiple of
// the required alignment at runtime.
VariableAllocaUsed = true;
VariableAllocaAlignBytes = AlignmentParam;
Variable *AlignAmount;
auto *TotalSizeR = legalizeToReg(TotalSize, Legal_Reg);
auto *T1 = I32Reg();
auto *T2 = I32Reg();
auto *T3 = I32Reg();
auto *T4 = I32Reg();
auto *T5 = I32Reg();
_addiu(T1, TotalSizeR, MIPS32_STACK_ALIGNMENT_BYTES - 1);
_addiu(T2, getZero(), -MIPS32_STACK_ALIGNMENT_BYTES);
_and(T3, T1, T2);
_subu(T4, SP, T3);
if (Instr->getAlignInBytes()) {
AlignAmount =
legalizeToReg(Ctx->getConstantInt32(-AlignmentParam), Legal_Reg);
_and(T5, T4, AlignAmount);
_mov(Dest, T5);
} else {
_mov(Dest, T4);
}
_mov(SP, Dest);
return;
}
}
void TargetMIPS32::lowerInt64Arithmetic(const InstArithmetic *Instr,
Variable *Dest, Operand *Src0,
Operand *Src1) {
InstArithmetic::OpKind Op = Instr->getOp();
auto *DestLo = llvm::cast<Variable>(loOperand(Dest));
auto *DestHi = llvm::cast<Variable>(hiOperand(Dest));
Variable *Src0LoR = nullptr;
Variable *Src1LoR = nullptr;
Variable *Src0HiR = nullptr;
Variable *Src1HiR = nullptr;
switch (Op) {
case InstArithmetic::_num:
llvm::report_fatal_error("Unknown arithmetic operator");
return;
case InstArithmetic::Add: {
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Carry = I32Reg(), *T_Lo = I32Reg(), *T_Hi = I32Reg(),
*T_Hi2 = I32Reg();
_addu(T_Lo, Src0LoR, Src1LoR);
_mov(DestLo, T_Lo);
_sltu(T_Carry, T_Lo, Src0LoR);
_addu(T_Hi, T_Carry, Src0HiR);
_addu(T_Hi2, Src1HiR, T_Hi);
_mov(DestHi, T_Hi2);
return;
}
case InstArithmetic::And: {
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Lo = I32Reg(), *T_Hi = I32Reg();
_and(T_Lo, Src0LoR, Src1LoR);
_mov(DestLo, T_Lo);
_and(T_Hi, Src0HiR, Src1HiR);
_mov(DestHi, T_Hi);
return;
}
case InstArithmetic::Sub: {
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Borrow = I32Reg(), *T_Lo = I32Reg(), *T_Hi = I32Reg(),
*T_Hi2 = I32Reg();
_subu(T_Lo, Src0LoR, Src1LoR);
_mov(DestLo, T_Lo);
_sltu(T_Borrow, Src0LoR, Src1LoR);
_addu(T_Hi, T_Borrow, Src1HiR);
_subu(T_Hi2, Src0HiR, T_Hi);
_mov(DestHi, T_Hi2);
return;
}
case InstArithmetic::Or: {
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Lo = I32Reg(), *T_Hi = I32Reg();
_or(T_Lo, Src0LoR, Src1LoR);
_mov(DestLo, T_Lo);
_or(T_Hi, Src0HiR, Src1HiR);
_mov(DestHi, T_Hi);
return;
}
case InstArithmetic::Xor: {
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Lo = I32Reg(), *T_Hi = I32Reg();
_xor(T_Lo, Src0LoR, Src1LoR);
_mov(DestLo, T_Lo);
_xor(T_Hi, Src0HiR, Src1HiR);
_mov(DestHi, T_Hi);
return;
}
case InstArithmetic::Mul: {
// TODO(rkotler): Make sure that mul has the side effect of clobbering
// LO, HI. Check for any other LO, HI quirkiness in this section.
Src0LoR = legalizeToReg(loOperand(Src0));
Src1LoR = legalizeToReg(loOperand(Src1));
Src0HiR = legalizeToReg(hiOperand(Src0));
Src1HiR = legalizeToReg(hiOperand(Src1));
auto *T_Lo = I32Reg(RegMIPS32::Reg_LO), *T_Hi = I32Reg(RegMIPS32::Reg_HI);
auto *T1 = I32Reg(), *T2 = I32Reg();
auto *TM1 = I32Reg(), *TM2 = I32Reg(), *TM3 = I32Reg(), *TM4 = I32Reg();
_multu(T_Lo, Src0LoR, Src1LoR);
Context.insert<InstFakeDef>(T_Hi, T_Lo);
_mflo(T1, T_Lo);
_mfhi(T2, T_Hi);
_mov(DestLo, T1);
_mul(TM1, Src0HiR, Src1LoR);
_mul(TM2, Src0LoR, Src1HiR);
_addu(TM3, TM1, T2);
_addu(TM4, TM3, TM2);
_mov(DestHi, TM4);
return;
}
case InstArithmetic::Shl: {
auto *T_Lo = I32Reg();
auto *T_Hi = I32Reg();
auto *T1_Lo = I32Reg();
auto *T1_Hi = I32Reg();
auto *T1 = I32Reg();
auto *T2 = I32Reg();
auto *T3 = I32Reg();
auto *T4 = I32Reg();
auto *T5 = I32Reg();
if (auto *Const = llvm::dyn_cast<ConstantInteger64&g