X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FIntrinsicLowering.cpp;h=bd3c8c71f5b221c1b5efc707b85a2a939c29ba39;hb=4a307ecce68f90e0eebf1ded52b947816cdc2304;hp=b5a03fc54d53e4aff1595a63a7803cf73067123b;hpb=b83eb6447ba155342598f0fabe1f08f5baa9164a;p=oota-llvm.git diff --git a/lib/CodeGen/IntrinsicLowering.cpp b/lib/CodeGen/IntrinsicLowering.cpp index b5a03fc54d5..bd3c8c71f5b 100644 --- a/lib/CodeGen/IntrinsicLowering.cpp +++ b/lib/CodeGen/IntrinsicLowering.cpp @@ -2,35 +2,36 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // -// This file implements the default intrinsic lowering implementation. +// This file implements the IntrinsicLowering class. // //===----------------------------------------------------------------------===// -#include "llvm/CodeGen/IntrinsicLowering.h" #include "llvm/Constants.h" #include "llvm/DerivedTypes.h" #include "llvm/Module.h" #include "llvm/Instructions.h" #include "llvm/Type.h" -#include - +#include "llvm/CodeGen/IntrinsicLowering.h" +#include "llvm/Support/Streams.h" +#include "llvm/Target/TargetData.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/STLExtras.h" using namespace llvm; template -static Function *EnsureFunctionExists(Module &M, const char *Name, - ArgIt ArgBegin, ArgIt ArgEnd, - const Type *RetTy) { - if (Function *F = M.getNamedFunction(Name)) return F; - // It doesn't already exist in the program, insert a new definition now. +static void EnsureFunctionExists(Module &M, const char *Name, + ArgIt ArgBegin, ArgIt ArgEnd, + const Type *RetTy) { + // Insert a correctly-typed definition now. std::vector ParamTys; for (ArgIt I = ArgBegin; I != ArgEnd; ++I) ParamTys.push_back(I->getType()); - return M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false)); + M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false)); } /// ReplaceCallWith - This function is used when we want to lower an intrinsic @@ -40,56 +41,35 @@ static Function *EnsureFunctionExists(Module &M, const char *Name, template static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI, ArgIt ArgBegin, ArgIt ArgEnd, - const Type *RetTy, Function *&FCache) { + const Type *RetTy, Constant *&FCache) { if (!FCache) { // If we haven't already looked up this function, check to see if the // program already contains a function with this name. Module *M = CI->getParent()->getParent()->getParent(); - FCache = M->getNamedFunction(NewFn); - if (!FCache) { - // It doesn't already exist in the program, insert a new definition now. - std::vector ParamTys; - for (ArgIt I = ArgBegin; I != ArgEnd; ++I) - ParamTys.push_back((*I)->getType()); - FCache = M->getOrInsertFunction(NewFn, - FunctionType::get(RetTy, ParamTys, false)); - } - } - - const FunctionType *FT = FCache->getFunctionType(); - std::vector Operands; - unsigned ArgNo = 0; - for (ArgIt I = ArgBegin; I != ArgEnd && ArgNo != FT->getNumParams(); - ++I, ++ArgNo) { - Value *Arg = *I; - if (Arg->getType() != FT->getParamType(ArgNo)) - Arg = new CastInst(Arg, FT->getParamType(ArgNo), Arg->getName(), CI); - Operands.push_back(Arg); - } - // Pass nulls into any additional arguments... - for (; ArgNo != FT->getNumParams(); ++ArgNo) - Operands.push_back(Constant::getNullValue(FT->getParamType(ArgNo))); - - std::string Name = CI->getName(); CI->setName(""); - if (FT->getReturnType() == Type::VoidTy) Name.clear(); - CallInst *NewCI = new CallInst(FCache, Operands, Name, CI); - if (!CI->use_empty()) { - Value *V = NewCI; - if (CI->getType() != NewCI->getType()) - V = new CastInst(NewCI, CI->getType(), Name, CI); - CI->replaceAllUsesWith(V); + // Get or insert the definition now. + std::vector ParamTys; + for (ArgIt I = ArgBegin; I != ArgEnd; ++I) + ParamTys.push_back((*I)->getType()); + FCache = M->getOrInsertFunction(NewFn, + FunctionType::get(RetTy, ParamTys, false)); } + + SmallVector Args(ArgBegin, ArgEnd); + CallInst *NewCI = CallInst::Create(FCache, Args.begin(), Args.end(), + CI->getName(), CI); + if (!CI->use_empty()) + CI->replaceAllUsesWith(NewCI); return NewCI; } -void DefaultIntrinsicLowering::AddPrototypes(Module &M) { +void IntrinsicLowering::AddPrototypes(Module &M) { for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) - if (I->isExternal() && !I->use_empty()) + if (I->isDeclaration() && !I->use_empty()) switch (I->getIntrinsicID()) { default: break; case Intrinsic::setjmp: EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(), - Type::IntTy); + Type::Int32Ty); break; case Intrinsic::longjmp: EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(), @@ -101,34 +81,84 @@ void DefaultIntrinsicLowering::AddPrototypes(Module &M) { break; case Intrinsic::memcpy_i32: case Intrinsic::memcpy_i64: - EnsureFunctionExists(M, "memcpy", I->arg_begin(), --I->arg_end(), - I->arg_begin()->getType()); + M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty), + PointerType::getUnqual(Type::Int8Ty), + PointerType::getUnqual(Type::Int8Ty), + TD.getIntPtrType(), (Type *)0); break; case Intrinsic::memmove_i32: case Intrinsic::memmove_i64: - EnsureFunctionExists(M, "memmove", I->arg_begin(), --I->arg_end(), - I->arg_begin()->getType()); + M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty), + PointerType::getUnqual(Type::Int8Ty), + PointerType::getUnqual(Type::Int8Ty), + TD.getIntPtrType(), (Type *)0); break; case Intrinsic::memset_i32: case Intrinsic::memset_i64: - M.getOrInsertFunction("memset", PointerType::get(Type::SByteTy), - PointerType::get(Type::SByteTy), - Type::IntTy, (--(--I->arg_end()))->getType(), - (Type *)0); + M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty), + PointerType::getUnqual(Type::Int8Ty), + Type::Int32Ty, + TD.getIntPtrType(), (Type *)0); break; - case Intrinsic::isunordered_f32: - case Intrinsic::isunordered_f64: - EnsureFunctionExists(M, "isunordered", I->arg_begin(), I->arg_end(), - Type::BoolTy); - break; - case Intrinsic::sqrt_f32: - case Intrinsic::sqrt_f64: - if(I->arg_begin()->getType() == Type::FloatTy) + case Intrinsic::sqrt: + switch((int)I->arg_begin()->getType()->getTypeID()) { + case Type::FloatTyID: EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(), Type::FloatTy); - else + case Type::DoubleTyID: EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(), Type::DoubleTy); + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + EnsureFunctionExists(M, "sqrtl", I->arg_begin(), I->arg_end(), + I->arg_begin()->getType()); + } + break; + case Intrinsic::sin: + switch((int)I->arg_begin()->getType()->getTypeID()) { + case Type::FloatTyID: + EnsureFunctionExists(M, "sinf", I->arg_begin(), I->arg_end(), + Type::FloatTy); + case Type::DoubleTyID: + EnsureFunctionExists(M, "sin", I->arg_begin(), I->arg_end(), + Type::DoubleTy); + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + EnsureFunctionExists(M, "sinl", I->arg_begin(), I->arg_end(), + I->arg_begin()->getType()); + } + break; + case Intrinsic::cos: + switch((int)I->arg_begin()->getType()->getTypeID()) { + case Type::FloatTyID: + EnsureFunctionExists(M, "cosf", I->arg_begin(), I->arg_end(), + Type::FloatTy); + case Type::DoubleTyID: + EnsureFunctionExists(M, "cos", I->arg_begin(), I->arg_end(), + Type::DoubleTy); + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + EnsureFunctionExists(M, "cosl", I->arg_begin(), I->arg_end(), + I->arg_begin()->getType()); + } + break; + case Intrinsic::pow: + switch((int)I->arg_begin()->getType()->getTypeID()) { + case Type::FloatTyID: + EnsureFunctionExists(M, "powf", I->arg_begin(), I->arg_end(), + Type::FloatTy); + case Type::DoubleTyID: + EnsureFunctionExists(M, "pow", I->arg_begin(), I->arg_end(), + Type::DoubleTy); + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + EnsureFunctionExists(M, "powl", I->arg_begin(), I->arg_end(), + I->arg_begin()->getType()); + } break; } } @@ -138,93 +168,84 @@ void DefaultIntrinsicLowering::AddPrototypes(Module &M) { static Value *LowerBSWAP(Value *V, Instruction *IP) { assert(V->getType()->isInteger() && "Can't bswap a non-integer type!"); - const Type *DestTy = V->getType(); - - // Force to unsigned so that the shift rights are logical. - if (DestTy->isSigned()) - V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP); - unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); switch(BitSize) { default: assert(0 && "Unhandled type size of value to byteswap!"); case 16: { - Value *Tmp1 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,8),"bswap.2",IP); - Value *Tmp2 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,8),"bswap.1",IP); - V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP); + Value *Tmp1 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),8),"bswap.2",IP); + Value *Tmp2 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),8),"bswap.1",IP); + V = BinaryOperator::CreateOr(Tmp1, Tmp2, "bswap.i16", IP); break; } case 32: { - Value *Tmp4 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,24),"bswap.4", IP); - Value *Tmp3 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,8),"bswap.3",IP); - Value *Tmp2 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,8),"bswap.2",IP); - Value *Tmp1 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,24),"bswap.1", IP); - Tmp3 = BinaryOperator::createAnd(Tmp3, - ConstantInt::get(Type::UIntTy, 0xFF0000), + Value *Tmp4 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),24),"bswap.4", IP); + Value *Tmp3 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),8),"bswap.3",IP); + Value *Tmp2 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),8),"bswap.2",IP); + Value *Tmp1 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),24),"bswap.1", IP); + Tmp3 = BinaryOperator::CreateAnd(Tmp3, + ConstantInt::get(Type::Int32Ty, 0xFF0000), "bswap.and3", IP); - Tmp2 = BinaryOperator::createAnd(Tmp2, - ConstantInt::get(Type::UIntTy, 0xFF00), + Tmp2 = BinaryOperator::CreateAnd(Tmp2, + ConstantInt::get(Type::Int32Ty, 0xFF00), "bswap.and2", IP); - Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP); - Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP); - V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP); + Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or1", IP); + Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or2", IP); + V = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.i32", IP); break; } case 64: { - Value *Tmp8 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,56),"bswap.8", IP); - Value *Tmp7 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,40),"bswap.7", IP); - Value *Tmp6 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,24),"bswap.6", IP); - Value *Tmp5 = new ShiftInst(Instruction::Shl, V, - ConstantInt::get(Type::UByteTy,8),"bswap.5",IP); - Value *Tmp4 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,8),"bswap.4",IP); - Value *Tmp3 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,24),"bswap.3", IP); - Value *Tmp2 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,40),"bswap.2", IP); - Value *Tmp1 = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy,56),"bswap.1", IP); - Tmp7 = BinaryOperator::createAnd(Tmp7, - ConstantInt::get(Type::ULongTy, + Value *Tmp8 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),56),"bswap.8", IP); + Value *Tmp7 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),40),"bswap.7", IP); + Value *Tmp6 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),24),"bswap.6", IP); + Value *Tmp5 = BinaryOperator::CreateShl(V, + ConstantInt::get(V->getType(),8),"bswap.5", IP); + Value* Tmp4 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),8),"bswap.4", IP); + Value* Tmp3 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),24),"bswap.3", IP); + Value* Tmp2 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),40),"bswap.2", IP); + Value* Tmp1 = BinaryOperator::CreateLShr(V, + ConstantInt::get(V->getType(),56),"bswap.1", IP); + Tmp7 = BinaryOperator::CreateAnd(Tmp7, + ConstantInt::get(Type::Int64Ty, 0xFF000000000000ULL), "bswap.and7", IP); - Tmp6 = BinaryOperator::createAnd(Tmp6, - ConstantInt::get(Type::ULongTy, 0xFF0000000000ULL), + Tmp6 = BinaryOperator::CreateAnd(Tmp6, + ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL), "bswap.and6", IP); - Tmp5 = BinaryOperator::createAnd(Tmp5, - ConstantInt::get(Type::ULongTy, 0xFF00000000ULL), + Tmp5 = BinaryOperator::CreateAnd(Tmp5, + ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL), "bswap.and5", IP); - Tmp4 = BinaryOperator::createAnd(Tmp4, - ConstantInt::get(Type::ULongTy, 0xFF000000ULL), + Tmp4 = BinaryOperator::CreateAnd(Tmp4, + ConstantInt::get(Type::Int64Ty, 0xFF000000ULL), "bswap.and4", IP); - Tmp3 = BinaryOperator::createAnd(Tmp3, - ConstantInt::get(Type::ULongTy, 0xFF0000ULL), + Tmp3 = BinaryOperator::CreateAnd(Tmp3, + ConstantInt::get(Type::Int64Ty, 0xFF0000ULL), "bswap.and3", IP); - Tmp2 = BinaryOperator::createAnd(Tmp2, - ConstantInt::get(Type::ULongTy, 0xFF00ULL), + Tmp2 = BinaryOperator::CreateAnd(Tmp2, + ConstantInt::get(Type::Int64Ty, 0xFF00ULL), "bswap.and2", IP); - Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP); - Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP); - Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP); - Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP); - Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP); - Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP); - V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP); + Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp7, "bswap.or1", IP); + Tmp6 = BinaryOperator::CreateOr(Tmp6, Tmp5, "bswap.or2", IP); + Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or3", IP); + Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or4", IP); + Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp6, "bswap.or5", IP); + Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.or6", IP); + V = BinaryOperator::CreateOr(Tmp8, Tmp4, "bswap.i64", IP); break; } } - - if (V->getType() != DestTy) - V = new CastInst(V, DestTy, V->getName(), IP); return V; } @@ -239,66 +260,402 @@ static Value *LowerCTPOP(Value *V, Instruction *IP) { 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL }; - const Type *DestTy = V->getType(); - - // Force to unsigned so that the shift rights are logical. - if (DestTy->isSigned()) - V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP); - unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); - for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) { - Value *MaskCst = - ConstantExpr::getCast(ConstantInt::get(Type::ULongTy, MaskValues[ct]), - V->getType()); - Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP); - Value *VShift = new ShiftInst(Instruction::Shr, V, - ConstantInt::get(Type::UByteTy, i), "ctpop.sh", IP); - Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP); - V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP); + unsigned WordSize = (BitSize + 63) / 64; + Value *Count = ConstantInt::get(V->getType(), 0); + + for (unsigned n = 0; n < WordSize; ++n) { + Value *PartValue = V; + for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); + i <<= 1, ++ct) { + Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]); + Value *LHS = BinaryOperator::CreateAnd( + PartValue, MaskCst, "cppop.and1", IP); + Value *VShift = BinaryOperator::CreateLShr(PartValue, + ConstantInt::get(V->getType(), i), "ctpop.sh", IP); + Value *RHS = BinaryOperator::CreateAnd(VShift, MaskCst, "cppop.and2", IP); + PartValue = BinaryOperator::CreateAdd(LHS, RHS, "ctpop.step", IP); + } + Count = BinaryOperator::CreateAdd(PartValue, Count, "ctpop.part", IP); + if (BitSize > 64) { + V = BinaryOperator::CreateLShr(V, ConstantInt::get(V->getType(), 64), + "ctpop.part.sh", IP); + BitSize -= 64; + } } - if (V->getType() != DestTy) - V = new CastInst(V, DestTy, V->getName(), IP); - return V; + return Count; } /// LowerCTLZ - Emit the code to lower ctlz of V before the specified /// instruction IP. static Value *LowerCTLZ(Value *V, Instruction *IP) { - const Type *DestTy = V->getType(); - - // Force to unsigned so that the shift rights are logical. - if (DestTy->isSigned()) - V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP); unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); - for (unsigned i = 1; i != BitSize; i <<= 1) { - Value *ShVal = ConstantInt::get(Type::UByteTy, i); - ShVal = new ShiftInst(Instruction::Shr, V, ShVal, "ctlz.sh", IP); - V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP); + for (unsigned i = 1; i < BitSize; i <<= 1) { + Value *ShVal = ConstantInt::get(V->getType(), i); + ShVal = BinaryOperator::CreateLShr(V, ShVal, "ctlz.sh", IP); + V = BinaryOperator::CreateOr(V, ShVal, "ctlz.step", IP); } - if (V->getType() != DestTy) - V = new CastInst(V, DestTy, V->getName(), IP); - - V = BinaryOperator::createNot(V, "", IP); + V = BinaryOperator::CreateNot(V, "", IP); return LowerCTPOP(V, IP); } +/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes +/// three integer arguments. The first argument is the Value from which the +/// bits will be selected. It may be of any bit width. The second and third +/// arguments specify a range of bits to select with the second argument +/// specifying the low bit and the third argument specifying the high bit. Both +/// must be type i32. The result is the corresponding selected bits from the +/// Value in the same width as the Value (first argument). If the low bit index +/// is higher than the high bit index then the inverse selection is done and +/// the bits are returned in inverse order. +/// @brief Lowering of llvm.part.select intrinsic. +static Instruction *LowerPartSelect(CallInst *CI) { + // Make sure we're dealing with a part select intrinsic here + Function *F = CI->getCalledFunction(); + const FunctionType *FT = F->getFunctionType(); + if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || + FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() || + !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger()) + return CI; + + // Get the intrinsic implementation function by converting all the . to _ + // in the intrinsic's function name and then reconstructing the function + // declaration. + std::string Name(F->getName()); + for (unsigned i = 4; i < Name.length(); ++i) + if (Name[i] == '.') + Name[i] = '_'; + Module* M = F->getParent(); + F = cast(M->getOrInsertFunction(Name, FT)); + F->setLinkage(GlobalValue::WeakLinkage); + + // If we haven't defined the impl function yet, do so now + if (F->isDeclaration()) { + + // Get the arguments to the function + Function::arg_iterator args = F->arg_begin(); + Value* Val = args++; Val->setName("Val"); + Value* Lo = args++; Lo->setName("Lo"); + Value* Hi = args++; Hi->setName("High"); + + // We want to select a range of bits here such that [Hi, Lo] is shifted + // down to the low bits. However, it is quite possible that Hi is smaller + // than Lo in which case the bits have to be reversed. + + // Create the blocks we will need for the two cases (forward, reverse) + BasicBlock* CurBB = BasicBlock::Create("entry", F); + BasicBlock *RevSize = BasicBlock::Create("revsize", CurBB->getParent()); + BasicBlock *FwdSize = BasicBlock::Create("fwdsize", CurBB->getParent()); + BasicBlock *Compute = BasicBlock::Create("compute", CurBB->getParent()); + BasicBlock *Reverse = BasicBlock::Create("reverse", CurBB->getParent()); + BasicBlock *RsltBlk = BasicBlock::Create("result", CurBB->getParent()); + + // Cast Hi and Lo to the size of Val so the widths are all the same + if (Hi->getType() != Val->getType()) + Hi = CastInst::CreateIntegerCast(Hi, Val->getType(), false, + "tmp", CurBB); + if (Lo->getType() != Val->getType()) + Lo = CastInst::CreateIntegerCast(Lo, Val->getType(), false, + "tmp", CurBB); + + // Compute a few things that both cases will need, up front. + Constant* Zero = ConstantInt::get(Val->getType(), 0); + Constant* One = ConstantInt::get(Val->getType(), 1); + Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType()); + + // Compare the Hi and Lo bit positions. This is used to determine + // which case we have (forward or reverse) + ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB); + BranchInst::Create(RevSize, FwdSize, Cmp, CurBB); + + // First, copmute the number of bits in the forward case. + Instruction* FBitSize = + BinaryOperator::CreateSub(Hi, Lo,"fbits", FwdSize); + BranchInst::Create(Compute, FwdSize); + + // Second, compute the number of bits in the reverse case. + Instruction* RBitSize = + BinaryOperator::CreateSub(Lo, Hi, "rbits", RevSize); + BranchInst::Create(Compute, RevSize); + + // Now, compute the bit range. Start by getting the bitsize and the shift + // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for + // the number of bits we want in the range. We shift the bits down to the + // least significant bits, apply the mask to zero out unwanted high bits, + // and we have computed the "forward" result. It may still need to be + // reversed. + + // Get the BitSize from one of the two subtractions + PHINode *BitSize = PHINode::Create(Val->getType(), "bits", Compute); + BitSize->reserveOperandSpace(2); + BitSize->addIncoming(FBitSize, FwdSize); + BitSize->addIncoming(RBitSize, RevSize); + + // Get the ShiftAmount as the smaller of Hi/Lo + PHINode *ShiftAmt = PHINode::Create(Val->getType(), "shiftamt", Compute); + ShiftAmt->reserveOperandSpace(2); + ShiftAmt->addIncoming(Lo, FwdSize); + ShiftAmt->addIncoming(Hi, RevSize); + + // Increment the bit size + Instruction *BitSizePlusOne = + BinaryOperator::CreateAdd(BitSize, One, "bits", Compute); + + // Create a Mask to zero out the high order bits. + Instruction* Mask = + BinaryOperator::CreateShl(AllOnes, BitSizePlusOne, "mask", Compute); + Mask = BinaryOperator::CreateNot(Mask, "mask", Compute); + + // Shift the bits down and apply the mask + Instruction* FRes = + BinaryOperator::CreateLShr(Val, ShiftAmt, "fres", Compute); + FRes = BinaryOperator::CreateAnd(FRes, Mask, "fres", Compute); + BranchInst::Create(Reverse, RsltBlk, Cmp, Compute); + + // In the Reverse block we have the mask already in FRes but we must reverse + // it by shifting FRes bits right and putting them in RRes by shifting them + // in from left. + + // First set up our loop counters + PHINode *Count = PHINode::Create(Val->getType(), "count", Reverse); + Count->reserveOperandSpace(2); + Count->addIncoming(BitSizePlusOne, Compute); + + // Next, get the value that we are shifting. + PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", Reverse); + BitsToShift->reserveOperandSpace(2); + BitsToShift->addIncoming(FRes, Compute); + + // Finally, get the result of the last computation + PHINode *RRes = PHINode::Create(Val->getType(), "rres", Reverse); + RRes->reserveOperandSpace(2); + RRes->addIncoming(Zero, Compute); + + // Decrement the counter + Instruction *Decr = BinaryOperator::CreateSub(Count, One, "decr", Reverse); + Count->addIncoming(Decr, Reverse); + + // Compute the Bit that we want to move + Instruction *Bit = + BinaryOperator::CreateAnd(BitsToShift, One, "bit", Reverse); + + // Compute the new value for next iteration. + Instruction *NewVal = + BinaryOperator::CreateLShr(BitsToShift, One, "rshift", Reverse); + BitsToShift->addIncoming(NewVal, Reverse); + + // Shift the bit into the low bits of the result. + Instruction *NewRes = + BinaryOperator::CreateShl(RRes, One, "lshift", Reverse); + NewRes = BinaryOperator::CreateOr(NewRes, Bit, "addbit", Reverse); + RRes->addIncoming(NewRes, Reverse); + + // Terminate loop if we've moved all the bits. + ICmpInst *Cond = + new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse); + BranchInst::Create(RsltBlk, Reverse, Cond, Reverse); + + // Finally, in the result block, select one of the two results with a PHI + // node and return the result; + CurBB = RsltBlk; + PHINode *BitSelect = PHINode::Create(Val->getType(), "part_select", CurBB); + BitSelect->reserveOperandSpace(2); + BitSelect->addIncoming(FRes, Compute); + BitSelect->addIncoming(NewRes, Reverse); + ReturnInst::Create(BitSelect, CurBB); + } + + // Return a call to the implementation function + Value *Args[] = { + CI->getOperand(1), + CI->getOperand(2), + CI->getOperand(3) + }; + return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI); +} + +/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes +/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High) +/// The first two arguments can be any bit width. The result is the same width +/// as %Value. The operation replaces bits between %Low and %High with the value +/// in %Replacement. If %Replacement is not the same width, it is truncated or +/// zero extended as appropriate to fit the bits being replaced. If %Low is +/// greater than %High then the inverse set of bits are replaced. +/// @brief Lowering of llvm.bit.part.set intrinsic. +static Instruction *LowerPartSet(CallInst *CI) { + // Make sure we're dealing with a part select intrinsic here + Function *F = CI->getCalledFunction(); + const FunctionType *FT = F->getFunctionType(); + if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || + FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() || + !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() || + !FT->getParamType(3)->isInteger()) + return CI; + + // Get the intrinsic implementation function by converting all the . to _ + // in the intrinsic's function name and then reconstructing the function + // declaration. + std::string Name(F->getName()); + for (unsigned i = 4; i < Name.length(); ++i) + if (Name[i] == '.') + Name[i] = '_'; + Module* M = F->getParent(); + F = cast(M->getOrInsertFunction(Name, FT)); + F->setLinkage(GlobalValue::WeakLinkage); + + // If we haven't defined the impl function yet, do so now + if (F->isDeclaration()) { + // Get the arguments for the function. + Function::arg_iterator args = F->arg_begin(); + Value* Val = args++; Val->setName("Val"); + Value* Rep = args++; Rep->setName("Rep"); + Value* Lo = args++; Lo->setName("Lo"); + Value* Hi = args++; Hi->setName("Hi"); + + // Get some types we need + const IntegerType* ValTy = cast(Val->getType()); + const IntegerType* RepTy = cast(Rep->getType()); + uint32_t ValBits = ValTy->getBitWidth(); + uint32_t RepBits = RepTy->getBitWidth(); + + // Constant Definitions + ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits); + ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy); + ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy); + ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1); + ConstantInt* ValOne = ConstantInt::get(ValTy, 1); + ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0); + ConstantInt* ValZero = ConstantInt::get(ValTy, 0); + + // Basic blocks we fill in below. + BasicBlock* entry = BasicBlock::Create("entry", F, 0); + BasicBlock* large = BasicBlock::Create("large", F, 0); + BasicBlock* small = BasicBlock::Create("small", F, 0); + BasicBlock* reverse = BasicBlock::Create("reverse", F, 0); + BasicBlock* result = BasicBlock::Create("result", F, 0); + + // BASIC BLOCK: entry + // First, get the number of bits that we're placing as an i32 + ICmpInst* is_forward = + new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry); + SelectInst* Hi_pn = SelectInst::Create(is_forward, Hi, Lo, "", entry); + SelectInst* Lo_pn = SelectInst::Create(is_forward, Lo, Hi, "", entry); + BinaryOperator* NumBits = BinaryOperator::CreateSub(Hi_pn, Lo_pn, "",entry); + NumBits = BinaryOperator::CreateAdd(NumBits, One, "", entry); + // Now, convert Lo and Hi to ValTy bit width + if (ValBits > 32) { + Lo = new ZExtInst(Lo_pn, ValTy, "", entry); + } else if (ValBits < 32) { + Lo = new TruncInst(Lo_pn, ValTy, "", entry); + } + // Determine if the replacement bits are larger than the number of bits we + // are replacing and deal with it. + ICmpInst* is_large = + new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry); + BranchInst::Create(large, small, is_large, entry); + + // BASIC BLOCK: large + Instruction* MaskBits = + BinaryOperator::CreateSub(RepBitWidth, NumBits, "", large); + MaskBits = CastInst::CreateIntegerCast(MaskBits, RepMask->getType(), + false, "", large); + BinaryOperator* Mask1 = + BinaryOperator::CreateLShr(RepMask, MaskBits, "", large); + BinaryOperator* Rep2 = BinaryOperator::CreateAnd(Mask1, Rep, "", large); + BranchInst::Create(small, large); + + // BASIC BLOCK: small + PHINode* Rep3 = PHINode::Create(RepTy, "", small); + Rep3->reserveOperandSpace(2); + Rep3->addIncoming(Rep2, large); + Rep3->addIncoming(Rep, entry); + Value* Rep4 = Rep3; + if (ValBits > RepBits) + Rep4 = new ZExtInst(Rep3, ValTy, "", small); + else if (ValBits < RepBits) + Rep4 = new TruncInst(Rep3, ValTy, "", small); + BranchInst::Create(result, reverse, is_forward, small); + + // BASIC BLOCK: reverse (reverses the bits of the replacement) + // Set up our loop counter as a PHI so we can decrement on each iteration. + // We will loop for the number of bits in the replacement value. + PHINode *Count = PHINode::Create(Type::Int32Ty, "count", reverse); + Count->reserveOperandSpace(2); + Count->addIncoming(NumBits, small); + + // Get the value that we are shifting bits out of as a PHI because + // we'll change this with each iteration. + PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", reverse); + BitsToShift->reserveOperandSpace(2); + BitsToShift->addIncoming(Rep4, small); + + // Get the result of the last computation or zero on first iteration + PHINode *RRes = PHINode::Create(Val->getType(), "rres", reverse); + RRes->reserveOperandSpace(2); + RRes->addIncoming(ValZero, small); + + // Decrement the loop counter by one + Instruction *Decr = BinaryOperator::CreateSub(Count, One, "", reverse); + Count->addIncoming(Decr, reverse); + + // Get the bit that we want to move into the result + Value *Bit = BinaryOperator::CreateAnd(BitsToShift, ValOne, "", reverse); + + // Compute the new value of the bits to shift for the next iteration. + Value *NewVal = BinaryOperator::CreateLShr(BitsToShift, ValOne,"", reverse); + BitsToShift->addIncoming(NewVal, reverse); + + // Shift the bit we extracted into the low bit of the result. + Instruction *NewRes = BinaryOperator::CreateShl(RRes, ValOne, "", reverse); + NewRes = BinaryOperator::CreateOr(NewRes, Bit, "", reverse); + RRes->addIncoming(NewRes, reverse); + + // Terminate loop if we've moved all the bits. + ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse); + BranchInst::Create(result, reverse, Cond, reverse); + + // BASIC BLOCK: result + PHINode *Rplcmnt = PHINode::Create(Val->getType(), "", result); + Rplcmnt->reserveOperandSpace(2); + Rplcmnt->addIncoming(NewRes, reverse); + Rplcmnt->addIncoming(Rep4, small); + Value* t0 = CastInst::CreateIntegerCast(NumBits,ValTy,false,"",result); + Value* t1 = BinaryOperator::CreateShl(ValMask, Lo, "", result); + Value* t2 = BinaryOperator::CreateNot(t1, "", result); + Value* t3 = BinaryOperator::CreateShl(t1, t0, "", result); + Value* t4 = BinaryOperator::CreateOr(t2, t3, "", result); + Value* t5 = BinaryOperator::CreateAnd(t4, Val, "", result); + Value* t6 = BinaryOperator::CreateShl(Rplcmnt, Lo, "", result); + Value* Rslt = BinaryOperator::CreateOr(t5, t6, "part_set", result); + ReturnInst::Create(Rslt, result); + } + + // Return a call to the implementation function + Value *Args[] = { + CI->getOperand(1), + CI->getOperand(2), + CI->getOperand(3), + CI->getOperand(4) + }; + return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI); +} -void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { +void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { Function *Callee = CI->getCalledFunction(); assert(Callee && "Cannot lower an indirect call!"); switch (Callee->getIntrinsicID()) { case Intrinsic::not_intrinsic: - std::cerr << "Cannot lower a call to a non-intrinsic function '" - << Callee->getName() << "'!\n"; + cerr << "Cannot lower a call to a non-intrinsic function '" + << Callee->getName() << "'!\n"; abort(); default: - std::cerr << "Error: Code generator does not support intrinsic function '" - << Callee->getName() << "'!\n"; + cerr << "Error: Code generator does not support intrinsic function '" + << Callee->getName() << "'!\n"; abort(); // The setjmp/longjmp intrinsics should only exist in the code if it was @@ -306,9 +663,9 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { // by the lowerinvoke pass. In both cases, the right thing to do is to // convert the call to an explicit setjmp or longjmp call. case Intrinsic::setjmp: { - static Function *SetjmpFCache = 0; + static Constant *SetjmpFCache = 0; Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(), - Type::IntTy, SetjmpFCache); + Type::Int32Ty, SetjmpFCache); if (CI->getType() != Type::VoidTy) CI->replaceAllUsesWith(V); break; @@ -319,7 +676,7 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { break; case Intrinsic::longjmp: { - static Function *LongjmpFCache = 0; + static Constant *LongjmpFCache = 0; ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(), Type::VoidTy, LongjmpFCache); break; @@ -327,52 +684,49 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { case Intrinsic::siglongjmp: { // Insert the call to abort - static Function *AbortFCache = 0; - ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), Type::VoidTy, - AbortFCache); + static Constant *AbortFCache = 0; + ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), + Type::VoidTy, AbortFCache); break; } - case Intrinsic::ctpop_i8: - case Intrinsic::ctpop_i16: - case Intrinsic::ctpop_i32: - case Intrinsic::ctpop_i64: + case Intrinsic::ctpop: CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI)); break; - case Intrinsic::bswap_i16: - case Intrinsic::bswap_i32: - case Intrinsic::bswap_i64: + case Intrinsic::bswap: CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI)); break; - case Intrinsic::ctlz_i8: - case Intrinsic::ctlz_i16: - case Intrinsic::ctlz_i32: - case Intrinsic::ctlz_i64: + case Intrinsic::ctlz: CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI)); break; - case Intrinsic::cttz_i8: - case Intrinsic::cttz_i16: - case Intrinsic::cttz_i32: - case Intrinsic::cttz_i64: { + case Intrinsic::cttz: { // cttz(x) -> ctpop(~X & (X-1)) Value *Src = CI->getOperand(1); - Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI); - Value *SrcM1 = ConstantInt::get(Src->getType(), 1); - SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI); - Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI); + Value *NotSrc = BinaryOperator::CreateNot(Src, Src->getName()+".not", CI); + Value *SrcM1 = ConstantInt::get(Src->getType(), 1); + SrcM1 = BinaryOperator::CreateSub(Src, SrcM1, "", CI); + Src = LowerCTPOP(BinaryOperator::CreateAnd(NotSrc, SrcM1, "", CI), CI); CI->replaceAllUsesWith(Src); break; } + case Intrinsic::part_select: + CI->replaceAllUsesWith(LowerPartSelect(CI)); + break; + + case Intrinsic::part_set: + CI->replaceAllUsesWith(LowerPartSet(CI)); + break; + case Intrinsic::stacksave: case Intrinsic::stackrestore: { static bool Warned = false; if (!Warned) - std::cerr << "WARNING: this target does not support the llvm.stack" - << (Callee->getIntrinsicID() == Intrinsic::stacksave ? - "save" : "restore") << " intrinsic.\n"; + cerr << "WARNING: this target does not support the llvm.stack" + << (Callee->getIntrinsicID() == Intrinsic::stacksave ? + "save" : "restore") << " intrinsic.\n"; Warned = true; if (Callee->getIntrinsicID() == Intrinsic::stacksave) CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); @@ -381,9 +735,9 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { case Intrinsic::returnaddress: case Intrinsic::frameaddress: - std::cerr << "WARNING: this target does not support the llvm." - << (Callee->getIntrinsicID() == Intrinsic::returnaddress ? - "return" : "frame") << "address intrinsic.\n"; + cerr << "WARNING: this target does not support the llvm." + << (Callee->getIntrinsicID() == Intrinsic::returnaddress ? + "return" : "frame") << "address intrinsic.\n"; CI->replaceAllUsesWith(ConstantPointerNull::get( cast(CI->getType()))); break; @@ -394,9 +748,9 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { case Intrinsic::pcmarker: break; // Simply strip out pcmarker on unsupported architectures case Intrinsic::readcyclecounter: { - std::cerr << "WARNING: this target does not support the llvm.readcyclecoun" - << "ter intrinsic. It is being lowered to a constant 0\n"; - CI->replaceAllUsesWith(ConstantInt::get(Type::ULongTy, 0)); + cerr << "WARNING: this target does not support the llvm.readcyclecoun" + << "ter intrinsic. It is being lowered to a constant 0\n"; + CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0)); break; } @@ -407,57 +761,107 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { case Intrinsic::dbg_declare: break; // Simply strip out debugging intrinsics + case Intrinsic::eh_exception: + case Intrinsic::eh_selector_i32: + case Intrinsic::eh_selector_i64: + CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); + break; + + case Intrinsic::eh_typeid_for_i32: + case Intrinsic::eh_typeid_for_i64: + // Return something different to eh_selector. + CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1)); + break; + + case Intrinsic::var_annotation: + break; // Strip out annotate intrinsic + case Intrinsic::memcpy_i32: case Intrinsic::memcpy_i64: { - // The memcpy intrinsic take an extra alignment argument that the memcpy - // libc function does not. - static Function *MemcpyFCache = 0; - ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1, - (*(CI->op_begin()+1))->getType(), MemcpyFCache); + static Constant *MemcpyFCache = 0; + Value *Size = CI->getOperand(3); + const Type *IntPtr = TD.getIntPtrType(); + if (Size->getType()->getPrimitiveSizeInBits() < + IntPtr->getPrimitiveSizeInBits()) + Size = new ZExtInst(Size, IntPtr, "", CI); + else if (Size->getType()->getPrimitiveSizeInBits() > + IntPtr->getPrimitiveSizeInBits()) + Size = new TruncInst(Size, IntPtr, "", CI); + Value *Ops[3]; + Ops[0] = CI->getOperand(1); + Ops[1] = CI->getOperand(2); + Ops[2] = Size; + ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(), + MemcpyFCache); break; } case Intrinsic::memmove_i32: case Intrinsic::memmove_i64: { - // The memmove intrinsic take an extra alignment argument that the memmove - // libc function does not. - static Function *MemmoveFCache = 0; - ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1, - (*(CI->op_begin()+1))->getType(), MemmoveFCache); + static Constant *MemmoveFCache = 0; + Value *Size = CI->getOperand(3); + const Type *IntPtr = TD.getIntPtrType(); + if (Size->getType()->getPrimitiveSizeInBits() < + IntPtr->getPrimitiveSizeInBits()) + Size = new ZExtInst(Size, IntPtr, "", CI); + else if (Size->getType()->getPrimitiveSizeInBits() > + IntPtr->getPrimitiveSizeInBits()) + Size = new TruncInst(Size, IntPtr, "", CI); + Value *Ops[3]; + Ops[0] = CI->getOperand(1); + Ops[1] = CI->getOperand(2); + Ops[2] = Size; + ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(), + MemmoveFCache); break; } case Intrinsic::memset_i32: case Intrinsic::memset_i64: { - // The memset intrinsic take an extra alignment argument that the memset - // libc function does not. - static Function *MemsetFCache = 0; - ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1, - (*(CI->op_begin()+1))->getType(), MemsetFCache); - break; - } - case Intrinsic::isunordered_f32: - case Intrinsic::isunordered_f64: { - Value *L = CI->getOperand(1); - Value *R = CI->getOperand(2); - - Value *LIsNan = new SetCondInst(Instruction::SetNE, L, L, "LIsNan", CI); - Value *RIsNan = new SetCondInst(Instruction::SetNE, R, R, "RIsNan", CI); - CI->replaceAllUsesWith( - BinaryOperator::create(Instruction::Or, LIsNan, RIsNan, - "isunordered", CI)); + static Constant *MemsetFCache = 0; + Value *Size = CI->getOperand(3); + const Type *IntPtr = TD.getIntPtrType(); + if (Size->getType()->getPrimitiveSizeInBits() < + IntPtr->getPrimitiveSizeInBits()) + Size = new ZExtInst(Size, IntPtr, "", CI); + else if (Size->getType()->getPrimitiveSizeInBits() > + IntPtr->getPrimitiveSizeInBits()) + Size = new TruncInst(Size, IntPtr, "", CI); + Value *Ops[3]; + Ops[0] = CI->getOperand(1); + // Extend the amount to i32. + Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI); + Ops[2] = Size; + ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(), + MemsetFCache); break; } - case Intrinsic::sqrt_f32: - case Intrinsic::sqrt_f64: { - static Function *sqrtFCache = 0; - static Function *sqrtfFCache = 0; - if(CI->getType() == Type::FloatTy) + case Intrinsic::sqrt: { + static Constant *sqrtfFCache = 0; + static Constant *sqrtFCache = 0; + static Constant *sqrtLDCache = 0; + switch (CI->getOperand(1)->getType()->getTypeID()) { + default: assert(0 && "Invalid type in sqrt"); abort(); + case Type::FloatTyID: ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(), - Type::FloatTy, sqrtfFCache); - else + Type::FloatTy, sqrtfFCache); + break; + case Type::DoubleTyID: ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(), - Type::DoubleTy, sqrtFCache); + Type::DoubleTy, sqrtFCache); + break; + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + ReplaceCallWith("sqrtl", CI, CI->op_begin()+1, CI->op_end(), + CI->getOperand(1)->getType(), sqrtLDCache); + break; + } break; } + case Intrinsic::flt_rounds: + // Lower to "round to the nearest" + if (CI->getType() != Type::VoidTy) + CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1)); + break; } assert(CI->use_empty() &&