//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Transforms/Utils/BuildLibCalls.h"
+#include "llvm/Transforms/Utils/Local.h"
using namespace llvm;
/// getPromotedType - Return the specified type promoted as it would be to pass
/// though a va_arg area.
-static const Type *getPromotedType(const Type *Ty) {
- if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
+static Type *getPromotedType(Type *Ty) {
+ if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
if (ITy->getBitWidth() < 32)
return Type::getInt32Ty(Ty->getContext());
}
return Ty;
}
-/// EnforceKnownAlignment - If the specified pointer points to an object that
-/// we control, modify the object's alignment to PrefAlign. This isn't
-/// often possible though. If alignment is important, a more reliable approach
-/// is to simply align all global variables and allocation instructions to
-/// their preferred alignment from the beginning.
-///
-static unsigned EnforceKnownAlignment(Value *V,
- unsigned Align, unsigned PrefAlign) {
-
- User *U = dyn_cast<User>(V);
- if (!U) return Align;
-
- switch (Operator::getOpcode(U)) {
- default: break;
- case Instruction::BitCast:
- return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
- case Instruction::GetElementPtr: {
- // If all indexes are zero, it is just the alignment of the base pointer.
- bool AllZeroOperands = true;
- for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
- if (!isa<Constant>(*i) ||
- !cast<Constant>(*i)->isNullValue()) {
- AllZeroOperands = false;
- break;
- }
-
- if (AllZeroOperands) {
- // Treat this like a bitcast.
- return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
- }
- return Align;
- }
- case Instruction::Alloca: {
- AllocaInst *AI = cast<AllocaInst>(V);
- // If there is a requested alignment and if this is an alloca, round up.
- if (AI->getAlignment() >= PrefAlign)
- return AI->getAlignment();
- AI->setAlignment(PrefAlign);
- return PrefAlign;
- }
- }
-
- if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
- // If there is a large requested alignment and we can, bump up the alignment
- // of the global.
- if (GV->isDeclaration()) return Align;
-
- if (GV->getAlignment() >= PrefAlign)
- return GV->getAlignment();
- // We can only increase the alignment of the global if it has no alignment
- // specified or if it is not assigned a section. If it is assigned a
- // section, the global could be densely packed with other objects in the
- // section, increasing the alignment could cause padding issues.
- if (!GV->hasSection() || GV->getAlignment() == 0)
- GV->setAlignment(PrefAlign);
- return GV->getAlignment();
- }
-
- return Align;
-}
-
-/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
-/// we can determine, return it, otherwise return 0. If PrefAlign is specified,
-/// and it is more than the alignment of the ultimate object, see if we can
-/// increase the alignment of the ultimate object, making this check succeed.
-unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
- unsigned PrefAlign) {
- assert(V->getType()->isPointerTy() &&
- "GetOrEnforceKnownAlignment expects a pointer!");
- unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 64;
- APInt Mask = APInt::getAllOnesValue(BitWidth);
- APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
- ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
- unsigned TrailZ = KnownZero.countTrailingOnes();
-
- // Avoid trouble with rediculously large TrailZ values, such as
- // those computed from a null pointer.
- TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
-
- unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
-
- // LLVM doesn't support alignments larger than this currently.
- Align = std::min(Align, +Value::MaximumAlignment);
-
- if (PrefAlign > Align)
- Align = EnforceKnownAlignment(V, Align, PrefAlign);
-
- // We don't need to make any adjustment.
- return Align;
-}
Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
- unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(0));
- unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(1));
+ unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD);
+ unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
unsigned MinAlign = std::min(DstAlign, SrcAlign);
unsigned CopyAlign = MI->getAlignment();
unsigned DstAddrSp =
cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
- const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
+ IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
// integer datatype.
Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
if (StrippedDest != MI->getArgOperand(0)) {
- const Type *SrcETy = cast<PointerType>(StrippedDest->getType())
+ Type *SrcETy = cast<PointerType>(StrippedDest->getType())
->getElementType();
if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
// The SrcETy might be something like {{{double}}} or [1 x double]. Rip
// down through these levels if so.
while (!SrcETy->isSingleValueType()) {
- if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
+ if (StructType *STy = dyn_cast<StructType>(SrcETy)) {
if (STy->getNumElements() == 1)
SrcETy = STy->getElementType(0);
else
break;
- } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
+ } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
if (ATy->getNumElements() == 1)
SrcETy = ATy->getElementType();
else
Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
- Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign);
- InsertNewInstBefore(L, *MI);
- InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign),
- *MI);
+ LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
+ L->setAlignment(SrcAlign);
+ StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
+ S->setAlignment(DstAlign);
// Set the size of the copy to 0, it will be deleted on the next iteration.
MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
}
Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
- unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
+ unsigned Alignment = getKnownAlignment(MI->getDest(), TD);
if (MI->getAlignment() < Alignment) {
MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
Alignment, false));
// memset(s,c,n) -> store s, c (for n=1,2,4,8)
if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
- const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
+ Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
Value *Dest = MI->getDest();
unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
// Extract the fill value and store.
uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
- InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
- Dest, false, Alignment), *MI);
+ StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
+ MI->isVolatile());
+ S->setAlignment(Alignment);
// Set the size of the copy to 0, it will be deleted on the next iteration.
MI->setLength(Constant::getNullValue(LenC->getType()));
if (GVSrc->isConstant()) {
Module *M = CI.getParent()->getParent()->getParent();
Intrinsic::ID MemCpyID = Intrinsic::memcpy;
- const Type *Tys[3] = { CI.getArgOperand(0)->getType(),
- CI.getArgOperand(1)->getType(),
- CI.getArgOperand(2)->getType() };
- CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
+ Type *Tys[3] = { CI.getArgOperand(0)->getType(),
+ CI.getArgOperand(1)->getType(),
+ CI.getArgOperand(2)->getType() };
+ CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
Changed = true;
}
}
// We need target data for just about everything so depend on it.
if (!TD) break;
- const Type *ReturnTy = CI.getType();
- bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
+ Type *ReturnTy = CI.getType();
+ uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL;
// Get to the real allocated thing and offset as fast as possible.
Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
-
+
+ uint64_t Offset = 0;
+ uint64_t Size = -1ULL;
+
+ // Try to look through constant GEPs.
+ if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) {
+ if (!GEP->hasAllConstantIndices()) break;
+
+ // Get the current byte offset into the thing. Use the original
+ // operand in case we're looking through a bitcast.
+ SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
+ if (!GEP->getPointerOperandType()->isPointerTy())
+ return 0;
+ Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops);
+
+ Op1 = GEP->getPointerOperand()->stripPointerCasts();
+
+ // Make sure we're not a constant offset from an external
+ // global.
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1))
+ if (!GV->hasDefinitiveInitializer()) break;
+ }
+
// If we've stripped down to a single global variable that we
// can know the size of then just return that.
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
if (GV->hasDefinitiveInitializer()) {
Constant *C = GV->getInitializer();
- uint64_t GlobalSize = TD->getTypeAllocSize(C->getType());
- return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize));
+ Size = TD->getTypeAllocSize(C->getType());
} else {
// Can't determine size of the GV.
- Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
+ Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow);
return ReplaceInstUsesWith(CI, RetVal);
}
} else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
// Get alloca size.
if (AI->getAllocatedType()->isSized()) {
- uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
+ Size = TD->getTypeAllocSize(AI->getAllocatedType());
if (AI->isArrayAllocation()) {
const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
if (!C) break;
- AllocaSize *= C->getZExtValue();
+ Size *= C->getZExtValue();
}
- return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize));
}
} else if (CallInst *MI = extractMallocCall(Op1)) {
- const Type* MallocType = getMallocAllocatedType(MI);
- // Get alloca size.
- if (MallocType && MallocType->isSized()) {
- if (Value *NElems = getMallocArraySize(MI, TD, true)) {
+ // Get allocation size.
+ Type* MallocType = getMallocAllocatedType(MI);
+ if (MallocType && MallocType->isSized())
+ if (Value *NElems = getMallocArraySize(MI, TD, true))
if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
- return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy,
- (NElements->getZExtValue() * TD->getTypeAllocSize(MallocType))));
- }
- }
- } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) {
- // Only handle constant GEPs here.
- if (CE->getOpcode() != Instruction::GetElementPtr) break;
- GEPOperator *GEP = cast<GEPOperator>(CE);
-
- // Make sure we're not a constant offset from an external
- // global.
- Value *Operand = GEP->getPointerOperand();
- Operand = Operand->stripPointerCasts();
- if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand))
- if (!GV->hasDefinitiveInitializer()) break;
-
- // Get what we're pointing to and its size.
- const PointerType *BaseType =
- cast<PointerType>(Operand->getType());
- uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType());
-
- // Get the current byte offset into the thing. Use the original
- // operand in case we're looking through a bitcast.
- SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end());
- const PointerType *OffsetType =
- cast<PointerType>(GEP->getPointerOperand()->getType());
- uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size());
-
- if (Size < Offset) {
- // Out of bound reference? Negative index normalized to large
- // index? Just return "I don't know".
- Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
- return ReplaceInstUsesWith(CI, RetVal);
- }
-
- Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset);
- return ReplaceInstUsesWith(CI, RetVal);
- }
+ Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
+ }
// Do not return "I don't know" here. Later optimization passes could
// make it possible to evaluate objectsize to a constant.
- break;
+ if (Size == -1ULL)
+ break;
+
+ if (Size < Offset) {
+ // Out of bound reference? Negative index normalized to large
+ // index? Just return "I don't know".
+ return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow));
+ }
+ return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
}
case Intrinsic::bswap:
// bswap(bswap(x)) -> x
case Intrinsic::cttz: {
// If all bits below the first known one are known zero,
// this value is constant.
- const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
+ IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
+ // FIXME: Try to simplify vectors of integers.
+ if (!IT) break;
uint32_t BitWidth = IT->getBitWidth();
APInt KnownZero(BitWidth, 0);
APInt KnownOne(BitWidth, 0);
case Intrinsic::ctlz: {
// If all bits above the first known one are known zero,
// this value is constant.
- const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
+ IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
+ // FIXME: Try to simplify vectors of integers.
+ if (!IT) break;
uint32_t BitWidth = IT->getBitWidth();
APInt KnownZero(BitWidth, 0);
APInt KnownOne(BitWidth, 0);
break;
case Intrinsic::uadd_with_overflow: {
Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
- const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
+ IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
uint32_t BitWidth = IT->getBitWidth();
APInt Mask = APInt::getSignBit(BitWidth);
APInt LHSKnownZero(BitWidth, 0);
if (LHSKnownNegative && RHSKnownNegative) {
// The sign bit is set in both cases: this MUST overflow.
// Create a simple add instruction, and insert it into the struct.
- Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
- Worklist.Add(Add);
+ Value *Add = Builder->CreateAdd(LHS, RHS);
+ Add->takeName(&CI);
Constant *V[] = {
- UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
+ UndefValue::get(LHS->getType()),
+ ConstantInt::getTrue(II->getContext())
};
- Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
+ StructType *ST = cast<StructType>(II->getType());
+ Constant *Struct = ConstantStruct::get(ST, V);
return InsertValueInst::Create(Struct, Add, 0);
}
-
+
if (LHSKnownPositive && RHSKnownPositive) {
// The sign bit is clear in both cases: this CANNOT overflow.
// Create a simple add instruction, and insert it into the struct.
- Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
- Worklist.Add(Add);
+ Value *Add = Builder->CreateNUWAdd(LHS, RHS);
+ Add->takeName(&CI);
Constant *V[] = {
UndefValue::get(LHS->getType()),
ConstantInt::getFalse(II->getContext())
};
- Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
+ StructType *ST = cast<StructType>(II->getType());
+ Constant *Struct = ConstantStruct::get(ST, V);
return InsertValueInst::Create(Struct, Add, 0);
}
}
UndefValue::get(II->getArgOperand(0)->getType()),
ConstantInt::getFalse(II->getContext())
};
- Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
+ Constant *Struct =
+ ConstantStruct::get(cast<StructType>(II->getType()), V);
return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
}
}
UndefValue::get(II->getArgOperand(0)->getType()),
ConstantInt::getFalse(II->getContext())
};
- Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
+ Constant *Struct =
+ ConstantStruct::get(cast<StructType>(II->getType()), V);
return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
}
}
break;
- case Intrinsic::umul_with_overflow:
+ case Intrinsic::umul_with_overflow: {
+ Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
+ unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
+ APInt Mask = APInt::getAllOnesValue(BitWidth);
+
+ APInt LHSKnownZero(BitWidth, 0);
+ APInt LHSKnownOne(BitWidth, 0);
+ ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
+ APInt RHSKnownZero(BitWidth, 0);
+ APInt RHSKnownOne(BitWidth, 0);
+ ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
+
+ // Get the largest possible values for each operand.
+ APInt LHSMax = ~LHSKnownZero;
+ APInt RHSMax = ~RHSKnownZero;
+
+ // If multiplying the maximum values does not overflow then we can turn
+ // this into a plain NUW mul.
+ bool Overflow;
+ LHSMax.umul_ov(RHSMax, Overflow);
+ if (!Overflow) {
+ Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
+ Constant *V[] = {
+ UndefValue::get(LHS->getType()),
+ Builder->getFalse()
+ };
+ Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
+ return InsertValueInst::Create(Struct, Mul, 0);
+ }
+ } // FALL THROUGH
case Intrinsic::smul_with_overflow:
// Canonicalize constants into the RHS.
if (isa<Constant>(II->getArgOperand(0)) &&
UndefValue::get(II->getArgOperand(0)->getType()),
ConstantInt::getFalse(II->getContext())
};
- Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
+ Constant *Struct =
+ ConstantStruct::get(cast<StructType>(II->getType()), V);
return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
}
}
break;
case Intrinsic::ppc_altivec_lvx:
case Intrinsic::ppc_altivec_lvxl:
- case Intrinsic::x86_sse_loadu_ps:
- case Intrinsic::x86_sse2_loadu_pd:
- case Intrinsic::x86_sse2_loadu_dq:
- // Turn PPC lvx -> load if the pointer is known aligned.
- // Turn X86 loadups -> load if the pointer is known aligned.
- if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
+ // Turn PPC lvx -> load if the pointer is known aligned.
+ if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
PointerType::getUnqual(II->getType()));
return new LoadInst(Ptr);
case Intrinsic::ppc_altivec_stvx:
case Intrinsic::ppc_altivec_stvxl:
// Turn stvx -> store if the pointer is known aligned.
- if (GetOrEnforceKnownAlignment(II->getArgOperand(1), 16) >= 16) {
- const Type *OpPtrTy =
+ if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
+ Type *OpPtrTy =
PointerType::getUnqual(II->getArgOperand(0)->getType());
Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
return new StoreInst(II->getArgOperand(0), Ptr);
case Intrinsic::x86_sse2_storeu_pd:
case Intrinsic::x86_sse2_storeu_dq:
// Turn X86 storeu -> store if the pointer is known aligned.
- if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
- const Type *OpPtrTy =
+ if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
+ Type *OpPtrTy =
PointerType::getUnqual(II->getArgOperand(1)->getType());
Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
return new StoreInst(II->getArgOperand(1), Ptr);
}
break;
-
- case Intrinsic::x86_sse_cvttss2si: {
- // These intrinsics only demands the 0th element of its input vector. If
+
+ case Intrinsic::x86_sse_cvtss2si:
+ case Intrinsic::x86_sse_cvtss2si64:
+ case Intrinsic::x86_sse_cvttss2si:
+ case Intrinsic::x86_sse_cvttss2si64:
+ case Intrinsic::x86_sse2_cvtsd2si:
+ case Intrinsic::x86_sse2_cvtsd2si64:
+ case Intrinsic::x86_sse2_cvttsd2si:
+ case Intrinsic::x86_sse2_cvttsd2si64: {
+ // These intrinsics only demand the 0th element of their input vectors. If
// we can simplify the input based on that, do so now.
unsigned VWidth =
cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
}
break;
}
-
+
+
+ case Intrinsic::x86_sse41_pmovsxbw:
+ case Intrinsic::x86_sse41_pmovsxwd:
+ case Intrinsic::x86_sse41_pmovsxdq:
+ case Intrinsic::x86_sse41_pmovzxbw:
+ case Intrinsic::x86_sse41_pmovzxwd:
+ case Intrinsic::x86_sse41_pmovzxdq: {
+ // pmov{s|z}x ignores the upper half of their input vectors.
+ unsigned VWidth =
+ cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
+ unsigned LowHalfElts = VWidth / 2;
+ APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
+ APInt UndefElts(VWidth, 0);
+ if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
+ InputDemandedElts,
+ UndefElts)) {
+ II->setArgOperand(0, TmpV);
+ return II;
+ }
+ break;
+ }
+
case Intrinsic::ppc_altivec_vperm:
// Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
if (ExtractedElts[Idx] == 0) {
ExtractedElts[Idx] =
- Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
- ConstantInt::get(Type::getInt32Ty(II->getContext()),
- Idx&15, false), "tmp");
+ Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
+ Builder->getInt32(Idx&15));
}
// Insert this value into the result vector.
Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
- ConstantInt::get(Type::getInt32Ty(II->getContext()),
- i, false), "tmp");
+ Builder->getInt32(i));
}
return CastInst::Create(Instruction::BitCast, Result, CI.getType());
}
case Intrinsic::arm_neon_vst2lane:
case Intrinsic::arm_neon_vst3lane:
case Intrinsic::arm_neon_vst4lane: {
- unsigned MemAlign = GetOrEnforceKnownAlignment(II->getArgOperand(0));
+ unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
unsigned AlignArg = II->getNumArgOperands() - 1;
ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
}
}
- // If the stack restore is in a return/unwind block and if there are no
- // allocas or calls between the restore and the return, nuke the restore.
- if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
+ // If the stack restore is in a return, resume, or unwind block and if there
+ // are no allocas or calls between the restore and the return, nuke the
+ // restore.
+ if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI) ||
+ isa<UnwindInst>(TI)))
return EraseInstFromFunction(CI);
break;
}
// The size of ByVal arguments is derived from the type, so we
// can't change to a type with a different size. If the size were
// passed explicitly we could avoid this check.
- if (!CS.paramHasAttr(ix, Attribute::ByVal))
+ if (!CS.isByValArgument(ix))
return true;
- const Type* SrcTy =
+ Type* SrcTy =
cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
- const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
+ Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
if (!SrcTy->isSized() || !DstTy->isSized())
return false;
if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
}
bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
+ if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
+ return true;
if (ConstantInt *SizeCI =
dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
if (SizeCI->isAllOnesValue())
return true;
- if (isString)
- return SizeCI->getZExtValue() >=
- GetStringLength(CI->getArgOperand(SizeArgOp));
+ if (isString) {
+ uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
+ // If the length is 0 we don't know how long it is and so we can't
+ // remove the check.
+ if (Len == 0) return false;
+ return SizeCI->getZExtValue() >= Len;
+ }
if (ConstantInt *Arg = dyn_cast<ConstantInt>(
CI->getArgOperand(SizeArgOp)))
return SizeCI->getZExtValue() >= Arg->getZExtValue();
return Simplifier.NewInstruction;
}
+static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
+ // Strip off at most one level of pointer casts, looking for an alloca. This
+ // is good enough in practice and simpler than handling any number of casts.
+ Value *Underlying = TrampMem->stripPointerCasts();
+ if (Underlying != TrampMem &&
+ (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
+ return 0;
+ if (!isa<AllocaInst>(Underlying))
+ return 0;
+
+ IntrinsicInst *InitTrampoline = 0;
+ for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
+ I != E; I++) {
+ IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
+ if (!II)
+ return 0;
+ if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
+ if (InitTrampoline)
+ // More than one init_trampoline writes to this value. Give up.
+ return 0;
+ InitTrampoline = II;
+ continue;
+ }
+ if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
+ // Allow any number of calls to adjust.trampoline.
+ continue;
+ return 0;
+ }
+
+ // No call to init.trampoline found.
+ if (!InitTrampoline)
+ return 0;
+
+ // Check that the alloca is being used in the expected way.
+ if (InitTrampoline->getOperand(0) != TrampMem)
+ return 0;
+
+ return InitTrampoline;
+}
+
+static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
+ Value *TrampMem) {
+ // Visit all the previous instructions in the basic block, and try to find a
+ // init.trampoline which has a direct path to the adjust.trampoline.
+ for (BasicBlock::iterator I = AdjustTramp,
+ E = AdjustTramp->getParent()->begin(); I != E; ) {
+ Instruction *Inst = --I;
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
+ if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
+ II->getOperand(0) == TrampMem)
+ return II;
+ if (Inst->mayWriteToMemory())
+ return 0;
+ }
+ return 0;
+}
+
+// Given a call to llvm.adjust.trampoline, find and return the corresponding
+// call to llvm.init.trampoline if the call to the trampoline can be optimized
+// to a direct call to a function. Otherwise return NULL.
+//
+static IntrinsicInst *FindInitTrampoline(Value *Callee) {
+ Callee = Callee->stripPointerCasts();
+ IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
+ if (!AdjustTramp ||
+ AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
+ return 0;
+
+ Value *TrampMem = AdjustTramp->getOperand(0);
+
+ if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
+ return IT;
+ if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
+ return IT;
+ return 0;
+}
+
// visitCallSite - Improvements for call and invoke instructions.
//
Instruction *InstCombiner::visitCallSite(CallSite CS) {
// If OldCall dues not return void then replaceAllUsesWith undef.
// This allows ValueHandlers and custom metadata to adjust itself.
if (!OldCall->getType()->isVoidTy())
- OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
+ ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
if (isa<CallInst>(OldCall))
return EraseInstFromFunction(*OldCall);
// If CS does not return void then replaceAllUsesWith undef.
// This allows ValueHandlers and custom metadata to adjust itself.
if (!CS.getInstruction()->getType()->isVoidTy())
- CS.getInstruction()->
- replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
+ ReplaceInstUsesWith(*CS.getInstruction(),
+ UndefValue::get(CS.getInstruction()->getType()));
if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
// Don't break the CFG, insert a dummy cond branch.
return EraseInstFromFunction(*CS.getInstruction());
}
- if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
- if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
- if (In->getIntrinsicID() == Intrinsic::init_trampoline)
- return transformCallThroughTrampoline(CS);
+ if (IntrinsicInst *II = FindInitTrampoline(Callee))
+ return transformCallThroughTrampoline(CS, II);
- const PointerType *PTy = cast<PointerType>(Callee->getType());
- const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
+ PointerType *PTy = cast<PointerType>(Callee->getType());
+ FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
if (FTy->isVarArg()) {
- int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
+ int ix = FTy->getNumParams();
// See if we can optimize any arguments passed through the varargs area of
// the call.
for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
// would cause a type conversion of one of our arguments, change this call to
// be a direct call with arguments casted to the appropriate types.
//
- const FunctionType *FT = Callee->getFunctionType();
- const Type *OldRetTy = Caller->getType();
- const Type *NewRetTy = FT->getReturnType();
+ FunctionType *FT = Callee->getFunctionType();
+ Type *OldRetTy = Caller->getType();
+ Type *NewRetTy = FT->getReturnType();
if (NewRetTy->isStructTy())
return false; // TODO: Handle multiple return values.
CallSite::arg_iterator AI = CS.arg_begin();
for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
- const Type *ParamTy = FT->getParamType(i);
- const Type *ActTy = (*AI)->getType();
+ Type *ParamTy = FT->getParamType(i);
+ Type *ActTy = (*AI)->getType();
if (!CastInst::isCastable(ActTy, ParamTy))
return false; // Cannot transform this parameter value.
- if (CallerPAL.getParamAttributes(i + 1)
- & Attribute::typeIncompatible(ParamTy))
+ unsigned Attrs = CallerPAL.getParamAttributes(i + 1);
+ if (Attrs & Attribute::typeIncompatible(ParamTy))
return false; // Attribute not compatible with transformed value.
+
+ // If the parameter is passed as a byval argument, then we have to have a
+ // sized type and the sized type has to have the same size as the old type.
+ if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
+ PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
+ if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
+ return false;
+
+ Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
+ if (TD->getTypeAllocSize(CurElTy) !=
+ TD->getTypeAllocSize(ParamPTy->getElementType()))
+ return false;
+ }
// Converting from one pointer type to another or between a pointer and an
// integer of the same size is safe even if we do not have a body.
if (Callee->isDeclaration() && !isConvertible) return false;
}
- if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
- Callee->isDeclaration())
- return false; // Do not delete arguments unless we have a function body.
-
+ if (Callee->isDeclaration()) {
+ // Do not delete arguments unless we have a function body.
+ if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
+ return false;
+
+ // If the callee is just a declaration, don't change the varargsness of the
+ // call. We don't want to introduce a varargs call where one doesn't
+ // already exist.
+ PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
+ if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
+ return false;
+ }
+
if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
!CallerPAL.isEmpty())
// In this case we have more arguments than the new function type, but we
return false;
}
+
// Okay, we decided that this is a safe thing to do: go ahead and start
- // inserting cast instructions as necessary...
+ // inserting cast instructions as necessary.
std::vector<Value*> Args;
Args.reserve(NumActualArgs);
SmallVector<AttributeWithIndex, 8> attrVec;
AI = CS.arg_begin();
for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
- const Type *ParamTy = FT->getParamType(i);
+ Type *ParamTy = FT->getParamType(i);
if ((*AI)->getType() == ParamTy) {
Args.push_back(*AI);
} else {
Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
false, ParamTy, false);
- Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
+ Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy));
}
// Add any parameter attributes.
} else {
// Add all of the arguments in their promoted form to the arg list.
for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
- const Type *PTy = getPromotedType((*AI)->getType());
+ Type *PTy = getPromotedType((*AI)->getType());
if (PTy != (*AI)->getType()) {
// Must promote to pass through va_arg area!
Instruction::CastOps opcode =
CastInst::getCastOpcode(*AI, false, PTy, false);
- Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
+ Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
} else {
Args.push_back(*AI);
}
Instruction *NC;
if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
- NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
- Args.begin(), Args.end(),
- Caller->getName(), Caller);
+ NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
+ II->getUnwindDest(), Args);
+ NC->takeName(II);
cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
} else {
- NC = CallInst::Create(Callee, Args.begin(), Args.end(),
- Caller->getName(), Caller);
CallInst *CI = cast<CallInst>(Caller);
+ NC = Builder->CreateCall(Callee, Args);
+ NC->takeName(CI);
if (CI->isTailCall())
cast<CallInst>(NC)->setTailCall();
cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
if (!NV->getType()->isVoidTy()) {
Instruction::CastOps opcode =
CastInst::getCastOpcode(NC, false, OldRetTy, false);
- NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
+ NV = NC = CastInst::Create(opcode, NC, OldRetTy);
+ NC->setDebugLoc(Caller->getDebugLoc());
// If this is an invoke instruction, we should insert it after the first
// non-phi, instruction in the normal successor block.
if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
- BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
+ BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
InsertNewInstBefore(NC, *I);
} else {
// Otherwise, it's a call, just insert cast right after the call.
}
if (!Caller->use_empty())
- Caller->replaceAllUsesWith(NV);
-
+ ReplaceInstUsesWith(*Caller, NV);
+
EraseInstFromFunction(*Caller);
return true;
}
-// transformCallThroughTrampoline - Turn a call to a function created by the
-// init_trampoline intrinsic into a direct call to the underlying function.
+// transformCallThroughTrampoline - Turn a call to a function created by
+// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
+// underlying function.
//
-Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
+Instruction *
+InstCombiner::transformCallThroughTrampoline(CallSite CS,
+ IntrinsicInst *Tramp) {
Value *Callee = CS.getCalledValue();
- const PointerType *PTy = cast<PointerType>(Callee->getType());
- const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
+ PointerType *PTy = cast<PointerType>(Callee->getType());
+ FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
const AttrListPtr &Attrs = CS.getAttributes();
// If the call already has the 'nest' attribute somewhere then give up -
if (Attrs.hasAttrSomewhere(Attribute::Nest))
return 0;
- IntrinsicInst *Tramp =
- cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
+ assert(Tramp &&
+ "transformCallThroughTrampoline called with incorrect CallSite.");
Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
- const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
- const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
+ PointerType *NestFPTy = cast<PointerType>(NestF->getType());
+ FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
const AttrListPtr &NestAttrs = NestF->getAttributes();
if (!NestAttrs.isEmpty()) {
unsigned NestIdx = 1;
- const Type *NestTy = 0;
+ Type *NestTy = 0;
Attributes NestAttr = Attribute::None;
// Look for a parameter marked with the 'nest' attribute.
// Add the chain argument and attributes.
Value *NestVal = Tramp->getArgOperand(2);
if (NestVal->getType() != NestTy)
- NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
+ NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
NewArgs.push_back(NestVal);
NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
}
// Handle this by synthesizing a new function type, equal to FTy
// with the chain parameter inserted.
- std::vector<const Type*> NewTypes;
+ std::vector<Type*> NewTypes;
NewTypes.reserve(FTy->getNumParams()+1);
// Insert the chain's type into the list of parameter types, which may
if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
NewCaller = InvokeInst::Create(NewCallee,
II->getNormalDest(), II->getUnwindDest(),
- NewArgs.begin(), NewArgs.end(),
- Caller->getName(), Caller);
+ NewArgs);
cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
} else {
- NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
- Caller->getName(), Caller);
+ NewCaller = CallInst::Create(NewCallee, NewArgs);
if (cast<CallInst>(Caller)->isTailCall())
cast<CallInst>(NewCaller)->setTailCall();
cast<CallInst>(NewCaller)->
setCallingConv(cast<CallInst>(Caller)->getCallingConv());
cast<CallInst>(NewCaller)->setAttributes(NewPAL);
}
- if (!Caller->getType()->isVoidTy())
- Caller->replaceAllUsesWith(NewCaller);
- Caller->eraseFromParent();
- Worklist.Remove(Caller);
- return 0;
+
+ return NewCaller;
}
}