-//===-- ConstantFolding.cpp - Analyze constant folding possibilities ------===//
+//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
//
// 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 family of functions determines the possibility of performing constant
-// folding.
+// This file defines routines for folding instructions into constants.
+//
+// Also, to supplement the basic VMCore ConstantExpr simplifications,
+// this file defines some additional folding routines that can make use of
+// TargetData information. These functions cannot go in VMCore due to library
+// dependency issues.
//
//===----------------------------------------------------------------------===//
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
+#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Target/TargetData.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringMap.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/MathExtras.h"
#include <cerrno>
// Constant Folding internal helper functions
//===----------------------------------------------------------------------===//
+/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
+/// TargetData. This always returns a non-null constant, but it may be a
+/// ConstantExpr if unfoldable.
+static Constant *FoldBitCast(Constant *C, const Type *DestTy,
+ const TargetData &TD) {
+
+ // This only handles casts to vectors currently.
+ const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
+ if (DestVTy == 0)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
+ // vector so the code below can handle it uniformly.
+ if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
+ Constant *Ops = C; // don't take the address of C!
+ return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD);
+ }
+
+ // If this is a bitcast from constant vector -> vector, fold it.
+ ConstantVector *CV = dyn_cast<ConstantVector>(C);
+ if (CV == 0)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // If the element types match, VMCore can fold it.
+ unsigned NumDstElt = DestVTy->getNumElements();
+ unsigned NumSrcElt = CV->getNumOperands();
+ if (NumDstElt == NumSrcElt)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ const Type *SrcEltTy = CV->getType()->getElementType();
+ const Type *DstEltTy = DestVTy->getElementType();
+
+ // Otherwise, we're changing the number of elements in a vector, which
+ // requires endianness information to do the right thing. For example,
+ // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
+ // folds to (little endian):
+ // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
+ // and to (big endian):
+ // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
+
+ // First thing is first. We only want to think about integer here, so if
+ // we have something in FP form, recast it as integer.
+ if (DstEltTy->isFloatingPointTy()) {
+ // Fold to an vector of integers with same size as our FP type.
+ unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
+ const Type *DestIVTy =
+ VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
+ // Recursively handle this integer conversion, if possible.
+ C = FoldBitCast(C, DestIVTy, TD);
+ if (!C) return ConstantExpr::getBitCast(C, DestTy);
+
+ // Finally, VMCore can handle this now that #elts line up.
+ return ConstantExpr::getBitCast(C, DestTy);
+ }
+
+ // Okay, we know the destination is integer, if the input is FP, convert
+ // it to integer first.
+ if (SrcEltTy->isFloatingPointTy()) {
+ unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
+ const Type *SrcIVTy =
+ VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
+ // Ask VMCore to do the conversion now that #elts line up.
+ C = ConstantExpr::getBitCast(C, SrcIVTy);
+ CV = dyn_cast<ConstantVector>(C);
+ if (!CV) // If VMCore wasn't able to fold it, bail out.
+ return C;
+ }
+
+ // Now we know that the input and output vectors are both integer vectors
+ // of the same size, and that their #elements is not the same. Do the
+ // conversion here, which depends on whether the input or output has
+ // more elements.
+ bool isLittleEndian = TD.isLittleEndian();
+
+ SmallVector<Constant*, 32> Result;
+ if (NumDstElt < NumSrcElt) {
+ // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
+ Constant *Zero = Constant::getNullValue(DstEltTy);
+ unsigned Ratio = NumSrcElt/NumDstElt;
+ unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
+ unsigned SrcElt = 0;
+ for (unsigned i = 0; i != NumDstElt; ++i) {
+ // Build each element of the result.
+ Constant *Elt = Zero;
+ unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
+ for (unsigned j = 0; j != Ratio; ++j) {
+ Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
+ if (!Src) // Reject constantexpr elements.
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // Zero extend the element to the right size.
+ Src = ConstantExpr::getZExt(Src, Elt->getType());
+
+ // Shift it to the right place, depending on endianness.
+ Src = ConstantExpr::getShl(Src,
+ ConstantInt::get(Src->getType(), ShiftAmt));
+ ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
+
+ // Mix it in.
+ Elt = ConstantExpr::getOr(Elt, Src);
+ }
+ Result.push_back(Elt);
+ }
+ } else {
+ // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
+ unsigned Ratio = NumDstElt/NumSrcElt;
+ unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
+
+ // Loop over each source value, expanding into multiple results.
+ for (unsigned i = 0; i != NumSrcElt; ++i) {
+ Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
+ if (!Src) // Reject constantexpr elements.
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
+ for (unsigned j = 0; j != Ratio; ++j) {
+ // Shift the piece of the value into the right place, depending on
+ // endianness.
+ Constant *Elt = ConstantExpr::getLShr(Src,
+ ConstantInt::get(Src->getType(), ShiftAmt));
+ ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
+
+ // Truncate and remember this piece.
+ Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
+ }
+ }
+ }
+
+ return ConstantVector::get(Result.data(), Result.size());
+}
+
+
/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
/// from a global, return the global and the constant. Because of
/// constantexprs, this function is recursive.
// Otherwise, add any offset that our operands provide.
gep_type_iterator GTI = gep_type_begin(CE);
- for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i, ++GTI) {
- ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(i));
+ for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
+ i != e; ++i, ++GTI) {
+ ConstantInt *CI = dyn_cast<ConstantInt>(*i);
if (!CI) return false; // Index isn't a simple constant?
if (CI->getZExtValue() == 0) continue; // Not adding anything.
Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
} else {
const SequentialType *SQT = cast<SequentialType>(*GTI);
- Offset += TD.getABITypeSize(SQT->getElementType())*CI->getSExtValue();
+ Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
}
}
return true;
return false;
}
+/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
+/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
+/// pointer to copy results into and BytesLeft is the number of bytes left in
+/// the CurPtr buffer. TD is the target data.
+static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
+ unsigned char *CurPtr, unsigned BytesLeft,
+ const TargetData &TD) {
+ assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
+ "Out of range access");
+
+ // If this element is zero or undefined, we can just return since *CurPtr is
+ // zero initialized.
+ if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
+ return true;
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
+ if (CI->getBitWidth() > 64 ||
+ (CI->getBitWidth() & 7) != 0)
+ return false;
+
+ uint64_t Val = CI->getZExtValue();
+ unsigned IntBytes = unsigned(CI->getBitWidth()/8);
+
+ for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
+ CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
+ ++ByteOffset;
+ }
+ return true;
+ }
+
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
+ if (CFP->getType()->isDoubleTy()) {
+ C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
+ return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
+ }
+ if (CFP->getType()->isFloatTy()){
+ C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
+ return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
+ }
+ return false;
+ }
+
+ if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
+ const StructLayout *SL = TD.getStructLayout(CS->getType());
+ unsigned Index = SL->getElementContainingOffset(ByteOffset);
+ uint64_t CurEltOffset = SL->getElementOffset(Index);
+ ByteOffset -= CurEltOffset;
+
+ while (1) {
+ // If the element access is to the element itself and not to tail padding,
+ // read the bytes from the element.
+ uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
+
+ if (ByteOffset < EltSize &&
+ !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
+ BytesLeft, TD))
+ return false;
+
+ ++Index;
+
+ // Check to see if we read from the last struct element, if so we're done.
+ if (Index == CS->getType()->getNumElements())
+ return true;
+
+ // If we read all of the bytes we needed from this element we're done.
+ uint64_t NextEltOffset = SL->getElementOffset(Index);
+
+ if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
+ return true;
+
+ // Move to the next element of the struct.
+ CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
+ BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
+ ByteOffset = 0;
+ CurEltOffset = NextEltOffset;
+ }
+ // not reached.
+ }
+
+ if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
+ uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
+ uint64_t Index = ByteOffset / EltSize;
+ uint64_t Offset = ByteOffset - Index * EltSize;
+ for (; Index != CA->getType()->getNumElements(); ++Index) {
+ if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
+ BytesLeft, TD))
+ return false;
+ if (EltSize >= BytesLeft)
+ return true;
+
+ Offset = 0;
+ BytesLeft -= EltSize;
+ CurPtr += EltSize;
+ }
+ return true;
+ }
+
+ if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
+ uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
+ uint64_t Index = ByteOffset / EltSize;
+ uint64_t Offset = ByteOffset - Index * EltSize;
+ for (; Index != CV->getType()->getNumElements(); ++Index) {
+ if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
+ BytesLeft, TD))
+ return false;
+ if (EltSize >= BytesLeft)
+ return true;
+
+ Offset = 0;
+ BytesLeft -= EltSize;
+ CurPtr += EltSize;
+ }
+ return true;
+ }
+
+ // Otherwise, unknown initializer type.
+ return false;
+}
+
+static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
+ const TargetData &TD) {
+ const Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
+ const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
+
+ // If this isn't an integer load we can't fold it directly.
+ if (!IntType) {
+ // If this is a float/double load, we can try folding it as an int32/64 load
+ // and then bitcast the result. This can be useful for union cases. Note
+ // that address spaces don't matter here since we're not going to result in
+ // an actual new load.
+ const Type *MapTy;
+ if (LoadTy->isFloatTy())
+ MapTy = Type::getInt32PtrTy(C->getContext());
+ else if (LoadTy->isDoubleTy())
+ MapTy = Type::getInt64PtrTy(C->getContext());
+ else if (LoadTy->isVectorTy()) {
+ MapTy = IntegerType::get(C->getContext(),
+ TD.getTypeAllocSizeInBits(LoadTy));
+ MapTy = PointerType::getUnqual(MapTy);
+ } else
+ return 0;
+
+ C = FoldBitCast(C, MapTy, TD);
+ if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
+ return FoldBitCast(Res, LoadTy, TD);
+ return 0;
+ }
+
+ unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
+ if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
+
+ GlobalValue *GVal;
+ int64_t Offset;
+ if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
+ return 0;
+
+ GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
+ if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
+ !GV->getInitializer()->getType()->isSized())
+ return 0;
+
+ // If we're loading off the beginning of the global, some bytes may be valid,
+ // but we don't try to handle this.
+ if (Offset < 0) return 0;
+
+ // If we're not accessing anything in this constant, the result is undefined.
+ if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
+ return UndefValue::get(IntType);
+
+ unsigned char RawBytes[32] = {0};
+ if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
+ BytesLoaded, TD))
+ return 0;
+
+ APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
+ for (unsigned i = 1; i != BytesLoaded; ++i) {
+ ResultVal <<= 8;
+ ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]);
+ }
+
+ return ConstantInt::get(IntType->getContext(), ResultVal);
+}
+
+/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
+/// produce if it is constant and determinable. If this is not determinable,
+/// return null.
+Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
+ const TargetData *TD) {
+ // First, try the easy cases:
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
+ if (GV->isConstant() && GV->hasDefinitiveInitializer())
+ return GV->getInitializer();
+
+ // If the loaded value isn't a constant expr, we can't handle it.
+ ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
+ if (!CE) return 0;
+
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
+ if (GV->isConstant() && GV->hasDefinitiveInitializer())
+ if (Constant *V =
+ ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
+ return V;
+ }
+
+ // Instead of loading constant c string, use corresponding integer value
+ // directly if string length is small enough.
+ std::string Str;
+ if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
+ unsigned StrLen = Str.length();
+ const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
+ unsigned NumBits = Ty->getPrimitiveSizeInBits();
+ // Replace LI with immediate integer store.
+ if ((NumBits >> 3) == StrLen + 1) {
+ APInt StrVal(NumBits, 0);
+ APInt SingleChar(NumBits, 0);
+ if (TD->isLittleEndian()) {
+ for (signed i = StrLen-1; i >= 0; i--) {
+ SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ } else {
+ for (unsigned i = 0; i < StrLen; i++) {
+ SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ // Append NULL at the end.
+ SingleChar = 0;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ return ConstantInt::get(CE->getContext(), StrVal);
+ }
+ }
+
+ // If this load comes from anywhere in a constant global, and if the global
+ // is all undef or zero, we know what it loads.
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){
+ if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
+ const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
+ if (GV->getInitializer()->isNullValue())
+ return Constant::getNullValue(ResTy);
+ if (isa<UndefValue>(GV->getInitializer()))
+ return UndefValue::get(ResTy);
+ }
+ }
+
+ // Try hard to fold loads from bitcasted strange and non-type-safe things. We
+ // currently don't do any of this for big endian systems. It can be
+ // generalized in the future if someone is interested.
+ if (TD && TD->isLittleEndian())
+ return FoldReinterpretLoadFromConstPtr(CE, *TD);
+ return 0;
+}
+
+static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
+ if (LI->isVolatile()) return 0;
+
+ if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
+ return ConstantFoldLoadFromConstPtr(C, TD);
+
+ return 0;
+}
/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
-/// Attempt to symbolically evaluate the result of a binary operator merging
+/// Attempt to symbolically evaluate the result of a binary operator merging
/// these together. If target data info is available, it is provided as TD,
/// otherwise TD is null.
static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
}
}
- // TODO: Fold icmp setne/seteq as well.
return 0;
}
+/// CastGEPIndices - If array indices are not pointer-sized integers,
+/// explicitly cast them so that they aren't implicitly casted by the
+/// getelementptr.
+static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps,
+ const Type *ResultTy,
+ const TargetData *TD) {
+ if (!TD) return 0;
+ const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
+
+ bool Any = false;
+ SmallVector<Constant*, 32> NewIdxs;
+ for (unsigned i = 1; i != NumOps; ++i) {
+ if ((i == 1 ||
+ !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
+ reinterpret_cast<Value *const *>(Ops+1),
+ i-1))) &&
+ Ops[i]->getType() != IntPtrTy) {
+ Any = true;
+ NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
+ true,
+ IntPtrTy,
+ true),
+ Ops[i], IntPtrTy));
+ } else
+ NewIdxs.push_back(Ops[i]);
+ }
+ if (!Any) return 0;
+
+ Constant *C =
+ ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size());
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
+ return C;
+}
+
/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
/// constant expression, do so.
-static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps,
+static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps,
const Type *ResultTy,
const TargetData *TD) {
Constant *Ptr = Ops[0];
- if (!cast<PointerType>(Ptr->getType())->getElementType()->isSized())
+ if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
return 0;
+
+ unsigned BitWidth =
+ TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext()));
+
+ // If this is a constant expr gep that is effectively computing an
+ // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
+ for (unsigned i = 1; i != NumOps; ++i)
+ if (!isa<ConstantInt>(Ops[i]))
+ return 0;
- if (TD && Ptr->isNullValue()) {
- // If this is a constant expr gep that is effectively computing an
- // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
- bool isFoldableGEP = true;
- for (unsigned i = 1; i != NumOps; ++i)
- if (!isa<ConstantInt>(Ops[i])) {
- isFoldableGEP = false;
+ APInt Offset = APInt(BitWidth,
+ TD->getIndexedOffset(Ptr->getType(),
+ (Value**)Ops+1, NumOps-1));
+ Ptr = cast<Constant>(Ptr->stripPointerCasts());
+
+ // If this is a GEP of a GEP, fold it all into a single GEP.
+ while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
+ SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
+
+ // Do not try the incorporate the sub-GEP if some index is not a number.
+ bool AllConstantInt = true;
+ for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
+ if (!isa<ConstantInt>(NestedOps[i])) {
+ AllConstantInt = false;
break;
}
- if (isFoldableGEP) {
- uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
- (Value**)Ops+1, NumOps-1);
- Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
- return ConstantExpr::getIntToPtr(C, ResultTy);
- }
+ if (!AllConstantInt)
+ break;
+
+ Ptr = cast<Constant>(GEP->getOperand(0));
+ Offset += APInt(BitWidth,
+ TD->getIndexedOffset(Ptr->getType(),
+ (Value**)NestedOps.data(),
+ NestedOps.size()));
+ Ptr = cast<Constant>(Ptr->stripPointerCasts());
}
-
- return 0;
+
+ // If the base value for this address is a literal integer value, fold the
+ // getelementptr to the resulting integer value casted to the pointer type.
+ APInt BasePtr(BitWidth, 0);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
+ if (CE->getOpcode() == Instruction::IntToPtr)
+ if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
+ BasePtr = Base->getValue();
+ BasePtr.zextOrTrunc(BitWidth);
+ }
+ if (Ptr->isNullValue() || BasePtr != 0) {
+ Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
+ return ConstantExpr::getIntToPtr(C, ResultTy);
+ }
+
+ // Otherwise form a regular getelementptr. Recompute the indices so that
+ // we eliminate over-indexing of the notional static type array bounds.
+ // This makes it easy to determine if the getelementptr is "inbounds".
+ // Also, this helps GlobalOpt do SROA on GlobalVariables.
+ const Type *Ty = Ptr->getType();
+ SmallVector<Constant*, 32> NewIdxs;
+ do {
+ if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
+ if (ATy->isPointerTy()) {
+ // The only pointer indexing we'll do is on the first index of the GEP.
+ if (!NewIdxs.empty())
+ break;
+
+ // Only handle pointers to sized types, not pointers to functions.
+ if (!ATy->getElementType()->isSized())
+ return 0;
+ }
+
+ // Determine which element of the array the offset points into.
+ APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
+ if (ElemSize == 0)
+ return 0;
+ APInt NewIdx = Offset.udiv(ElemSize);
+ Offset -= NewIdx * ElemSize;
+ NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Ty->getContext()),
+ NewIdx));
+ Ty = ATy->getElementType();
+ } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
+ // Determine which field of the struct the offset points into. The
+ // getZExtValue is at least as safe as the StructLayout API because we
+ // know the offset is within the struct at this point.
+ const StructLayout &SL = *TD->getStructLayout(STy);
+ unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
+ NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
+ ElIdx));
+ Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
+ Ty = STy->getTypeAtIndex(ElIdx);
+ } else {
+ // We've reached some non-indexable type.
+ break;
+ }
+ } while (Ty != cast<PointerType>(ResultTy)->getElementType());
+
+ // If we haven't used up the entire offset by descending the static
+ // type, then the offset is pointing into the middle of an indivisible
+ // member, so we can't simplify it.
+ if (Offset != 0)
+ return 0;
+
+ // Create a GEP.
+ Constant *C =
+ ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
+ assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
+ "Computed GetElementPtr has unexpected type!");
+
+ // If we ended up indexing a member with a type that doesn't match
+ // the type of what the original indices indexed, add a cast.
+ if (Ty != cast<PointerType>(ResultTy)->getElementType())
+ C = FoldBitCast(C, ResultTy, *TD);
+
+ return C;
}
+
//===----------------------------------------------------------------------===//
// Constant Folding public APIs
//===----------------------------------------------------------------------===//
Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
if (PHINode *PN = dyn_cast<PHINode>(I)) {
if (PN->getNumIncomingValues() == 0)
- return Constant::getNullValue(PN->getType());
+ return UndefValue::get(PN->getType());
Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
if (Result == 0) return 0;
// Scan the operand list, checking to see if they are all constants, if so,
// hand off to ConstantFoldInstOperands.
SmallVector<Constant*, 8> Ops;
- for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
- if (Constant *Op = dyn_cast<Constant>(I->getOperand(i)))
+ for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
+ if (Constant *Op = dyn_cast<Constant>(*i))
Ops.push_back(Op);
else
return 0; // All operands not constant!
if (const CmpInst *CI = dyn_cast<CmpInst>(I))
- return ConstantFoldCompareInstOperands(CI->getPredicate(),
- &Ops[0], Ops.size(), TD);
- else
- return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
- &Ops[0], Ops.size(), TD);
+ return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
+ TD);
+
+ if (const LoadInst *LI = dyn_cast<LoadInst>(I))
+ return ConstantFoldLoadInst(LI, TD);
+
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ Ops.data(), Ops.size(), TD);
+}
+
+/// ConstantFoldConstantExpression - Attempt to fold the constant expression
+/// using the specified TargetData. If successful, the constant result is
+/// result is returned, if not, null is returned.
+Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
+ const TargetData *TD) {
+ SmallVector<Constant*, 8> Ops;
+ for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i) {
+ Constant *NewC = cast<Constant>(*i);
+ // Recursively fold the ConstantExpr's operands.
+ if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
+ NewC = ConstantFoldConstantExpression(NewCE, TD);
+ Ops.push_back(NewC);
+ }
+
+ if (CE->isCompare())
+ return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
+ TD);
+ return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
+ Ops.data(), Ops.size(), TD);
}
/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
/// attempting to fold instructions like loads and stores, which have no
/// constant expression form.
///
+/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
+/// information, due to only being passed an opcode and operands. Constant
+/// folding using this function strips this information.
+///
Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
Constant* const* Ops, unsigned NumOps,
const TargetData *TD) {
switch (Opcode) {
default: return 0;
+ case Instruction::ICmp:
+ case Instruction::FCmp: assert(0 && "Invalid for compares");
case Instruction::Call:
if (Function *F = dyn_cast<Function>(Ops[0]))
if (canConstantFoldCallTo(F))
return ConstantFoldCall(F, Ops+1, NumOps-1);
return 0;
- case Instruction::ICmp:
- case Instruction::FCmp:
- assert(0 &&"This function is invalid for compares: no predicate specified");
case Instruction::PtrToInt:
// If the input is a inttoptr, eliminate the pair. This requires knowing
// the width of a pointer, so it can't be done in ConstantExpr::getCast.
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
if (TD && CE->getOpcode() == Instruction::IntToPtr) {
Constant *Input = CE->getOperand(0);
- unsigned InWidth = Input->getType()->getPrimitiveSizeInBits();
- Constant *Mask =
- ConstantInt::get(APInt::getLowBitsSet(InWidth,
- TD->getPointerSizeInBits()));
- Input = ConstantExpr::getAnd(Input, Mask);
+ unsigned InWidth = Input->getType()->getScalarSizeInBits();
+ if (TD->getPointerSizeInBits() < InWidth) {
+ Constant *Mask =
+ ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
+ TD->getPointerSizeInBits()));
+ Input = ConstantExpr::getAnd(Input, Mask);
+ }
// Do a zext or trunc to get to the dest size.
return ConstantExpr::getIntegerCast(Input, DestTy, false);
}
}
- // FALL THROUGH.
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
case Instruction::IntToPtr:
+ // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
+ // the int size is >= the ptr size. This requires knowing the width of a
+ // pointer, so it can't be done in ConstantExpr::getCast.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
+ if (TD &&
+ TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
+ CE->getOpcode() == Instruction::PtrToInt)
+ return FoldBitCast(CE->getOperand(0), DestTy, *TD);
+
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::SIToFP:
case Instruction::FPToUI:
case Instruction::FPToSI:
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
case Instruction::BitCast:
- return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
+ if (TD)
+ return FoldBitCast(Ops[0], DestTy, *TD);
+ return ConstantExpr::getBitCast(Ops[0], DestTy);
case Instruction::Select:
return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
case Instruction::ExtractElement:
case Instruction::ShuffleVector:
return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
case Instruction::GetElementPtr:
+ if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD))
+ return C;
if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD))
return C;
/// returns a constant expression of the specified operands.
///
Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
- Constant*const * Ops,
- unsigned NumOps,
+ Constant *Ops0, Constant *Ops1,
const TargetData *TD) {
// fold: icmp (inttoptr x), null -> icmp x, 0
// fold: icmp (ptrtoint x), 0 -> icmp x, null
- // fold: icmp (inttoptr x), (inttoptr y) -> icmp x, y
+ // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
// fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
//
// ConstantExpr::getCompare cannot do this, because it doesn't have TD
// around to know if bit truncation is happening.
- if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops[0])) {
- if (TD && Ops[1]->isNullValue()) {
- const Type *IntPtrTy = TD->getIntPtrType();
+ if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
+ if (TD && Ops1->isNullValue()) {
+ const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
if (CE0->getOpcode() == Instruction::IntToPtr) {
// Convert the integer value to the right size to ensure we get the
// proper extension or truncation.
Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
IntPtrTy, false);
- Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
- return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, TD);
+ Constant *Null = Constant::getNullValue(C->getType());
+ return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
}
// Only do this transformation if the int is intptrty in size, otherwise
if (CE0->getOpcode() == Instruction::PtrToInt &&
CE0->getType() == IntPtrTy) {
Constant *C = CE0->getOperand(0);
- Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
- // FIXME!
- return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, TD);
+ Constant *Null = Constant::getNullValue(C->getType());
+ return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
}
}
- if (TD && isa<ConstantExpr>(Ops[1]) &&
- cast<ConstantExpr>(Ops[1])->getOpcode() == CE0->getOpcode()) {
- const Type *IntPtrTy = TD->getIntPtrType();
- // Only do this transformation if the int is intptrty in size, otherwise
- // there is a truncation or extension that we aren't modeling.
- if ((CE0->getOpcode() == Instruction::IntToPtr &&
- CE0->getOperand(0)->getType() == IntPtrTy &&
- CE0->getOperand(1)->getType() == IntPtrTy) ||
- (CE0->getOpcode() == Instruction::PtrToInt &&
- CE0->getType() == IntPtrTy &&
- CE0->getOperand(0)->getType() == CE0->getOperand(1)->getType())) {
- Constant *NewOps[] = {
- CE0->getOperand(0), cast<ConstantExpr>(Ops[1])->getOperand(0)
- };
- return ConstantFoldCompareInstOperands(Predicate, NewOps, 2, TD);
+ if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
+ if (TD && CE0->getOpcode() == CE1->getOpcode()) {
+ const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
+
+ if (CE0->getOpcode() == Instruction::IntToPtr) {
+ // Convert the integer value to the right size to ensure we get the
+ // proper extension or truncation.
+ Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
+ IntPtrTy, false);
+ Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
+ IntPtrTy, false);
+ return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
+ }
+
+ // Only do this transformation if the int is intptrty in size, otherwise
+ // there is a truncation or extension that we aren't modeling.
+ if ((CE0->getOpcode() == Instruction::PtrToInt &&
+ CE0->getType() == IntPtrTy &&
+ CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
+ return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
+ CE1->getOperand(0), TD);
}
}
+
+ // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
+ // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
+ if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
+ CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
+ Constant *LHS =
+ ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
+ Constant *RHS =
+ ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
+ unsigned OpC =
+ Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
+ Constant *Ops[] = { LHS, RHS };
+ return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD);
+ }
}
- return ConstantExpr::getCompare(Predicate, Ops[0], Ops[1]);
+
+ return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
}
C = UndefValue::get(ATy->getElementType());
else
return 0;
- } else if (const VectorType *PTy = dyn_cast<VectorType>(*I)) {
- if (CI->getZExtValue() >= PTy->getNumElements())
+ } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
+ if (CI->getZExtValue() >= VTy->getNumElements())
return 0;
if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
C = CP->getOperand(CI->getZExtValue());
else if (isa<ConstantAggregateZero>(C))
- C = Constant::getNullValue(PTy->getElementType());
+ C = Constant::getNullValue(VTy->getElementType());
else if (isa<UndefValue>(C))
- C = UndefValue::get(PTy->getElementType());
+ C = UndefValue::get(VTy->getElementType());
else
return 0;
} else {
/// canConstantFoldCallTo - Return true if its even possible to fold a call to
/// the specified function.
bool
-llvm::canConstantFoldCallTo(Function *F) {
+llvm::canConstantFoldCallTo(const Function *F) {
switch (F->getIntrinsicID()) {
case Intrinsic::sqrt:
case Intrinsic::powi:
case Intrinsic::ctpop:
case Intrinsic::ctlz:
case Intrinsic::cttz:
+ case Intrinsic::uadd_with_overflow:
+ case Intrinsic::usub_with_overflow:
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::ssub_with_overflow:
+ case Intrinsic::convert_from_fp16:
+ case Intrinsic::convert_to_fp16:
return true;
- default: break;
+ default:
+ return false;
+ case 0: break;
}
- const ValueName *NameVal = F->getValueName();
- if (NameVal == 0) return false;
- const char *Str = NameVal->getKeyData();
- unsigned Len = NameVal->getKeyLength();
+ if (!F->hasName()) return false;
+ StringRef Name = F->getName();
// In these cases, the check of the length is required. We don't want to
// return true for a name like "cos\0blah" which strcmp would return equal to
// "cos", but has length 8.
- switch (Str[0]) {
+ switch (Name[0]) {
default: return false;
case 'a':
- if (Len == 4)
- return !strcmp(Str, "acos") || !strcmp(Str, "asin") ||
- !strcmp(Str, "atan");
- else if (Len == 5)
- return !strcmp(Str, "atan2");
- return false;
+ return Name == "acos" || Name == "asin" ||
+ Name == "atan" || Name == "atan2";
case 'c':
- if (Len == 3)
- return !strcmp(Str, "cos");
- else if (Len == 4)
- return !strcmp(Str, "ceil") || !strcmp(Str, "cosf") ||
- !strcmp(Str, "cosh");
- return false;
+ return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
case 'e':
- if (Len == 3)
- return !strcmp(Str, "exp");
- return false;
+ return Name == "exp";
case 'f':
- if (Len == 4)
- return !strcmp(Str, "fabs") || !strcmp(Str, "fmod");
- else if (Len == 5)
- return !strcmp(Str, "floor");
- return false;
- break;
+ return Name == "fabs" || Name == "fmod" || Name == "floor";
case 'l':
- if (Len == 3 && !strcmp(Str, "log"))
- return true;
- if (Len == 5 && !strcmp(Str, "log10"))
- return true;
- return false;
+ return Name == "log" || Name == "log10";
case 'p':
- if (Len == 3 && !strcmp(Str, "pow"))
- return true;
- return false;
+ return Name == "pow";
case 's':
- if (Len == 3)
- return !strcmp(Str, "sin");
- if (Len == 4)
- return !strcmp(Str, "sinh") || !strcmp(Str, "sqrt");
- if (Len == 5)
- return !strcmp(Str, "sqrtf");
- return false;
+ return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
+ Name == "sinf" || Name == "sqrtf";
case 't':
- if (Len == 3 && !strcmp(Str, "tan"))
- return true;
- else if (Len == 4 && !strcmp(Str, "tanh"))
- return true;
- return false;
+ return Name == "tan" || Name == "tanh";
}
}
const Type *Ty) {
errno = 0;
V = NativeFP(V);
- if (errno == 0) {
- if (Ty==Type::FloatTy)
- return ConstantFP::get(Ty, APFloat((float)V));
- else if (Ty==Type::DoubleTy)
- return ConstantFP::get(Ty, APFloat(V));
- else
- assert(0);
+ if (errno != 0) {
+ errno = 0;
+ return 0;
}
- errno = 0;
- return 0;
+
+ if (Ty->isFloatTy())
+ return ConstantFP::get(Ty->getContext(), APFloat((float)V));
+ if (Ty->isDoubleTy())
+ return ConstantFP::get(Ty->getContext(), APFloat(V));
+ llvm_unreachable("Can only constant fold float/double");
+ return 0; // dummy return to suppress warning
}
static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
- double V, double W,
- const Type *Ty) {
+ double V, double W, const Type *Ty) {
errno = 0;
V = NativeFP(V, W);
- if (errno == 0) {
- if (Ty==Type::FloatTy)
- return ConstantFP::get(Ty, APFloat((float)V));
- else if (Ty==Type::DoubleTy)
- return ConstantFP::get(Ty, APFloat(V));
- else
- assert(0);
+ if (errno != 0) {
+ errno = 0;
+ return 0;
}
- errno = 0;
- return 0;
+
+ if (Ty->isFloatTy())
+ return ConstantFP::get(Ty->getContext(), APFloat((float)V));
+ if (Ty->isDoubleTy())
+ return ConstantFP::get(Ty->getContext(), APFloat(V));
+ llvm_unreachable("Can only constant fold float/double");
+ return 0; // dummy return to suppress warning
}
/// ConstantFoldCall - Attempt to constant fold a call to the specified function
/// with the specified arguments, returning null if unsuccessful.
-
Constant *
llvm::ConstantFoldCall(Function *F,
- Constant* const* Operands, unsigned NumOperands) {
- const ValueName *NameVal = F->getValueName();
- if (NameVal == 0) return 0;
- const char *Str = NameVal->getKeyData();
- unsigned Len = NameVal->getKeyLength();
-
+ Constant *const *Operands, unsigned NumOperands) {
+ if (!F->hasName()) return 0;
+ StringRef Name = F->getName();
+
const Type *Ty = F->getReturnType();
if (NumOperands == 1) {
if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
- if (Ty!=Type::FloatTy && Ty!=Type::DoubleTy)
+ if (Name == "llvm.convert.to.fp16") {
+ APFloat Val(Op->getValueAPF());
+
+ bool lost = false;
+ Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
+
+ return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
+ }
+
+ if (!Ty->isFloatTy() && !Ty->isDoubleTy())
return 0;
/// Currently APFloat versions of these functions do not exist, so we use
/// the host native double versions. Float versions are not called
/// directly but for all these it is true (float)(f((double)arg)) ==
/// f(arg). Long double not supported yet.
- double V = Ty==Type::FloatTy ? (double)Op->getValueAPF().convertToFloat():
+ double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
Op->getValueAPF().convertToDouble();
- switch (Str[0]) {
+ switch (Name[0]) {
case 'a':
- if (Len == 4 && !strcmp(Str, "acos"))
+ if (Name == "acos")
return ConstantFoldFP(acos, V, Ty);
- else if (Len == 4 && !strcmp(Str, "asin"))
+ else if (Name == "asin")
return ConstantFoldFP(asin, V, Ty);
- else if (Len == 4 && !strcmp(Str, "atan"))
+ else if (Name == "atan")
return ConstantFoldFP(atan, V, Ty);
break;
case 'c':
- if (Len == 4 && !strcmp(Str, "ceil"))
+ if (Name == "ceil")
return ConstantFoldFP(ceil, V, Ty);
- else if (Len == 3 && !strcmp(Str, "cos"))
+ else if (Name == "cos")
return ConstantFoldFP(cos, V, Ty);
- else if (Len == 4 && !strcmp(Str, "cosh"))
+ else if (Name == "cosh")
return ConstantFoldFP(cosh, V, Ty);
+ else if (Name == "cosf")
+ return ConstantFoldFP(cos, V, Ty);
break;
case 'e':
- if (Len == 3 && !strcmp(Str, "exp"))
+ if (Name == "exp")
return ConstantFoldFP(exp, V, Ty);
break;
case 'f':
- if (Len == 4 && !strcmp(Str, "fabs"))
+ if (Name == "fabs")
return ConstantFoldFP(fabs, V, Ty);
- else if (Len == 5 && !strcmp(Str, "floor"))
+ else if (Name == "floor")
return ConstantFoldFP(floor, V, Ty);
break;
case 'l':
- if (Len == 3 && !strcmp(Str, "log") && V > 0)
+ if (Name == "log" && V > 0)
return ConstantFoldFP(log, V, Ty);
- else if (Len == 5 && !strcmp(Str, "log10") && V > 0)
+ else if (Name == "log10" && V > 0)
return ConstantFoldFP(log10, V, Ty);
- else if (!strcmp(Str, "llvm.sqrt.f32") ||
- !strcmp(Str, "llvm.sqrt.f64")) {
+ else if (Name == "llvm.sqrt.f32" ||
+ Name == "llvm.sqrt.f64") {
if (V >= -0.0)
return ConstantFoldFP(sqrt, V, Ty);
else // Undefined
- return ConstantFP::get(Ty, Ty==Type::FloatTy ? APFloat(0.0f) :
- APFloat(0.0));
+ return Constant::getNullValue(Ty);
}
break;
case 's':
- if (Len == 3 && !strcmp(Str, "sin"))
+ if (Name == "sin")
return ConstantFoldFP(sin, V, Ty);
- else if (Len == 4 && !strcmp(Str, "sinh"))
+ else if (Name == "sinh")
return ConstantFoldFP(sinh, V, Ty);
- else if (Len == 4 && !strcmp(Str, "sqrt") && V >= 0)
+ else if (Name == "sqrt" && V >= 0)
return ConstantFoldFP(sqrt, V, Ty);
- else if (Len == 5 && !strcmp(Str, "sqrtf") && V >= 0)
+ else if (Name == "sqrtf" && V >= 0)
return ConstantFoldFP(sqrt, V, Ty);
+ else if (Name == "sinf")
+ return ConstantFoldFP(sin, V, Ty);
break;
case 't':
- if (Len == 3 && !strcmp(Str, "tan"))
+ if (Name == "tan")
return ConstantFoldFP(tan, V, Ty);
- else if (Len == 4 && !strcmp(Str, "tanh"))
+ else if (Name == "tanh")
return ConstantFoldFP(tanh, V, Ty);
break;
default:
break;
}
- } else if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
- if (Len > 11 && !memcmp(Str, "llvm.bswap", 10))
- return ConstantInt::get(Op->getValue().byteSwap());
- else if (Len > 11 && !memcmp(Str, "llvm.ctpop", 10))
+ return 0;
+ }
+
+
+ if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
+ if (Name.startswith("llvm.bswap"))
+ return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
+ else if (Name.startswith("llvm.ctpop"))
return ConstantInt::get(Ty, Op->getValue().countPopulation());
- else if (Len > 10 && !memcmp(Str, "llvm.cttz", 9))
+ else if (Name.startswith("llvm.cttz"))
return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
- else if (Len > 10 && !memcmp(Str, "llvm.ctlz", 9))
+ else if (Name.startswith("llvm.ctlz"))
return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
+ else if (Name == "llvm.convert.from.fp16") {
+ APFloat Val(Op->getValue());
+
+ bool lost = false;
+ APFloat::opStatus status =
+ Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
+
+ // Conversion is always precise.
+ status = status;
+ assert(status == APFloat::opOK && !lost &&
+ "Precision lost during fp16 constfolding");
+
+ return ConstantFP::get(F->getContext(), Val);
+ }
+ return 0;
+ }
+
+ if (isa<UndefValue>(Operands[0])) {
+ if (Name.startswith("llvm.bswap"))
+ return Operands[0];
+ return 0;
}
- } else if (NumOperands == 2) {
+
+ return 0;
+ }
+
+ if (NumOperands == 2) {
if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
- if (Ty!=Type::FloatTy && Ty!=Type::DoubleTy)
+ if (!Ty->isFloatTy() && !Ty->isDoubleTy())
return 0;
- double Op1V = Ty==Type::FloatTy ?
- (double)Op1->getValueAPF().convertToFloat():
+ double Op1V = Ty->isFloatTy() ?
+ (double)Op1->getValueAPF().convertToFloat() :
Op1->getValueAPF().convertToDouble();
if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
- double Op2V = Ty==Type::FloatTy ?
+ if (Op2->getType() != Op1->getType())
+ return 0;
+
+ double Op2V = Ty->isFloatTy() ?
(double)Op2->getValueAPF().convertToFloat():
Op2->getValueAPF().convertToDouble();
- if (Len == 3 && !strcmp(Str, "pow")) {
+ if (Name == "pow")
return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
- } else if (Len == 4 && !strcmp(Str, "fmod")) {
+ if (Name == "fmod")
return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
- } else if (Len == 5 && !strcmp(Str, "atan2")) {
+ if (Name == "atan2")
return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
- }
} else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
- if (!strcmp(Str, "llvm.powi.f32")) {
- return ConstantFP::get(Ty, APFloat((float)std::pow((float)Op1V,
- (int)Op2C->getZExtValue())));
- } else if (!strcmp(Str, "llvm.powi.f64")) {
- return ConstantFP::get(Ty, APFloat((double)std::pow((double)Op1V,
- (int)Op2C->getZExtValue())));
+ if (Name == "llvm.powi.f32")
+ return ConstantFP::get(F->getContext(),
+ APFloat((float)std::pow((float)Op1V,
+ (int)Op2C->getZExtValue())));
+ if (Name == "llvm.powi.f64")
+ return ConstantFP::get(F->getContext(),
+ APFloat((double)std::pow((double)Op1V,
+ (int)Op2C->getZExtValue())));
+ }
+ return 0;
+ }
+
+
+ if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
+ if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
+ switch (F->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::uadd_with_overflow: {
+ Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
+ Constant *Ops[] = {
+ Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow.
+ };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::usub_with_overflow: {
+ Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
+ Constant *Ops[] = {
+ Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow.
+ };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::sadd_with_overflow: {
+ Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
+ Constant *Overflow = ConstantExpr::getSelect(
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT,
+ ConstantInt::get(Op1->getType(), 0), Op1),
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2),
+ ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow.
+
+ Constant *Ops[] = { Res, Overflow };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::ssub_with_overflow: {
+ Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
+ Constant *Overflow = ConstantExpr::getSelect(
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT,
+ ConstantInt::get(Op2->getType(), 0), Op2),
+ ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1),
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow.
+
+ Constant *Ops[] = { Res, Overflow };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
}
}
+
+ return 0;
}
+ return 0;
}
return 0;
}