1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines routines for folding instructions into constants.
12 // Also, to supplement the basic IR ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // DataLayout information. These functions cannot go in IR due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringMap.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/Config/config.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Target/TargetLibraryInfo.h"
46 //===----------------------------------------------------------------------===//
47 // Constant Folding internal helper functions
48 //===----------------------------------------------------------------------===//
50 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
51 /// DataLayout. This always returns a non-null constant, but it may be a
52 /// ConstantExpr if unfoldable.
53 static Constant *FoldBitCast(Constant *C, Type *DestTy,
54 const DataLayout &TD) {
55 // Catch the obvious splat cases.
56 if (C->isNullValue() && !DestTy->isX86_MMXTy())
57 return Constant::getNullValue(DestTy);
58 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
59 return Constant::getAllOnesValue(DestTy);
61 // Handle a vector->integer cast.
62 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
63 VectorType *VTy = dyn_cast<VectorType>(C->getType());
65 return ConstantExpr::getBitCast(C, DestTy);
67 unsigned NumSrcElts = VTy->getNumElements();
68 Type *SrcEltTy = VTy->getElementType();
70 // If the vector is a vector of floating point, convert it to vector of int
71 // to simplify things.
72 if (SrcEltTy->isFloatingPointTy()) {
73 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
75 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
76 // Ask IR to do the conversion now that #elts line up.
77 C = ConstantExpr::getBitCast(C, SrcIVTy);
80 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
82 return ConstantExpr::getBitCast(C, DestTy);
84 // Now that we know that the input value is a vector of integers, just shift
85 // and insert them into our result.
86 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
87 APInt Result(IT->getBitWidth(), 0);
88 for (unsigned i = 0; i != NumSrcElts; ++i) {
90 if (TD.isLittleEndian())
91 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
93 Result |= CDV->getElementAsInteger(i);
96 return ConstantInt::get(IT, Result);
99 // The code below only handles casts to vectors currently.
100 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
102 return ConstantExpr::getBitCast(C, DestTy);
104 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
105 // vector so the code below can handle it uniformly.
106 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
107 Constant *Ops = C; // don't take the address of C!
108 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
111 // If this is a bitcast from constant vector -> vector, fold it.
112 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
113 return ConstantExpr::getBitCast(C, DestTy);
115 // If the element types match, IR can fold it.
116 unsigned NumDstElt = DestVTy->getNumElements();
117 unsigned NumSrcElt = C->getType()->getVectorNumElements();
118 if (NumDstElt == NumSrcElt)
119 return ConstantExpr::getBitCast(C, DestTy);
121 Type *SrcEltTy = C->getType()->getVectorElementType();
122 Type *DstEltTy = DestVTy->getElementType();
124 // Otherwise, we're changing the number of elements in a vector, which
125 // requires endianness information to do the right thing. For example,
126 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
127 // folds to (little endian):
128 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
129 // and to (big endian):
130 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
132 // First thing is first. We only want to think about integer here, so if
133 // we have something in FP form, recast it as integer.
134 if (DstEltTy->isFloatingPointTy()) {
135 // Fold to an vector of integers with same size as our FP type.
136 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
138 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
139 // Recursively handle this integer conversion, if possible.
140 C = FoldBitCast(C, DestIVTy, TD);
142 // Finally, IR can handle this now that #elts line up.
143 return ConstantExpr::getBitCast(C, DestTy);
146 // Okay, we know the destination is integer, if the input is FP, convert
147 // it to integer first.
148 if (SrcEltTy->isFloatingPointTy()) {
149 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
151 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
152 // Ask IR to do the conversion now that #elts line up.
153 C = ConstantExpr::getBitCast(C, SrcIVTy);
154 // If IR wasn't able to fold it, bail out.
155 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
156 !isa<ConstantDataVector>(C))
160 // Now we know that the input and output vectors are both integer vectors
161 // of the same size, and that their #elements is not the same. Do the
162 // conversion here, which depends on whether the input or output has
164 bool isLittleEndian = TD.isLittleEndian();
166 SmallVector<Constant*, 32> Result;
167 if (NumDstElt < NumSrcElt) {
168 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
169 Constant *Zero = Constant::getNullValue(DstEltTy);
170 unsigned Ratio = NumSrcElt/NumDstElt;
171 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
173 for (unsigned i = 0; i != NumDstElt; ++i) {
174 // Build each element of the result.
175 Constant *Elt = Zero;
176 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
177 for (unsigned j = 0; j != Ratio; ++j) {
178 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
179 if (!Src) // Reject constantexpr elements.
180 return ConstantExpr::getBitCast(C, DestTy);
182 // Zero extend the element to the right size.
183 Src = ConstantExpr::getZExt(Src, Elt->getType());
185 // Shift it to the right place, depending on endianness.
186 Src = ConstantExpr::getShl(Src,
187 ConstantInt::get(Src->getType(), ShiftAmt));
188 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
191 Elt = ConstantExpr::getOr(Elt, Src);
193 Result.push_back(Elt);
195 return ConstantVector::get(Result);
198 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
199 unsigned Ratio = NumDstElt/NumSrcElt;
200 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
202 // Loop over each source value, expanding into multiple results.
203 for (unsigned i = 0; i != NumSrcElt; ++i) {
204 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
205 if (!Src) // Reject constantexpr elements.
206 return ConstantExpr::getBitCast(C, DestTy);
208 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
209 for (unsigned j = 0; j != Ratio; ++j) {
210 // Shift the piece of the value into the right place, depending on
212 Constant *Elt = ConstantExpr::getLShr(Src,
213 ConstantInt::get(Src->getType(), ShiftAmt));
214 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
216 // Truncate and remember this piece.
217 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
221 return ConstantVector::get(Result);
225 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
226 /// from a global, return the global and the constant. Because of
227 /// constantexprs, this function is recursive.
228 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
229 APInt &Offset, const DataLayout &TD) {
230 // Trivial case, constant is the global.
231 if ((GV = dyn_cast<GlobalValue>(C))) {
232 unsigned BitWidth = TD.getPointerTypeSizeInBits(GV->getType());
233 Offset = APInt(BitWidth, 0);
237 // Otherwise, if this isn't a constant expr, bail out.
238 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
239 if (!CE) return false;
241 // Look through ptr->int and ptr->ptr casts.
242 if (CE->getOpcode() == Instruction::PtrToInt ||
243 CE->getOpcode() == Instruction::BitCast)
244 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
246 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
247 GEPOperator *GEP = dyn_cast<GEPOperator>(CE);
251 unsigned BitWidth = TD.getPointerTypeSizeInBits(GEP->getType());
252 APInt TmpOffset(BitWidth, 0);
254 // If the base isn't a global+constant, we aren't either.
255 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, TD))
258 // Otherwise, add any offset that our operands provide.
259 if (!GEP->accumulateConstantOffset(TD, TmpOffset))
266 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
267 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
268 /// pointer to copy results into and BytesLeft is the number of bytes left in
269 /// the CurPtr buffer. TD is the target data.
270 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
271 unsigned char *CurPtr, unsigned BytesLeft,
272 const DataLayout &TD) {
273 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
274 "Out of range access");
276 // If this element is zero or undefined, we can just return since *CurPtr is
278 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
281 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
282 if (CI->getBitWidth() > 64 ||
283 (CI->getBitWidth() & 7) != 0)
286 uint64_t Val = CI->getZExtValue();
287 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
289 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
291 if (!TD.isLittleEndian())
292 n = IntBytes - n - 1;
293 CurPtr[i] = (unsigned char)(Val >> (n * 8));
299 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
300 if (CFP->getType()->isDoubleTy()) {
301 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
302 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
304 if (CFP->getType()->isFloatTy()){
305 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
306 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
308 if (CFP->getType()->isHalfTy()){
309 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), TD);
310 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
315 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
316 const StructLayout *SL = TD.getStructLayout(CS->getType());
317 unsigned Index = SL->getElementContainingOffset(ByteOffset);
318 uint64_t CurEltOffset = SL->getElementOffset(Index);
319 ByteOffset -= CurEltOffset;
322 // If the element access is to the element itself and not to tail padding,
323 // read the bytes from the element.
324 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
326 if (ByteOffset < EltSize &&
327 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
333 // Check to see if we read from the last struct element, if so we're done.
334 if (Index == CS->getType()->getNumElements())
337 // If we read all of the bytes we needed from this element we're done.
338 uint64_t NextEltOffset = SL->getElementOffset(Index);
340 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
343 // Move to the next element of the struct.
344 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
345 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
347 CurEltOffset = NextEltOffset;
352 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
353 isa<ConstantDataSequential>(C)) {
354 Type *EltTy = C->getType()->getSequentialElementType();
355 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
356 uint64_t Index = ByteOffset / EltSize;
357 uint64_t Offset = ByteOffset - Index * EltSize;
359 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
360 NumElts = AT->getNumElements();
362 NumElts = C->getType()->getVectorNumElements();
364 for (; Index != NumElts; ++Index) {
365 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
369 uint64_t BytesWritten = EltSize - Offset;
370 assert(BytesWritten <= EltSize && "Not indexing into this element?");
371 if (BytesWritten >= BytesLeft)
375 BytesLeft -= BytesWritten;
376 CurPtr += BytesWritten;
381 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
382 if (CE->getOpcode() == Instruction::IntToPtr &&
383 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getType())) {
384 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
389 // Otherwise, unknown initializer type.
393 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
394 const DataLayout &TD) {
395 PointerType *PTy = cast<PointerType>(C->getType());
396 Type *LoadTy = PTy->getElementType();
397 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
399 // If this isn't an integer load we can't fold it directly.
401 unsigned AS = PTy->getAddressSpace();
403 // If this is a float/double load, we can try folding it as an int32/64 load
404 // and then bitcast the result. This can be useful for union cases. Note
405 // that address spaces don't matter here since we're not going to result in
406 // an actual new load.
408 if (LoadTy->isHalfTy())
409 MapTy = Type::getInt16PtrTy(C->getContext(), AS);
410 else if (LoadTy->isFloatTy())
411 MapTy = Type::getInt32PtrTy(C->getContext(), AS);
412 else if (LoadTy->isDoubleTy())
413 MapTy = Type::getInt64PtrTy(C->getContext(), AS);
414 else if (LoadTy->isVectorTy()) {
415 MapTy = PointerType::getIntNPtrTy(C->getContext(),
416 TD.getTypeAllocSizeInBits(LoadTy),
421 C = FoldBitCast(C, MapTy, TD);
422 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
423 return FoldBitCast(Res, LoadTy, TD);
427 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
428 if (BytesLoaded > 32 || BytesLoaded == 0)
433 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
436 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
437 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
438 !GV->getInitializer()->getType()->isSized())
441 // If we're loading off the beginning of the global, some bytes may be valid,
442 // but we don't try to handle this.
443 if (Offset.isNegative())
446 // If we're not accessing anything in this constant, the result is undefined.
447 if (Offset.getZExtValue() >=
448 TD.getTypeAllocSize(GV->getInitializer()->getType()))
449 return UndefValue::get(IntType);
451 unsigned char RawBytes[32] = {0};
452 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
456 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
457 if (TD.isLittleEndian()) {
458 ResultVal = RawBytes[BytesLoaded - 1];
459 for (unsigned i = 1; i != BytesLoaded; ++i) {
461 ResultVal |= RawBytes[BytesLoaded - 1 - i];
464 ResultVal = RawBytes[0];
465 for (unsigned i = 1; i != BytesLoaded; ++i) {
467 ResultVal |= RawBytes[i];
471 return ConstantInt::get(IntType->getContext(), ResultVal);
474 static Constant *ConstantFoldLoadThroughBitcast(ConstantExpr *CE,
475 const DataLayout *DL) {
478 auto *DestPtrTy = dyn_cast<PointerType>(CE->getType());
481 Type *DestTy = DestPtrTy->getElementType();
483 Constant *C = ConstantFoldLoadFromConstPtr(CE->getOperand(0), DL);
488 Type *SrcTy = C->getType();
490 // If the type sizes are the same and a cast is legal, just directly
491 // cast the constant.
492 if (DL->getTypeSizeInBits(DestTy) == DL->getTypeSizeInBits(SrcTy)) {
493 Instruction::CastOps Cast = Instruction::BitCast;
494 // If we are going from a pointer to int or vice versa, we spell the cast
496 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
497 Cast = Instruction::IntToPtr;
498 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
499 Cast = Instruction::PtrToInt;
501 if (CastInst::castIsValid(Cast, C, DestTy))
502 return ConstantExpr::getCast(Cast, C, DestTy);
505 // If this isn't an aggregate type, there is nothing we can do to drill down
506 // and find a bitcastable constant.
507 if (!SrcTy->isAggregateType())
510 // We're simulating a load through a pointer that was bitcast to point to
511 // a different type, so we can try to walk down through the initial
512 // elements of an aggregate to see if some part of th e aggregate is
513 // castable to implement the "load" semantic model.
514 C = C->getAggregateElement(0u);
520 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
521 /// produce if it is constant and determinable. If this is not determinable,
523 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
524 const DataLayout *TD) {
525 // First, try the easy cases:
526 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
527 if (GV->isConstant() && GV->hasDefinitiveInitializer())
528 return GV->getInitializer();
530 // If the loaded value isn't a constant expr, we can't handle it.
531 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
535 if (CE->getOpcode() == Instruction::GetElementPtr) {
536 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
537 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
539 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
545 if (CE->getOpcode() == Instruction::BitCast)
546 if (Constant *LoadedC = ConstantFoldLoadThroughBitcast(CE, TD))
549 // Instead of loading constant c string, use corresponding integer value
550 // directly if string length is small enough.
552 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
553 unsigned StrLen = Str.size();
554 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
555 unsigned NumBits = Ty->getPrimitiveSizeInBits();
556 // Replace load with immediate integer if the result is an integer or fp
558 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
559 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
560 APInt StrVal(NumBits, 0);
561 APInt SingleChar(NumBits, 0);
562 if (TD->isLittleEndian()) {
563 for (signed i = StrLen-1; i >= 0; i--) {
564 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
565 StrVal = (StrVal << 8) | SingleChar;
568 for (unsigned i = 0; i < StrLen; i++) {
569 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
570 StrVal = (StrVal << 8) | SingleChar;
572 // Append NULL at the end.
574 StrVal = (StrVal << 8) | SingleChar;
577 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
578 if (Ty->isFloatingPointTy())
579 Res = ConstantExpr::getBitCast(Res, Ty);
584 // If this load comes from anywhere in a constant global, and if the global
585 // is all undef or zero, we know what it loads.
586 if (GlobalVariable *GV =
587 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
588 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
589 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
590 if (GV->getInitializer()->isNullValue())
591 return Constant::getNullValue(ResTy);
592 if (isa<UndefValue>(GV->getInitializer()))
593 return UndefValue::get(ResTy);
597 // Try hard to fold loads from bitcasted strange and non-type-safe things.
599 return FoldReinterpretLoadFromConstPtr(CE, *TD);
603 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
604 if (LI->isVolatile()) return nullptr;
606 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
607 return ConstantFoldLoadFromConstPtr(C, TD);
612 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
613 /// Attempt to symbolically evaluate the result of a binary operator merging
614 /// these together. If target data info is available, it is provided as DL,
615 /// otherwise DL is null.
616 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
617 Constant *Op1, const DataLayout *DL){
620 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
621 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
625 if (Opc == Instruction::And && DL) {
626 unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType());
627 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
628 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
629 computeKnownBits(Op0, KnownZero0, KnownOne0, DL);
630 computeKnownBits(Op1, KnownZero1, KnownOne1, DL);
631 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
632 // All the bits of Op0 that the 'and' could be masking are already zero.
635 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
636 // All the bits of Op1 that the 'and' could be masking are already zero.
640 APInt KnownZero = KnownZero0 | KnownZero1;
641 APInt KnownOne = KnownOne0 & KnownOne1;
642 if ((KnownZero | KnownOne).isAllOnesValue()) {
643 return ConstantInt::get(Op0->getType(), KnownOne);
647 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
648 // constant. This happens frequently when iterating over a global array.
649 if (Opc == Instruction::Sub && DL) {
650 GlobalValue *GV1, *GV2;
653 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL))
654 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) &&
656 unsigned OpSize = DL->getTypeSizeInBits(Op0->getType());
658 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
659 // PtrToInt may change the bitwidth so we have convert to the right size
661 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
662 Offs2.zextOrTrunc(OpSize));
669 /// CastGEPIndices - If array indices are not pointer-sized integers,
670 /// explicitly cast them so that they aren't implicitly casted by the
672 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
673 Type *ResultTy, const DataLayout *TD,
674 const TargetLibraryInfo *TLI) {
678 Type *IntPtrTy = TD->getIntPtrType(ResultTy);
681 SmallVector<Constant*, 32> NewIdxs;
682 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
684 !isa<StructType>(GetElementPtrInst::getIndexedType(
686 Ops.slice(1, i - 1)))) &&
687 Ops[i]->getType() != IntPtrTy) {
689 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
695 NewIdxs.push_back(Ops[i]);
701 Constant *C = ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
702 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
703 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
710 /// Strip the pointer casts, but preserve the address space information.
711 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
712 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
713 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
714 Ptr = Ptr->stripPointerCasts();
715 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
717 // Preserve the address space number of the pointer.
718 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
719 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
720 OldPtrTy->getAddressSpace());
721 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
726 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
727 /// constant expression, do so.
728 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
729 Type *ResultTy, const DataLayout *TD,
730 const TargetLibraryInfo *TLI) {
731 Constant *Ptr = Ops[0];
732 if (!TD || !Ptr->getType()->getPointerElementType()->isSized() ||
733 !Ptr->getType()->isPointerTy())
736 Type *IntPtrTy = TD->getIntPtrType(Ptr->getType());
737 Type *ResultElementTy = ResultTy->getPointerElementType();
739 // If this is a constant expr gep that is effectively computing an
740 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
741 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
742 if (!isa<ConstantInt>(Ops[i])) {
744 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
745 // "inttoptr (sub (ptrtoint Ptr), V)"
746 if (Ops.size() == 2 && ResultElementTy->isIntegerTy(8)) {
747 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
748 assert((!CE || CE->getType() == IntPtrTy) &&
749 "CastGEPIndices didn't canonicalize index types!");
750 if (CE && CE->getOpcode() == Instruction::Sub &&
751 CE->getOperand(0)->isNullValue()) {
752 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
753 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
754 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
755 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
756 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
763 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
765 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
766 makeArrayRef((Value *const*)
769 Ptr = StripPtrCastKeepAS(Ptr);
771 // If this is a GEP of a GEP, fold it all into a single GEP.
772 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
773 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
775 // Do not try the incorporate the sub-GEP if some index is not a number.
776 bool AllConstantInt = true;
777 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
778 if (!isa<ConstantInt>(NestedOps[i])) {
779 AllConstantInt = false;
785 Ptr = cast<Constant>(GEP->getOperand(0));
786 Offset += APInt(BitWidth,
787 TD->getIndexedOffset(Ptr->getType(), NestedOps));
788 Ptr = StripPtrCastKeepAS(Ptr);
791 // If the base value for this address is a literal integer value, fold the
792 // getelementptr to the resulting integer value casted to the pointer type.
793 APInt BasePtr(BitWidth, 0);
794 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
795 if (CE->getOpcode() == Instruction::IntToPtr) {
796 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
797 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
801 if (Ptr->isNullValue() || BasePtr != 0) {
802 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
803 return ConstantExpr::getIntToPtr(C, ResultTy);
806 // Otherwise form a regular getelementptr. Recompute the indices so that
807 // we eliminate over-indexing of the notional static type array bounds.
808 // This makes it easy to determine if the getelementptr is "inbounds".
809 // Also, this helps GlobalOpt do SROA on GlobalVariables.
810 Type *Ty = Ptr->getType();
811 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
812 SmallVector<Constant *, 32> NewIdxs;
815 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
816 if (ATy->isPointerTy()) {
817 // The only pointer indexing we'll do is on the first index of the GEP.
818 if (!NewIdxs.empty())
821 // Only handle pointers to sized types, not pointers to functions.
822 if (!ATy->getElementType()->isSized())
826 // Determine which element of the array the offset points into.
827 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
829 // The element size is 0. This may be [0 x Ty]*, so just use a zero
830 // index for this level and proceed to the next level to see if it can
831 // accommodate the offset.
832 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
834 // The element size is non-zero divide the offset by the element
835 // size (rounding down), to compute the index at this level.
836 APInt NewIdx = Offset.udiv(ElemSize);
837 Offset -= NewIdx * ElemSize;
838 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
840 Ty = ATy->getElementType();
841 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
842 // If we end up with an offset that isn't valid for this struct type, we
843 // can't re-form this GEP in a regular form, so bail out. The pointer
844 // operand likely went through casts that are necessary to make the GEP
846 const StructLayout &SL = *TD->getStructLayout(STy);
847 if (Offset.uge(SL.getSizeInBytes()))
850 // Determine which field of the struct the offset points into. The
851 // getZExtValue is fine as we've already ensured that the offset is
852 // within the range representable by the StructLayout API.
853 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
854 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
856 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
857 Ty = STy->getTypeAtIndex(ElIdx);
859 // We've reached some non-indexable type.
862 } while (Ty != ResultElementTy);
864 // If we haven't used up the entire offset by descending the static
865 // type, then the offset is pointing into the middle of an indivisible
866 // member, so we can't simplify it.
871 Constant *C = ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
872 assert(C->getType()->getPointerElementType() == Ty &&
873 "Computed GetElementPtr has unexpected type!");
875 // If we ended up indexing a member with a type that doesn't match
876 // the type of what the original indices indexed, add a cast.
877 if (Ty != ResultElementTy)
878 C = FoldBitCast(C, ResultTy, *TD);
885 //===----------------------------------------------------------------------===//
886 // Constant Folding public APIs
887 //===----------------------------------------------------------------------===//
889 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
890 /// If successful, the constant result is returned, if not, null is returned.
891 /// Note that this fails if not all of the operands are constant. Otherwise,
892 /// this function can only fail when attempting to fold instructions like loads
893 /// and stores, which have no constant expression form.
894 Constant *llvm::ConstantFoldInstruction(Instruction *I,
895 const DataLayout *TD,
896 const TargetLibraryInfo *TLI) {
897 // Handle PHI nodes quickly here...
898 if (PHINode *PN = dyn_cast<PHINode>(I)) {
899 Constant *CommonValue = nullptr;
901 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
902 Value *Incoming = PN->getIncomingValue(i);
903 // If the incoming value is undef then skip it. Note that while we could
904 // skip the value if it is equal to the phi node itself we choose not to
905 // because that would break the rule that constant folding only applies if
906 // all operands are constants.
907 if (isa<UndefValue>(Incoming))
909 // If the incoming value is not a constant, then give up.
910 Constant *C = dyn_cast<Constant>(Incoming);
913 // Fold the PHI's operands.
914 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
915 C = ConstantFoldConstantExpression(NewC, TD, TLI);
916 // If the incoming value is a different constant to
917 // the one we saw previously, then give up.
918 if (CommonValue && C != CommonValue)
924 // If we reach here, all incoming values are the same constant or undef.
925 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
928 // Scan the operand list, checking to see if they are all constants, if so,
929 // hand off to ConstantFoldInstOperands.
930 SmallVector<Constant*, 8> Ops;
931 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
932 Constant *Op = dyn_cast<Constant>(*i);
934 return nullptr; // All operands not constant!
936 // Fold the Instruction's operands.
937 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
938 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
943 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
944 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
947 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
948 return ConstantFoldLoadInst(LI, TD);
950 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) {
951 return ConstantExpr::getInsertValue(
952 cast<Constant>(IVI->getAggregateOperand()),
953 cast<Constant>(IVI->getInsertedValueOperand()),
957 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) {
958 return ConstantExpr::getExtractValue(
959 cast<Constant>(EVI->getAggregateOperand()),
963 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
967 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD,
968 const TargetLibraryInfo *TLI,
969 SmallPtrSet<ConstantExpr *, 4> &FoldedOps) {
970 SmallVector<Constant *, 8> Ops;
971 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e;
973 Constant *NewC = cast<Constant>(*i);
974 // Recursively fold the ConstantExpr's operands. If we have already folded
975 // a ConstantExpr, we don't have to process it again.
976 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) {
977 if (FoldedOps.insert(NewCE))
978 NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps);
984 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
986 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
989 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
990 /// using the specified DataLayout. If successful, the constant result is
991 /// result is returned, if not, null is returned.
992 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
993 const DataLayout *TD,
994 const TargetLibraryInfo *TLI) {
995 SmallPtrSet<ConstantExpr *, 4> FoldedOps;
996 return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps);
999 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
1000 /// specified opcode and operands. If successful, the constant result is
1001 /// returned, if not, null is returned. Note that this function can fail when
1002 /// attempting to fold instructions like loads and stores, which have no
1003 /// constant expression form.
1005 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
1006 /// information, due to only being passed an opcode and operands. Constant
1007 /// folding using this function strips this information.
1009 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
1010 ArrayRef<Constant *> Ops,
1011 const DataLayout *TD,
1012 const TargetLibraryInfo *TLI) {
1013 // Handle easy binops first.
1014 if (Instruction::isBinaryOp(Opcode)) {
1015 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) {
1016 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
1020 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
1024 default: return nullptr;
1025 case Instruction::ICmp:
1026 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1027 case Instruction::Call:
1028 if (Function *F = dyn_cast<Function>(Ops.back()))
1029 if (canConstantFoldCallTo(F))
1030 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
1032 case Instruction::PtrToInt:
1033 // If the input is a inttoptr, eliminate the pair. This requires knowing
1034 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1035 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1036 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
1037 Constant *Input = CE->getOperand(0);
1038 unsigned InWidth = Input->getType()->getScalarSizeInBits();
1039 unsigned PtrWidth = TD->getPointerTypeSizeInBits(CE->getType());
1040 if (PtrWidth < InWidth) {
1042 ConstantInt::get(CE->getContext(),
1043 APInt::getLowBitsSet(InWidth, PtrWidth));
1044 Input = ConstantExpr::getAnd(Input, Mask);
1046 // Do a zext or trunc to get to the dest size.
1047 return ConstantExpr::getIntegerCast(Input, DestTy, false);
1050 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1051 case Instruction::IntToPtr:
1052 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1053 // the int size is >= the ptr size and the address spaces are the same.
1054 // This requires knowing the width of a pointer, so it can't be done in
1055 // ConstantExpr::getCast.
1056 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1057 if (TD && CE->getOpcode() == Instruction::PtrToInt) {
1058 Constant *SrcPtr = CE->getOperand(0);
1059 unsigned SrcPtrSize = TD->getPointerTypeSizeInBits(SrcPtr->getType());
1060 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1062 if (MidIntSize >= SrcPtrSize) {
1063 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1064 if (SrcAS == DestTy->getPointerAddressSpace())
1065 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
1070 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1071 case Instruction::Trunc:
1072 case Instruction::ZExt:
1073 case Instruction::SExt:
1074 case Instruction::FPTrunc:
1075 case Instruction::FPExt:
1076 case Instruction::UIToFP:
1077 case Instruction::SIToFP:
1078 case Instruction::FPToUI:
1079 case Instruction::FPToSI:
1080 case Instruction::AddrSpaceCast:
1081 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1082 case Instruction::BitCast:
1084 return FoldBitCast(Ops[0], DestTy, *TD);
1085 return ConstantExpr::getBitCast(Ops[0], DestTy);
1086 case Instruction::Select:
1087 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1088 case Instruction::ExtractElement:
1089 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1090 case Instruction::InsertElement:
1091 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1092 case Instruction::ShuffleVector:
1093 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1094 case Instruction::GetElementPtr:
1095 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
1097 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
1100 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
1104 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
1105 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
1106 /// returns a constant expression of the specified operands.
1108 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1109 Constant *Ops0, Constant *Ops1,
1110 const DataLayout *TD,
1111 const TargetLibraryInfo *TLI) {
1112 // fold: icmp (inttoptr x), null -> icmp x, 0
1113 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1114 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1115 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1117 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
1118 // around to know if bit truncation is happening.
1119 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1120 if (TD && Ops1->isNullValue()) {
1121 if (CE0->getOpcode() == Instruction::IntToPtr) {
1122 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1123 // Convert the integer value to the right size to ensure we get the
1124 // proper extension or truncation.
1125 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1127 Constant *Null = Constant::getNullValue(C->getType());
1128 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1131 // Only do this transformation if the int is intptrty in size, otherwise
1132 // there is a truncation or extension that we aren't modeling.
1133 if (CE0->getOpcode() == Instruction::PtrToInt) {
1134 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1135 if (CE0->getType() == IntPtrTy) {
1136 Constant *C = CE0->getOperand(0);
1137 Constant *Null = Constant::getNullValue(C->getType());
1138 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1143 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1144 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1145 if (CE0->getOpcode() == Instruction::IntToPtr) {
1146 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1148 // Convert the integer value to the right size to ensure we get the
1149 // proper extension or truncation.
1150 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1152 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1154 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1157 // Only do this transformation if the int is intptrty in size, otherwise
1158 // there is a truncation or extension that we aren't modeling.
1159 if (CE0->getOpcode() == Instruction::PtrToInt) {
1160 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1161 if (CE0->getType() == IntPtrTy &&
1162 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1163 return ConstantFoldCompareInstOperands(Predicate,
1173 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1174 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1175 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1176 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1178 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1181 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1184 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1185 Constant *Ops[] = { LHS, RHS };
1186 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1190 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1194 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1195 /// getelementptr constantexpr, return the constant value being addressed by the
1196 /// constant expression, or null if something is funny and we can't decide.
1197 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1199 if (!CE->getOperand(1)->isNullValue())
1200 return nullptr; // Do not allow stepping over the value!
1202 // Loop over all of the operands, tracking down which value we are
1204 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1205 C = C->getAggregateElement(CE->getOperand(i));
1212 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1213 /// indices (with an *implied* zero pointer index that is not in the list),
1214 /// return the constant value being addressed by a virtual load, or null if
1215 /// something is funny and we can't decide.
1216 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1217 ArrayRef<Constant*> Indices) {
1218 // Loop over all of the operands, tracking down which value we are
1220 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1221 C = C->getAggregateElement(Indices[i]);
1229 //===----------------------------------------------------------------------===//
1230 // Constant Folding for Calls
1233 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1234 /// the specified function.
1235 bool llvm::canConstantFoldCallTo(const Function *F) {
1236 switch (F->getIntrinsicID()) {
1237 case Intrinsic::fabs:
1238 case Intrinsic::log:
1239 case Intrinsic::log2:
1240 case Intrinsic::log10:
1241 case Intrinsic::exp:
1242 case Intrinsic::exp2:
1243 case Intrinsic::floor:
1244 case Intrinsic::ceil:
1245 case Intrinsic::sqrt:
1246 case Intrinsic::pow:
1247 case Intrinsic::powi:
1248 case Intrinsic::bswap:
1249 case Intrinsic::ctpop:
1250 case Intrinsic::ctlz:
1251 case Intrinsic::cttz:
1252 case Intrinsic::fma:
1253 case Intrinsic::fmuladd:
1254 case Intrinsic::copysign:
1255 case Intrinsic::round:
1256 case Intrinsic::sadd_with_overflow:
1257 case Intrinsic::uadd_with_overflow:
1258 case Intrinsic::ssub_with_overflow:
1259 case Intrinsic::usub_with_overflow:
1260 case Intrinsic::smul_with_overflow:
1261 case Intrinsic::umul_with_overflow:
1262 case Intrinsic::convert_from_fp16:
1263 case Intrinsic::convert_to_fp16:
1264 case Intrinsic::x86_sse_cvtss2si:
1265 case Intrinsic::x86_sse_cvtss2si64:
1266 case Intrinsic::x86_sse_cvttss2si:
1267 case Intrinsic::x86_sse_cvttss2si64:
1268 case Intrinsic::x86_sse2_cvtsd2si:
1269 case Intrinsic::x86_sse2_cvtsd2si64:
1270 case Intrinsic::x86_sse2_cvttsd2si:
1271 case Intrinsic::x86_sse2_cvttsd2si64:
1280 StringRef Name = F->getName();
1282 // In these cases, the check of the length is required. We don't want to
1283 // return true for a name like "cos\0blah" which strcmp would return equal to
1284 // "cos", but has length 8.
1286 default: return false;
1288 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2";
1290 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1292 return Name == "exp" || Name == "exp2";
1294 return Name == "fabs" || Name == "fmod" || Name == "floor";
1296 return Name == "log" || Name == "log10";
1298 return Name == "pow";
1300 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1301 Name == "sinf" || Name == "sqrtf";
1303 return Name == "tan" || Name == "tanh";
1307 static Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1308 if (Ty->isHalfTy()) {
1311 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1312 return ConstantFP::get(Ty->getContext(), APF);
1314 if (Ty->isFloatTy())
1315 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1316 if (Ty->isDoubleTy())
1317 return ConstantFP::get(Ty->getContext(), APFloat(V));
1318 llvm_unreachable("Can only constant fold half/float/double");
1323 /// llvm_fenv_clearexcept - Clear the floating-point exception state.
1324 static inline void llvm_fenv_clearexcept() {
1325 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1326 feclearexcept(FE_ALL_EXCEPT);
1331 /// llvm_fenv_testexcept - Test if a floating-point exception was raised.
1332 static inline bool llvm_fenv_testexcept() {
1333 int errno_val = errno;
1334 if (errno_val == ERANGE || errno_val == EDOM)
1336 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1337 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1344 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1346 llvm_fenv_clearexcept();
1348 if (llvm_fenv_testexcept()) {
1349 llvm_fenv_clearexcept();
1353 return GetConstantFoldFPValue(V, Ty);
1356 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1357 double V, double W, Type *Ty) {
1358 llvm_fenv_clearexcept();
1360 if (llvm_fenv_testexcept()) {
1361 llvm_fenv_clearexcept();
1365 return GetConstantFoldFPValue(V, Ty);
1368 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1369 /// conversion of a constant floating point. If roundTowardZero is false, the
1370 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1371 /// the behavior of the non-truncating SSE instructions in the default rounding
1372 /// mode. The desired integer type Ty is used to select how many bits are
1373 /// available for the result. Returns null if the conversion cannot be
1374 /// performed, otherwise returns the Constant value resulting from the
1376 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1377 bool roundTowardZero, Type *Ty) {
1378 // All of these conversion intrinsics form an integer of at most 64bits.
1379 unsigned ResultWidth = Ty->getIntegerBitWidth();
1380 assert(ResultWidth <= 64 &&
1381 "Can only constant fold conversions to 64 and 32 bit ints");
1384 bool isExact = false;
1385 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1386 : APFloat::rmNearestTiesToEven;
1387 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1388 /*isSigned=*/true, mode,
1390 if (status != APFloat::opOK && status != APFloat::opInexact)
1392 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1395 static double getValueAsDouble(ConstantFP *Op) {
1396 Type *Ty = Op->getType();
1398 if (Ty->isFloatTy())
1399 return Op->getValueAPF().convertToFloat();
1401 if (Ty->isDoubleTy())
1402 return Op->getValueAPF().convertToDouble();
1405 APFloat APF = Op->getValueAPF();
1406 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1407 return APF.convertToDouble();
1410 static Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID,
1411 Type *Ty, ArrayRef<Constant *> Operands,
1412 const TargetLibraryInfo *TLI) {
1413 if (Operands.size() == 1) {
1414 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1415 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1416 APFloat Val(Op->getValueAPF());
1419 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1421 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1424 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1427 if (IntrinsicID == Intrinsic::round) {
1428 APFloat V = Op->getValueAPF();
1429 V.roundToIntegral(APFloat::rmNearestTiesToAway);
1430 return ConstantFP::get(Ty->getContext(), V);
1433 /// We only fold functions with finite arguments. Folding NaN and inf is
1434 /// likely to be aborted with an exception anyway, and some host libms
1435 /// have known errors raising exceptions.
1436 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1439 /// Currently APFloat versions of these functions do not exist, so we use
1440 /// the host native double versions. Float versions are not called
1441 /// directly but for all these it is true (float)(f((double)arg)) ==
1442 /// f(arg). Long double not supported yet.
1443 double V = getValueAsDouble(Op);
1445 switch (IntrinsicID) {
1447 case Intrinsic::fabs:
1448 return ConstantFoldFP(fabs, V, Ty);
1450 case Intrinsic::log2:
1451 return ConstantFoldFP(log2, V, Ty);
1454 case Intrinsic::log:
1455 return ConstantFoldFP(log, V, Ty);
1458 case Intrinsic::log10:
1459 return ConstantFoldFP(log10, V, Ty);
1462 case Intrinsic::exp:
1463 return ConstantFoldFP(exp, V, Ty);
1466 case Intrinsic::exp2:
1467 return ConstantFoldFP(exp2, V, Ty);
1469 case Intrinsic::floor:
1470 return ConstantFoldFP(floor, V, Ty);
1471 case Intrinsic::ceil:
1472 return ConstantFoldFP(ceil, V, Ty);
1480 if (Name == "acos" && TLI->has(LibFunc::acos))
1481 return ConstantFoldFP(acos, V, Ty);
1482 else if (Name == "asin" && TLI->has(LibFunc::asin))
1483 return ConstantFoldFP(asin, V, Ty);
1484 else if (Name == "atan" && TLI->has(LibFunc::atan))
1485 return ConstantFoldFP(atan, V, Ty);
1488 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1489 return ConstantFoldFP(ceil, V, Ty);
1490 else if (Name == "cos" && TLI->has(LibFunc::cos))
1491 return ConstantFoldFP(cos, V, Ty);
1492 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1493 return ConstantFoldFP(cosh, V, Ty);
1494 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1495 return ConstantFoldFP(cos, V, Ty);
1498 if (Name == "exp" && TLI->has(LibFunc::exp))
1499 return ConstantFoldFP(exp, V, Ty);
1501 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1502 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1504 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1508 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1509 return ConstantFoldFP(fabs, V, Ty);
1510 else if (Name == "floor" && TLI->has(LibFunc::floor))
1511 return ConstantFoldFP(floor, V, Ty);
1514 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1515 return ConstantFoldFP(log, V, Ty);
1516 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1517 return ConstantFoldFP(log10, V, Ty);
1518 else if (IntrinsicID == Intrinsic::sqrt &&
1519 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1521 return ConstantFoldFP(sqrt, V, Ty);
1523 return Constant::getNullValue(Ty);
1527 if (Name == "sin" && TLI->has(LibFunc::sin))
1528 return ConstantFoldFP(sin, V, Ty);
1529 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1530 return ConstantFoldFP(sinh, V, Ty);
1531 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1532 return ConstantFoldFP(sqrt, V, Ty);
1533 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1534 return ConstantFoldFP(sqrt, V, Ty);
1535 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1536 return ConstantFoldFP(sin, V, Ty);
1539 if (Name == "tan" && TLI->has(LibFunc::tan))
1540 return ConstantFoldFP(tan, V, Ty);
1541 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1542 return ConstantFoldFP(tanh, V, Ty);
1550 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1551 switch (IntrinsicID) {
1552 case Intrinsic::bswap:
1553 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1554 case Intrinsic::ctpop:
1555 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1556 case Intrinsic::convert_from_fp16: {
1557 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1560 APFloat::opStatus status =
1561 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1563 // Conversion is always precise.
1565 assert(status == APFloat::opOK && !lost &&
1566 "Precision lost during fp16 constfolding");
1568 return ConstantFP::get(Ty->getContext(), Val);
1575 // Support ConstantVector in case we have an Undef in the top.
1576 if (isa<ConstantVector>(Operands[0]) ||
1577 isa<ConstantDataVector>(Operands[0])) {
1578 Constant *Op = cast<Constant>(Operands[0]);
1579 switch (IntrinsicID) {
1581 case Intrinsic::x86_sse_cvtss2si:
1582 case Intrinsic::x86_sse_cvtss2si64:
1583 case Intrinsic::x86_sse2_cvtsd2si:
1584 case Intrinsic::x86_sse2_cvtsd2si64:
1585 if (ConstantFP *FPOp =
1586 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1587 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1588 /*roundTowardZero=*/false, Ty);
1589 case Intrinsic::x86_sse_cvttss2si:
1590 case Intrinsic::x86_sse_cvttss2si64:
1591 case Intrinsic::x86_sse2_cvttsd2si:
1592 case Intrinsic::x86_sse2_cvttsd2si64:
1593 if (ConstantFP *FPOp =
1594 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1595 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1596 /*roundTowardZero=*/true, Ty);
1600 if (isa<UndefValue>(Operands[0])) {
1601 if (IntrinsicID == Intrinsic::bswap)
1609 if (Operands.size() == 2) {
1610 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1611 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1613 double Op1V = getValueAsDouble(Op1);
1615 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1616 if (Op2->getType() != Op1->getType())
1619 double Op2V = getValueAsDouble(Op2);
1620 if (IntrinsicID == Intrinsic::pow) {
1621 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1623 if (IntrinsicID == Intrinsic::copysign) {
1624 APFloat V1 = Op1->getValueAPF();
1625 APFloat V2 = Op2->getValueAPF();
1627 return ConstantFP::get(Ty->getContext(), V1);
1631 if (Name == "pow" && TLI->has(LibFunc::pow))
1632 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1633 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1634 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1635 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1636 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1637 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1638 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1639 return ConstantFP::get(Ty->getContext(),
1640 APFloat((float)std::pow((float)Op1V,
1641 (int)Op2C->getZExtValue())));
1642 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1643 return ConstantFP::get(Ty->getContext(),
1644 APFloat((float)std::pow((float)Op1V,
1645 (int)Op2C->getZExtValue())));
1646 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1647 return ConstantFP::get(Ty->getContext(),
1648 APFloat((double)std::pow((double)Op1V,
1649 (int)Op2C->getZExtValue())));
1654 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1655 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1656 switch (IntrinsicID) {
1658 case Intrinsic::sadd_with_overflow:
1659 case Intrinsic::uadd_with_overflow:
1660 case Intrinsic::ssub_with_overflow:
1661 case Intrinsic::usub_with_overflow:
1662 case Intrinsic::smul_with_overflow:
1663 case Intrinsic::umul_with_overflow: {
1666 switch (IntrinsicID) {
1667 default: llvm_unreachable("Invalid case");
1668 case Intrinsic::sadd_with_overflow:
1669 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1671 case Intrinsic::uadd_with_overflow:
1672 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1674 case Intrinsic::ssub_with_overflow:
1675 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1677 case Intrinsic::usub_with_overflow:
1678 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1680 case Intrinsic::smul_with_overflow:
1681 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1683 case Intrinsic::umul_with_overflow:
1684 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1688 ConstantInt::get(Ty->getContext(), Res),
1689 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1691 return ConstantStruct::get(cast<StructType>(Ty), Ops);
1693 case Intrinsic::cttz:
1694 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1695 return UndefValue::get(Ty);
1696 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1697 case Intrinsic::ctlz:
1698 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1699 return UndefValue::get(Ty);
1700 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1709 if (Operands.size() != 3)
1712 if (const ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1713 if (const ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1714 if (const ConstantFP *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1715 switch (IntrinsicID) {
1717 case Intrinsic::fma:
1718 case Intrinsic::fmuladd: {
1719 APFloat V = Op1->getValueAPF();
1720 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
1722 APFloat::rmNearestTiesToEven);
1723 if (s != APFloat::opInvalidOp)
1724 return ConstantFP::get(Ty->getContext(), V);
1736 static Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
1738 ArrayRef<Constant *> Operands,
1739 const TargetLibraryInfo *TLI) {
1740 SmallVector<Constant *, 4> Result(VTy->getNumElements());
1741 SmallVector<Constant *, 4> Lane(Operands.size());
1742 Type *Ty = VTy->getElementType();
1744 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1745 // Gather a column of constants.
1746 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
1747 Constant *Agg = Operands[J]->getAggregateElement(I);
1754 // Use the regular scalar folding to simplify this column.
1755 Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI);
1761 return ConstantVector::get(Result);
1764 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1765 /// with the specified arguments, returning null if unsuccessful.
1767 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1768 const TargetLibraryInfo *TLI) {
1771 StringRef Name = F->getName();
1773 Type *Ty = F->getReturnType();
1775 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
1776 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands, TLI);
1778 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);