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/APFloat.h"
21 #include "llvm/ADT/APInt.h"
22 #include "llvm/ADT/ArrayRef.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/Config/config.h"
30 #include "llvm/IR/Constant.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/GlobalValue.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/InstrTypes.h"
38 #include "llvm/IR/Instruction.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/IR/Value.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/MathExtras.h"
57 //===----------------------------------------------------------------------===//
58 // Constant Folding internal helper functions
59 //===----------------------------------------------------------------------===//
61 static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
62 Constant *C, Type *SrcEltTy,
64 const DataLayout &DL) {
65 // Now that we know that the input value is a vector of integers, just shift
66 // and insert them into our result.
67 unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
68 for (unsigned i = 0; i != NumSrcElts; ++i) {
70 if (DL.isLittleEndian())
71 Element = C->getAggregateElement(NumSrcElts - i - 1);
73 Element = C->getAggregateElement(i);
75 if (Element && isa<UndefValue>(Element)) {
80 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
82 return ConstantExpr::getBitCast(C, DestTy);
85 Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth());
91 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
92 /// This always returns a non-null constant, but it may be a
93 /// ConstantExpr if unfoldable.
94 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
95 // Catch the obvious splat cases.
96 if (C->isNullValue() && !DestTy->isX86_MMXTy())
97 return Constant::getNullValue(DestTy);
98 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
99 !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
100 return Constant::getAllOnesValue(DestTy);
102 if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
103 // Handle a vector->scalar integer/fp cast.
104 if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
105 unsigned NumSrcElts = VTy->getNumElements();
106 Type *SrcEltTy = VTy->getElementType();
108 // If the vector is a vector of floating point, convert it to vector of int
109 // to simplify things.
110 if (SrcEltTy->isFloatingPointTy()) {
111 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
113 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
114 // Ask IR to do the conversion now that #elts line up.
115 C = ConstantExpr::getBitCast(C, SrcIVTy);
118 APInt Result(DL.getTypeSizeInBits(DestTy), 0);
119 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
120 SrcEltTy, NumSrcElts, DL))
123 if (isa<IntegerType>(DestTy))
124 return ConstantInt::get(DestTy, Result);
126 APFloat FP(DestTy->getFltSemantics(), Result);
127 return ConstantFP::get(DestTy->getContext(), FP);
131 // The code below only handles casts to vectors currently.
132 auto *DestVTy = dyn_cast<VectorType>(DestTy);
134 return ConstantExpr::getBitCast(C, DestTy);
136 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
137 // vector so the code below can handle it uniformly.
138 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
139 Constant *Ops = C; // don't take the address of C!
140 return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
143 // If this is a bitcast from constant vector -> vector, fold it.
144 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
145 return ConstantExpr::getBitCast(C, DestTy);
147 // If the element types match, IR can fold it.
148 unsigned NumDstElt = DestVTy->getNumElements();
149 unsigned NumSrcElt = C->getType()->getVectorNumElements();
150 if (NumDstElt == NumSrcElt)
151 return ConstantExpr::getBitCast(C, DestTy);
153 Type *SrcEltTy = C->getType()->getVectorElementType();
154 Type *DstEltTy = DestVTy->getElementType();
156 // Otherwise, we're changing the number of elements in a vector, which
157 // requires endianness information to do the right thing. For example,
158 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
159 // folds to (little endian):
160 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
161 // and to (big endian):
162 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
164 // First thing is first. We only want to think about integer here, so if
165 // we have something in FP form, recast it as integer.
166 if (DstEltTy->isFloatingPointTy()) {
167 // Fold to an vector of integers with same size as our FP type.
168 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
170 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
171 // Recursively handle this integer conversion, if possible.
172 C = FoldBitCast(C, DestIVTy, DL);
174 // Finally, IR can handle this now that #elts line up.
175 return ConstantExpr::getBitCast(C, DestTy);
178 // Okay, we know the destination is integer, if the input is FP, convert
179 // it to integer first.
180 if (SrcEltTy->isFloatingPointTy()) {
181 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
183 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
184 // Ask IR to do the conversion now that #elts line up.
185 C = ConstantExpr::getBitCast(C, SrcIVTy);
186 // If IR wasn't able to fold it, bail out.
187 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
188 !isa<ConstantDataVector>(C))
192 // Now we know that the input and output vectors are both integer vectors
193 // of the same size, and that their #elements is not the same. Do the
194 // conversion here, which depends on whether the input or output has
196 bool isLittleEndian = DL.isLittleEndian();
198 SmallVector<Constant*, 32> Result;
199 if (NumDstElt < NumSrcElt) {
200 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
201 Constant *Zero = Constant::getNullValue(DstEltTy);
202 unsigned Ratio = NumSrcElt/NumDstElt;
203 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
205 for (unsigned i = 0; i != NumDstElt; ++i) {
206 // Build each element of the result.
207 Constant *Elt = Zero;
208 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
209 for (unsigned j = 0; j != Ratio; ++j) {
210 Constant *Src = C->getAggregateElement(SrcElt++);
211 if (Src && isa<UndefValue>(Src))
212 Src = Constant::getNullValue(C->getType()->getVectorElementType());
214 Src = dyn_cast_or_null<ConstantInt>(Src);
215 if (!Src) // Reject constantexpr elements.
216 return ConstantExpr::getBitCast(C, DestTy);
218 // Zero extend the element to the right size.
219 Src = ConstantExpr::getZExt(Src, Elt->getType());
221 // Shift it to the right place, depending on endianness.
222 Src = ConstantExpr::getShl(Src,
223 ConstantInt::get(Src->getType(), ShiftAmt));
224 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
227 Elt = ConstantExpr::getOr(Elt, Src);
229 Result.push_back(Elt);
231 return ConstantVector::get(Result);
234 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
235 unsigned Ratio = NumDstElt/NumSrcElt;
236 unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
238 // Loop over each source value, expanding into multiple results.
239 for (unsigned i = 0; i != NumSrcElt; ++i) {
240 auto *Element = C->getAggregateElement(i);
242 if (!Element) // Reject constantexpr elements.
243 return ConstantExpr::getBitCast(C, DestTy);
245 if (isa<UndefValue>(Element)) {
246 // Correctly Propagate undef values.
247 Result.append(Ratio, UndefValue::get(DstEltTy));
251 auto *Src = dyn_cast<ConstantInt>(Element);
253 return ConstantExpr::getBitCast(C, DestTy);
255 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
256 for (unsigned j = 0; j != Ratio; ++j) {
257 // Shift the piece of the value into the right place, depending on
259 Constant *Elt = ConstantExpr::getLShr(Src,
260 ConstantInt::get(Src->getType(), ShiftAmt));
261 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
263 // Truncate the element to an integer with the same pointer size and
264 // convert the element back to a pointer using a inttoptr.
265 if (DstEltTy->isPointerTy()) {
266 IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
267 Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
268 Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
272 // Truncate and remember this piece.
273 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
277 return ConstantVector::get(Result);
280 } // end anonymous namespace
282 /// If this constant is a constant offset from a global, return the global and
283 /// the constant. Because of constantexprs, this function is recursive.
284 bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
285 APInt &Offset, const DataLayout &DL) {
286 // Trivial case, constant is the global.
287 if ((GV = dyn_cast<GlobalValue>(C))) {
288 unsigned BitWidth = DL.getPointerTypeSizeInBits(GV->getType());
289 Offset = APInt(BitWidth, 0);
293 // Otherwise, if this isn't a constant expr, bail out.
294 auto *CE = dyn_cast<ConstantExpr>(C);
295 if (!CE) return false;
297 // Look through ptr->int and ptr->ptr casts.
298 if (CE->getOpcode() == Instruction::PtrToInt ||
299 CE->getOpcode() == Instruction::BitCast)
300 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
302 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
303 auto *GEP = dyn_cast<GEPOperator>(CE);
307 unsigned BitWidth = DL.getPointerTypeSizeInBits(GEP->getType());
308 APInt TmpOffset(BitWidth, 0);
310 // If the base isn't a global+constant, we aren't either.
311 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
314 // Otherwise, add any offset that our operands provide.
315 if (!GEP->accumulateConstantOffset(DL, TmpOffset))
324 /// Recursive helper to read bits out of global. C is the constant being copied
325 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
326 /// results into and BytesLeft is the number of bytes left in
327 /// the CurPtr buffer. DL is the DataLayout.
328 bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
329 unsigned BytesLeft, const DataLayout &DL) {
330 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
331 "Out of range access");
333 // If this element is zero or undefined, we can just return since *CurPtr is
335 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
338 if (auto *CI = dyn_cast<ConstantInt>(C)) {
339 if (CI->getBitWidth() > 64 ||
340 (CI->getBitWidth() & 7) != 0)
343 uint64_t Val = CI->getZExtValue();
344 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
346 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
348 if (!DL.isLittleEndian())
349 n = IntBytes - n - 1;
350 CurPtr[i] = (unsigned char)(Val >> (n * 8));
356 if (auto *CFP = dyn_cast<ConstantFP>(C)) {
357 if (CFP->getType()->isDoubleTy()) {
358 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
359 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
361 if (CFP->getType()->isFloatTy()){
362 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
363 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
365 if (CFP->getType()->isHalfTy()){
366 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
367 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
372 if (auto *CS = dyn_cast<ConstantStruct>(C)) {
373 const StructLayout *SL = DL.getStructLayout(CS->getType());
374 unsigned Index = SL->getElementContainingOffset(ByteOffset);
375 uint64_t CurEltOffset = SL->getElementOffset(Index);
376 ByteOffset -= CurEltOffset;
379 // If the element access is to the element itself and not to tail padding,
380 // read the bytes from the element.
381 uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
383 if (ByteOffset < EltSize &&
384 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
390 // Check to see if we read from the last struct element, if so we're done.
391 if (Index == CS->getType()->getNumElements())
394 // If we read all of the bytes we needed from this element we're done.
395 uint64_t NextEltOffset = SL->getElementOffset(Index);
397 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
400 // Move to the next element of the struct.
401 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
402 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
404 CurEltOffset = NextEltOffset;
409 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
410 isa<ConstantDataSequential>(C)) {
411 Type *EltTy = C->getType()->getSequentialElementType();
412 uint64_t EltSize = DL.getTypeAllocSize(EltTy);
413 uint64_t Index = ByteOffset / EltSize;
414 uint64_t Offset = ByteOffset - Index * EltSize;
416 if (auto *AT = dyn_cast<ArrayType>(C->getType()))
417 NumElts = AT->getNumElements();
419 NumElts = C->getType()->getVectorNumElements();
421 for (; Index != NumElts; ++Index) {
422 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
426 uint64_t BytesWritten = EltSize - Offset;
427 assert(BytesWritten <= EltSize && "Not indexing into this element?");
428 if (BytesWritten >= BytesLeft)
432 BytesLeft -= BytesWritten;
433 CurPtr += BytesWritten;
438 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
439 if (CE->getOpcode() == Instruction::IntToPtr &&
440 CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
441 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
446 // Otherwise, unknown initializer type.
450 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
451 const DataLayout &DL) {
452 auto *PTy = cast<PointerType>(C->getType());
453 auto *IntType = dyn_cast<IntegerType>(LoadTy);
455 // If this isn't an integer load we can't fold it directly.
457 unsigned AS = PTy->getAddressSpace();
459 // If this is a float/double load, we can try folding it as an int32/64 load
460 // and then bitcast the result. This can be useful for union cases. Note
461 // that address spaces don't matter here since we're not going to result in
462 // an actual new load.
464 if (LoadTy->isHalfTy())
465 MapTy = Type::getInt16Ty(C->getContext());
466 else if (LoadTy->isFloatTy())
467 MapTy = Type::getInt32Ty(C->getContext());
468 else if (LoadTy->isDoubleTy())
469 MapTy = Type::getInt64Ty(C->getContext());
470 else if (LoadTy->isVectorTy()) {
471 MapTy = PointerType::getIntNTy(C->getContext(),
472 DL.getTypeAllocSizeInBits(LoadTy));
476 C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
477 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL))
478 return FoldBitCast(Res, LoadTy, DL);
482 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
483 if (BytesLoaded > 32 || BytesLoaded == 0)
488 if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL))
491 auto *GV = dyn_cast<GlobalVariable>(GVal);
492 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
493 !GV->getInitializer()->getType()->isSized())
496 int64_t Offset = OffsetAI.getSExtValue();
497 int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
499 // If we're not accessing anything in this constant, the result is undefined.
500 if (Offset + BytesLoaded <= 0)
501 return UndefValue::get(IntType);
503 // If we're not accessing anything in this constant, the result is undefined.
504 if (Offset >= InitializerSize)
505 return UndefValue::get(IntType);
507 unsigned char RawBytes[32] = {0};
508 unsigned char *CurPtr = RawBytes;
509 unsigned BytesLeft = BytesLoaded;
511 // If we're loading off the beginning of the global, some bytes may be valid.
518 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL))
521 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
522 if (DL.isLittleEndian()) {
523 ResultVal = RawBytes[BytesLoaded - 1];
524 for (unsigned i = 1; i != BytesLoaded; ++i) {
526 ResultVal |= RawBytes[BytesLoaded - 1 - i];
529 ResultVal = RawBytes[0];
530 for (unsigned i = 1; i != BytesLoaded; ++i) {
532 ResultVal |= RawBytes[i];
536 return ConstantInt::get(IntType->getContext(), ResultVal);
539 Constant *ConstantFoldLoadThroughBitcast(ConstantExpr *CE, Type *DestTy,
540 const DataLayout &DL) {
541 auto *SrcPtr = CE->getOperand(0);
542 auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType());
545 Type *SrcTy = SrcPtrTy->getPointerElementType();
547 Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL);
552 Type *SrcTy = C->getType();
554 // If the type sizes are the same and a cast is legal, just directly
555 // cast the constant.
556 if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
557 Instruction::CastOps Cast = Instruction::BitCast;
558 // If we are going from a pointer to int or vice versa, we spell the cast
560 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
561 Cast = Instruction::IntToPtr;
562 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
563 Cast = Instruction::PtrToInt;
565 if (CastInst::castIsValid(Cast, C, DestTy))
566 return ConstantExpr::getCast(Cast, C, DestTy);
569 // If this isn't an aggregate type, there is nothing we can do to drill down
570 // and find a bitcastable constant.
571 if (!SrcTy->isAggregateType())
574 // We're simulating a load through a pointer that was bitcast to point to
575 // a different type, so we can try to walk down through the initial
576 // elements of an aggregate to see if some part of th e aggregate is
577 // castable to implement the "load" semantic model.
578 C = C->getAggregateElement(0u);
584 } // end anonymous namespace
586 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
587 const DataLayout &DL) {
588 // First, try the easy cases:
589 if (auto *GV = dyn_cast<GlobalVariable>(C))
590 if (GV->isConstant() && GV->hasDefinitiveInitializer())
591 return GV->getInitializer();
593 if (auto *GA = dyn_cast<GlobalAlias>(C))
594 if (GA->getAliasee() && !GA->isInterposable())
595 return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL);
597 // If the loaded value isn't a constant expr, we can't handle it.
598 auto *CE = dyn_cast<ConstantExpr>(C);
602 if (CE->getOpcode() == Instruction::GetElementPtr) {
603 if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
604 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
606 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
612 if (CE->getOpcode() == Instruction::BitCast)
613 if (Constant *LoadedC = ConstantFoldLoadThroughBitcast(CE, Ty, DL))
616 // Instead of loading constant c string, use corresponding integer value
617 // directly if string length is small enough.
619 if (getConstantStringInfo(CE, Str) && !Str.empty()) {
620 size_t StrLen = Str.size();
621 unsigned NumBits = Ty->getPrimitiveSizeInBits();
622 // Replace load with immediate integer if the result is an integer or fp
624 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
625 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
626 APInt StrVal(NumBits, 0);
627 APInt SingleChar(NumBits, 0);
628 if (DL.isLittleEndian()) {
629 for (unsigned char C : reverse(Str.bytes())) {
630 SingleChar = static_cast<uint64_t>(C);
631 StrVal = (StrVal << 8) | SingleChar;
634 for (unsigned char C : Str.bytes()) {
635 SingleChar = static_cast<uint64_t>(C);
636 StrVal = (StrVal << 8) | SingleChar;
638 // Append NULL at the end.
640 StrVal = (StrVal << 8) | SingleChar;
643 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
644 if (Ty->isFloatingPointTy())
645 Res = ConstantExpr::getBitCast(Res, Ty);
650 // If this load comes from anywhere in a constant global, and if the global
651 // is all undef or zero, we know what it loads.
652 if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
653 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
654 if (GV->getInitializer()->isNullValue())
655 return Constant::getNullValue(Ty);
656 if (isa<UndefValue>(GV->getInitializer()))
657 return UndefValue::get(Ty);
661 // Try hard to fold loads from bitcasted strange and non-type-safe things.
662 return FoldReinterpretLoadFromConstPtr(CE, Ty, DL);
667 Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) {
668 if (LI->isVolatile()) return nullptr;
670 if (auto *C = dyn_cast<Constant>(LI->getOperand(0)))
671 return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL);
676 /// One of Op0/Op1 is a constant expression.
677 /// Attempt to symbolically evaluate the result of a binary operator merging
678 /// these together. If target data info is available, it is provided as DL,
679 /// otherwise DL is null.
680 Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
681 const DataLayout &DL) {
684 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
685 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
688 if (Opc == Instruction::And) {
689 unsigned BitWidth = DL.getTypeSizeInBits(Op0->getType()->getScalarType());
690 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
691 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
692 computeKnownBits(Op0, KnownZero0, KnownOne0, DL);
693 computeKnownBits(Op1, KnownZero1, KnownOne1, DL);
694 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
695 // All the bits of Op0 that the 'and' could be masking are already zero.
698 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
699 // All the bits of Op1 that the 'and' could be masking are already zero.
703 APInt KnownZero = KnownZero0 | KnownZero1;
704 APInt KnownOne = KnownOne0 & KnownOne1;
705 if ((KnownZero | KnownOne).isAllOnesValue()) {
706 return ConstantInt::get(Op0->getType(), KnownOne);
710 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
711 // constant. This happens frequently when iterating over a global array.
712 if (Opc == Instruction::Sub) {
713 GlobalValue *GV1, *GV2;
716 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
717 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
718 unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
720 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
721 // PtrToInt may change the bitwidth so we have convert to the right size
723 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
724 Offs2.zextOrTrunc(OpSize));
731 /// If array indices are not pointer-sized integers, explicitly cast them so
732 /// that they aren't implicitly casted by the getelementptr.
733 Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
734 Type *ResultTy, Optional<unsigned> InRangeIndex,
735 const DataLayout &DL, const TargetLibraryInfo *TLI) {
736 Type *IntPtrTy = DL.getIntPtrType(ResultTy);
737 Type *IntPtrScalarTy = IntPtrTy->getScalarType();
740 SmallVector<Constant*, 32> NewIdxs;
741 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
743 !isa<StructType>(GetElementPtrInst::getIndexedType(
744 SrcElemTy, Ops.slice(1, i - 1)))) &&
745 Ops[i]->getType()->getScalarType() != IntPtrScalarTy) {
747 Type *NewType = Ops[i]->getType()->isVectorTy()
749 : IntPtrTy->getScalarType();
750 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
756 NewIdxs.push_back(Ops[i]);
762 Constant *C = ConstantExpr::getGetElementPtr(
763 SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex);
764 if (Constant *Folded = ConstantFoldConstant(C, DL, TLI))
770 /// Strip the pointer casts, but preserve the address space information.
771 Constant* StripPtrCastKeepAS(Constant* Ptr, Type *&ElemTy) {
772 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
773 auto *OldPtrTy = cast<PointerType>(Ptr->getType());
774 Ptr = Ptr->stripPointerCasts();
775 auto *NewPtrTy = cast<PointerType>(Ptr->getType());
777 ElemTy = NewPtrTy->getPointerElementType();
779 // Preserve the address space number of the pointer.
780 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
781 NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace());
782 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
787 /// If we can symbolically evaluate the GEP constant expression, do so.
788 Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
789 ArrayRef<Constant *> Ops,
790 const DataLayout &DL,
791 const TargetLibraryInfo *TLI) {
792 const GEPOperator *InnermostGEP = GEP;
793 bool InBounds = GEP->isInBounds();
795 Type *SrcElemTy = GEP->getSourceElementType();
796 Type *ResElemTy = GEP->getResultElementType();
797 Type *ResTy = GEP->getType();
798 if (!SrcElemTy->isSized())
801 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy,
802 GEP->getInRangeIndex(), DL, TLI))
805 Constant *Ptr = Ops[0];
806 if (!Ptr->getType()->isPointerTy())
809 Type *IntPtrTy = DL.getIntPtrType(Ptr->getType());
811 // If this is a constant expr gep that is effectively computing an
812 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
813 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
814 if (!isa<ConstantInt>(Ops[i])) {
816 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
817 // "inttoptr (sub (ptrtoint Ptr), V)"
818 if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
819 auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
820 assert((!CE || CE->getType() == IntPtrTy) &&
821 "CastGEPIndices didn't canonicalize index types!");
822 if (CE && CE->getOpcode() == Instruction::Sub &&
823 CE->getOperand(0)->isNullValue()) {
824 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
825 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
826 Res = ConstantExpr::getIntToPtr(Res, ResTy);
827 if (auto *FoldedRes = ConstantFoldConstant(Res, DL, TLI))
835 unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy);
838 DL.getIndexedOffsetInType(
840 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
841 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
843 // If this is a GEP of a GEP, fold it all into a single GEP.
844 while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
846 InBounds &= GEP->isInBounds();
848 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
850 // Do not try the incorporate the sub-GEP if some index is not a number.
851 bool AllConstantInt = true;
852 for (Value *NestedOp : NestedOps)
853 if (!isa<ConstantInt>(NestedOp)) {
854 AllConstantInt = false;
860 Ptr = cast<Constant>(GEP->getOperand(0));
861 SrcElemTy = GEP->getSourceElementType();
862 Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
863 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
866 // If the base value for this address is a literal integer value, fold the
867 // getelementptr to the resulting integer value casted to the pointer type.
868 APInt BasePtr(BitWidth, 0);
869 if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
870 if (CE->getOpcode() == Instruction::IntToPtr) {
871 if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
872 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
876 auto *PTy = cast<PointerType>(Ptr->getType());
877 if ((Ptr->isNullValue() || BasePtr != 0) &&
878 !DL.isNonIntegralPointerType(PTy)) {
879 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
880 return ConstantExpr::getIntToPtr(C, ResTy);
883 // Otherwise form a regular getelementptr. Recompute the indices so that
884 // we eliminate over-indexing of the notional static type array bounds.
885 // This makes it easy to determine if the getelementptr is "inbounds".
886 // Also, this helps GlobalOpt do SROA on GlobalVariables.
888 SmallVector<Constant *, 32> NewIdxs;
891 if (!Ty->isStructTy()) {
892 if (Ty->isPointerTy()) {
893 // The only pointer indexing we'll do is on the first index of the GEP.
894 if (!NewIdxs.empty())
899 // Only handle pointers to sized types, not pointers to functions.
902 } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) {
903 Ty = ATy->getElementType();
905 // We've reached some non-indexable type.
909 // Determine which element of the array the offset points into.
910 APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty));
912 // The element size is 0. This may be [0 x Ty]*, so just use a zero
913 // index for this level and proceed to the next level to see if it can
914 // accommodate the offset.
915 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
917 // The element size is non-zero divide the offset by the element
918 // size (rounding down), to compute the index at this level.
920 APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow);
923 Offset -= NewIdx * ElemSize;
924 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
927 auto *STy = cast<StructType>(Ty);
928 // If we end up with an offset that isn't valid for this struct type, we
929 // can't re-form this GEP in a regular form, so bail out. The pointer
930 // operand likely went through casts that are necessary to make the GEP
932 const StructLayout &SL = *DL.getStructLayout(STy);
933 if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes()))
936 // Determine which field of the struct the offset points into. The
937 // getZExtValue is fine as we've already ensured that the offset is
938 // within the range representable by the StructLayout API.
939 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
940 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
942 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
943 Ty = STy->getTypeAtIndex(ElIdx);
945 } while (Ty != ResElemTy);
947 // If we haven't used up the entire offset by descending the static
948 // type, then the offset is pointing into the middle of an indivisible
949 // member, so we can't simplify it.
953 // Preserve the inrange index from the innermost GEP if possible. We must
954 // have calculated the same indices up to and including the inrange index.
955 Optional<unsigned> InRangeIndex;
956 if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex())
957 if (SrcElemTy == InnermostGEP->getSourceElementType() &&
958 NewIdxs.size() > *LastIRIndex) {
959 InRangeIndex = LastIRIndex;
960 for (unsigned I = 0; I <= *LastIRIndex; ++I)
961 if (NewIdxs[I] != InnermostGEP->getOperand(I + 1)) {
968 Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs,
969 InBounds, InRangeIndex);
970 assert(C->getType()->getPointerElementType() == Ty &&
971 "Computed GetElementPtr has unexpected type!");
973 // If we ended up indexing a member with a type that doesn't match
974 // the type of what the original indices indexed, add a cast.
976 C = FoldBitCast(C, ResTy, DL);
981 /// Attempt to constant fold an instruction with the
982 /// specified opcode and operands. If successful, the constant result is
983 /// returned, if not, null is returned. Note that this function can fail when
984 /// attempting to fold instructions like loads and stores, which have no
985 /// constant expression form.
987 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/inrange
988 /// etc information, due to only being passed an opcode and operands. Constant
989 /// folding using this function strips this information.
991 Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
992 ArrayRef<Constant *> Ops,
993 const DataLayout &DL,
994 const TargetLibraryInfo *TLI) {
995 Type *DestTy = InstOrCE->getType();
997 // Handle easy binops first.
998 if (Instruction::isBinaryOp(Opcode))
999 return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
1001 if (Instruction::isCast(Opcode))
1002 return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
1004 if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
1005 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1008 return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0],
1009 Ops.slice(1), GEP->isInBounds(),
1010 GEP->getInRangeIndex());
1013 if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))
1014 return CE->getWithOperands(Ops);
1017 default: return nullptr;
1018 case Instruction::ICmp:
1019 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1020 case Instruction::Call:
1021 if (auto *F = dyn_cast<Function>(Ops.back()))
1022 if (canConstantFoldCallTo(F))
1023 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
1025 case Instruction::Select:
1026 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1027 case Instruction::ExtractElement:
1028 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1029 case Instruction::InsertElement:
1030 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1031 case Instruction::ShuffleVector:
1032 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1036 } // end anonymous namespace
1038 //===----------------------------------------------------------------------===//
1039 // Constant Folding public APIs
1040 //===----------------------------------------------------------------------===//
1045 ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1046 const TargetLibraryInfo *TLI,
1047 SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1048 if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))
1051 SmallVector<Constant *, 8> Ops;
1052 for (const Use &NewU : C->operands()) {
1053 auto *NewC = cast<Constant>(&NewU);
1054 // Recursively fold the ConstantExpr's operands. If we have already folded
1055 // a ConstantExpr, we don't have to process it again.
1056 if (isa<ConstantVector>(NewC) || isa<ConstantExpr>(NewC)) {
1057 auto It = FoldedOps.find(NewC);
1058 if (It == FoldedOps.end()) {
1060 ConstantFoldConstantImpl(NewC, DL, TLI, FoldedOps)) {
1062 FoldedOps.insert({NewC, FoldedC});
1064 FoldedOps.insert({NewC, NewC});
1070 Ops.push_back(NewC);
1073 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1074 if (CE->isCompare())
1075 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1078 return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI);
1081 assert(isa<ConstantVector>(C));
1082 return ConstantVector::get(Ops);
1085 } // end anonymous namespace
1087 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
1088 const TargetLibraryInfo *TLI) {
1089 // Handle PHI nodes quickly here...
1090 if (auto *PN = dyn_cast<PHINode>(I)) {
1091 Constant *CommonValue = nullptr;
1093 SmallDenseMap<Constant *, Constant *> FoldedOps;
1094 for (Value *Incoming : PN->incoming_values()) {
1095 // If the incoming value is undef then skip it. Note that while we could
1096 // skip the value if it is equal to the phi node itself we choose not to
1097 // because that would break the rule that constant folding only applies if
1098 // all operands are constants.
1099 if (isa<UndefValue>(Incoming))
1101 // If the incoming value is not a constant, then give up.
1102 auto *C = dyn_cast<Constant>(Incoming);
1105 // Fold the PHI's operands.
1106 if (auto *FoldedC = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps))
1108 // If the incoming value is a different constant to
1109 // the one we saw previously, then give up.
1110 if (CommonValue && C != CommonValue)
1115 // If we reach here, all incoming values are the same constant or undef.
1116 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
1119 // Scan the operand list, checking to see if they are all constants, if so,
1120 // hand off to ConstantFoldInstOperandsImpl.
1121 if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1124 SmallDenseMap<Constant *, Constant *> FoldedOps;
1125 SmallVector<Constant *, 8> Ops;
1126 for (const Use &OpU : I->operands()) {
1127 auto *Op = cast<Constant>(&OpU);
1128 // Fold the Instruction's operands.
1129 if (auto *FoldedOp = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps))
1135 if (const auto *CI = dyn_cast<CmpInst>(I))
1136 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1139 if (const auto *LI = dyn_cast<LoadInst>(I))
1140 return ConstantFoldLoadInst(LI, DL);
1142 if (auto *IVI = dyn_cast<InsertValueInst>(I)) {
1143 return ConstantExpr::getInsertValue(
1144 cast<Constant>(IVI->getAggregateOperand()),
1145 cast<Constant>(IVI->getInsertedValueOperand()),
1149 if (auto *EVI = dyn_cast<ExtractValueInst>(I)) {
1150 return ConstantExpr::getExtractValue(
1151 cast<Constant>(EVI->getAggregateOperand()),
1155 return ConstantFoldInstOperands(I, Ops, DL, TLI);
1158 Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1159 const TargetLibraryInfo *TLI) {
1160 SmallDenseMap<Constant *, Constant *> FoldedOps;
1161 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1164 Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1165 ArrayRef<Constant *> Ops,
1166 const DataLayout &DL,
1167 const TargetLibraryInfo *TLI) {
1168 return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI);
1171 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1172 Constant *Ops0, Constant *Ops1,
1173 const DataLayout &DL,
1174 const TargetLibraryInfo *TLI) {
1175 // fold: icmp (inttoptr x), null -> icmp x, 0
1176 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1177 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1178 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1180 // FIXME: The following comment is out of data and the DataLayout is here now.
1181 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1182 // around to know if bit truncation is happening.
1183 if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1184 if (Ops1->isNullValue()) {
1185 if (CE0->getOpcode() == Instruction::IntToPtr) {
1186 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1187 // Convert the integer value to the right size to ensure we get the
1188 // proper extension or truncation.
1189 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1191 Constant *Null = Constant::getNullValue(C->getType());
1192 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1195 // Only do this transformation if the int is intptrty in size, otherwise
1196 // there is a truncation or extension that we aren't modeling.
1197 if (CE0->getOpcode() == Instruction::PtrToInt) {
1198 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1199 if (CE0->getType() == IntPtrTy) {
1200 Constant *C = CE0->getOperand(0);
1201 Constant *Null = Constant::getNullValue(C->getType());
1202 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1207 if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1208 if (CE0->getOpcode() == CE1->getOpcode()) {
1209 if (CE0->getOpcode() == Instruction::IntToPtr) {
1210 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1212 // Convert the integer value to the right size to ensure we get the
1213 // proper extension or truncation.
1214 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1216 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1218 return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1221 // Only do this transformation if the int is intptrty in size, otherwise
1222 // there is a truncation or extension that we aren't modeling.
1223 if (CE0->getOpcode() == Instruction::PtrToInt) {
1224 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1225 if (CE0->getType() == IntPtrTy &&
1226 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1227 return ConstantFoldCompareInstOperands(
1228 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1234 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1235 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1236 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1237 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1238 Constant *LHS = ConstantFoldCompareInstOperands(
1239 Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1240 Constant *RHS = ConstantFoldCompareInstOperands(
1241 Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1243 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1244 return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1248 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1251 Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1253 const DataLayout &DL) {
1254 assert(Instruction::isBinaryOp(Opcode));
1255 if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1256 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1259 return ConstantExpr::get(Opcode, LHS, RHS);
1262 Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1263 Type *DestTy, const DataLayout &DL) {
1264 assert(Instruction::isCast(Opcode));
1267 llvm_unreachable("Missing case");
1268 case Instruction::PtrToInt:
1269 // If the input is a inttoptr, eliminate the pair. This requires knowing
1270 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1271 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1272 if (CE->getOpcode() == Instruction::IntToPtr) {
1273 Constant *Input = CE->getOperand(0);
1274 unsigned InWidth = Input->getType()->getScalarSizeInBits();
1275 unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1276 if (PtrWidth < InWidth) {
1278 ConstantInt::get(CE->getContext(),
1279 APInt::getLowBitsSet(InWidth, PtrWidth));
1280 Input = ConstantExpr::getAnd(Input, Mask);
1282 // Do a zext or trunc to get to the dest size.
1283 return ConstantExpr::getIntegerCast(Input, DestTy, false);
1286 return ConstantExpr::getCast(Opcode, C, DestTy);
1287 case Instruction::IntToPtr:
1288 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1289 // the int size is >= the ptr size and the address spaces are the same.
1290 // This requires knowing the width of a pointer, so it can't be done in
1291 // ConstantExpr::getCast.
1292 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1293 if (CE->getOpcode() == Instruction::PtrToInt) {
1294 Constant *SrcPtr = CE->getOperand(0);
1295 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1296 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1298 if (MidIntSize >= SrcPtrSize) {
1299 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1300 if (SrcAS == DestTy->getPointerAddressSpace())
1301 return FoldBitCast(CE->getOperand(0), DestTy, DL);
1306 return ConstantExpr::getCast(Opcode, C, DestTy);
1307 case Instruction::Trunc:
1308 case Instruction::ZExt:
1309 case Instruction::SExt:
1310 case Instruction::FPTrunc:
1311 case Instruction::FPExt:
1312 case Instruction::UIToFP:
1313 case Instruction::SIToFP:
1314 case Instruction::FPToUI:
1315 case Instruction::FPToSI:
1316 case Instruction::AddrSpaceCast:
1317 return ConstantExpr::getCast(Opcode, C, DestTy);
1318 case Instruction::BitCast:
1319 return FoldBitCast(C, DestTy, DL);
1323 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1325 if (!CE->getOperand(1)->isNullValue())
1326 return nullptr; // Do not allow stepping over the value!
1328 // Loop over all of the operands, tracking down which value we are
1330 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1331 C = C->getAggregateElement(CE->getOperand(i));
1339 llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1340 ArrayRef<Constant *> Indices) {
1341 // Loop over all of the operands, tracking down which value we are
1343 for (Constant *Index : Indices) {
1344 C = C->getAggregateElement(Index);
1351 //===----------------------------------------------------------------------===//
1352 // Constant Folding for Calls
1355 bool llvm::canConstantFoldCallTo(const Function *F) {
1356 switch (F->getIntrinsicID()) {
1357 case Intrinsic::fabs:
1358 case Intrinsic::minnum:
1359 case Intrinsic::maxnum:
1360 case Intrinsic::log:
1361 case Intrinsic::log2:
1362 case Intrinsic::log10:
1363 case Intrinsic::exp:
1364 case Intrinsic::exp2:
1365 case Intrinsic::floor:
1366 case Intrinsic::ceil:
1367 case Intrinsic::sqrt:
1368 case Intrinsic::sin:
1369 case Intrinsic::cos:
1370 case Intrinsic::trunc:
1371 case Intrinsic::rint:
1372 case Intrinsic::nearbyint:
1373 case Intrinsic::pow:
1374 case Intrinsic::powi:
1375 case Intrinsic::bswap:
1376 case Intrinsic::ctpop:
1377 case Intrinsic::ctlz:
1378 case Intrinsic::cttz:
1379 case Intrinsic::fma:
1380 case Intrinsic::fmuladd:
1381 case Intrinsic::copysign:
1382 case Intrinsic::round:
1383 case Intrinsic::masked_load:
1384 case Intrinsic::sadd_with_overflow:
1385 case Intrinsic::uadd_with_overflow:
1386 case Intrinsic::ssub_with_overflow:
1387 case Intrinsic::usub_with_overflow:
1388 case Intrinsic::smul_with_overflow:
1389 case Intrinsic::umul_with_overflow:
1390 case Intrinsic::convert_from_fp16:
1391 case Intrinsic::convert_to_fp16:
1392 case Intrinsic::bitreverse:
1393 case Intrinsic::x86_sse_cvtss2si:
1394 case Intrinsic::x86_sse_cvtss2si64:
1395 case Intrinsic::x86_sse_cvttss2si:
1396 case Intrinsic::x86_sse_cvttss2si64:
1397 case Intrinsic::x86_sse2_cvtsd2si:
1398 case Intrinsic::x86_sse2_cvtsd2si64:
1399 case Intrinsic::x86_sse2_cvttsd2si:
1400 case Intrinsic::x86_sse2_cvttsd2si64:
1409 StringRef Name = F->getName();
1411 // In these cases, the check of the length is required. We don't want to
1412 // return true for a name like "cos\0blah" which strcmp would return equal to
1413 // "cos", but has length 8.
1418 return Name == "acos" || Name == "asin" || Name == "atan" ||
1419 Name == "atan2" || Name == "acosf" || Name == "asinf" ||
1420 Name == "atanf" || Name == "atan2f";
1422 return Name == "ceil" || Name == "cos" || Name == "cosh" ||
1423 Name == "ceilf" || Name == "cosf" || Name == "coshf";
1425 return Name == "exp" || Name == "exp2" || Name == "expf" || Name == "exp2f";
1427 return Name == "fabs" || Name == "floor" || Name == "fmod" ||
1428 Name == "fabsf" || Name == "floorf" || Name == "fmodf";
1430 return Name == "log" || Name == "log10" || Name == "logf" ||
1433 return Name == "pow" || Name == "powf";
1435 return Name == "round" || Name == "roundf";
1437 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1438 Name == "sinf" || Name == "sinhf" || Name == "sqrtf";
1440 return Name == "tan" || Name == "tanh" || Name == "tanf" || Name == "tanhf";
1446 Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1447 if (Ty->isHalfTy()) {
1450 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &unused);
1451 return ConstantFP::get(Ty->getContext(), APF);
1453 if (Ty->isFloatTy())
1454 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1455 if (Ty->isDoubleTy())
1456 return ConstantFP::get(Ty->getContext(), APFloat(V));
1457 llvm_unreachable("Can only constant fold half/float/double");
1460 /// Clear the floating-point exception state.
1461 inline void llvm_fenv_clearexcept() {
1462 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1463 feclearexcept(FE_ALL_EXCEPT);
1468 /// Test if a floating-point exception was raised.
1469 inline bool llvm_fenv_testexcept() {
1470 int errno_val = errno;
1471 if (errno_val == ERANGE || errno_val == EDOM)
1473 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1474 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1480 Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) {
1481 llvm_fenv_clearexcept();
1483 if (llvm_fenv_testexcept()) {
1484 llvm_fenv_clearexcept();
1488 return GetConstantFoldFPValue(V, Ty);
1491 Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
1492 double W, Type *Ty) {
1493 llvm_fenv_clearexcept();
1495 if (llvm_fenv_testexcept()) {
1496 llvm_fenv_clearexcept();
1500 return GetConstantFoldFPValue(V, Ty);
1503 /// Attempt to fold an SSE floating point to integer conversion of a constant
1504 /// floating point. If roundTowardZero is false, the default IEEE rounding is
1505 /// used (toward nearest, ties to even). This matches the behavior of the
1506 /// non-truncating SSE instructions in the default rounding mode. The desired
1507 /// integer type Ty is used to select how many bits are available for the
1508 /// result. Returns null if the conversion cannot be performed, otherwise
1509 /// returns the Constant value resulting from the conversion.
1510 Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
1512 // All of these conversion intrinsics form an integer of at most 64bits.
1513 unsigned ResultWidth = Ty->getIntegerBitWidth();
1514 assert(ResultWidth <= 64 &&
1515 "Can only constant fold conversions to 64 and 32 bit ints");
1518 bool isExact = false;
1519 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1520 : APFloat::rmNearestTiesToEven;
1521 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1522 /*isSigned=*/true, mode,
1524 if (status != APFloat::opOK &&
1525 (!roundTowardZero || status != APFloat::opInexact))
1527 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1530 double getValueAsDouble(ConstantFP *Op) {
1531 Type *Ty = Op->getType();
1533 if (Ty->isFloatTy())
1534 return Op->getValueAPF().convertToFloat();
1536 if (Ty->isDoubleTy())
1537 return Op->getValueAPF().convertToDouble();
1540 APFloat APF = Op->getValueAPF();
1541 APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
1542 return APF.convertToDouble();
1545 Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
1546 ArrayRef<Constant *> Operands,
1547 const TargetLibraryInfo *TLI) {
1548 if (Operands.size() == 1) {
1549 if (isa<UndefValue>(Operands[0])) {
1550 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN
1551 if (IntrinsicID == Intrinsic::cos)
1552 return Constant::getNullValue(Ty);
1554 if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1555 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1556 APFloat Val(Op->getValueAPF());
1559 Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);
1561 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1564 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1567 if (IntrinsicID == Intrinsic::round) {
1568 APFloat V = Op->getValueAPF();
1569 V.roundToIntegral(APFloat::rmNearestTiesToAway);
1570 return ConstantFP::get(Ty->getContext(), V);
1573 if (IntrinsicID == Intrinsic::floor) {
1574 APFloat V = Op->getValueAPF();
1575 V.roundToIntegral(APFloat::rmTowardNegative);
1576 return ConstantFP::get(Ty->getContext(), V);
1579 if (IntrinsicID == Intrinsic::ceil) {
1580 APFloat V = Op->getValueAPF();
1581 V.roundToIntegral(APFloat::rmTowardPositive);
1582 return ConstantFP::get(Ty->getContext(), V);
1585 if (IntrinsicID == Intrinsic::trunc) {
1586 APFloat V = Op->getValueAPF();
1587 V.roundToIntegral(APFloat::rmTowardZero);
1588 return ConstantFP::get(Ty->getContext(), V);
1591 if (IntrinsicID == Intrinsic::rint) {
1592 APFloat V = Op->getValueAPF();
1593 V.roundToIntegral(APFloat::rmNearestTiesToEven);
1594 return ConstantFP::get(Ty->getContext(), V);
1597 if (IntrinsicID == Intrinsic::nearbyint) {
1598 APFloat V = Op->getValueAPF();
1599 V.roundToIntegral(APFloat::rmNearestTiesToEven);
1600 return ConstantFP::get(Ty->getContext(), V);
1603 /// We only fold functions with finite arguments. Folding NaN and inf is
1604 /// likely to be aborted with an exception anyway, and some host libms
1605 /// have known errors raising exceptions.
1606 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1609 /// Currently APFloat versions of these functions do not exist, so we use
1610 /// the host native double versions. Float versions are not called
1611 /// directly but for all these it is true (float)(f((double)arg)) ==
1612 /// f(arg). Long double not supported yet.
1613 double V = getValueAsDouble(Op);
1615 switch (IntrinsicID) {
1617 case Intrinsic::fabs:
1618 return ConstantFoldFP(fabs, V, Ty);
1619 case Intrinsic::log2:
1620 return ConstantFoldFP(Log2, V, Ty);
1621 case Intrinsic::log:
1622 return ConstantFoldFP(log, V, Ty);
1623 case Intrinsic::log10:
1624 return ConstantFoldFP(log10, V, Ty);
1625 case Intrinsic::exp:
1626 return ConstantFoldFP(exp, V, Ty);
1627 case Intrinsic::exp2:
1628 return ConstantFoldFP(exp2, V, Ty);
1629 case Intrinsic::sin:
1630 return ConstantFoldFP(sin, V, Ty);
1631 case Intrinsic::cos:
1632 return ConstantFoldFP(cos, V, Ty);
1640 if ((Name == "acos" && TLI->has(LibFunc::acos)) ||
1641 (Name == "acosf" && TLI->has(LibFunc::acosf)))
1642 return ConstantFoldFP(acos, V, Ty);
1643 else if ((Name == "asin" && TLI->has(LibFunc::asin)) ||
1644 (Name == "asinf" && TLI->has(LibFunc::asinf)))
1645 return ConstantFoldFP(asin, V, Ty);
1646 else if ((Name == "atan" && TLI->has(LibFunc::atan)) ||
1647 (Name == "atanf" && TLI->has(LibFunc::atanf)))
1648 return ConstantFoldFP(atan, V, Ty);
1651 if ((Name == "ceil" && TLI->has(LibFunc::ceil)) ||
1652 (Name == "ceilf" && TLI->has(LibFunc::ceilf)))
1653 return ConstantFoldFP(ceil, V, Ty);
1654 else if ((Name == "cos" && TLI->has(LibFunc::cos)) ||
1655 (Name == "cosf" && TLI->has(LibFunc::cosf)))
1656 return ConstantFoldFP(cos, V, Ty);
1657 else if ((Name == "cosh" && TLI->has(LibFunc::cosh)) ||
1658 (Name == "coshf" && TLI->has(LibFunc::coshf)))
1659 return ConstantFoldFP(cosh, V, Ty);
1662 if ((Name == "exp" && TLI->has(LibFunc::exp)) ||
1663 (Name == "expf" && TLI->has(LibFunc::expf)))
1664 return ConstantFoldFP(exp, V, Ty);
1665 if ((Name == "exp2" && TLI->has(LibFunc::exp2)) ||
1666 (Name == "exp2f" && TLI->has(LibFunc::exp2f)))
1667 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1669 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1672 if ((Name == "fabs" && TLI->has(LibFunc::fabs)) ||
1673 (Name == "fabsf" && TLI->has(LibFunc::fabsf)))
1674 return ConstantFoldFP(fabs, V, Ty);
1675 else if ((Name == "floor" && TLI->has(LibFunc::floor)) ||
1676 (Name == "floorf" && TLI->has(LibFunc::floorf)))
1677 return ConstantFoldFP(floor, V, Ty);
1680 if ((Name == "log" && V > 0 && TLI->has(LibFunc::log)) ||
1681 (Name == "logf" && V > 0 && TLI->has(LibFunc::logf)))
1682 return ConstantFoldFP(log, V, Ty);
1683 else if ((Name == "log10" && V > 0 && TLI->has(LibFunc::log10)) ||
1684 (Name == "log10f" && V > 0 && TLI->has(LibFunc::log10f)))
1685 return ConstantFoldFP(log10, V, Ty);
1686 else if (IntrinsicID == Intrinsic::sqrt &&
1687 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1689 return ConstantFoldFP(sqrt, V, Ty);
1691 // Unlike the sqrt definitions in C/C++, POSIX, and IEEE-754 - which
1692 // all guarantee or favor returning NaN - the square root of a
1693 // negative number is not defined for the LLVM sqrt intrinsic.
1694 // This is because the intrinsic should only be emitted in place of
1695 // libm's sqrt function when using "no-nans-fp-math".
1696 return UndefValue::get(Ty);
1701 if ((Name == "round" && TLI->has(LibFunc::round)) ||
1702 (Name == "roundf" && TLI->has(LibFunc::roundf)))
1703 return ConstantFoldFP(round, V, Ty);
1705 if ((Name == "sin" && TLI->has(LibFunc::sin)) ||
1706 (Name == "sinf" && TLI->has(LibFunc::sinf)))
1707 return ConstantFoldFP(sin, V, Ty);
1708 else if ((Name == "sinh" && TLI->has(LibFunc::sinh)) ||
1709 (Name == "sinhf" && TLI->has(LibFunc::sinhf)))
1710 return ConstantFoldFP(sinh, V, Ty);
1711 else if ((Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt)) ||
1712 (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf)))
1713 return ConstantFoldFP(sqrt, V, Ty);
1716 if ((Name == "tan" && TLI->has(LibFunc::tan)) ||
1717 (Name == "tanf" && TLI->has(LibFunc::tanf)))
1718 return ConstantFoldFP(tan, V, Ty);
1719 else if ((Name == "tanh" && TLI->has(LibFunc::tanh)) ||
1720 (Name == "tanhf" && TLI->has(LibFunc::tanhf)))
1721 return ConstantFoldFP(tanh, V, Ty);
1729 if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
1730 switch (IntrinsicID) {
1731 case Intrinsic::bswap:
1732 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1733 case Intrinsic::ctpop:
1734 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1735 case Intrinsic::bitreverse:
1736 return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
1737 case Intrinsic::convert_from_fp16: {
1738 APFloat Val(APFloat::IEEEhalf(), Op->getValue());
1741 APFloat::opStatus status = Val.convert(
1742 Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
1744 // Conversion is always precise.
1746 assert(status == APFloat::opOK && !lost &&
1747 "Precision lost during fp16 constfolding");
1749 return ConstantFP::get(Ty->getContext(), Val);
1756 // Support ConstantVector in case we have an Undef in the top.
1757 if (isa<ConstantVector>(Operands[0]) ||
1758 isa<ConstantDataVector>(Operands[0])) {
1759 auto *Op = cast<Constant>(Operands[0]);
1760 switch (IntrinsicID) {
1762 case Intrinsic::x86_sse_cvtss2si:
1763 case Intrinsic::x86_sse_cvtss2si64:
1764 case Intrinsic::x86_sse2_cvtsd2si:
1765 case Intrinsic::x86_sse2_cvtsd2si64:
1766 if (ConstantFP *FPOp =
1767 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1768 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1769 /*roundTowardZero=*/false, Ty);
1770 case Intrinsic::x86_sse_cvttss2si:
1771 case Intrinsic::x86_sse_cvttss2si64:
1772 case Intrinsic::x86_sse2_cvttsd2si:
1773 case Intrinsic::x86_sse2_cvttsd2si64:
1774 if (ConstantFP *FPOp =
1775 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1776 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1777 /*roundTowardZero=*/true, Ty);
1781 if (isa<UndefValue>(Operands[0])) {
1782 if (IntrinsicID == Intrinsic::bswap)
1790 if (Operands.size() == 2) {
1791 if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1792 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1794 double Op1V = getValueAsDouble(Op1);
1796 if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1797 if (Op2->getType() != Op1->getType())
1800 double Op2V = getValueAsDouble(Op2);
1801 if (IntrinsicID == Intrinsic::pow) {
1802 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1804 if (IntrinsicID == Intrinsic::copysign) {
1805 APFloat V1 = Op1->getValueAPF();
1806 const APFloat &V2 = Op2->getValueAPF();
1808 return ConstantFP::get(Ty->getContext(), V1);
1811 if (IntrinsicID == Intrinsic::minnum) {
1812 const APFloat &C1 = Op1->getValueAPF();
1813 const APFloat &C2 = Op2->getValueAPF();
1814 return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
1817 if (IntrinsicID == Intrinsic::maxnum) {
1818 const APFloat &C1 = Op1->getValueAPF();
1819 const APFloat &C2 = Op2->getValueAPF();
1820 return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
1825 if ((Name == "pow" && TLI->has(LibFunc::pow)) ||
1826 (Name == "powf" && TLI->has(LibFunc::powf)))
1827 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1828 if ((Name == "fmod" && TLI->has(LibFunc::fmod)) ||
1829 (Name == "fmodf" && TLI->has(LibFunc::fmodf)))
1830 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1831 if ((Name == "atan2" && TLI->has(LibFunc::atan2)) ||
1832 (Name == "atan2f" && TLI->has(LibFunc::atan2f)))
1833 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1834 } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1835 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1836 return ConstantFP::get(Ty->getContext(),
1837 APFloat((float)std::pow((float)Op1V,
1838 (int)Op2C->getZExtValue())));
1839 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1840 return ConstantFP::get(Ty->getContext(),
1841 APFloat((float)std::pow((float)Op1V,
1842 (int)Op2C->getZExtValue())));
1843 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1844 return ConstantFP::get(Ty->getContext(),
1845 APFloat((double)std::pow((double)Op1V,
1846 (int)Op2C->getZExtValue())));
1851 if (auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1852 if (auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1853 switch (IntrinsicID) {
1855 case Intrinsic::sadd_with_overflow:
1856 case Intrinsic::uadd_with_overflow:
1857 case Intrinsic::ssub_with_overflow:
1858 case Intrinsic::usub_with_overflow:
1859 case Intrinsic::smul_with_overflow:
1860 case Intrinsic::umul_with_overflow: {
1863 switch (IntrinsicID) {
1864 default: llvm_unreachable("Invalid case");
1865 case Intrinsic::sadd_with_overflow:
1866 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1868 case Intrinsic::uadd_with_overflow:
1869 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1871 case Intrinsic::ssub_with_overflow:
1872 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1874 case Intrinsic::usub_with_overflow:
1875 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1877 case Intrinsic::smul_with_overflow:
1878 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1880 case Intrinsic::umul_with_overflow:
1881 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1885 ConstantInt::get(Ty->getContext(), Res),
1886 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1888 return ConstantStruct::get(cast<StructType>(Ty), Ops);
1890 case Intrinsic::cttz:
1891 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1892 return UndefValue::get(Ty);
1893 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1894 case Intrinsic::ctlz:
1895 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1896 return UndefValue::get(Ty);
1897 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1906 if (Operands.size() != 3)
1909 if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1910 if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1911 if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1912 switch (IntrinsicID) {
1914 case Intrinsic::fma:
1915 case Intrinsic::fmuladd: {
1916 APFloat V = Op1->getValueAPF();
1917 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
1919 APFloat::rmNearestTiesToEven);
1920 if (s != APFloat::opInvalidOp)
1921 return ConstantFP::get(Ty->getContext(), V);
1933 Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
1934 VectorType *VTy, ArrayRef<Constant *> Operands,
1935 const DataLayout &DL,
1936 const TargetLibraryInfo *TLI) {
1937 SmallVector<Constant *, 4> Result(VTy->getNumElements());
1938 SmallVector<Constant *, 4> Lane(Operands.size());
1939 Type *Ty = VTy->getElementType();
1941 if (IntrinsicID == Intrinsic::masked_load) {
1942 auto *SrcPtr = Operands[0];
1943 auto *Mask = Operands[2];
1944 auto *Passthru = Operands[3];
1946 Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL);
1948 SmallVector<Constant *, 32> NewElements;
1949 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1950 auto *MaskElt = Mask->getAggregateElement(I);
1953 auto *PassthruElt = Passthru->getAggregateElement(I);
1954 auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
1955 if (isa<UndefValue>(MaskElt)) {
1957 NewElements.push_back(PassthruElt);
1959 NewElements.push_back(VecElt);
1963 if (MaskElt->isNullValue()) {
1966 NewElements.push_back(PassthruElt);
1967 } else if (MaskElt->isOneValue()) {
1970 NewElements.push_back(VecElt);
1975 if (NewElements.size() != VTy->getNumElements())
1977 return ConstantVector::get(NewElements);
1980 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1981 // Gather a column of constants.
1982 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
1983 Constant *Agg = Operands[J]->getAggregateElement(I);
1990 // Use the regular scalar folding to simplify this column.
1991 Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI);
1997 return ConstantVector::get(Result);
2000 } // end anonymous namespace
2003 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
2004 const TargetLibraryInfo *TLI) {
2007 StringRef Name = F->getName();
2009 Type *Ty = F->getReturnType();
2011 if (auto *VTy = dyn_cast<VectorType>(Ty))
2012 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
2013 F->getParent()->getDataLayout(), TLI);
2015 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);
2018 bool llvm::isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI) {
2019 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
2020 // (and to some extent ConstantFoldScalarCall).
2021 Function *F = CS.getCalledFunction();
2026 if (!TLI || !TLI->getLibFunc(*F, Func))
2029 if (CS.getNumArgOperands() == 1) {
2030 if (ConstantFP *OpC = dyn_cast<ConstantFP>(CS.getArgOperand(0))) {
2031 const APFloat &Op = OpC->getValueAPF();
2036 case LibFunc::log2l:
2038 case LibFunc::log2f:
2039 case LibFunc::log10l:
2040 case LibFunc::log10:
2041 case LibFunc::log10f:
2042 return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
2047 // FIXME: These boundaries are slightly conservative.
2048 if (OpC->getType()->isDoubleTy())
2049 return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan &&
2050 Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan;
2051 if (OpC->getType()->isFloatTy())
2052 return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan &&
2053 Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan;
2056 case LibFunc::exp2l:
2058 case LibFunc::exp2f:
2059 // FIXME: These boundaries are slightly conservative.
2060 if (OpC->getType()->isDoubleTy())
2061 return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan &&
2062 Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan;
2063 if (OpC->getType()->isFloatTy())
2064 return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan &&
2065 Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan;
2074 return !Op.isInfinity();
2078 case LibFunc::tanf: {
2079 // FIXME: Stop using the host math library.
2080 // FIXME: The computation isn't done in the right precision.
2081 Type *Ty = OpC->getType();
2082 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2083 double OpV = getValueAsDouble(OpC);
2084 return ConstantFoldFP(tan, OpV, Ty) != nullptr;
2089 case LibFunc::asinl:
2091 case LibFunc::asinf:
2092 case LibFunc::acosl:
2094 case LibFunc::acosf:
2095 return Op.compare(APFloat(Op.getSemantics(), "-1")) !=
2096 APFloat::cmpLessThan &&
2097 Op.compare(APFloat(Op.getSemantics(), "1")) !=
2098 APFloat::cmpGreaterThan;
2102 case LibFunc::sinhf:
2103 case LibFunc::coshf:
2104 case LibFunc::sinhl:
2105 case LibFunc::coshl:
2106 // FIXME: These boundaries are slightly conservative.
2107 if (OpC->getType()->isDoubleTy())
2108 return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan &&
2109 Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan;
2110 if (OpC->getType()->isFloatTy())
2111 return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan &&
2112 Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan;
2115 case LibFunc::sqrtl:
2117 case LibFunc::sqrtf:
2118 return Op.isNaN() || Op.isZero() || !Op.isNegative();
2120 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
2128 if (CS.getNumArgOperands() == 2) {
2129 ConstantFP *Op0C = dyn_cast<ConstantFP>(CS.getArgOperand(0));
2130 ConstantFP *Op1C = dyn_cast<ConstantFP>(CS.getArgOperand(1));
2132 const APFloat &Op0 = Op0C->getValueAPF();
2133 const APFloat &Op1 = Op1C->getValueAPF();
2138 case LibFunc::powf: {
2139 // FIXME: Stop using the host math library.
2140 // FIXME: The computation isn't done in the right precision.
2141 Type *Ty = Op0C->getType();
2142 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2143 if (Ty == Op1C->getType()) {
2144 double Op0V = getValueAsDouble(Op0C);
2145 double Op1V = getValueAsDouble(Op1C);
2146 return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr;
2152 case LibFunc::fmodl:
2154 case LibFunc::fmodf:
2155 return Op0.isNaN() || Op1.isNaN() ||
2156 (!Op0.isInfinity() && !Op1.isZero());