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/SmallVector.h"
26 #include "llvm/ADT/StringRef.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/KnownBits.h"
46 #include "llvm/Support/MathExtras.h"
58 //===----------------------------------------------------------------------===//
59 // Constant Folding internal helper functions
60 //===----------------------------------------------------------------------===//
62 static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
63 Constant *C, Type *SrcEltTy,
65 const DataLayout &DL) {
66 // Now that we know that the input value is a vector of integers, just shift
67 // and insert them into our result.
68 unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
69 for (unsigned i = 0; i != NumSrcElts; ++i) {
71 if (DL.isLittleEndian())
72 Element = C->getAggregateElement(NumSrcElts - i - 1);
74 Element = C->getAggregateElement(i);
76 if (Element && isa<UndefValue>(Element)) {
81 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
83 return ConstantExpr::getBitCast(C, DestTy);
86 Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth());
92 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
93 /// This always returns a non-null constant, but it may be a
94 /// ConstantExpr if unfoldable.
95 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
96 // Catch the obvious splat cases.
97 if (C->isNullValue() && !DestTy->isX86_MMXTy())
98 return Constant::getNullValue(DestTy);
99 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
100 !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
101 return Constant::getAllOnesValue(DestTy);
103 if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
104 // Handle a vector->scalar integer/fp cast.
105 if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
106 unsigned NumSrcElts = VTy->getNumElements();
107 Type *SrcEltTy = VTy->getElementType();
109 // If the vector is a vector of floating point, convert it to vector of int
110 // to simplify things.
111 if (SrcEltTy->isFloatingPointTy()) {
112 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
114 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
115 // Ask IR to do the conversion now that #elts line up.
116 C = ConstantExpr::getBitCast(C, SrcIVTy);
119 APInt Result(DL.getTypeSizeInBits(DestTy), 0);
120 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
121 SrcEltTy, NumSrcElts, DL))
124 if (isa<IntegerType>(DestTy))
125 return ConstantInt::get(DestTy, Result);
127 APFloat FP(DestTy->getFltSemantics(), Result);
128 return ConstantFP::get(DestTy->getContext(), FP);
132 // The code below only handles casts to vectors currently.
133 auto *DestVTy = dyn_cast<VectorType>(DestTy);
135 return ConstantExpr::getBitCast(C, DestTy);
137 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
138 // vector so the code below can handle it uniformly.
139 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
140 Constant *Ops = C; // don't take the address of C!
141 return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
144 // If this is a bitcast from constant vector -> vector, fold it.
145 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
146 return ConstantExpr::getBitCast(C, DestTy);
148 // If the element types match, IR can fold it.
149 unsigned NumDstElt = DestVTy->getNumElements();
150 unsigned NumSrcElt = C->getType()->getVectorNumElements();
151 if (NumDstElt == NumSrcElt)
152 return ConstantExpr::getBitCast(C, DestTy);
154 Type *SrcEltTy = C->getType()->getVectorElementType();
155 Type *DstEltTy = DestVTy->getElementType();
157 // Otherwise, we're changing the number of elements in a vector, which
158 // requires endianness information to do the right thing. For example,
159 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
160 // folds to (little endian):
161 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
162 // and to (big endian):
163 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
165 // First thing is first. We only want to think about integer here, so if
166 // we have something in FP form, recast it as integer.
167 if (DstEltTy->isFloatingPointTy()) {
168 // Fold to an vector of integers with same size as our FP type.
169 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
171 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
172 // Recursively handle this integer conversion, if possible.
173 C = FoldBitCast(C, DestIVTy, DL);
175 // Finally, IR can handle this now that #elts line up.
176 return ConstantExpr::getBitCast(C, DestTy);
179 // Okay, we know the destination is integer, if the input is FP, convert
180 // it to integer first.
181 if (SrcEltTy->isFloatingPointTy()) {
182 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
184 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
185 // Ask IR to do the conversion now that #elts line up.
186 C = ConstantExpr::getBitCast(C, SrcIVTy);
187 // If IR wasn't able to fold it, bail out.
188 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
189 !isa<ConstantDataVector>(C))
193 // Now we know that the input and output vectors are both integer vectors
194 // of the same size, and that their #elements is not the same. Do the
195 // conversion here, which depends on whether the input or output has
197 bool isLittleEndian = DL.isLittleEndian();
199 SmallVector<Constant*, 32> Result;
200 if (NumDstElt < NumSrcElt) {
201 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
202 Constant *Zero = Constant::getNullValue(DstEltTy);
203 unsigned Ratio = NumSrcElt/NumDstElt;
204 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
206 for (unsigned i = 0; i != NumDstElt; ++i) {
207 // Build each element of the result.
208 Constant *Elt = Zero;
209 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
210 for (unsigned j = 0; j != Ratio; ++j) {
211 Constant *Src = C->getAggregateElement(SrcElt++);
212 if (Src && isa<UndefValue>(Src))
213 Src = Constant::getNullValue(C->getType()->getVectorElementType());
215 Src = dyn_cast_or_null<ConstantInt>(Src);
216 if (!Src) // Reject constantexpr elements.
217 return ConstantExpr::getBitCast(C, DestTy);
219 // Zero extend the element to the right size.
220 Src = ConstantExpr::getZExt(Src, Elt->getType());
222 // Shift it to the right place, depending on endianness.
223 Src = ConstantExpr::getShl(Src,
224 ConstantInt::get(Src->getType(), ShiftAmt));
225 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
228 Elt = ConstantExpr::getOr(Elt, Src);
230 Result.push_back(Elt);
232 return ConstantVector::get(Result);
235 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
236 unsigned Ratio = NumDstElt/NumSrcElt;
237 unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
239 // Loop over each source value, expanding into multiple results.
240 for (unsigned i = 0; i != NumSrcElt; ++i) {
241 auto *Element = C->getAggregateElement(i);
243 if (!Element) // Reject constantexpr elements.
244 return ConstantExpr::getBitCast(C, DestTy);
246 if (isa<UndefValue>(Element)) {
247 // Correctly Propagate undef values.
248 Result.append(Ratio, UndefValue::get(DstEltTy));
252 auto *Src = dyn_cast<ConstantInt>(Element);
254 return ConstantExpr::getBitCast(C, DestTy);
256 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
257 for (unsigned j = 0; j != Ratio; ++j) {
258 // Shift the piece of the value into the right place, depending on
260 Constant *Elt = ConstantExpr::getLShr(Src,
261 ConstantInt::get(Src->getType(), ShiftAmt));
262 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
264 // Truncate the element to an integer with the same pointer size and
265 // convert the element back to a pointer using a inttoptr.
266 if (DstEltTy->isPointerTy()) {
267 IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
268 Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
269 Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
273 // Truncate and remember this piece.
274 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
278 return ConstantVector::get(Result);
281 } // end anonymous namespace
283 /// If this constant is a constant offset from a global, return the global and
284 /// the constant. Because of constantexprs, this function is recursive.
285 bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
286 APInt &Offset, const DataLayout &DL) {
287 // Trivial case, constant is the global.
288 if ((GV = dyn_cast<GlobalValue>(C))) {
289 unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());
290 Offset = APInt(BitWidth, 0);
294 // Otherwise, if this isn't a constant expr, bail out.
295 auto *CE = dyn_cast<ConstantExpr>(C);
296 if (!CE) return false;
298 // Look through ptr->int and ptr->ptr casts.
299 if (CE->getOpcode() == Instruction::PtrToInt ||
300 CE->getOpcode() == Instruction::BitCast)
301 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
303 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
304 auto *GEP = dyn_cast<GEPOperator>(CE);
308 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
309 APInt TmpOffset(BitWidth, 0);
311 // If the base isn't a global+constant, we aren't either.
312 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
315 // Otherwise, add any offset that our operands provide.
316 if (!GEP->accumulateConstantOffset(DL, TmpOffset))
323 Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
324 const DataLayout &DL) {
326 Type *SrcTy = C->getType();
328 // If the type sizes are the same and a cast is legal, just directly
329 // cast the constant.
330 if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
331 Instruction::CastOps Cast = Instruction::BitCast;
332 // If we are going from a pointer to int or vice versa, we spell the cast
334 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
335 Cast = Instruction::IntToPtr;
336 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
337 Cast = Instruction::PtrToInt;
339 if (CastInst::castIsValid(Cast, C, DestTy))
340 return ConstantExpr::getCast(Cast, C, DestTy);
343 // If this isn't an aggregate type, there is nothing we can do to drill down
344 // and find a bitcastable constant.
345 if (!SrcTy->isAggregateType())
348 // We're simulating a load through a pointer that was bitcast to point to
349 // a different type, so we can try to walk down through the initial
350 // elements of an aggregate to see if some part of the aggregate is
351 // castable to implement the "load" semantic model.
352 if (SrcTy->isStructTy()) {
353 // Struct types might have leading zero-length elements like [0 x i32],
354 // which are certainly not what we are looking for, so skip them.
358 ElemC = C->getAggregateElement(Elem++);
359 } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()) == 0);
362 C = C->getAggregateElement(0u);
371 /// Recursive helper to read bits out of global. C is the constant being copied
372 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
373 /// results into and BytesLeft is the number of bytes left in
374 /// the CurPtr buffer. DL is the DataLayout.
375 bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
376 unsigned BytesLeft, const DataLayout &DL) {
377 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
378 "Out of range access");
380 // If this element is zero or undefined, we can just return since *CurPtr is
382 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
385 if (auto *CI = dyn_cast<ConstantInt>(C)) {
386 if (CI->getBitWidth() > 64 ||
387 (CI->getBitWidth() & 7) != 0)
390 uint64_t Val = CI->getZExtValue();
391 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
393 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
395 if (!DL.isLittleEndian())
396 n = IntBytes - n - 1;
397 CurPtr[i] = (unsigned char)(Val >> (n * 8));
403 if (auto *CFP = dyn_cast<ConstantFP>(C)) {
404 if (CFP->getType()->isDoubleTy()) {
405 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
406 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
408 if (CFP->getType()->isFloatTy()){
409 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
410 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
412 if (CFP->getType()->isHalfTy()){
413 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
414 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
419 if (auto *CS = dyn_cast<ConstantStruct>(C)) {
420 const StructLayout *SL = DL.getStructLayout(CS->getType());
421 unsigned Index = SL->getElementContainingOffset(ByteOffset);
422 uint64_t CurEltOffset = SL->getElementOffset(Index);
423 ByteOffset -= CurEltOffset;
426 // If the element access is to the element itself and not to tail padding,
427 // read the bytes from the element.
428 uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
430 if (ByteOffset < EltSize &&
431 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
437 // Check to see if we read from the last struct element, if so we're done.
438 if (Index == CS->getType()->getNumElements())
441 // If we read all of the bytes we needed from this element we're done.
442 uint64_t NextEltOffset = SL->getElementOffset(Index);
444 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
447 // Move to the next element of the struct.
448 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
449 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
451 CurEltOffset = NextEltOffset;
456 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
457 isa<ConstantDataSequential>(C)) {
458 Type *EltTy = C->getType()->getSequentialElementType();
459 uint64_t EltSize = DL.getTypeAllocSize(EltTy);
460 uint64_t Index = ByteOffset / EltSize;
461 uint64_t Offset = ByteOffset - Index * EltSize;
463 if (auto *AT = dyn_cast<ArrayType>(C->getType()))
464 NumElts = AT->getNumElements();
466 NumElts = C->getType()->getVectorNumElements();
468 for (; Index != NumElts; ++Index) {
469 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
473 uint64_t BytesWritten = EltSize - Offset;
474 assert(BytesWritten <= EltSize && "Not indexing into this element?");
475 if (BytesWritten >= BytesLeft)
479 BytesLeft -= BytesWritten;
480 CurPtr += BytesWritten;
485 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
486 if (CE->getOpcode() == Instruction::IntToPtr &&
487 CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
488 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
493 // Otherwise, unknown initializer type.
497 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
498 const DataLayout &DL) {
499 auto *PTy = cast<PointerType>(C->getType());
500 auto *IntType = dyn_cast<IntegerType>(LoadTy);
502 // If this isn't an integer load we can't fold it directly.
504 unsigned AS = PTy->getAddressSpace();
506 // If this is a float/double load, we can try folding it as an int32/64 load
507 // and then bitcast the result. This can be useful for union cases. Note
508 // that address spaces don't matter here since we're not going to result in
509 // an actual new load.
511 if (LoadTy->isHalfTy())
512 MapTy = Type::getInt16Ty(C->getContext());
513 else if (LoadTy->isFloatTy())
514 MapTy = Type::getInt32Ty(C->getContext());
515 else if (LoadTy->isDoubleTy())
516 MapTy = Type::getInt64Ty(C->getContext());
517 else if (LoadTy->isVectorTy()) {
518 MapTy = PointerType::getIntNTy(C->getContext(),
519 DL.getTypeAllocSizeInBits(LoadTy));
523 C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
524 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL))
525 return FoldBitCast(Res, LoadTy, DL);
529 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
530 if (BytesLoaded > 32 || BytesLoaded == 0)
535 if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL))
538 auto *GV = dyn_cast<GlobalVariable>(GVal);
539 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
540 !GV->getInitializer()->getType()->isSized())
543 int64_t Offset = OffsetAI.getSExtValue();
544 int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
546 // If we're not accessing anything in this constant, the result is undefined.
547 if (Offset + BytesLoaded <= 0)
548 return UndefValue::get(IntType);
550 // If we're not accessing anything in this constant, the result is undefined.
551 if (Offset >= InitializerSize)
552 return UndefValue::get(IntType);
554 unsigned char RawBytes[32] = {0};
555 unsigned char *CurPtr = RawBytes;
556 unsigned BytesLeft = BytesLoaded;
558 // If we're loading off the beginning of the global, some bytes may be valid.
565 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL))
568 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
569 if (DL.isLittleEndian()) {
570 ResultVal = RawBytes[BytesLoaded - 1];
571 for (unsigned i = 1; i != BytesLoaded; ++i) {
573 ResultVal |= RawBytes[BytesLoaded - 1 - i];
576 ResultVal = RawBytes[0];
577 for (unsigned i = 1; i != BytesLoaded; ++i) {
579 ResultVal |= RawBytes[i];
583 return ConstantInt::get(IntType->getContext(), ResultVal);
586 Constant *ConstantFoldLoadThroughBitcastExpr(ConstantExpr *CE, Type *DestTy,
587 const DataLayout &DL) {
588 auto *SrcPtr = CE->getOperand(0);
589 auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType());
592 Type *SrcTy = SrcPtrTy->getPointerElementType();
594 Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL);
598 return llvm::ConstantFoldLoadThroughBitcast(C, DestTy, DL);
601 } // end anonymous namespace
603 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
604 const DataLayout &DL) {
605 // First, try the easy cases:
606 if (auto *GV = dyn_cast<GlobalVariable>(C))
607 if (GV->isConstant() && GV->hasDefinitiveInitializer())
608 return GV->getInitializer();
610 if (auto *GA = dyn_cast<GlobalAlias>(C))
611 if (GA->getAliasee() && !GA->isInterposable())
612 return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL);
614 // If the loaded value isn't a constant expr, we can't handle it.
615 auto *CE = dyn_cast<ConstantExpr>(C);
619 if (CE->getOpcode() == Instruction::GetElementPtr) {
620 if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
621 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
623 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
629 if (CE->getOpcode() == Instruction::BitCast)
630 if (Constant *LoadedC = ConstantFoldLoadThroughBitcastExpr(CE, Ty, DL))
633 // Instead of loading constant c string, use corresponding integer value
634 // directly if string length is small enough.
636 if (getConstantStringInfo(CE, Str) && !Str.empty()) {
637 size_t StrLen = Str.size();
638 unsigned NumBits = Ty->getPrimitiveSizeInBits();
639 // Replace load with immediate integer if the result is an integer or fp
641 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
642 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
643 APInt StrVal(NumBits, 0);
644 APInt SingleChar(NumBits, 0);
645 if (DL.isLittleEndian()) {
646 for (unsigned char C : reverse(Str.bytes())) {
647 SingleChar = static_cast<uint64_t>(C);
648 StrVal = (StrVal << 8) | SingleChar;
651 for (unsigned char C : Str.bytes()) {
652 SingleChar = static_cast<uint64_t>(C);
653 StrVal = (StrVal << 8) | SingleChar;
655 // Append NULL at the end.
657 StrVal = (StrVal << 8) | SingleChar;
660 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
661 if (Ty->isFloatingPointTy())
662 Res = ConstantExpr::getBitCast(Res, Ty);
667 // If this load comes from anywhere in a constant global, and if the global
668 // is all undef or zero, we know what it loads.
669 if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
670 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
671 if (GV->getInitializer()->isNullValue())
672 return Constant::getNullValue(Ty);
673 if (isa<UndefValue>(GV->getInitializer()))
674 return UndefValue::get(Ty);
678 // Try hard to fold loads from bitcasted strange and non-type-safe things.
679 return FoldReinterpretLoadFromConstPtr(CE, Ty, DL);
684 Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) {
685 if (LI->isVolatile()) return nullptr;
687 if (auto *C = dyn_cast<Constant>(LI->getOperand(0)))
688 return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL);
693 /// One of Op0/Op1 is a constant expression.
694 /// Attempt to symbolically evaluate the result of a binary operator merging
695 /// these together. If target data info is available, it is provided as DL,
696 /// otherwise DL is null.
697 Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
698 const DataLayout &DL) {
701 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
702 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
705 if (Opc == Instruction::And) {
706 KnownBits Known0 = computeKnownBits(Op0, DL);
707 KnownBits Known1 = computeKnownBits(Op1, DL);
708 if ((Known1.One | Known0.Zero).isAllOnesValue()) {
709 // All the bits of Op0 that the 'and' could be masking are already zero.
712 if ((Known0.One | Known1.Zero).isAllOnesValue()) {
713 // All the bits of Op1 that the 'and' could be masking are already zero.
717 Known0.Zero |= Known1.Zero;
718 Known0.One &= Known1.One;
719 if (Known0.isConstant())
720 return ConstantInt::get(Op0->getType(), Known0.getConstant());
723 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
724 // constant. This happens frequently when iterating over a global array.
725 if (Opc == Instruction::Sub) {
726 GlobalValue *GV1, *GV2;
729 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
730 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
731 unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
733 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
734 // PtrToInt may change the bitwidth so we have convert to the right size
736 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
737 Offs2.zextOrTrunc(OpSize));
744 /// If array indices are not pointer-sized integers, explicitly cast them so
745 /// that they aren't implicitly casted by the getelementptr.
746 Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
747 Type *ResultTy, Optional<unsigned> InRangeIndex,
748 const DataLayout &DL, const TargetLibraryInfo *TLI) {
749 Type *IntPtrTy = DL.getIntPtrType(ResultTy);
750 Type *IntPtrScalarTy = IntPtrTy->getScalarType();
753 SmallVector<Constant*, 32> NewIdxs;
754 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
756 !isa<StructType>(GetElementPtrInst::getIndexedType(
757 SrcElemTy, Ops.slice(1, i - 1)))) &&
758 Ops[i]->getType()->getScalarType() != IntPtrScalarTy) {
760 Type *NewType = Ops[i]->getType()->isVectorTy()
762 : IntPtrTy->getScalarType();
763 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
769 NewIdxs.push_back(Ops[i]);
775 Constant *C = ConstantExpr::getGetElementPtr(
776 SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex);
777 if (Constant *Folded = ConstantFoldConstant(C, DL, TLI))
783 /// Strip the pointer casts, but preserve the address space information.
784 Constant* StripPtrCastKeepAS(Constant* Ptr, Type *&ElemTy) {
785 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
786 auto *OldPtrTy = cast<PointerType>(Ptr->getType());
787 Ptr = Ptr->stripPointerCasts();
788 auto *NewPtrTy = cast<PointerType>(Ptr->getType());
790 ElemTy = NewPtrTy->getPointerElementType();
792 // Preserve the address space number of the pointer.
793 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
794 NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace());
795 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
800 /// If we can symbolically evaluate the GEP constant expression, do so.
801 Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
802 ArrayRef<Constant *> Ops,
803 const DataLayout &DL,
804 const TargetLibraryInfo *TLI) {
805 const GEPOperator *InnermostGEP = GEP;
806 bool InBounds = GEP->isInBounds();
808 Type *SrcElemTy = GEP->getSourceElementType();
809 Type *ResElemTy = GEP->getResultElementType();
810 Type *ResTy = GEP->getType();
811 if (!SrcElemTy->isSized())
814 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy,
815 GEP->getInRangeIndex(), DL, TLI))
818 Constant *Ptr = Ops[0];
819 if (!Ptr->getType()->isPointerTy())
822 Type *IntPtrTy = DL.getIntPtrType(Ptr->getType());
824 // If this is a constant expr gep that is effectively computing an
825 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
826 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
827 if (!isa<ConstantInt>(Ops[i])) {
829 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
830 // "inttoptr (sub (ptrtoint Ptr), V)"
831 if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
832 auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
833 assert((!CE || CE->getType() == IntPtrTy) &&
834 "CastGEPIndices didn't canonicalize index types!");
835 if (CE && CE->getOpcode() == Instruction::Sub &&
836 CE->getOperand(0)->isNullValue()) {
837 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
838 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
839 Res = ConstantExpr::getIntToPtr(Res, ResTy);
840 if (auto *FoldedRes = ConstantFoldConstant(Res, DL, TLI))
848 unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy);
851 DL.getIndexedOffsetInType(
853 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
854 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
856 // If this is a GEP of a GEP, fold it all into a single GEP.
857 while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
859 InBounds &= GEP->isInBounds();
861 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
863 // Do not try the incorporate the sub-GEP if some index is not a number.
864 bool AllConstantInt = true;
865 for (Value *NestedOp : NestedOps)
866 if (!isa<ConstantInt>(NestedOp)) {
867 AllConstantInt = false;
873 Ptr = cast<Constant>(GEP->getOperand(0));
874 SrcElemTy = GEP->getSourceElementType();
875 Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
876 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
879 // If the base value for this address is a literal integer value, fold the
880 // getelementptr to the resulting integer value casted to the pointer type.
881 APInt BasePtr(BitWidth, 0);
882 if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
883 if (CE->getOpcode() == Instruction::IntToPtr) {
884 if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
885 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
889 auto *PTy = cast<PointerType>(Ptr->getType());
890 if ((Ptr->isNullValue() || BasePtr != 0) &&
891 !DL.isNonIntegralPointerType(PTy)) {
892 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
893 return ConstantExpr::getIntToPtr(C, ResTy);
896 // Otherwise form a regular getelementptr. Recompute the indices so that
897 // we eliminate over-indexing of the notional static type array bounds.
898 // This makes it easy to determine if the getelementptr is "inbounds".
899 // Also, this helps GlobalOpt do SROA on GlobalVariables.
901 SmallVector<Constant *, 32> NewIdxs;
904 if (!Ty->isStructTy()) {
905 if (Ty->isPointerTy()) {
906 // The only pointer indexing we'll do is on the first index of the GEP.
907 if (!NewIdxs.empty())
912 // Only handle pointers to sized types, not pointers to functions.
915 } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) {
916 Ty = ATy->getElementType();
918 // We've reached some non-indexable type.
922 // Determine which element of the array the offset points into.
923 APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty));
925 // The element size is 0. This may be [0 x Ty]*, so just use a zero
926 // index for this level and proceed to the next level to see if it can
927 // accommodate the offset.
928 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
930 // The element size is non-zero divide the offset by the element
931 // size (rounding down), to compute the index at this level.
933 APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow);
936 Offset -= NewIdx * ElemSize;
937 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
940 auto *STy = cast<StructType>(Ty);
941 // If we end up with an offset that isn't valid for this struct type, we
942 // can't re-form this GEP in a regular form, so bail out. The pointer
943 // operand likely went through casts that are necessary to make the GEP
945 const StructLayout &SL = *DL.getStructLayout(STy);
946 if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes()))
949 // Determine which field of the struct the offset points into. The
950 // getZExtValue is fine as we've already ensured that the offset is
951 // within the range representable by the StructLayout API.
952 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
953 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
955 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
956 Ty = STy->getTypeAtIndex(ElIdx);
958 } while (Ty != ResElemTy);
960 // If we haven't used up the entire offset by descending the static
961 // type, then the offset is pointing into the middle of an indivisible
962 // member, so we can't simplify it.
966 // Preserve the inrange index from the innermost GEP if possible. We must
967 // have calculated the same indices up to and including the inrange index.
968 Optional<unsigned> InRangeIndex;
969 if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex())
970 if (SrcElemTy == InnermostGEP->getSourceElementType() &&
971 NewIdxs.size() > *LastIRIndex) {
972 InRangeIndex = LastIRIndex;
973 for (unsigned I = 0; I <= *LastIRIndex; ++I)
974 if (NewIdxs[I] != InnermostGEP->getOperand(I + 1))
979 Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs,
980 InBounds, InRangeIndex);
981 assert(C->getType()->getPointerElementType() == Ty &&
982 "Computed GetElementPtr has unexpected type!");
984 // If we ended up indexing a member with a type that doesn't match
985 // the type of what the original indices indexed, add a cast.
987 C = FoldBitCast(C, ResTy, DL);
992 /// Attempt to constant fold an instruction with the
993 /// specified opcode and operands. If successful, the constant result is
994 /// returned, if not, null is returned. Note that this function can fail when
995 /// attempting to fold instructions like loads and stores, which have no
996 /// constant expression form.
997 Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
998 ArrayRef<Constant *> Ops,
999 const DataLayout &DL,
1000 const TargetLibraryInfo *TLI) {
1001 Type *DestTy = InstOrCE->getType();
1003 // Handle easy binops first.
1004 if (Instruction::isBinaryOp(Opcode))
1005 return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
1007 if (Instruction::isCast(Opcode))
1008 return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
1010 if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
1011 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1014 return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0],
1015 Ops.slice(1), GEP->isInBounds(),
1016 GEP->getInRangeIndex());
1019 if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))
1020 return CE->getWithOperands(Ops);
1023 default: return nullptr;
1024 case Instruction::ICmp:
1025 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1026 case Instruction::Call:
1027 if (auto *F = dyn_cast<Function>(Ops.back())) {
1028 ImmutableCallSite CS(cast<CallInst>(InstOrCE));
1029 if (canConstantFoldCallTo(CS, F))
1030 return ConstantFoldCall(CS, F, Ops.slice(0, Ops.size() - 1), TLI);
1033 case Instruction::Select:
1034 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1035 case Instruction::ExtractElement:
1036 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1037 case Instruction::InsertElement:
1038 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1039 case Instruction::ShuffleVector:
1040 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1044 } // end anonymous namespace
1046 //===----------------------------------------------------------------------===//
1047 // Constant Folding public APIs
1048 //===----------------------------------------------------------------------===//
1053 ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1054 const TargetLibraryInfo *TLI,
1055 SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1056 if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))
1059 SmallVector<Constant *, 8> Ops;
1060 for (const Use &NewU : C->operands()) {
1061 auto *NewC = cast<Constant>(&NewU);
1062 // Recursively fold the ConstantExpr's operands. If we have already folded
1063 // a ConstantExpr, we don't have to process it again.
1064 if (isa<ConstantVector>(NewC) || isa<ConstantExpr>(NewC)) {
1065 auto It = FoldedOps.find(NewC);
1066 if (It == FoldedOps.end()) {
1068 ConstantFoldConstantImpl(NewC, DL, TLI, FoldedOps)) {
1069 FoldedOps.insert({NewC, FoldedC});
1072 FoldedOps.insert({NewC, NewC});
1078 Ops.push_back(NewC);
1081 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1082 if (CE->isCompare())
1083 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1086 return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI);
1089 assert(isa<ConstantVector>(C));
1090 return ConstantVector::get(Ops);
1093 } // end anonymous namespace
1095 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
1096 const TargetLibraryInfo *TLI) {
1097 // Handle PHI nodes quickly here...
1098 if (auto *PN = dyn_cast<PHINode>(I)) {
1099 Constant *CommonValue = nullptr;
1101 SmallDenseMap<Constant *, Constant *> FoldedOps;
1102 for (Value *Incoming : PN->incoming_values()) {
1103 // If the incoming value is undef then skip it. Note that while we could
1104 // skip the value if it is equal to the phi node itself we choose not to
1105 // because that would break the rule that constant folding only applies if
1106 // all operands are constants.
1107 if (isa<UndefValue>(Incoming))
1109 // If the incoming value is not a constant, then give up.
1110 auto *C = dyn_cast<Constant>(Incoming);
1113 // Fold the PHI's operands.
1114 if (auto *FoldedC = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps))
1116 // If the incoming value is a different constant to
1117 // the one we saw previously, then give up.
1118 if (CommonValue && C != CommonValue)
1123 // If we reach here, all incoming values are the same constant or undef.
1124 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
1127 // Scan the operand list, checking to see if they are all constants, if so,
1128 // hand off to ConstantFoldInstOperandsImpl.
1129 if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1132 SmallDenseMap<Constant *, Constant *> FoldedOps;
1133 SmallVector<Constant *, 8> Ops;
1134 for (const Use &OpU : I->operands()) {
1135 auto *Op = cast<Constant>(&OpU);
1136 // Fold the Instruction's operands.
1137 if (auto *FoldedOp = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps))
1143 if (const auto *CI = dyn_cast<CmpInst>(I))
1144 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1147 if (const auto *LI = dyn_cast<LoadInst>(I))
1148 return ConstantFoldLoadInst(LI, DL);
1150 if (auto *IVI = dyn_cast<InsertValueInst>(I)) {
1151 return ConstantExpr::getInsertValue(
1152 cast<Constant>(IVI->getAggregateOperand()),
1153 cast<Constant>(IVI->getInsertedValueOperand()),
1157 if (auto *EVI = dyn_cast<ExtractValueInst>(I)) {
1158 return ConstantExpr::getExtractValue(
1159 cast<Constant>(EVI->getAggregateOperand()),
1163 return ConstantFoldInstOperands(I, Ops, DL, TLI);
1166 Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1167 const TargetLibraryInfo *TLI) {
1168 SmallDenseMap<Constant *, Constant *> FoldedOps;
1169 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1172 Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1173 ArrayRef<Constant *> Ops,
1174 const DataLayout &DL,
1175 const TargetLibraryInfo *TLI) {
1176 return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI);
1179 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1180 Constant *Ops0, Constant *Ops1,
1181 const DataLayout &DL,
1182 const TargetLibraryInfo *TLI) {
1183 // fold: icmp (inttoptr x), null -> icmp x, 0
1184 // fold: icmp null, (inttoptr x) -> icmp 0, x
1185 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1186 // fold: icmp 0, (ptrtoint x) -> icmp null, x
1187 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1188 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1190 // FIXME: The following comment is out of data and the DataLayout is here now.
1191 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1192 // around to know if bit truncation is happening.
1193 if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1194 if (Ops1->isNullValue()) {
1195 if (CE0->getOpcode() == Instruction::IntToPtr) {
1196 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1197 // Convert the integer value to the right size to ensure we get the
1198 // proper extension or truncation.
1199 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1201 Constant *Null = Constant::getNullValue(C->getType());
1202 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1205 // Only do this transformation if the int is intptrty in size, otherwise
1206 // there is a truncation or extension that we aren't modeling.
1207 if (CE0->getOpcode() == Instruction::PtrToInt) {
1208 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1209 if (CE0->getType() == IntPtrTy) {
1210 Constant *C = CE0->getOperand(0);
1211 Constant *Null = Constant::getNullValue(C->getType());
1212 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1217 if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1218 if (CE0->getOpcode() == CE1->getOpcode()) {
1219 if (CE0->getOpcode() == Instruction::IntToPtr) {
1220 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1222 // Convert the integer value to the right size to ensure we get the
1223 // proper extension or truncation.
1224 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1226 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1228 return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1231 // Only do this transformation if the int is intptrty in size, otherwise
1232 // there is a truncation or extension that we aren't modeling.
1233 if (CE0->getOpcode() == Instruction::PtrToInt) {
1234 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1235 if (CE0->getType() == IntPtrTy &&
1236 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1237 return ConstantFoldCompareInstOperands(
1238 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1244 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1245 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1246 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1247 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1248 Constant *LHS = ConstantFoldCompareInstOperands(
1249 Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1250 Constant *RHS = ConstantFoldCompareInstOperands(
1251 Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1253 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1254 return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1256 } else if (isa<ConstantExpr>(Ops1)) {
1257 // If RHS is a constant expression, but the left side isn't, swap the
1258 // operands and try again.
1259 Predicate = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)Predicate);
1260 return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI);
1263 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1266 Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1268 const DataLayout &DL) {
1269 assert(Instruction::isBinaryOp(Opcode));
1270 if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1271 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1274 return ConstantExpr::get(Opcode, LHS, RHS);
1277 Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1278 Type *DestTy, const DataLayout &DL) {
1279 assert(Instruction::isCast(Opcode));
1282 llvm_unreachable("Missing case");
1283 case Instruction::PtrToInt:
1284 // If the input is a inttoptr, eliminate the pair. This requires knowing
1285 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1286 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1287 if (CE->getOpcode() == Instruction::IntToPtr) {
1288 Constant *Input = CE->getOperand(0);
1289 unsigned InWidth = Input->getType()->getScalarSizeInBits();
1290 unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1291 if (PtrWidth < InWidth) {
1293 ConstantInt::get(CE->getContext(),
1294 APInt::getLowBitsSet(InWidth, PtrWidth));
1295 Input = ConstantExpr::getAnd(Input, Mask);
1297 // Do a zext or trunc to get to the dest size.
1298 return ConstantExpr::getIntegerCast(Input, DestTy, false);
1301 return ConstantExpr::getCast(Opcode, C, DestTy);
1302 case Instruction::IntToPtr:
1303 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1304 // the int size is >= the ptr size and the address spaces are the same.
1305 // This requires knowing the width of a pointer, so it can't be done in
1306 // ConstantExpr::getCast.
1307 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1308 if (CE->getOpcode() == Instruction::PtrToInt) {
1309 Constant *SrcPtr = CE->getOperand(0);
1310 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1311 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1313 if (MidIntSize >= SrcPtrSize) {
1314 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1315 if (SrcAS == DestTy->getPointerAddressSpace())
1316 return FoldBitCast(CE->getOperand(0), DestTy, DL);
1321 return ConstantExpr::getCast(Opcode, C, DestTy);
1322 case Instruction::Trunc:
1323 case Instruction::ZExt:
1324 case Instruction::SExt:
1325 case Instruction::FPTrunc:
1326 case Instruction::FPExt:
1327 case Instruction::UIToFP:
1328 case Instruction::SIToFP:
1329 case Instruction::FPToUI:
1330 case Instruction::FPToSI:
1331 case Instruction::AddrSpaceCast:
1332 return ConstantExpr::getCast(Opcode, C, DestTy);
1333 case Instruction::BitCast:
1334 return FoldBitCast(C, DestTy, DL);
1338 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1340 if (!CE->getOperand(1)->isNullValue())
1341 return nullptr; // Do not allow stepping over the value!
1343 // Loop over all of the operands, tracking down which value we are
1345 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1346 C = C->getAggregateElement(CE->getOperand(i));
1354 llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1355 ArrayRef<Constant *> Indices) {
1356 // Loop over all of the operands, tracking down which value we are
1358 for (Constant *Index : Indices) {
1359 C = C->getAggregateElement(Index);
1366 //===----------------------------------------------------------------------===//
1367 // Constant Folding for Calls
1370 bool llvm::canConstantFoldCallTo(ImmutableCallSite CS, const Function *F) {
1371 if (CS.isNoBuiltin() || CS.isStrictFP())
1373 switch (F->getIntrinsicID()) {
1374 case Intrinsic::fabs:
1375 case Intrinsic::minnum:
1376 case Intrinsic::maxnum:
1377 case Intrinsic::minimum:
1378 case Intrinsic::maximum:
1379 case Intrinsic::log:
1380 case Intrinsic::log2:
1381 case Intrinsic::log10:
1382 case Intrinsic::exp:
1383 case Intrinsic::exp2:
1384 case Intrinsic::floor:
1385 case Intrinsic::ceil:
1386 case Intrinsic::sqrt:
1387 case Intrinsic::sin:
1388 case Intrinsic::cos:
1389 case Intrinsic::trunc:
1390 case Intrinsic::rint:
1391 case Intrinsic::nearbyint:
1392 case Intrinsic::pow:
1393 case Intrinsic::powi:
1394 case Intrinsic::bswap:
1395 case Intrinsic::ctpop:
1396 case Intrinsic::ctlz:
1397 case Intrinsic::cttz:
1398 case Intrinsic::fshl:
1399 case Intrinsic::fshr:
1400 case Intrinsic::fma:
1401 case Intrinsic::fmuladd:
1402 case Intrinsic::copysign:
1403 case Intrinsic::launder_invariant_group:
1404 case Intrinsic::strip_invariant_group:
1405 case Intrinsic::round:
1406 case Intrinsic::masked_load:
1407 case Intrinsic::sadd_with_overflow:
1408 case Intrinsic::uadd_with_overflow:
1409 case Intrinsic::ssub_with_overflow:
1410 case Intrinsic::usub_with_overflow:
1411 case Intrinsic::smul_with_overflow:
1412 case Intrinsic::umul_with_overflow:
1413 case Intrinsic::sadd_sat:
1414 case Intrinsic::uadd_sat:
1415 case Intrinsic::ssub_sat:
1416 case Intrinsic::usub_sat:
1417 case Intrinsic::convert_from_fp16:
1418 case Intrinsic::convert_to_fp16:
1419 case Intrinsic::bitreverse:
1420 case Intrinsic::x86_sse_cvtss2si:
1421 case Intrinsic::x86_sse_cvtss2si64:
1422 case Intrinsic::x86_sse_cvttss2si:
1423 case Intrinsic::x86_sse_cvttss2si64:
1424 case Intrinsic::x86_sse2_cvtsd2si:
1425 case Intrinsic::x86_sse2_cvtsd2si64:
1426 case Intrinsic::x86_sse2_cvttsd2si:
1427 case Intrinsic::x86_sse2_cvttsd2si64:
1428 case Intrinsic::x86_avx512_vcvtss2si32:
1429 case Intrinsic::x86_avx512_vcvtss2si64:
1430 case Intrinsic::x86_avx512_cvttss2si:
1431 case Intrinsic::x86_avx512_cvttss2si64:
1432 case Intrinsic::x86_avx512_vcvtsd2si32:
1433 case Intrinsic::x86_avx512_vcvtsd2si64:
1434 case Intrinsic::x86_avx512_cvttsd2si:
1435 case Intrinsic::x86_avx512_cvttsd2si64:
1436 case Intrinsic::x86_avx512_vcvtss2usi32:
1437 case Intrinsic::x86_avx512_vcvtss2usi64:
1438 case Intrinsic::x86_avx512_cvttss2usi:
1439 case Intrinsic::x86_avx512_cvttss2usi64:
1440 case Intrinsic::x86_avx512_vcvtsd2usi32:
1441 case Intrinsic::x86_avx512_vcvtsd2usi64:
1442 case Intrinsic::x86_avx512_cvttsd2usi:
1443 case Intrinsic::x86_avx512_cvttsd2usi64:
1444 case Intrinsic::is_constant:
1448 case Intrinsic::not_intrinsic: break;
1453 StringRef Name = F->getName();
1455 // In these cases, the check of the length is required. We don't want to
1456 // return true for a name like "cos\0blah" which strcmp would return equal to
1457 // "cos", but has length 8.
1462 return Name == "acos" || Name == "asin" || Name == "atan" ||
1463 Name == "atan2" || Name == "acosf" || Name == "asinf" ||
1464 Name == "atanf" || Name == "atan2f";
1466 return Name == "ceil" || Name == "cos" || Name == "cosh" ||
1467 Name == "ceilf" || Name == "cosf" || Name == "coshf";
1469 return Name == "exp" || Name == "exp2" || Name == "expf" || Name == "exp2f";
1471 return Name == "fabs" || Name == "floor" || Name == "fmod" ||
1472 Name == "fabsf" || Name == "floorf" || Name == "fmodf";
1474 return Name == "log" || Name == "log10" || Name == "logf" ||
1477 return Name == "pow" || Name == "powf";
1479 return Name == "round" || Name == "roundf";
1481 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1482 Name == "sinf" || Name == "sinhf" || Name == "sqrtf";
1484 return Name == "tan" || Name == "tanh" || Name == "tanf" || Name == "tanhf";
1487 // Check for various function names that get used for the math functions
1488 // when the header files are preprocessed with the macro
1489 // __FINITE_MATH_ONLY__ enabled.
1490 // The '12' here is the length of the shortest name that can match.
1491 // We need to check the size before looking at Name[1] and Name[2]
1492 // so we may as well check a limit that will eliminate mismatches.
1493 if (Name.size() < 12 || Name[1] != '_')
1499 return Name == "__acos_finite" || Name == "__acosf_finite" ||
1500 Name == "__asin_finite" || Name == "__asinf_finite" ||
1501 Name == "__atan2_finite" || Name == "__atan2f_finite";
1503 return Name == "__cosh_finite" || Name == "__coshf_finite";
1505 return Name == "__exp_finite" || Name == "__expf_finite" ||
1506 Name == "__exp2_finite" || Name == "__exp2f_finite";
1508 return Name == "__log_finite" || Name == "__logf_finite" ||
1509 Name == "__log10_finite" || Name == "__log10f_finite";
1511 return Name == "__pow_finite" || Name == "__powf_finite";
1513 return Name == "__sinh_finite" || Name == "__sinhf_finite";
1520 Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1521 if (Ty->isHalfTy()) {
1524 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &unused);
1525 return ConstantFP::get(Ty->getContext(), APF);
1527 if (Ty->isFloatTy())
1528 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1529 if (Ty->isDoubleTy())
1530 return ConstantFP::get(Ty->getContext(), APFloat(V));
1531 llvm_unreachable("Can only constant fold half/float/double");
1534 /// Clear the floating-point exception state.
1535 inline void llvm_fenv_clearexcept() {
1536 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1537 feclearexcept(FE_ALL_EXCEPT);
1542 /// Test if a floating-point exception was raised.
1543 inline bool llvm_fenv_testexcept() {
1544 int errno_val = errno;
1545 if (errno_val == ERANGE || errno_val == EDOM)
1547 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1548 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1554 Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) {
1555 llvm_fenv_clearexcept();
1557 if (llvm_fenv_testexcept()) {
1558 llvm_fenv_clearexcept();
1562 return GetConstantFoldFPValue(V, Ty);
1565 Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
1566 double W, Type *Ty) {
1567 llvm_fenv_clearexcept();
1569 if (llvm_fenv_testexcept()) {
1570 llvm_fenv_clearexcept();
1574 return GetConstantFoldFPValue(V, Ty);
1577 /// Attempt to fold an SSE floating point to integer conversion of a constant
1578 /// floating point. If roundTowardZero is false, the default IEEE rounding is
1579 /// used (toward nearest, ties to even). This matches the behavior of the
1580 /// non-truncating SSE instructions in the default rounding mode. The desired
1581 /// integer type Ty is used to select how many bits are available for the
1582 /// result. Returns null if the conversion cannot be performed, otherwise
1583 /// returns the Constant value resulting from the conversion.
1584 Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
1585 Type *Ty, bool IsSigned) {
1586 // All of these conversion intrinsics form an integer of at most 64bits.
1587 unsigned ResultWidth = Ty->getIntegerBitWidth();
1588 assert(ResultWidth <= 64 &&
1589 "Can only constant fold conversions to 64 and 32 bit ints");
1592 bool isExact = false;
1593 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1594 : APFloat::rmNearestTiesToEven;
1595 APFloat::opStatus status =
1596 Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth,
1597 IsSigned, mode, &isExact);
1598 if (status != APFloat::opOK &&
1599 (!roundTowardZero || status != APFloat::opInexact))
1601 return ConstantInt::get(Ty, UIntVal, IsSigned);
1604 double getValueAsDouble(ConstantFP *Op) {
1605 Type *Ty = Op->getType();
1607 if (Ty->isFloatTy())
1608 return Op->getValueAPF().convertToFloat();
1610 if (Ty->isDoubleTy())
1611 return Op->getValueAPF().convertToDouble();
1614 APFloat APF = Op->getValueAPF();
1615 APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
1616 return APF.convertToDouble();
1619 static bool isManifestConstant(const Constant *c) {
1620 if (isa<ConstantData>(c)) {
1622 } else if (isa<ConstantAggregate>(c) || isa<ConstantExpr>(c)) {
1623 for (const Value *subc : c->operand_values()) {
1624 if (!isManifestConstant(cast<Constant>(subc)))
1632 static bool getConstIntOrUndef(Value *Op, const APInt *&C) {
1633 if (auto *CI = dyn_cast<ConstantInt>(Op)) {
1634 C = &CI->getValue();
1637 if (isa<UndefValue>(Op)) {
1644 Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
1645 ArrayRef<Constant *> Operands,
1646 const TargetLibraryInfo *TLI,
1647 ImmutableCallSite CS) {
1648 if (Operands.size() == 1) {
1649 if (IntrinsicID == Intrinsic::is_constant) {
1650 // We know we have a "Constant" argument. But we want to only
1651 // return true for manifest constants, not those that depend on
1652 // constants with unknowable values, e.g. GlobalValue or BlockAddress.
1653 if (isManifestConstant(Operands[0]))
1654 return ConstantInt::getTrue(Ty->getContext());
1657 if (isa<UndefValue>(Operands[0])) {
1658 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.
1659 // ctpop() is between 0 and bitwidth, pick 0 for undef.
1660 if (IntrinsicID == Intrinsic::cos ||
1661 IntrinsicID == Intrinsic::ctpop)
1662 return Constant::getNullValue(Ty);
1663 if (IntrinsicID == Intrinsic::bswap ||
1664 IntrinsicID == Intrinsic::bitreverse ||
1665 IntrinsicID == Intrinsic::launder_invariant_group ||
1666 IntrinsicID == Intrinsic::strip_invariant_group)
1670 if (isa<ConstantPointerNull>(Operands[0])) {
1671 // launder(null) == null == strip(null) iff in addrspace 0
1672 if (IntrinsicID == Intrinsic::launder_invariant_group ||
1673 IntrinsicID == Intrinsic::strip_invariant_group) {
1674 // If instruction is not yet put in a basic block (e.g. when cloning
1675 // a function during inlining), CS caller may not be available.
1676 // So check CS's BB first before querying CS.getCaller.
1677 const Function *Caller = CS.getParent() ? CS.getCaller() : nullptr;
1679 !NullPointerIsDefined(
1680 Caller, Operands[0]->getType()->getPointerAddressSpace())) {
1687 if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1688 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1689 APFloat Val(Op->getValueAPF());
1692 Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);
1694 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1697 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1700 if (IntrinsicID == Intrinsic::round) {
1701 APFloat V = Op->getValueAPF();
1702 V.roundToIntegral(APFloat::rmNearestTiesToAway);
1703 return ConstantFP::get(Ty->getContext(), V);
1706 if (IntrinsicID == Intrinsic::floor) {
1707 APFloat V = Op->getValueAPF();
1708 V.roundToIntegral(APFloat::rmTowardNegative);
1709 return ConstantFP::get(Ty->getContext(), V);
1712 if (IntrinsicID == Intrinsic::ceil) {
1713 APFloat V = Op->getValueAPF();
1714 V.roundToIntegral(APFloat::rmTowardPositive);
1715 return ConstantFP::get(Ty->getContext(), V);
1718 if (IntrinsicID == Intrinsic::trunc) {
1719 APFloat V = Op->getValueAPF();
1720 V.roundToIntegral(APFloat::rmTowardZero);
1721 return ConstantFP::get(Ty->getContext(), V);
1724 if (IntrinsicID == Intrinsic::rint) {
1725 APFloat V = Op->getValueAPF();
1726 V.roundToIntegral(APFloat::rmNearestTiesToEven);
1727 return ConstantFP::get(Ty->getContext(), V);
1730 if (IntrinsicID == Intrinsic::nearbyint) {
1731 APFloat V = Op->getValueAPF();
1732 V.roundToIntegral(APFloat::rmNearestTiesToEven);
1733 return ConstantFP::get(Ty->getContext(), V);
1736 /// We only fold functions with finite arguments. Folding NaN and inf is
1737 /// likely to be aborted with an exception anyway, and some host libms
1738 /// have known errors raising exceptions.
1739 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1742 /// Currently APFloat versions of these functions do not exist, so we use
1743 /// the host native double versions. Float versions are not called
1744 /// directly but for all these it is true (float)(f((double)arg)) ==
1745 /// f(arg). Long double not supported yet.
1746 double V = getValueAsDouble(Op);
1748 switch (IntrinsicID) {
1750 case Intrinsic::fabs:
1751 return ConstantFoldFP(fabs, V, Ty);
1752 case Intrinsic::log2:
1753 return ConstantFoldFP(Log2, V, Ty);
1754 case Intrinsic::log:
1755 return ConstantFoldFP(log, V, Ty);
1756 case Intrinsic::log10:
1757 return ConstantFoldFP(log10, V, Ty);
1758 case Intrinsic::exp:
1759 return ConstantFoldFP(exp, V, Ty);
1760 case Intrinsic::exp2:
1761 return ConstantFoldFP(exp2, V, Ty);
1762 case Intrinsic::sin:
1763 return ConstantFoldFP(sin, V, Ty);
1764 case Intrinsic::cos:
1765 return ConstantFoldFP(cos, V, Ty);
1766 case Intrinsic::sqrt:
1767 return ConstantFoldFP(sqrt, V, Ty);
1773 char NameKeyChar = Name[0];
1774 if (Name[0] == '_' && Name.size() > 2 && Name[1] == '_')
1775 NameKeyChar = Name[2];
1777 switch (NameKeyChar) {
1779 if ((Name == "acos" && TLI->has(LibFunc_acos)) ||
1780 (Name == "acosf" && TLI->has(LibFunc_acosf)) ||
1781 (Name == "__acos_finite" && TLI->has(LibFunc_acos_finite)) ||
1782 (Name == "__acosf_finite" && TLI->has(LibFunc_acosf_finite)))
1783 return ConstantFoldFP(acos, V, Ty);
1784 else if ((Name == "asin" && TLI->has(LibFunc_asin)) ||
1785 (Name == "asinf" && TLI->has(LibFunc_asinf)) ||
1786 (Name == "__asin_finite" && TLI->has(LibFunc_asin_finite)) ||
1787 (Name == "__asinf_finite" && TLI->has(LibFunc_asinf_finite)))
1788 return ConstantFoldFP(asin, V, Ty);
1789 else if ((Name == "atan" && TLI->has(LibFunc_atan)) ||
1790 (Name == "atanf" && TLI->has(LibFunc_atanf)))
1791 return ConstantFoldFP(atan, V, Ty);
1794 if ((Name == "ceil" && TLI->has(LibFunc_ceil)) ||
1795 (Name == "ceilf" && TLI->has(LibFunc_ceilf)))
1796 return ConstantFoldFP(ceil, V, Ty);
1797 else if ((Name == "cos" && TLI->has(LibFunc_cos)) ||
1798 (Name == "cosf" && TLI->has(LibFunc_cosf)))
1799 return ConstantFoldFP(cos, V, Ty);
1800 else if ((Name == "cosh" && TLI->has(LibFunc_cosh)) ||
1801 (Name == "coshf" && TLI->has(LibFunc_coshf)) ||
1802 (Name == "__cosh_finite" && TLI->has(LibFunc_cosh_finite)) ||
1803 (Name == "__coshf_finite" && TLI->has(LibFunc_coshf_finite)))
1804 return ConstantFoldFP(cosh, V, Ty);
1807 if ((Name == "exp" && TLI->has(LibFunc_exp)) ||
1808 (Name == "expf" && TLI->has(LibFunc_expf)) ||
1809 (Name == "__exp_finite" && TLI->has(LibFunc_exp_finite)) ||
1810 (Name == "__expf_finite" && TLI->has(LibFunc_expf_finite)))
1811 return ConstantFoldFP(exp, V, Ty);
1812 if ((Name == "exp2" && TLI->has(LibFunc_exp2)) ||
1813 (Name == "exp2f" && TLI->has(LibFunc_exp2f)) ||
1814 (Name == "__exp2_finite" && TLI->has(LibFunc_exp2_finite)) ||
1815 (Name == "__exp2f_finite" && TLI->has(LibFunc_exp2f_finite)))
1816 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1818 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1821 if ((Name == "fabs" && TLI->has(LibFunc_fabs)) ||
1822 (Name == "fabsf" && TLI->has(LibFunc_fabsf)))
1823 return ConstantFoldFP(fabs, V, Ty);
1824 else if ((Name == "floor" && TLI->has(LibFunc_floor)) ||
1825 (Name == "floorf" && TLI->has(LibFunc_floorf)))
1826 return ConstantFoldFP(floor, V, Ty);
1829 if ((Name == "log" && V > 0 && TLI->has(LibFunc_log)) ||
1830 (Name == "logf" && V > 0 && TLI->has(LibFunc_logf)) ||
1831 (Name == "__log_finite" && V > 0 &&
1832 TLI->has(LibFunc_log_finite)) ||
1833 (Name == "__logf_finite" && V > 0 &&
1834 TLI->has(LibFunc_logf_finite)))
1835 return ConstantFoldFP(log, V, Ty);
1836 else if ((Name == "log10" && V > 0 && TLI->has(LibFunc_log10)) ||
1837 (Name == "log10f" && V > 0 && TLI->has(LibFunc_log10f)) ||
1838 (Name == "__log10_finite" && V > 0 &&
1839 TLI->has(LibFunc_log10_finite)) ||
1840 (Name == "__log10f_finite" && V > 0 &&
1841 TLI->has(LibFunc_log10f_finite)))
1842 return ConstantFoldFP(log10, V, Ty);
1845 if ((Name == "round" && TLI->has(LibFunc_round)) ||
1846 (Name == "roundf" && TLI->has(LibFunc_roundf)))
1847 return ConstantFoldFP(round, V, Ty);
1850 if ((Name == "sin" && TLI->has(LibFunc_sin)) ||
1851 (Name == "sinf" && TLI->has(LibFunc_sinf)))
1852 return ConstantFoldFP(sin, V, Ty);
1853 else if ((Name == "sinh" && TLI->has(LibFunc_sinh)) ||
1854 (Name == "sinhf" && TLI->has(LibFunc_sinhf)) ||
1855 (Name == "__sinh_finite" && TLI->has(LibFunc_sinh_finite)) ||
1856 (Name == "__sinhf_finite" && TLI->has(LibFunc_sinhf_finite)))
1857 return ConstantFoldFP(sinh, V, Ty);
1858 else if ((Name == "sqrt" && V >= 0 && TLI->has(LibFunc_sqrt)) ||
1859 (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc_sqrtf)))
1860 return ConstantFoldFP(sqrt, V, Ty);
1863 if ((Name == "tan" && TLI->has(LibFunc_tan)) ||
1864 (Name == "tanf" && TLI->has(LibFunc_tanf)))
1865 return ConstantFoldFP(tan, V, Ty);
1866 else if ((Name == "tanh" && TLI->has(LibFunc_tanh)) ||
1867 (Name == "tanhf" && TLI->has(LibFunc_tanhf)))
1868 return ConstantFoldFP(tanh, V, Ty);
1876 if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
1877 switch (IntrinsicID) {
1878 case Intrinsic::bswap:
1879 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1880 case Intrinsic::ctpop:
1881 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1882 case Intrinsic::bitreverse:
1883 return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
1884 case Intrinsic::convert_from_fp16: {
1885 APFloat Val(APFloat::IEEEhalf(), Op->getValue());
1888 APFloat::opStatus status = Val.convert(
1889 Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
1891 // Conversion is always precise.
1893 assert(status == APFloat::opOK && !lost &&
1894 "Precision lost during fp16 constfolding");
1896 return ConstantFP::get(Ty->getContext(), Val);
1903 // Support ConstantVector in case we have an Undef in the top.
1904 if (isa<ConstantVector>(Operands[0]) ||
1905 isa<ConstantDataVector>(Operands[0])) {
1906 auto *Op = cast<Constant>(Operands[0]);
1907 switch (IntrinsicID) {
1909 case Intrinsic::x86_sse_cvtss2si:
1910 case Intrinsic::x86_sse_cvtss2si64:
1911 case Intrinsic::x86_sse2_cvtsd2si:
1912 case Intrinsic::x86_sse2_cvtsd2si64:
1913 if (ConstantFP *FPOp =
1914 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1915 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1916 /*roundTowardZero=*/false, Ty,
1919 case Intrinsic::x86_sse_cvttss2si:
1920 case Intrinsic::x86_sse_cvttss2si64:
1921 case Intrinsic::x86_sse2_cvttsd2si:
1922 case Intrinsic::x86_sse2_cvttsd2si64:
1923 if (ConstantFP *FPOp =
1924 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1925 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1926 /*roundTowardZero=*/true, Ty,
1935 if (Operands.size() == 2) {
1936 if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1937 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1939 double Op1V = getValueAsDouble(Op1);
1941 if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1942 if (Op2->getType() != Op1->getType())
1945 double Op2V = getValueAsDouble(Op2);
1946 if (IntrinsicID == Intrinsic::pow) {
1947 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1949 if (IntrinsicID == Intrinsic::copysign) {
1950 APFloat V1 = Op1->getValueAPF();
1951 const APFloat &V2 = Op2->getValueAPF();
1953 return ConstantFP::get(Ty->getContext(), V1);
1956 if (IntrinsicID == Intrinsic::minnum) {
1957 const APFloat &C1 = Op1->getValueAPF();
1958 const APFloat &C2 = Op2->getValueAPF();
1959 return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
1962 if (IntrinsicID == Intrinsic::maxnum) {
1963 const APFloat &C1 = Op1->getValueAPF();
1964 const APFloat &C2 = Op2->getValueAPF();
1965 return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
1968 if (IntrinsicID == Intrinsic::minimum) {
1969 const APFloat &C1 = Op1->getValueAPF();
1970 const APFloat &C2 = Op2->getValueAPF();
1971 return ConstantFP::get(Ty->getContext(), minimum(C1, C2));
1974 if (IntrinsicID == Intrinsic::maximum) {
1975 const APFloat &C1 = Op1->getValueAPF();
1976 const APFloat &C2 = Op2->getValueAPF();
1977 return ConstantFP::get(Ty->getContext(), maximum(C1, C2));
1982 if ((Name == "pow" && TLI->has(LibFunc_pow)) ||
1983 (Name == "powf" && TLI->has(LibFunc_powf)) ||
1984 (Name == "__pow_finite" && TLI->has(LibFunc_pow_finite)) ||
1985 (Name == "__powf_finite" && TLI->has(LibFunc_powf_finite)))
1986 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1987 if ((Name == "fmod" && TLI->has(LibFunc_fmod)) ||
1988 (Name == "fmodf" && TLI->has(LibFunc_fmodf)))
1989 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1990 if ((Name == "atan2" && TLI->has(LibFunc_atan2)) ||
1991 (Name == "atan2f" && TLI->has(LibFunc_atan2f)) ||
1992 (Name == "__atan2_finite" && TLI->has(LibFunc_atan2_finite)) ||
1993 (Name == "__atan2f_finite" && TLI->has(LibFunc_atan2f_finite)))
1994 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1995 } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1996 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1997 return ConstantFP::get(Ty->getContext(),
1998 APFloat((float)std::pow((float)Op1V,
1999 (int)Op2C->getZExtValue())));
2000 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
2001 return ConstantFP::get(Ty->getContext(),
2002 APFloat((float)std::pow((float)Op1V,
2003 (int)Op2C->getZExtValue())));
2004 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
2005 return ConstantFP::get(Ty->getContext(),
2006 APFloat((double)std::pow((double)Op1V,
2007 (int)Op2C->getZExtValue())));
2012 if (Operands[0]->getType()->isIntegerTy() &&
2013 Operands[1]->getType()->isIntegerTy()) {
2014 const APInt *C0, *C1;
2015 if (!getConstIntOrUndef(Operands[0], C0) ||
2016 !getConstIntOrUndef(Operands[1], C1))
2019 switch (IntrinsicID) {
2021 case Intrinsic::smul_with_overflow:
2022 case Intrinsic::umul_with_overflow:
2023 // Even if both operands are undef, we cannot fold muls to undef
2024 // in the general case. For example, on i2 there are no inputs
2025 // that would produce { i2 -1, i1 true } as the result.
2027 return Constant::getNullValue(Ty);
2029 case Intrinsic::sadd_with_overflow:
2030 case Intrinsic::uadd_with_overflow:
2031 case Intrinsic::ssub_with_overflow:
2032 case Intrinsic::usub_with_overflow: {
2034 return UndefValue::get(Ty);
2038 switch (IntrinsicID) {
2039 default: llvm_unreachable("Invalid case");
2040 case Intrinsic::sadd_with_overflow:
2041 Res = C0->sadd_ov(*C1, Overflow);
2043 case Intrinsic::uadd_with_overflow:
2044 Res = C0->uadd_ov(*C1, Overflow);
2046 case Intrinsic::ssub_with_overflow:
2047 Res = C0->ssub_ov(*C1, Overflow);
2049 case Intrinsic::usub_with_overflow:
2050 Res = C0->usub_ov(*C1, Overflow);
2052 case Intrinsic::smul_with_overflow:
2053 Res = C0->smul_ov(*C1, Overflow);
2055 case Intrinsic::umul_with_overflow:
2056 Res = C0->umul_ov(*C1, Overflow);
2060 ConstantInt::get(Ty->getContext(), Res),
2061 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
2063 return ConstantStruct::get(cast<StructType>(Ty), Ops);
2065 case Intrinsic::uadd_sat:
2066 case Intrinsic::sadd_sat:
2068 return UndefValue::get(Ty);
2070 return Constant::getAllOnesValue(Ty);
2071 if (IntrinsicID == Intrinsic::uadd_sat)
2072 return ConstantInt::get(Ty, C0->uadd_sat(*C1));
2074 return ConstantInt::get(Ty, C0->sadd_sat(*C1));
2075 case Intrinsic::usub_sat:
2076 case Intrinsic::ssub_sat:
2078 return UndefValue::get(Ty);
2080 return Constant::getNullValue(Ty);
2081 if (IntrinsicID == Intrinsic::usub_sat)
2082 return ConstantInt::get(Ty, C0->usub_sat(*C1));
2084 return ConstantInt::get(Ty, C0->ssub_sat(*C1));
2085 case Intrinsic::cttz:
2086 case Intrinsic::ctlz:
2087 assert(C1 && "Must be constant int");
2089 // cttz(0, 1) and ctlz(0, 1) are undef.
2090 if (C1->isOneValue() && (!C0 || C0->isNullValue()))
2091 return UndefValue::get(Ty);
2093 return Constant::getNullValue(Ty);
2094 if (IntrinsicID == Intrinsic::cttz)
2095 return ConstantInt::get(Ty, C0->countTrailingZeros());
2097 return ConstantInt::get(Ty, C0->countLeadingZeros());
2103 // Support ConstantVector in case we have an Undef in the top.
2104 if ((isa<ConstantVector>(Operands[0]) ||
2105 isa<ConstantDataVector>(Operands[0])) &&
2106 // Check for default rounding mode.
2107 // FIXME: Support other rounding modes?
2108 isa<ConstantInt>(Operands[1]) &&
2109 cast<ConstantInt>(Operands[1])->getValue() == 4) {
2110 auto *Op = cast<Constant>(Operands[0]);
2111 switch (IntrinsicID) {
2113 case Intrinsic::x86_avx512_vcvtss2si32:
2114 case Intrinsic::x86_avx512_vcvtss2si64:
2115 case Intrinsic::x86_avx512_vcvtsd2si32:
2116 case Intrinsic::x86_avx512_vcvtsd2si64:
2117 if (ConstantFP *FPOp =
2118 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2119 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2120 /*roundTowardZero=*/false, Ty,
2123 case Intrinsic::x86_avx512_vcvtss2usi32:
2124 case Intrinsic::x86_avx512_vcvtss2usi64:
2125 case Intrinsic::x86_avx512_vcvtsd2usi32:
2126 case Intrinsic::x86_avx512_vcvtsd2usi64:
2127 if (ConstantFP *FPOp =
2128 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2129 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2130 /*roundTowardZero=*/false, Ty,
2133 case Intrinsic::x86_avx512_cvttss2si:
2134 case Intrinsic::x86_avx512_cvttss2si64:
2135 case Intrinsic::x86_avx512_cvttsd2si:
2136 case Intrinsic::x86_avx512_cvttsd2si64:
2137 if (ConstantFP *FPOp =
2138 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2139 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2140 /*roundTowardZero=*/true, Ty,
2143 case Intrinsic::x86_avx512_cvttss2usi:
2144 case Intrinsic::x86_avx512_cvttss2usi64:
2145 case Intrinsic::x86_avx512_cvttsd2usi:
2146 case Intrinsic::x86_avx512_cvttsd2usi64:
2147 if (ConstantFP *FPOp =
2148 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2149 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2150 /*roundTowardZero=*/true, Ty,
2158 if (Operands.size() != 3)
2161 if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
2162 if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
2163 if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
2164 switch (IntrinsicID) {
2166 case Intrinsic::fma:
2167 case Intrinsic::fmuladd: {
2168 APFloat V = Op1->getValueAPF();
2169 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
2171 APFloat::rmNearestTiesToEven);
2172 if (s != APFloat::opInvalidOp)
2173 return ConstantFP::get(Ty->getContext(), V);
2182 if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
2183 const APInt *C0, *C1, *C2;
2184 if (!getConstIntOrUndef(Operands[0], C0) ||
2185 !getConstIntOrUndef(Operands[1], C1) ||
2186 !getConstIntOrUndef(Operands[2], C2))
2189 bool IsRight = IntrinsicID == Intrinsic::fshr;
2191 return Operands[IsRight ? 1 : 0];
2193 return UndefValue::get(Ty);
2195 // The shift amount is interpreted as modulo the bitwidth. If the shift
2196 // amount is effectively 0, avoid UB due to oversized inverse shift below.
2197 unsigned BitWidth = C2->getBitWidth();
2198 unsigned ShAmt = C2->urem(BitWidth);
2200 return Operands[IsRight ? 1 : 0];
2202 // (C0 << ShlAmt) | (C1 >> LshrAmt)
2203 unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;
2204 unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;
2206 return ConstantInt::get(Ty, C1->lshr(LshrAmt));
2208 return ConstantInt::get(Ty, C0->shl(ShlAmt));
2209 return ConstantInt::get(Ty, C0->shl(ShlAmt) | C1->lshr(LshrAmt));
2215 Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
2216 VectorType *VTy, ArrayRef<Constant *> Operands,
2217 const DataLayout &DL,
2218 const TargetLibraryInfo *TLI,
2219 ImmutableCallSite CS) {
2220 SmallVector<Constant *, 4> Result(VTy->getNumElements());
2221 SmallVector<Constant *, 4> Lane(Operands.size());
2222 Type *Ty = VTy->getElementType();
2224 if (IntrinsicID == Intrinsic::masked_load) {
2225 auto *SrcPtr = Operands[0];
2226 auto *Mask = Operands[2];
2227 auto *Passthru = Operands[3];
2229 Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL);
2231 SmallVector<Constant *, 32> NewElements;
2232 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2233 auto *MaskElt = Mask->getAggregateElement(I);
2236 auto *PassthruElt = Passthru->getAggregateElement(I);
2237 auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
2238 if (isa<UndefValue>(MaskElt)) {
2240 NewElements.push_back(PassthruElt);
2242 NewElements.push_back(VecElt);
2246 if (MaskElt->isNullValue()) {
2249 NewElements.push_back(PassthruElt);
2250 } else if (MaskElt->isOneValue()) {
2253 NewElements.push_back(VecElt);
2258 if (NewElements.size() != VTy->getNumElements())
2260 return ConstantVector::get(NewElements);
2263 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2264 // Gather a column of constants.
2265 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
2266 // These intrinsics use a scalar type for their second argument.
2268 (IntrinsicID == Intrinsic::cttz || IntrinsicID == Intrinsic::ctlz ||
2269 IntrinsicID == Intrinsic::powi)) {
2270 Lane[J] = Operands[J];
2274 Constant *Agg = Operands[J]->getAggregateElement(I);
2281 // Use the regular scalar folding to simplify this column.
2282 Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, CS);
2288 return ConstantVector::get(Result);
2291 } // end anonymous namespace
2294 llvm::ConstantFoldCall(ImmutableCallSite CS, Function *F,
2295 ArrayRef<Constant *> Operands,
2296 const TargetLibraryInfo *TLI) {
2297 if (CS.isNoBuiltin() || CS.isStrictFP())
2301 StringRef Name = F->getName();
2303 Type *Ty = F->getReturnType();
2305 if (auto *VTy = dyn_cast<VectorType>(Ty))
2306 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
2307 F->getParent()->getDataLayout(), TLI, CS);
2309 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI, CS);
2312 bool llvm::isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI) {
2313 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
2314 // (and to some extent ConstantFoldScalarCall).
2315 if (CS.isNoBuiltin() || CS.isStrictFP())
2317 Function *F = CS.getCalledFunction();
2322 if (!TLI || !TLI->getLibFunc(*F, Func))
2325 if (CS.getNumArgOperands() == 1) {
2326 if (ConstantFP *OpC = dyn_cast<ConstantFP>(CS.getArgOperand(0))) {
2327 const APFloat &Op = OpC->getValueAPF();
2335 case LibFunc_log10l:
2337 case LibFunc_log10f:
2338 return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
2343 // FIXME: These boundaries are slightly conservative.
2344 if (OpC->getType()->isDoubleTy())
2345 return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan &&
2346 Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan;
2347 if (OpC->getType()->isFloatTy())
2348 return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan &&
2349 Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan;
2355 // FIXME: These boundaries are slightly conservative.
2356 if (OpC->getType()->isDoubleTy())
2357 return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan &&
2358 Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan;
2359 if (OpC->getType()->isFloatTy())
2360 return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan &&
2361 Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan;
2370 return !Op.isInfinity();
2374 case LibFunc_tanf: {
2375 // FIXME: Stop using the host math library.
2376 // FIXME: The computation isn't done in the right precision.
2377 Type *Ty = OpC->getType();
2378 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2379 double OpV = getValueAsDouble(OpC);
2380 return ConstantFoldFP(tan, OpV, Ty) != nullptr;
2391 return Op.compare(APFloat(Op.getSemantics(), "-1")) !=
2392 APFloat::cmpLessThan &&
2393 Op.compare(APFloat(Op.getSemantics(), "1")) !=
2394 APFloat::cmpGreaterThan;
2402 // FIXME: These boundaries are slightly conservative.
2403 if (OpC->getType()->isDoubleTy())
2404 return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan &&
2405 Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan;
2406 if (OpC->getType()->isFloatTy())
2407 return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan &&
2408 Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan;
2414 return Op.isNaN() || Op.isZero() || !Op.isNegative();
2416 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
2424 if (CS.getNumArgOperands() == 2) {
2425 ConstantFP *Op0C = dyn_cast<ConstantFP>(CS.getArgOperand(0));
2426 ConstantFP *Op1C = dyn_cast<ConstantFP>(CS.getArgOperand(1));
2428 const APFloat &Op0 = Op0C->getValueAPF();
2429 const APFloat &Op1 = Op1C->getValueAPF();
2434 case LibFunc_powf: {
2435 // FIXME: Stop using the host math library.
2436 // FIXME: The computation isn't done in the right precision.
2437 Type *Ty = Op0C->getType();
2438 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2439 if (Ty == Op1C->getType()) {
2440 double Op0V = getValueAsDouble(Op0C);
2441 double Op1V = getValueAsDouble(Op1C);
2442 return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr;
2451 return Op0.isNaN() || Op1.isNaN() ||
2452 (!Op0.isInfinity() && !Op1.isZero());