1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines routines for folding instructions into constants.
12 // Also, to supplement the basic IR ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // DataLayout information. These functions cannot go in IR due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringMap.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/FEnv.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Target/TargetLibraryInfo.h"
41 //===----------------------------------------------------------------------===//
42 // Constant Folding internal helper functions
43 //===----------------------------------------------------------------------===//
45 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
46 /// DataLayout. This always returns a non-null constant, but it may be a
47 /// ConstantExpr if unfoldable.
48 static Constant *FoldBitCast(Constant *C, Type *DestTy,
49 const DataLayout &TD) {
50 // Catch the obvious splat cases.
51 if (C->isNullValue() && !DestTy->isX86_MMXTy())
52 return Constant::getNullValue(DestTy);
53 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
54 return Constant::getAllOnesValue(DestTy);
56 // Handle a vector->integer cast.
57 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
58 VectorType *VTy = dyn_cast<VectorType>(C->getType());
60 return ConstantExpr::getBitCast(C, DestTy);
62 unsigned NumSrcElts = VTy->getNumElements();
63 Type *SrcEltTy = VTy->getElementType();
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (SrcEltTy->isFloatingPointTy()) {
68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
71 // Ask IR to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
75 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
77 return ConstantExpr::getBitCast(C, DestTy);
79 // Now that we know that the input value is a vector of integers, just shift
80 // and insert them into our result.
81 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
82 APInt Result(IT->getBitWidth(), 0);
83 for (unsigned i = 0; i != NumSrcElts; ++i) {
85 if (TD.isLittleEndian())
86 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
88 Result |= CDV->getElementAsInteger(i);
91 return ConstantInt::get(IT, Result);
94 // The code below only handles casts to vectors currently.
95 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
97 return ConstantExpr::getBitCast(C, DestTy);
99 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
100 // vector so the code below can handle it uniformly.
101 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
102 Constant *Ops = C; // don't take the address of C!
103 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
106 // If this is a bitcast from constant vector -> vector, fold it.
107 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
108 return ConstantExpr::getBitCast(C, DestTy);
110 // If the element types match, IR can fold it.
111 unsigned NumDstElt = DestVTy->getNumElements();
112 unsigned NumSrcElt = C->getType()->getVectorNumElements();
113 if (NumDstElt == NumSrcElt)
114 return ConstantExpr::getBitCast(C, DestTy);
116 Type *SrcEltTy = C->getType()->getVectorElementType();
117 Type *DstEltTy = DestVTy->getElementType();
119 // Otherwise, we're changing the number of elements in a vector, which
120 // requires endianness information to do the right thing. For example,
121 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
122 // folds to (little endian):
123 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
124 // and to (big endian):
125 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
127 // First thing is first. We only want to think about integer here, so if
128 // we have something in FP form, recast it as integer.
129 if (DstEltTy->isFloatingPointTy()) {
130 // Fold to an vector of integers with same size as our FP type.
131 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
133 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
134 // Recursively handle this integer conversion, if possible.
135 C = FoldBitCast(C, DestIVTy, TD);
137 // Finally, IR can handle this now that #elts line up.
138 return ConstantExpr::getBitCast(C, DestTy);
141 // Okay, we know the destination is integer, if the input is FP, convert
142 // it to integer first.
143 if (SrcEltTy->isFloatingPointTy()) {
144 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
146 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
147 // Ask IR to do the conversion now that #elts line up.
148 C = ConstantExpr::getBitCast(C, SrcIVTy);
149 // If IR wasn't able to fold it, bail out.
150 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
151 !isa<ConstantDataVector>(C))
155 // Now we know that the input and output vectors are both integer vectors
156 // of the same size, and that their #elements is not the same. Do the
157 // conversion here, which depends on whether the input or output has
159 bool isLittleEndian = TD.isLittleEndian();
161 SmallVector<Constant*, 32> Result;
162 if (NumDstElt < NumSrcElt) {
163 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
164 Constant *Zero = Constant::getNullValue(DstEltTy);
165 unsigned Ratio = NumSrcElt/NumDstElt;
166 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
168 for (unsigned i = 0; i != NumDstElt; ++i) {
169 // Build each element of the result.
170 Constant *Elt = Zero;
171 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
172 for (unsigned j = 0; j != Ratio; ++j) {
173 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
174 if (!Src) // Reject constantexpr elements.
175 return ConstantExpr::getBitCast(C, DestTy);
177 // Zero extend the element to the right size.
178 Src = ConstantExpr::getZExt(Src, Elt->getType());
180 // Shift it to the right place, depending on endianness.
181 Src = ConstantExpr::getShl(Src,
182 ConstantInt::get(Src->getType(), ShiftAmt));
183 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
186 Elt = ConstantExpr::getOr(Elt, Src);
188 Result.push_back(Elt);
190 return ConstantVector::get(Result);
193 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
194 unsigned Ratio = NumDstElt/NumSrcElt;
195 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
197 // Loop over each source value, expanding into multiple results.
198 for (unsigned i = 0; i != NumSrcElt; ++i) {
199 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
200 if (!Src) // Reject constantexpr elements.
201 return ConstantExpr::getBitCast(C, DestTy);
203 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
204 for (unsigned j = 0; j != Ratio; ++j) {
205 // Shift the piece of the value into the right place, depending on
207 Constant *Elt = ConstantExpr::getLShr(Src,
208 ConstantInt::get(Src->getType(), ShiftAmt));
209 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
211 // Truncate and remember this piece.
212 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
216 return ConstantVector::get(Result);
220 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
221 /// from a global, return the global and the constant. Because of
222 /// constantexprs, this function is recursive.
223 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
224 APInt &Offset, const DataLayout &TD) {
225 // Trivial case, constant is the global.
226 if ((GV = dyn_cast<GlobalValue>(C))) {
227 unsigned BitWidth = TD.getPointerTypeSizeInBits(GV->getType());
228 Offset = APInt(BitWidth, 0);
232 // Otherwise, if this isn't a constant expr, bail out.
233 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
234 if (!CE) return false;
236 // Look through ptr->int and ptr->ptr casts.
237 if (CE->getOpcode() == Instruction::PtrToInt ||
238 CE->getOpcode() == Instruction::BitCast)
239 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
241 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
242 GEPOperator *GEP = dyn_cast<GEPOperator>(CE);
246 unsigned BitWidth = TD.getPointerTypeSizeInBits(GEP->getType());
247 APInt TmpOffset(BitWidth, 0);
249 // If the base isn't a global+constant, we aren't either.
250 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, TD))
253 // Otherwise, add any offset that our operands provide.
254 if (!GEP->accumulateConstantOffset(TD, TmpOffset))
261 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
262 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
263 /// pointer to copy results into and BytesLeft is the number of bytes left in
264 /// the CurPtr buffer. TD is the target data.
265 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
266 unsigned char *CurPtr, unsigned BytesLeft,
267 const DataLayout &TD) {
268 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
269 "Out of range access");
271 // If this element is zero or undefined, we can just return since *CurPtr is
273 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
276 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
277 if (CI->getBitWidth() > 64 ||
278 (CI->getBitWidth() & 7) != 0)
281 uint64_t Val = CI->getZExtValue();
282 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
284 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
286 if (!TD.isLittleEndian())
287 n = IntBytes - n - 1;
288 CurPtr[i] = (unsigned char)(Val >> (n * 8));
294 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
295 if (CFP->getType()->isDoubleTy()) {
296 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
297 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
299 if (CFP->getType()->isFloatTy()){
300 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
301 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
303 if (CFP->getType()->isHalfTy()){
304 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), TD);
305 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
310 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
311 const StructLayout *SL = TD.getStructLayout(CS->getType());
312 unsigned Index = SL->getElementContainingOffset(ByteOffset);
313 uint64_t CurEltOffset = SL->getElementOffset(Index);
314 ByteOffset -= CurEltOffset;
317 // If the element access is to the element itself and not to tail padding,
318 // read the bytes from the element.
319 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
321 if (ByteOffset < EltSize &&
322 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
328 // Check to see if we read from the last struct element, if so we're done.
329 if (Index == CS->getType()->getNumElements())
332 // If we read all of the bytes we needed from this element we're done.
333 uint64_t NextEltOffset = SL->getElementOffset(Index);
335 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
338 // Move to the next element of the struct.
339 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
340 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
342 CurEltOffset = NextEltOffset;
347 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
348 isa<ConstantDataSequential>(C)) {
349 Type *EltTy = C->getType()->getSequentialElementType();
350 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
351 uint64_t Index = ByteOffset / EltSize;
352 uint64_t Offset = ByteOffset - Index * EltSize;
354 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
355 NumElts = AT->getNumElements();
357 NumElts = C->getType()->getVectorNumElements();
359 for (; Index != NumElts; ++Index) {
360 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
364 uint64_t BytesWritten = EltSize - Offset;
365 assert(BytesWritten <= EltSize && "Not indexing into this element?");
366 if (BytesWritten >= BytesLeft)
370 BytesLeft -= BytesWritten;
371 CurPtr += BytesWritten;
376 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
377 if (CE->getOpcode() == Instruction::IntToPtr &&
378 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getType())) {
379 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
384 // Otherwise, unknown initializer type.
388 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
389 const DataLayout &TD) {
390 PointerType *PTy = cast<PointerType>(C->getType());
391 Type *LoadTy = PTy->getElementType();
392 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
394 // If this isn't an integer load we can't fold it directly.
396 unsigned AS = PTy->getAddressSpace();
398 // If this is a float/double load, we can try folding it as an int32/64 load
399 // and then bitcast the result. This can be useful for union cases. Note
400 // that address spaces don't matter here since we're not going to result in
401 // an actual new load.
403 if (LoadTy->isHalfTy())
404 MapTy = Type::getInt16PtrTy(C->getContext(), AS);
405 else if (LoadTy->isFloatTy())
406 MapTy = Type::getInt32PtrTy(C->getContext(), AS);
407 else if (LoadTy->isDoubleTy())
408 MapTy = Type::getInt64PtrTy(C->getContext(), AS);
409 else if (LoadTy->isVectorTy()) {
410 MapTy = PointerType::getIntNPtrTy(C->getContext(),
411 TD.getTypeAllocSizeInBits(LoadTy),
416 C = FoldBitCast(C, MapTy, TD);
417 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
418 return FoldBitCast(Res, LoadTy, TD);
422 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
423 if (BytesLoaded > 32 || BytesLoaded == 0)
428 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
431 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
432 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
433 !GV->getInitializer()->getType()->isSized())
436 // If we're loading off the beginning of the global, some bytes may be valid,
437 // but we don't try to handle this.
438 if (Offset.isNegative())
441 // If we're not accessing anything in this constant, the result is undefined.
442 if (Offset.getZExtValue() >=
443 TD.getTypeAllocSize(GV->getInitializer()->getType()))
444 return UndefValue::get(IntType);
446 unsigned char RawBytes[32] = {0};
447 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
451 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
452 if (TD.isLittleEndian()) {
453 ResultVal = RawBytes[BytesLoaded - 1];
454 for (unsigned i = 1; i != BytesLoaded; ++i) {
456 ResultVal |= RawBytes[BytesLoaded - 1 - i];
459 ResultVal = RawBytes[0];
460 for (unsigned i = 1; i != BytesLoaded; ++i) {
462 ResultVal |= RawBytes[i];
466 return ConstantInt::get(IntType->getContext(), ResultVal);
469 static Constant *ConstantFoldLoadThroughBitcast(ConstantExpr *CE,
470 const DataLayout *DL) {
473 auto *DestPtrTy = dyn_cast<PointerType>(CE->getType());
476 Type *DestTy = DestPtrTy->getElementType();
478 Constant *C = ConstantFoldLoadFromConstPtr(CE->getOperand(0), DL);
483 Type *SrcTy = C->getType();
485 // If the type sizes are the same and a cast is legal, just directly
486 // cast the constant.
487 if (DL->getTypeSizeInBits(DestTy) == DL->getTypeSizeInBits(SrcTy)) {
488 Instruction::CastOps Cast = Instruction::BitCast;
489 // If we are going from a pointer to int or vice versa, we spell the cast
491 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
492 Cast = Instruction::IntToPtr;
493 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
494 Cast = Instruction::PtrToInt;
496 if (CastInst::castIsValid(Cast, C, DestTy))
497 return ConstantExpr::getCast(Cast, C, DestTy);
500 // If this isn't an aggregate type, there is nothing we can do to drill down
501 // and find a bitcastable constant.
502 if (!SrcTy->isAggregateType())
505 // We're simulating a load through a pointer that was bitcast to point to
506 // a different type, so we can try to walk down through the initial
507 // elements of an aggregate to see if some part of th e aggregate is
508 // castable to implement the "load" semantic model.
509 C = C->getAggregateElement(0u);
515 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
516 /// produce if it is constant and determinable. If this is not determinable,
518 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
519 const DataLayout *TD) {
520 // First, try the easy cases:
521 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
522 if (GV->isConstant() && GV->hasDefinitiveInitializer())
523 return GV->getInitializer();
525 // If the loaded value isn't a constant expr, we can't handle it.
526 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
530 if (CE->getOpcode() == Instruction::GetElementPtr) {
531 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
532 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
534 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
540 if (CE->getOpcode() == Instruction::BitCast)
541 if (Constant *LoadedC = ConstantFoldLoadThroughBitcast(CE, TD))
544 // Instead of loading constant c string, use corresponding integer value
545 // directly if string length is small enough.
547 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
548 unsigned StrLen = Str.size();
549 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
550 unsigned NumBits = Ty->getPrimitiveSizeInBits();
551 // Replace load with immediate integer if the result is an integer or fp
553 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
554 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
555 APInt StrVal(NumBits, 0);
556 APInt SingleChar(NumBits, 0);
557 if (TD->isLittleEndian()) {
558 for (signed i = StrLen-1; i >= 0; i--) {
559 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
560 StrVal = (StrVal << 8) | SingleChar;
563 for (unsigned i = 0; i < StrLen; i++) {
564 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
565 StrVal = (StrVal << 8) | SingleChar;
567 // Append NULL at the end.
569 StrVal = (StrVal << 8) | SingleChar;
572 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
573 if (Ty->isFloatingPointTy())
574 Res = ConstantExpr::getBitCast(Res, Ty);
579 // If this load comes from anywhere in a constant global, and if the global
580 // is all undef or zero, we know what it loads.
581 if (GlobalVariable *GV =
582 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
583 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
584 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
585 if (GV->getInitializer()->isNullValue())
586 return Constant::getNullValue(ResTy);
587 if (isa<UndefValue>(GV->getInitializer()))
588 return UndefValue::get(ResTy);
592 // Try hard to fold loads from bitcasted strange and non-type-safe things.
594 return FoldReinterpretLoadFromConstPtr(CE, *TD);
598 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
599 if (LI->isVolatile()) return nullptr;
601 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
602 return ConstantFoldLoadFromConstPtr(C, TD);
607 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
608 /// Attempt to symbolically evaluate the result of a binary operator merging
609 /// these together. If target data info is available, it is provided as DL,
610 /// otherwise DL is null.
611 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
612 Constant *Op1, const DataLayout *DL){
615 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
616 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
620 if (Opc == Instruction::And && DL) {
621 unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType());
622 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
623 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
624 computeKnownBits(Op0, KnownZero0, KnownOne0, DL);
625 computeKnownBits(Op1, KnownZero1, KnownOne1, DL);
626 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
627 // All the bits of Op0 that the 'and' could be masking are already zero.
630 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
631 // All the bits of Op1 that the 'and' could be masking are already zero.
635 APInt KnownZero = KnownZero0 | KnownZero1;
636 APInt KnownOne = KnownOne0 & KnownOne1;
637 if ((KnownZero | KnownOne).isAllOnesValue()) {
638 return ConstantInt::get(Op0->getType(), KnownOne);
642 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
643 // constant. This happens frequently when iterating over a global array.
644 if (Opc == Instruction::Sub && DL) {
645 GlobalValue *GV1, *GV2;
648 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL))
649 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) &&
651 unsigned OpSize = DL->getTypeSizeInBits(Op0->getType());
653 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
654 // PtrToInt may change the bitwidth so we have convert to the right size
656 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
657 Offs2.zextOrTrunc(OpSize));
664 /// CastGEPIndices - If array indices are not pointer-sized integers,
665 /// explicitly cast them so that they aren't implicitly casted by the
667 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
668 Type *ResultTy, const DataLayout *TD,
669 const TargetLibraryInfo *TLI) {
673 Type *IntPtrTy = TD->getIntPtrType(ResultTy);
676 SmallVector<Constant*, 32> NewIdxs;
677 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
679 !isa<StructType>(GetElementPtrInst::getIndexedType(
681 Ops.slice(1, i - 1)))) &&
682 Ops[i]->getType() != IntPtrTy) {
684 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
690 NewIdxs.push_back(Ops[i]);
696 Constant *C = ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
697 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
698 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
705 /// Strip the pointer casts, but preserve the address space information.
706 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
707 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
708 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
709 Ptr = cast<Constant>(Ptr->stripPointerCasts());
710 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
712 // Preserve the address space number of the pointer.
713 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
714 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
715 OldPtrTy->getAddressSpace());
716 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
721 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
722 /// constant expression, do so.
723 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
724 Type *ResultTy, const DataLayout *TD,
725 const TargetLibraryInfo *TLI) {
726 Constant *Ptr = Ops[0];
727 if (!TD || !Ptr->getType()->getPointerElementType()->isSized() ||
728 !Ptr->getType()->isPointerTy())
731 Type *IntPtrTy = TD->getIntPtrType(Ptr->getType());
732 Type *ResultElementTy = ResultTy->getPointerElementType();
734 // If this is a constant expr gep that is effectively computing an
735 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
736 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
737 if (!isa<ConstantInt>(Ops[i])) {
739 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
740 // "inttoptr (sub (ptrtoint Ptr), V)"
741 if (Ops.size() == 2 && ResultElementTy->isIntegerTy(8)) {
742 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
743 assert((!CE || CE->getType() == IntPtrTy) &&
744 "CastGEPIndices didn't canonicalize index types!");
745 if (CE && CE->getOpcode() == Instruction::Sub &&
746 CE->getOperand(0)->isNullValue()) {
747 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
748 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
749 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
750 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
751 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
758 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
760 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
761 makeArrayRef((Value *const*)
764 Ptr = StripPtrCastKeepAS(Ptr);
766 // If this is a GEP of a GEP, fold it all into a single GEP.
767 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
768 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
770 // Do not try the incorporate the sub-GEP if some index is not a number.
771 bool AllConstantInt = true;
772 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
773 if (!isa<ConstantInt>(NestedOps[i])) {
774 AllConstantInt = false;
780 Ptr = cast<Constant>(GEP->getOperand(0));
781 Offset += APInt(BitWidth,
782 TD->getIndexedOffset(Ptr->getType(), NestedOps));
783 Ptr = StripPtrCastKeepAS(Ptr);
786 // If the base value for this address is a literal integer value, fold the
787 // getelementptr to the resulting integer value casted to the pointer type.
788 APInt BasePtr(BitWidth, 0);
789 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
790 if (CE->getOpcode() == Instruction::IntToPtr) {
791 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
792 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
796 if (Ptr->isNullValue() || BasePtr != 0) {
797 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
798 return ConstantExpr::getIntToPtr(C, ResultTy);
801 // Otherwise form a regular getelementptr. Recompute the indices so that
802 // we eliminate over-indexing of the notional static type array bounds.
803 // This makes it easy to determine if the getelementptr is "inbounds".
804 // Also, this helps GlobalOpt do SROA on GlobalVariables.
805 Type *Ty = Ptr->getType();
806 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
807 SmallVector<Constant *, 32> NewIdxs;
810 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
811 if (ATy->isPointerTy()) {
812 // The only pointer indexing we'll do is on the first index of the GEP.
813 if (!NewIdxs.empty())
816 // Only handle pointers to sized types, not pointers to functions.
817 if (!ATy->getElementType()->isSized())
821 // Determine which element of the array the offset points into.
822 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
824 // The element size is 0. This may be [0 x Ty]*, so just use a zero
825 // index for this level and proceed to the next level to see if it can
826 // accommodate the offset.
827 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
829 // The element size is non-zero divide the offset by the element
830 // size (rounding down), to compute the index at this level.
831 APInt NewIdx = Offset.udiv(ElemSize);
832 Offset -= NewIdx * ElemSize;
833 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
835 Ty = ATy->getElementType();
836 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
837 // If we end up with an offset that isn't valid for this struct type, we
838 // can't re-form this GEP in a regular form, so bail out. The pointer
839 // operand likely went through casts that are necessary to make the GEP
841 const StructLayout &SL = *TD->getStructLayout(STy);
842 if (Offset.uge(SL.getSizeInBytes()))
845 // Determine which field of the struct the offset points into. The
846 // getZExtValue is fine as we've already ensured that the offset is
847 // within the range representable by the StructLayout API.
848 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
849 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
851 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
852 Ty = STy->getTypeAtIndex(ElIdx);
854 // We've reached some non-indexable type.
857 } while (Ty != ResultElementTy);
859 // If we haven't used up the entire offset by descending the static
860 // type, then the offset is pointing into the middle of an indivisible
861 // member, so we can't simplify it.
866 Constant *C = ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
867 assert(C->getType()->getPointerElementType() == Ty &&
868 "Computed GetElementPtr has unexpected type!");
870 // If we ended up indexing a member with a type that doesn't match
871 // the type of what the original indices indexed, add a cast.
872 if (Ty != ResultElementTy)
873 C = FoldBitCast(C, ResultTy, *TD);
880 //===----------------------------------------------------------------------===//
881 // Constant Folding public APIs
882 //===----------------------------------------------------------------------===//
884 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
885 /// If successful, the constant result is returned, if not, null is returned.
886 /// Note that this fails if not all of the operands are constant. Otherwise,
887 /// this function can only fail when attempting to fold instructions like loads
888 /// and stores, which have no constant expression form.
889 Constant *llvm::ConstantFoldInstruction(Instruction *I,
890 const DataLayout *TD,
891 const TargetLibraryInfo *TLI) {
892 // Handle PHI nodes quickly here...
893 if (PHINode *PN = dyn_cast<PHINode>(I)) {
894 Constant *CommonValue = nullptr;
896 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
897 Value *Incoming = PN->getIncomingValue(i);
898 // If the incoming value is undef then skip it. Note that while we could
899 // skip the value if it is equal to the phi node itself we choose not to
900 // because that would break the rule that constant folding only applies if
901 // all operands are constants.
902 if (isa<UndefValue>(Incoming))
904 // If the incoming value is not a constant, then give up.
905 Constant *C = dyn_cast<Constant>(Incoming);
908 // Fold the PHI's operands.
909 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
910 C = ConstantFoldConstantExpression(NewC, TD, TLI);
911 // If the incoming value is a different constant to
912 // the one we saw previously, then give up.
913 if (CommonValue && C != CommonValue)
919 // If we reach here, all incoming values are the same constant or undef.
920 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
923 // Scan the operand list, checking to see if they are all constants, if so,
924 // hand off to ConstantFoldInstOperands.
925 SmallVector<Constant*, 8> Ops;
926 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
927 Constant *Op = dyn_cast<Constant>(*i);
929 return nullptr; // All operands not constant!
931 // Fold the Instruction's operands.
932 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
933 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
938 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
939 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
942 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
943 return ConstantFoldLoadInst(LI, TD);
945 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) {
946 return ConstantExpr::getInsertValue(
947 cast<Constant>(IVI->getAggregateOperand()),
948 cast<Constant>(IVI->getInsertedValueOperand()),
952 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) {
953 return ConstantExpr::getExtractValue(
954 cast<Constant>(EVI->getAggregateOperand()),
958 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
962 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD,
963 const TargetLibraryInfo *TLI,
964 SmallPtrSet<ConstantExpr *, 4> &FoldedOps) {
965 SmallVector<Constant *, 8> Ops;
966 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e;
968 Constant *NewC = cast<Constant>(*i);
969 // Recursively fold the ConstantExpr's operands. If we have already folded
970 // a ConstantExpr, we don't have to process it again.
971 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) {
972 if (FoldedOps.insert(NewCE))
973 NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps);
979 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
981 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
984 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
985 /// using the specified DataLayout. If successful, the constant result is
986 /// result is returned, if not, null is returned.
987 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
988 const DataLayout *TD,
989 const TargetLibraryInfo *TLI) {
990 SmallPtrSet<ConstantExpr *, 4> FoldedOps;
991 return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps);
994 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
995 /// specified opcode and operands. If successful, the constant result is
996 /// returned, if not, null is returned. Note that this function can fail when
997 /// attempting to fold instructions like loads and stores, which have no
998 /// constant expression form.
1000 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
1001 /// information, due to only being passed an opcode and operands. Constant
1002 /// folding using this function strips this information.
1004 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
1005 ArrayRef<Constant *> Ops,
1006 const DataLayout *TD,
1007 const TargetLibraryInfo *TLI) {
1008 // Handle easy binops first.
1009 if (Instruction::isBinaryOp(Opcode)) {
1010 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) {
1011 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
1015 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
1019 default: return nullptr;
1020 case Instruction::ICmp:
1021 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1022 case Instruction::Call:
1023 if (Function *F = dyn_cast<Function>(Ops.back()))
1024 if (canConstantFoldCallTo(F))
1025 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
1027 case Instruction::PtrToInt:
1028 // If the input is a inttoptr, eliminate the pair. This requires knowing
1029 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1030 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1031 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
1032 Constant *Input = CE->getOperand(0);
1033 unsigned InWidth = Input->getType()->getScalarSizeInBits();
1034 unsigned PtrWidth = TD->getPointerTypeSizeInBits(CE->getType());
1035 if (PtrWidth < InWidth) {
1037 ConstantInt::get(CE->getContext(),
1038 APInt::getLowBitsSet(InWidth, PtrWidth));
1039 Input = ConstantExpr::getAnd(Input, Mask);
1041 // Do a zext or trunc to get to the dest size.
1042 return ConstantExpr::getIntegerCast(Input, DestTy, false);
1045 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1046 case Instruction::IntToPtr:
1047 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1048 // the int size is >= the ptr size and the address spaces are the same.
1049 // This requires knowing the width of a pointer, so it can't be done in
1050 // ConstantExpr::getCast.
1051 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1052 if (TD && CE->getOpcode() == Instruction::PtrToInt) {
1053 Constant *SrcPtr = CE->getOperand(0);
1054 unsigned SrcPtrSize = TD->getPointerTypeSizeInBits(SrcPtr->getType());
1055 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1057 if (MidIntSize >= SrcPtrSize) {
1058 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1059 if (SrcAS == DestTy->getPointerAddressSpace())
1060 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
1065 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1066 case Instruction::Trunc:
1067 case Instruction::ZExt:
1068 case Instruction::SExt:
1069 case Instruction::FPTrunc:
1070 case Instruction::FPExt:
1071 case Instruction::UIToFP:
1072 case Instruction::SIToFP:
1073 case Instruction::FPToUI:
1074 case Instruction::FPToSI:
1075 case Instruction::AddrSpaceCast:
1076 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1077 case Instruction::BitCast:
1079 return FoldBitCast(Ops[0], DestTy, *TD);
1080 return ConstantExpr::getBitCast(Ops[0], DestTy);
1081 case Instruction::Select:
1082 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1083 case Instruction::ExtractElement:
1084 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1085 case Instruction::InsertElement:
1086 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1087 case Instruction::ShuffleVector:
1088 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1089 case Instruction::GetElementPtr:
1090 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
1092 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
1095 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
1099 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
1100 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
1101 /// returns a constant expression of the specified operands.
1103 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1104 Constant *Ops0, Constant *Ops1,
1105 const DataLayout *TD,
1106 const TargetLibraryInfo *TLI) {
1107 // fold: icmp (inttoptr x), null -> icmp x, 0
1108 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1109 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1110 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1112 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
1113 // around to know if bit truncation is happening.
1114 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1115 if (TD && Ops1->isNullValue()) {
1116 if (CE0->getOpcode() == Instruction::IntToPtr) {
1117 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1118 // Convert the integer value to the right size to ensure we get the
1119 // proper extension or truncation.
1120 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1122 Constant *Null = Constant::getNullValue(C->getType());
1123 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1126 // Only do this transformation if the int is intptrty in size, otherwise
1127 // there is a truncation or extension that we aren't modeling.
1128 if (CE0->getOpcode() == Instruction::PtrToInt) {
1129 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1130 if (CE0->getType() == IntPtrTy) {
1131 Constant *C = CE0->getOperand(0);
1132 Constant *Null = Constant::getNullValue(C->getType());
1133 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1138 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1139 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1140 if (CE0->getOpcode() == Instruction::IntToPtr) {
1141 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1143 // Convert the integer value to the right size to ensure we get the
1144 // proper extension or truncation.
1145 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1147 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1149 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1152 // Only do this transformation if the int is intptrty in size, otherwise
1153 // there is a truncation or extension that we aren't modeling.
1154 if (CE0->getOpcode() == Instruction::PtrToInt) {
1155 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1156 if (CE0->getType() == IntPtrTy &&
1157 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1158 return ConstantFoldCompareInstOperands(Predicate,
1168 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1169 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1170 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1171 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1173 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1176 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1179 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1180 Constant *Ops[] = { LHS, RHS };
1181 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1185 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1189 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1190 /// getelementptr constantexpr, return the constant value being addressed by the
1191 /// constant expression, or null if something is funny and we can't decide.
1192 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1194 if (!CE->getOperand(1)->isNullValue())
1195 return nullptr; // Do not allow stepping over the value!
1197 // Loop over all of the operands, tracking down which value we are
1199 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1200 C = C->getAggregateElement(CE->getOperand(i));
1207 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1208 /// indices (with an *implied* zero pointer index that is not in the list),
1209 /// return the constant value being addressed by a virtual load, or null if
1210 /// something is funny and we can't decide.
1211 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1212 ArrayRef<Constant*> Indices) {
1213 // Loop over all of the operands, tracking down which value we are
1215 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1216 C = C->getAggregateElement(Indices[i]);
1224 //===----------------------------------------------------------------------===//
1225 // Constant Folding for Calls
1228 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1229 /// the specified function.
1230 bool llvm::canConstantFoldCallTo(const Function *F) {
1231 switch (F->getIntrinsicID()) {
1232 case Intrinsic::fabs:
1233 case Intrinsic::log:
1234 case Intrinsic::log2:
1235 case Intrinsic::log10:
1236 case Intrinsic::exp:
1237 case Intrinsic::exp2:
1238 case Intrinsic::floor:
1239 case Intrinsic::ceil:
1240 case Intrinsic::sqrt:
1241 case Intrinsic::pow:
1242 case Intrinsic::powi:
1243 case Intrinsic::bswap:
1244 case Intrinsic::ctpop:
1245 case Intrinsic::ctlz:
1246 case Intrinsic::cttz:
1247 case Intrinsic::fma:
1248 case Intrinsic::fmuladd:
1249 case Intrinsic::copysign:
1250 case Intrinsic::round:
1251 case Intrinsic::sadd_with_overflow:
1252 case Intrinsic::uadd_with_overflow:
1253 case Intrinsic::ssub_with_overflow:
1254 case Intrinsic::usub_with_overflow:
1255 case Intrinsic::smul_with_overflow:
1256 case Intrinsic::umul_with_overflow:
1257 case Intrinsic::convert_from_fp16:
1258 case Intrinsic::convert_to_fp16:
1259 case Intrinsic::x86_sse_cvtss2si:
1260 case Intrinsic::x86_sse_cvtss2si64:
1261 case Intrinsic::x86_sse_cvttss2si:
1262 case Intrinsic::x86_sse_cvttss2si64:
1263 case Intrinsic::x86_sse2_cvtsd2si:
1264 case Intrinsic::x86_sse2_cvtsd2si64:
1265 case Intrinsic::x86_sse2_cvttsd2si:
1266 case Intrinsic::x86_sse2_cvttsd2si64:
1275 StringRef Name = F->getName();
1277 // In these cases, the check of the length is required. We don't want to
1278 // return true for a name like "cos\0blah" which strcmp would return equal to
1279 // "cos", but has length 8.
1281 default: return false;
1283 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2";
1285 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1287 return Name == "exp" || Name == "exp2";
1289 return Name == "fabs" || Name == "fmod" || Name == "floor";
1291 return Name == "log" || Name == "log10";
1293 return Name == "pow";
1295 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1296 Name == "sinf" || Name == "sqrtf";
1298 return Name == "tan" || Name == "tanh";
1302 static Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1303 if (Ty->isHalfTy()) {
1306 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1307 return ConstantFP::get(Ty->getContext(), APF);
1309 if (Ty->isFloatTy())
1310 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1311 if (Ty->isDoubleTy())
1312 return ConstantFP::get(Ty->getContext(), APFloat(V));
1313 llvm_unreachable("Can only constant fold half/float/double");
1317 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1319 sys::llvm_fenv_clearexcept();
1321 if (sys::llvm_fenv_testexcept()) {
1322 sys::llvm_fenv_clearexcept();
1326 return GetConstantFoldFPValue(V, Ty);
1329 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1330 double V, double W, Type *Ty) {
1331 sys::llvm_fenv_clearexcept();
1333 if (sys::llvm_fenv_testexcept()) {
1334 sys::llvm_fenv_clearexcept();
1338 return GetConstantFoldFPValue(V, Ty);
1341 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1342 /// conversion of a constant floating point. If roundTowardZero is false, the
1343 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1344 /// the behavior of the non-truncating SSE instructions in the default rounding
1345 /// mode. The desired integer type Ty is used to select how many bits are
1346 /// available for the result. Returns null if the conversion cannot be
1347 /// performed, otherwise returns the Constant value resulting from the
1349 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1350 bool roundTowardZero, Type *Ty) {
1351 // All of these conversion intrinsics form an integer of at most 64bits.
1352 unsigned ResultWidth = Ty->getIntegerBitWidth();
1353 assert(ResultWidth <= 64 &&
1354 "Can only constant fold conversions to 64 and 32 bit ints");
1357 bool isExact = false;
1358 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1359 : APFloat::rmNearestTiesToEven;
1360 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1361 /*isSigned=*/true, mode,
1363 if (status != APFloat::opOK && status != APFloat::opInexact)
1365 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1368 static double getValueAsDouble(ConstantFP *Op) {
1369 Type *Ty = Op->getType();
1371 if (Ty->isFloatTy())
1372 return Op->getValueAPF().convertToFloat();
1374 if (Ty->isDoubleTy())
1375 return Op->getValueAPF().convertToDouble();
1378 APFloat APF = Op->getValueAPF();
1379 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1380 return APF.convertToDouble();
1383 static Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID,
1384 Type *Ty, ArrayRef<Constant *> Operands,
1385 const TargetLibraryInfo *TLI) {
1386 if (Operands.size() == 1) {
1387 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1388 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1389 APFloat Val(Op->getValueAPF());
1392 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1394 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1397 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1400 if (IntrinsicID == Intrinsic::round) {
1401 APFloat V = Op->getValueAPF();
1402 V.roundToIntegral(APFloat::rmNearestTiesToAway);
1403 return ConstantFP::get(Ty->getContext(), V);
1406 /// We only fold functions with finite arguments. Folding NaN and inf is
1407 /// likely to be aborted with an exception anyway, and some host libms
1408 /// have known errors raising exceptions.
1409 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1412 /// Currently APFloat versions of these functions do not exist, so we use
1413 /// the host native double versions. Float versions are not called
1414 /// directly but for all these it is true (float)(f((double)arg)) ==
1415 /// f(arg). Long double not supported yet.
1416 double V = getValueAsDouble(Op);
1418 switch (IntrinsicID) {
1420 case Intrinsic::fabs:
1421 return ConstantFoldFP(fabs, V, Ty);
1423 case Intrinsic::log2:
1424 return ConstantFoldFP(log2, V, Ty);
1427 case Intrinsic::log:
1428 return ConstantFoldFP(log, V, Ty);
1431 case Intrinsic::log10:
1432 return ConstantFoldFP(log10, V, Ty);
1435 case Intrinsic::exp:
1436 return ConstantFoldFP(exp, V, Ty);
1439 case Intrinsic::exp2:
1440 return ConstantFoldFP(exp2, V, Ty);
1442 case Intrinsic::floor:
1443 return ConstantFoldFP(floor, V, Ty);
1444 case Intrinsic::ceil:
1445 return ConstantFoldFP(ceil, V, Ty);
1453 if (Name == "acos" && TLI->has(LibFunc::acos))
1454 return ConstantFoldFP(acos, V, Ty);
1455 else if (Name == "asin" && TLI->has(LibFunc::asin))
1456 return ConstantFoldFP(asin, V, Ty);
1457 else if (Name == "atan" && TLI->has(LibFunc::atan))
1458 return ConstantFoldFP(atan, V, Ty);
1461 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1462 return ConstantFoldFP(ceil, V, Ty);
1463 else if (Name == "cos" && TLI->has(LibFunc::cos))
1464 return ConstantFoldFP(cos, V, Ty);
1465 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1466 return ConstantFoldFP(cosh, V, Ty);
1467 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1468 return ConstantFoldFP(cos, V, Ty);
1471 if (Name == "exp" && TLI->has(LibFunc::exp))
1472 return ConstantFoldFP(exp, V, Ty);
1474 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1475 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1477 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1481 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1482 return ConstantFoldFP(fabs, V, Ty);
1483 else if (Name == "floor" && TLI->has(LibFunc::floor))
1484 return ConstantFoldFP(floor, V, Ty);
1487 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1488 return ConstantFoldFP(log, V, Ty);
1489 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1490 return ConstantFoldFP(log10, V, Ty);
1491 else if (IntrinsicID == Intrinsic::sqrt &&
1492 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1494 return ConstantFoldFP(sqrt, V, Ty);
1496 return Constant::getNullValue(Ty);
1500 if (Name == "sin" && TLI->has(LibFunc::sin))
1501 return ConstantFoldFP(sin, V, Ty);
1502 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1503 return ConstantFoldFP(sinh, V, Ty);
1504 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1505 return ConstantFoldFP(sqrt, V, Ty);
1506 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1507 return ConstantFoldFP(sqrt, V, Ty);
1508 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1509 return ConstantFoldFP(sin, V, Ty);
1512 if (Name == "tan" && TLI->has(LibFunc::tan))
1513 return ConstantFoldFP(tan, V, Ty);
1514 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1515 return ConstantFoldFP(tanh, V, Ty);
1523 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1524 switch (IntrinsicID) {
1525 case Intrinsic::bswap:
1526 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1527 case Intrinsic::ctpop:
1528 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1529 case Intrinsic::convert_from_fp16: {
1530 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1533 APFloat::opStatus status =
1534 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1536 // Conversion is always precise.
1538 assert(status == APFloat::opOK && !lost &&
1539 "Precision lost during fp16 constfolding");
1541 return ConstantFP::get(Ty->getContext(), Val);
1548 // Support ConstantVector in case we have an Undef in the top.
1549 if (isa<ConstantVector>(Operands[0]) ||
1550 isa<ConstantDataVector>(Operands[0])) {
1551 Constant *Op = cast<Constant>(Operands[0]);
1552 switch (IntrinsicID) {
1554 case Intrinsic::x86_sse_cvtss2si:
1555 case Intrinsic::x86_sse_cvtss2si64:
1556 case Intrinsic::x86_sse2_cvtsd2si:
1557 case Intrinsic::x86_sse2_cvtsd2si64:
1558 if (ConstantFP *FPOp =
1559 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1560 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1561 /*roundTowardZero=*/false, Ty);
1562 case Intrinsic::x86_sse_cvttss2si:
1563 case Intrinsic::x86_sse_cvttss2si64:
1564 case Intrinsic::x86_sse2_cvttsd2si:
1565 case Intrinsic::x86_sse2_cvttsd2si64:
1566 if (ConstantFP *FPOp =
1567 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1568 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1569 /*roundTowardZero=*/true, Ty);
1573 if (isa<UndefValue>(Operands[0])) {
1574 if (IntrinsicID == Intrinsic::bswap)
1582 if (Operands.size() == 2) {
1583 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1584 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1586 double Op1V = getValueAsDouble(Op1);
1588 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1589 if (Op2->getType() != Op1->getType())
1592 double Op2V = getValueAsDouble(Op2);
1593 if (IntrinsicID == Intrinsic::pow) {
1594 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1596 if (IntrinsicID == Intrinsic::copysign) {
1597 APFloat V1 = Op1->getValueAPF();
1598 APFloat V2 = Op2->getValueAPF();
1600 return ConstantFP::get(Ty->getContext(), V1);
1604 if (Name == "pow" && TLI->has(LibFunc::pow))
1605 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1606 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1607 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1608 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1609 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1610 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1611 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1612 return ConstantFP::get(Ty->getContext(),
1613 APFloat((float)std::pow((float)Op1V,
1614 (int)Op2C->getZExtValue())));
1615 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1616 return ConstantFP::get(Ty->getContext(),
1617 APFloat((float)std::pow((float)Op1V,
1618 (int)Op2C->getZExtValue())));
1619 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1620 return ConstantFP::get(Ty->getContext(),
1621 APFloat((double)std::pow((double)Op1V,
1622 (int)Op2C->getZExtValue())));
1627 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1628 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1629 switch (IntrinsicID) {
1631 case Intrinsic::sadd_with_overflow:
1632 case Intrinsic::uadd_with_overflow:
1633 case Intrinsic::ssub_with_overflow:
1634 case Intrinsic::usub_with_overflow:
1635 case Intrinsic::smul_with_overflow:
1636 case Intrinsic::umul_with_overflow: {
1639 switch (IntrinsicID) {
1640 default: llvm_unreachable("Invalid case");
1641 case Intrinsic::sadd_with_overflow:
1642 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1644 case Intrinsic::uadd_with_overflow:
1645 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1647 case Intrinsic::ssub_with_overflow:
1648 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1650 case Intrinsic::usub_with_overflow:
1651 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1653 case Intrinsic::smul_with_overflow:
1654 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1656 case Intrinsic::umul_with_overflow:
1657 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1661 ConstantInt::get(Ty->getContext(), Res),
1662 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1664 return ConstantStruct::get(cast<StructType>(Ty), Ops);
1666 case Intrinsic::cttz:
1667 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1668 return UndefValue::get(Ty);
1669 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1670 case Intrinsic::ctlz:
1671 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1672 return UndefValue::get(Ty);
1673 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1682 if (Operands.size() != 3)
1685 if (const ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1686 if (const ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1687 if (const ConstantFP *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1688 switch (IntrinsicID) {
1690 case Intrinsic::fma:
1691 case Intrinsic::fmuladd: {
1692 APFloat V = Op1->getValueAPF();
1693 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
1695 APFloat::rmNearestTiesToEven);
1696 if (s != APFloat::opInvalidOp)
1697 return ConstantFP::get(Ty->getContext(), V);
1709 static Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
1711 ArrayRef<Constant *> Operands,
1712 const TargetLibraryInfo *TLI) {
1713 SmallVector<Constant *, 4> Result(VTy->getNumElements());
1714 SmallVector<Constant *, 4> Lane(Operands.size());
1715 Type *Ty = VTy->getElementType();
1717 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1718 // Gather a column of constants.
1719 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
1720 Constant *Agg = Operands[J]->getAggregateElement(I);
1727 // Use the regular scalar folding to simplify this column.
1728 Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI);
1734 return ConstantVector::get(Result);
1737 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1738 /// with the specified arguments, returning null if unsuccessful.
1740 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1741 const TargetLibraryInfo *TLI) {
1744 StringRef Name = F->getName();
1746 Type *Ty = F->getReturnType();
1748 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
1749 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands, TLI);
1751 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);