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/GlobalVariable.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/FEnv.h"
34 #include "llvm/Support/GetElementPtrTypeIterator.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 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
470 /// produce if it is constant and determinable. If this is not determinable,
472 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
473 const DataLayout *TD) {
474 // First, try the easy cases:
475 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
476 if (GV->isConstant() && GV->hasDefinitiveInitializer())
477 return GV->getInitializer();
479 // If the loaded value isn't a constant expr, we can't handle it.
480 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
484 if (CE->getOpcode() == Instruction::GetElementPtr) {
485 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
486 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
488 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
494 // Instead of loading constant c string, use corresponding integer value
495 // directly if string length is small enough.
497 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
498 unsigned StrLen = Str.size();
499 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
500 unsigned NumBits = Ty->getPrimitiveSizeInBits();
501 // Replace load with immediate integer if the result is an integer or fp
503 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
504 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
505 APInt StrVal(NumBits, 0);
506 APInt SingleChar(NumBits, 0);
507 if (TD->isLittleEndian()) {
508 for (signed i = StrLen-1; i >= 0; i--) {
509 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
510 StrVal = (StrVal << 8) | SingleChar;
513 for (unsigned i = 0; i < StrLen; i++) {
514 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
515 StrVal = (StrVal << 8) | SingleChar;
517 // Append NULL at the end.
519 StrVal = (StrVal << 8) | SingleChar;
522 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
523 if (Ty->isFloatingPointTy())
524 Res = ConstantExpr::getBitCast(Res, Ty);
529 // If this load comes from anywhere in a constant global, and if the global
530 // is all undef or zero, we know what it loads.
531 if (GlobalVariable *GV =
532 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
533 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
534 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
535 if (GV->getInitializer()->isNullValue())
536 return Constant::getNullValue(ResTy);
537 if (isa<UndefValue>(GV->getInitializer()))
538 return UndefValue::get(ResTy);
542 // Try hard to fold loads from bitcasted strange and non-type-safe things.
544 return FoldReinterpretLoadFromConstPtr(CE, *TD);
548 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
549 if (LI->isVolatile()) return 0;
551 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
552 return ConstantFoldLoadFromConstPtr(C, TD);
557 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
558 /// Attempt to symbolically evaluate the result of a binary operator merging
559 /// these together. If target data info is available, it is provided as DL,
560 /// otherwise DL is null.
561 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
562 Constant *Op1, const DataLayout *DL){
565 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
566 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
570 if (Opc == Instruction::And && DL) {
571 unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType());
572 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
573 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
574 ComputeMaskedBits(Op0, KnownZero0, KnownOne0, DL);
575 ComputeMaskedBits(Op1, KnownZero1, KnownOne1, DL);
576 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
577 // All the bits of Op0 that the 'and' could be masking are already zero.
580 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
581 // All the bits of Op1 that the 'and' could be masking are already zero.
585 APInt KnownZero = KnownZero0 | KnownZero1;
586 APInt KnownOne = KnownOne0 & KnownOne1;
587 if ((KnownZero | KnownOne).isAllOnesValue()) {
588 return ConstantInt::get(Op0->getType(), KnownOne);
592 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
593 // constant. This happens frequently when iterating over a global array.
594 if (Opc == Instruction::Sub && DL) {
595 GlobalValue *GV1, *GV2;
598 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL))
599 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) &&
601 unsigned OpSize = DL->getTypeSizeInBits(Op0->getType());
603 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
604 // PtrToInt may change the bitwidth so we have convert to the right size
606 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
607 Offs2.zextOrTrunc(OpSize));
614 /// CastGEPIndices - If array indices are not pointer-sized integers,
615 /// explicitly cast them so that they aren't implicitly casted by the
617 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
618 Type *ResultTy, const DataLayout *TD,
619 const TargetLibraryInfo *TLI) {
623 Type *IntPtrTy = TD->getIntPtrType(ResultTy);
626 SmallVector<Constant*, 32> NewIdxs;
627 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
629 !isa<StructType>(GetElementPtrInst::getIndexedType(
631 Ops.slice(1, i - 1)))) &&
632 Ops[i]->getType() != IntPtrTy) {
634 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
640 NewIdxs.push_back(Ops[i]);
646 Constant *C = ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
647 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
648 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
655 /// Strip the pointer casts, but preserve the address space information.
656 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
657 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
658 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
659 Ptr = cast<Constant>(Ptr->stripPointerCasts());
660 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
662 // Preserve the address space number of the pointer.
663 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
664 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
665 OldPtrTy->getAddressSpace());
666 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
671 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
672 /// constant expression, do so.
673 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
674 Type *ResultTy, const DataLayout *TD,
675 const TargetLibraryInfo *TLI) {
676 Constant *Ptr = Ops[0];
677 if (!TD || !Ptr->getType()->getPointerElementType()->isSized() ||
678 !Ptr->getType()->isPointerTy())
681 Type *IntPtrTy = TD->getIntPtrType(Ptr->getType());
682 Type *ResultElementTy = ResultTy->getPointerElementType();
684 // If this is a constant expr gep that is effectively computing an
685 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
686 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
687 if (!isa<ConstantInt>(Ops[i])) {
689 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
690 // "inttoptr (sub (ptrtoint Ptr), V)"
691 if (Ops.size() == 2 && ResultElementTy->isIntegerTy(8)) {
692 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
693 assert((CE == 0 || CE->getType() == IntPtrTy) &&
694 "CastGEPIndices didn't canonicalize index types!");
695 if (CE && CE->getOpcode() == Instruction::Sub &&
696 CE->getOperand(0)->isNullValue()) {
697 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
698 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
699 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
700 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
701 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
708 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
710 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
711 makeArrayRef((Value *const*)
714 Ptr = StripPtrCastKeepAS(Ptr);
716 // If this is a GEP of a GEP, fold it all into a single GEP.
717 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
718 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
720 // Do not try the incorporate the sub-GEP if some index is not a number.
721 bool AllConstantInt = true;
722 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
723 if (!isa<ConstantInt>(NestedOps[i])) {
724 AllConstantInt = false;
730 Ptr = cast<Constant>(GEP->getOperand(0));
731 Offset += APInt(BitWidth,
732 TD->getIndexedOffset(Ptr->getType(), NestedOps));
733 Ptr = StripPtrCastKeepAS(Ptr);
736 // If the base value for this address is a literal integer value, fold the
737 // getelementptr to the resulting integer value casted to the pointer type.
738 APInt BasePtr(BitWidth, 0);
739 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
740 if (CE->getOpcode() == Instruction::IntToPtr) {
741 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
742 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
746 if (Ptr->isNullValue() || BasePtr != 0) {
747 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
748 return ConstantExpr::getIntToPtr(C, ResultTy);
751 // Otherwise form a regular getelementptr. Recompute the indices so that
752 // we eliminate over-indexing of the notional static type array bounds.
753 // This makes it easy to determine if the getelementptr is "inbounds".
754 // Also, this helps GlobalOpt do SROA on GlobalVariables.
755 Type *Ty = Ptr->getType();
756 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
757 SmallVector<Constant *, 32> NewIdxs;
760 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
761 if (ATy->isPointerTy()) {
762 // The only pointer indexing we'll do is on the first index of the GEP.
763 if (!NewIdxs.empty())
766 // Only handle pointers to sized types, not pointers to functions.
767 if (!ATy->getElementType()->isSized())
771 // Determine which element of the array the offset points into.
772 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
774 // The element size is 0. This may be [0 x Ty]*, so just use a zero
775 // index for this level and proceed to the next level to see if it can
776 // accommodate the offset.
777 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
779 // The element size is non-zero divide the offset by the element
780 // size (rounding down), to compute the index at this level.
781 APInt NewIdx = Offset.udiv(ElemSize);
782 Offset -= NewIdx * ElemSize;
783 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
785 Ty = ATy->getElementType();
786 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
787 // If we end up with an offset that isn't valid for this struct type, we
788 // can't re-form this GEP in a regular form, so bail out. The pointer
789 // operand likely went through casts that are necessary to make the GEP
791 const StructLayout &SL = *TD->getStructLayout(STy);
792 if (Offset.uge(SL.getSizeInBytes()))
795 // Determine which field of the struct the offset points into. The
796 // getZExtValue is fine as we've already ensured that the offset is
797 // within the range representable by the StructLayout API.
798 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
799 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
801 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
802 Ty = STy->getTypeAtIndex(ElIdx);
804 // We've reached some non-indexable type.
807 } while (Ty != ResultElementTy);
809 // If we haven't used up the entire offset by descending the static
810 // type, then the offset is pointing into the middle of an indivisible
811 // member, so we can't simplify it.
816 Constant *C = ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
817 assert(C->getType()->getPointerElementType() == Ty &&
818 "Computed GetElementPtr has unexpected type!");
820 // If we ended up indexing a member with a type that doesn't match
821 // the type of what the original indices indexed, add a cast.
822 if (Ty != ResultElementTy)
823 C = FoldBitCast(C, ResultTy, *TD);
830 //===----------------------------------------------------------------------===//
831 // Constant Folding public APIs
832 //===----------------------------------------------------------------------===//
834 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
835 /// If successful, the constant result is returned, if not, null is returned.
836 /// Note that this fails if not all of the operands are constant. Otherwise,
837 /// this function can only fail when attempting to fold instructions like loads
838 /// and stores, which have no constant expression form.
839 Constant *llvm::ConstantFoldInstruction(Instruction *I,
840 const DataLayout *TD,
841 const TargetLibraryInfo *TLI) {
842 // Handle PHI nodes quickly here...
843 if (PHINode *PN = dyn_cast<PHINode>(I)) {
844 Constant *CommonValue = 0;
846 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
847 Value *Incoming = PN->getIncomingValue(i);
848 // If the incoming value is undef then skip it. Note that while we could
849 // skip the value if it is equal to the phi node itself we choose not to
850 // because that would break the rule that constant folding only applies if
851 // all operands are constants.
852 if (isa<UndefValue>(Incoming))
854 // If the incoming value is not a constant, then give up.
855 Constant *C = dyn_cast<Constant>(Incoming);
858 // Fold the PHI's operands.
859 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
860 C = ConstantFoldConstantExpression(NewC, TD, TLI);
861 // If the incoming value is a different constant to
862 // the one we saw previously, then give up.
863 if (CommonValue && C != CommonValue)
869 // If we reach here, all incoming values are the same constant or undef.
870 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
873 // Scan the operand list, checking to see if they are all constants, if so,
874 // hand off to ConstantFoldInstOperands.
875 SmallVector<Constant*, 8> Ops;
876 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
877 Constant *Op = dyn_cast<Constant>(*i);
879 return 0; // All operands not constant!
881 // Fold the Instruction's operands.
882 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
883 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
888 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
889 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
892 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
893 return ConstantFoldLoadInst(LI, TD);
895 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) {
896 return ConstantExpr::getInsertValue(
897 cast<Constant>(IVI->getAggregateOperand()),
898 cast<Constant>(IVI->getInsertedValueOperand()),
902 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) {
903 return ConstantExpr::getExtractValue(
904 cast<Constant>(EVI->getAggregateOperand()),
908 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
912 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD,
913 const TargetLibraryInfo *TLI,
914 SmallPtrSet<ConstantExpr *, 4> &FoldedOps) {
915 SmallVector<Constant *, 8> Ops;
916 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e;
918 Constant *NewC = cast<Constant>(*i);
919 // Recursively fold the ConstantExpr's operands. If we have already folded
920 // a ConstantExpr, we don't have to process it again.
921 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) {
922 if (FoldedOps.insert(NewCE))
923 NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps);
929 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
931 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
934 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
935 /// using the specified DataLayout. If successful, the constant result is
936 /// result is returned, if not, null is returned.
937 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
938 const DataLayout *TD,
939 const TargetLibraryInfo *TLI) {
940 SmallPtrSet<ConstantExpr *, 4> FoldedOps;
941 return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps);
944 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
945 /// specified opcode and operands. If successful, the constant result is
946 /// returned, if not, null is returned. Note that this function can fail when
947 /// attempting to fold instructions like loads and stores, which have no
948 /// constant expression form.
950 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
951 /// information, due to only being passed an opcode and operands. Constant
952 /// folding using this function strips this information.
954 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
955 ArrayRef<Constant *> Ops,
956 const DataLayout *TD,
957 const TargetLibraryInfo *TLI) {
958 // Handle easy binops first.
959 if (Instruction::isBinaryOp(Opcode)) {
960 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) {
961 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
965 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
970 case Instruction::ICmp:
971 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
972 case Instruction::Call:
973 if (Function *F = dyn_cast<Function>(Ops.back()))
974 if (canConstantFoldCallTo(F))
975 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
977 case Instruction::PtrToInt:
978 // If the input is a inttoptr, eliminate the pair. This requires knowing
979 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
980 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
981 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
982 Constant *Input = CE->getOperand(0);
983 unsigned InWidth = Input->getType()->getScalarSizeInBits();
984 unsigned PtrWidth = TD->getPointerTypeSizeInBits(CE->getType());
985 if (PtrWidth < InWidth) {
987 ConstantInt::get(CE->getContext(),
988 APInt::getLowBitsSet(InWidth, PtrWidth));
989 Input = ConstantExpr::getAnd(Input, Mask);
991 // Do a zext or trunc to get to the dest size.
992 return ConstantExpr::getIntegerCast(Input, DestTy, false);
995 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
996 case Instruction::IntToPtr:
997 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
998 // the int size is >= the ptr size and the address spaces are the same.
999 // This requires knowing the width of a pointer, so it can't be done in
1000 // ConstantExpr::getCast.
1001 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1002 if (TD && CE->getOpcode() == Instruction::PtrToInt) {
1003 Constant *SrcPtr = CE->getOperand(0);
1004 unsigned SrcPtrSize = TD->getPointerTypeSizeInBits(SrcPtr->getType());
1005 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1007 if (MidIntSize >= SrcPtrSize) {
1008 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1009 if (SrcAS == DestTy->getPointerAddressSpace())
1010 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
1015 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1016 case Instruction::Trunc:
1017 case Instruction::ZExt:
1018 case Instruction::SExt:
1019 case Instruction::FPTrunc:
1020 case Instruction::FPExt:
1021 case Instruction::UIToFP:
1022 case Instruction::SIToFP:
1023 case Instruction::FPToUI:
1024 case Instruction::FPToSI:
1025 case Instruction::AddrSpaceCast:
1026 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1027 case Instruction::BitCast:
1029 return FoldBitCast(Ops[0], DestTy, *TD);
1030 return ConstantExpr::getBitCast(Ops[0], DestTy);
1031 case Instruction::Select:
1032 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1033 case Instruction::ExtractElement:
1034 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1035 case Instruction::InsertElement:
1036 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1037 case Instruction::ShuffleVector:
1038 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1039 case Instruction::GetElementPtr:
1040 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
1042 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
1045 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
1049 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
1050 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
1051 /// returns a constant expression of the specified operands.
1053 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1054 Constant *Ops0, Constant *Ops1,
1055 const DataLayout *TD,
1056 const TargetLibraryInfo *TLI) {
1057 // fold: icmp (inttoptr x), null -> icmp x, 0
1058 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1059 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1060 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1062 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
1063 // around to know if bit truncation is happening.
1064 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1065 if (TD && Ops1->isNullValue()) {
1066 if (CE0->getOpcode() == Instruction::IntToPtr) {
1067 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1068 // Convert the integer value to the right size to ensure we get the
1069 // proper extension or truncation.
1070 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1072 Constant *Null = Constant::getNullValue(C->getType());
1073 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1076 // Only do this transformation if the int is intptrty in size, otherwise
1077 // there is a truncation or extension that we aren't modeling.
1078 if (CE0->getOpcode() == Instruction::PtrToInt) {
1079 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1080 if (CE0->getType() == IntPtrTy) {
1081 Constant *C = CE0->getOperand(0);
1082 Constant *Null = Constant::getNullValue(C->getType());
1083 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1088 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1089 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1090 if (CE0->getOpcode() == Instruction::IntToPtr) {
1091 Type *IntPtrTy = TD->getIntPtrType(CE0->getType());
1093 // Convert the integer value to the right size to ensure we get the
1094 // proper extension or truncation.
1095 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1097 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1099 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1102 // Only do this transformation if the int is intptrty in size, otherwise
1103 // there is a truncation or extension that we aren't modeling.
1104 if (CE0->getOpcode() == Instruction::PtrToInt) {
1105 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType());
1106 if (CE0->getType() == IntPtrTy &&
1107 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1108 return ConstantFoldCompareInstOperands(Predicate,
1118 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1119 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1120 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1121 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1123 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1126 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1129 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1130 Constant *Ops[] = { LHS, RHS };
1131 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1135 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1139 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1140 /// getelementptr constantexpr, return the constant value being addressed by the
1141 /// constant expression, or null if something is funny and we can't decide.
1142 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1144 if (!CE->getOperand(1)->isNullValue())
1145 return 0; // Do not allow stepping over the value!
1147 // Loop over all of the operands, tracking down which value we are
1149 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1150 C = C->getAggregateElement(CE->getOperand(i));
1157 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1158 /// indices (with an *implied* zero pointer index that is not in the list),
1159 /// return the constant value being addressed by a virtual load, or null if
1160 /// something is funny and we can't decide.
1161 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1162 ArrayRef<Constant*> Indices) {
1163 // Loop over all of the operands, tracking down which value we are
1165 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1166 C = C->getAggregateElement(Indices[i]);
1174 //===----------------------------------------------------------------------===//
1175 // Constant Folding for Calls
1178 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1179 /// the specified function.
1180 bool llvm::canConstantFoldCallTo(const Function *F) {
1181 switch (F->getIntrinsicID()) {
1182 case Intrinsic::fabs:
1183 case Intrinsic::log:
1184 case Intrinsic::log2:
1185 case Intrinsic::log10:
1186 case Intrinsic::exp:
1187 case Intrinsic::exp2:
1188 case Intrinsic::floor:
1189 case Intrinsic::sqrt:
1190 case Intrinsic::pow:
1191 case Intrinsic::powi:
1192 case Intrinsic::bswap:
1193 case Intrinsic::ctpop:
1194 case Intrinsic::ctlz:
1195 case Intrinsic::cttz:
1196 case Intrinsic::sadd_with_overflow:
1197 case Intrinsic::uadd_with_overflow:
1198 case Intrinsic::ssub_with_overflow:
1199 case Intrinsic::usub_with_overflow:
1200 case Intrinsic::smul_with_overflow:
1201 case Intrinsic::umul_with_overflow:
1202 case Intrinsic::convert_from_fp16:
1203 case Intrinsic::convert_to_fp16:
1204 case Intrinsic::x86_sse_cvtss2si:
1205 case Intrinsic::x86_sse_cvtss2si64:
1206 case Intrinsic::x86_sse_cvttss2si:
1207 case Intrinsic::x86_sse_cvttss2si64:
1208 case Intrinsic::x86_sse2_cvtsd2si:
1209 case Intrinsic::x86_sse2_cvtsd2si64:
1210 case Intrinsic::x86_sse2_cvttsd2si:
1211 case Intrinsic::x86_sse2_cvttsd2si64:
1220 StringRef Name = F->getName();
1222 // In these cases, the check of the length is required. We don't want to
1223 // return true for a name like "cos\0blah" which strcmp would return equal to
1224 // "cos", but has length 8.
1226 default: return false;
1228 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2";
1230 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1232 return Name == "exp" || Name == "exp2";
1234 return Name == "fabs" || Name == "fmod" || Name == "floor";
1236 return Name == "log" || Name == "log10";
1238 return Name == "pow";
1240 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1241 Name == "sinf" || Name == "sqrtf";
1243 return Name == "tan" || Name == "tanh";
1247 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1249 sys::llvm_fenv_clearexcept();
1251 if (sys::llvm_fenv_testexcept()) {
1252 sys::llvm_fenv_clearexcept();
1256 if (Ty->isHalfTy()) {
1259 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1260 return ConstantFP::get(Ty->getContext(), APF);
1262 if (Ty->isFloatTy())
1263 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1264 if (Ty->isDoubleTy())
1265 return ConstantFP::get(Ty->getContext(), APFloat(V));
1266 llvm_unreachable("Can only constant fold half/float/double");
1269 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1270 double V, double W, Type *Ty) {
1271 sys::llvm_fenv_clearexcept();
1273 if (sys::llvm_fenv_testexcept()) {
1274 sys::llvm_fenv_clearexcept();
1278 if (Ty->isHalfTy()) {
1281 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1282 return ConstantFP::get(Ty->getContext(), APF);
1284 if (Ty->isFloatTy())
1285 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1286 if (Ty->isDoubleTy())
1287 return ConstantFP::get(Ty->getContext(), APFloat(V));
1288 llvm_unreachable("Can only constant fold half/float/double");
1291 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1292 /// conversion of a constant floating point. If roundTowardZero is false, the
1293 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1294 /// the behavior of the non-truncating SSE instructions in the default rounding
1295 /// mode. The desired integer type Ty is used to select how many bits are
1296 /// available for the result. Returns null if the conversion cannot be
1297 /// performed, otherwise returns the Constant value resulting from the
1299 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1300 bool roundTowardZero, Type *Ty) {
1301 // All of these conversion intrinsics form an integer of at most 64bits.
1302 unsigned ResultWidth = Ty->getIntegerBitWidth();
1303 assert(ResultWidth <= 64 &&
1304 "Can only constant fold conversions to 64 and 32 bit ints");
1307 bool isExact = false;
1308 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1309 : APFloat::rmNearestTiesToEven;
1310 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1311 /*isSigned=*/true, mode,
1313 if (status != APFloat::opOK && status != APFloat::opInexact)
1315 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1318 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1319 /// with the specified arguments, returning null if unsuccessful.
1321 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1322 const TargetLibraryInfo *TLI) {
1325 StringRef Name = F->getName();
1327 Type *Ty = F->getReturnType();
1328 if (Operands.size() == 1) {
1329 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1330 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1331 APFloat Val(Op->getValueAPF());
1334 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1336 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1341 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1344 /// We only fold functions with finite arguments. Folding NaN and inf is
1345 /// likely to be aborted with an exception anyway, and some host libms
1346 /// have known errors raising exceptions.
1347 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1350 /// Currently APFloat versions of these functions do not exist, so we use
1351 /// the host native double versions. Float versions are not called
1352 /// directly but for all these it is true (float)(f((double)arg)) ==
1353 /// f(arg). Long double not supported yet.
1355 if (Ty->isFloatTy())
1356 V = Op->getValueAPF().convertToFloat();
1357 else if (Ty->isDoubleTy())
1358 V = Op->getValueAPF().convertToDouble();
1361 APFloat APF = Op->getValueAPF();
1362 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1363 V = APF.convertToDouble();
1366 switch (F->getIntrinsicID()) {
1368 case Intrinsic::fabs:
1369 return ConstantFoldFP(fabs, V, Ty);
1371 case Intrinsic::log2:
1372 return ConstantFoldFP(log2, V, Ty);
1375 case Intrinsic::log:
1376 return ConstantFoldFP(log, V, Ty);
1379 case Intrinsic::log10:
1380 return ConstantFoldFP(log10, V, Ty);
1383 case Intrinsic::exp:
1384 return ConstantFoldFP(exp, V, Ty);
1387 case Intrinsic::exp2:
1388 return ConstantFoldFP(exp2, V, Ty);
1390 case Intrinsic::floor:
1391 return ConstantFoldFP(floor, V, Ty);
1396 if (Name == "acos" && TLI->has(LibFunc::acos))
1397 return ConstantFoldFP(acos, V, Ty);
1398 else if (Name == "asin" && TLI->has(LibFunc::asin))
1399 return ConstantFoldFP(asin, V, Ty);
1400 else if (Name == "atan" && TLI->has(LibFunc::atan))
1401 return ConstantFoldFP(atan, V, Ty);
1404 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1405 return ConstantFoldFP(ceil, V, Ty);
1406 else if (Name == "cos" && TLI->has(LibFunc::cos))
1407 return ConstantFoldFP(cos, V, Ty);
1408 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1409 return ConstantFoldFP(cosh, V, Ty);
1410 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1411 return ConstantFoldFP(cos, V, Ty);
1414 if (Name == "exp" && TLI->has(LibFunc::exp))
1415 return ConstantFoldFP(exp, V, Ty);
1417 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1418 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1420 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1424 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1425 return ConstantFoldFP(fabs, V, Ty);
1426 else if (Name == "floor" && TLI->has(LibFunc::floor))
1427 return ConstantFoldFP(floor, V, Ty);
1430 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1431 return ConstantFoldFP(log, V, Ty);
1432 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1433 return ConstantFoldFP(log10, V, Ty);
1434 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1435 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1437 return ConstantFoldFP(sqrt, V, Ty);
1439 return Constant::getNullValue(Ty);
1443 if (Name == "sin" && TLI->has(LibFunc::sin))
1444 return ConstantFoldFP(sin, V, Ty);
1445 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1446 return ConstantFoldFP(sinh, V, Ty);
1447 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1448 return ConstantFoldFP(sqrt, V, Ty);
1449 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1450 return ConstantFoldFP(sqrt, V, Ty);
1451 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1452 return ConstantFoldFP(sin, V, Ty);
1455 if (Name == "tan" && TLI->has(LibFunc::tan))
1456 return ConstantFoldFP(tan, V, Ty);
1457 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1458 return ConstantFoldFP(tanh, V, Ty);
1466 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1467 switch (F->getIntrinsicID()) {
1468 case Intrinsic::bswap:
1469 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1470 case Intrinsic::ctpop:
1471 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1472 case Intrinsic::convert_from_fp16: {
1473 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1476 APFloat::opStatus status =
1477 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1479 // Conversion is always precise.
1481 assert(status == APFloat::opOK && !lost &&
1482 "Precision lost during fp16 constfolding");
1484 return ConstantFP::get(F->getContext(), Val);
1491 // Support ConstantVector in case we have an Undef in the top.
1492 if (isa<ConstantVector>(Operands[0]) ||
1493 isa<ConstantDataVector>(Operands[0])) {
1494 Constant *Op = cast<Constant>(Operands[0]);
1495 switch (F->getIntrinsicID()) {
1497 case Intrinsic::x86_sse_cvtss2si:
1498 case Intrinsic::x86_sse_cvtss2si64:
1499 case Intrinsic::x86_sse2_cvtsd2si:
1500 case Intrinsic::x86_sse2_cvtsd2si64:
1501 if (ConstantFP *FPOp =
1502 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1503 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1504 /*roundTowardZero=*/false, Ty);
1505 case Intrinsic::x86_sse_cvttss2si:
1506 case Intrinsic::x86_sse_cvttss2si64:
1507 case Intrinsic::x86_sse2_cvttsd2si:
1508 case Intrinsic::x86_sse2_cvttsd2si64:
1509 if (ConstantFP *FPOp =
1510 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1511 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1512 /*roundTowardZero=*/true, Ty);
1516 if (isa<UndefValue>(Operands[0])) {
1517 if (F->getIntrinsicID() == Intrinsic::bswap)
1525 if (Operands.size() == 2) {
1526 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1527 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1530 if (Ty->isFloatTy())
1531 Op1V = Op1->getValueAPF().convertToFloat();
1532 else if (Ty->isDoubleTy())
1533 Op1V = Op1->getValueAPF().convertToDouble();
1536 APFloat APF = Op1->getValueAPF();
1537 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1538 Op1V = APF.convertToDouble();
1541 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1542 if (Op2->getType() != Op1->getType())
1546 if (Ty->isFloatTy())
1547 Op2V = Op2->getValueAPF().convertToFloat();
1548 else if (Ty->isDoubleTy())
1549 Op2V = Op2->getValueAPF().convertToDouble();
1552 APFloat APF = Op2->getValueAPF();
1553 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1554 Op2V = APF.convertToDouble();
1557 if (F->getIntrinsicID() == Intrinsic::pow) {
1558 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1562 if (Name == "pow" && TLI->has(LibFunc::pow))
1563 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1564 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1565 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1566 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1567 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1568 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1569 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isHalfTy())
1570 return ConstantFP::get(F->getContext(),
1571 APFloat((float)std::pow((float)Op1V,
1572 (int)Op2C->getZExtValue())));
1573 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1574 return ConstantFP::get(F->getContext(),
1575 APFloat((float)std::pow((float)Op1V,
1576 (int)Op2C->getZExtValue())));
1577 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1578 return ConstantFP::get(F->getContext(),
1579 APFloat((double)std::pow((double)Op1V,
1580 (int)Op2C->getZExtValue())));
1585 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1586 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1587 switch (F->getIntrinsicID()) {
1589 case Intrinsic::sadd_with_overflow:
1590 case Intrinsic::uadd_with_overflow:
1591 case Intrinsic::ssub_with_overflow:
1592 case Intrinsic::usub_with_overflow:
1593 case Intrinsic::smul_with_overflow:
1594 case Intrinsic::umul_with_overflow: {
1597 switch (F->getIntrinsicID()) {
1598 default: llvm_unreachable("Invalid case");
1599 case Intrinsic::sadd_with_overflow:
1600 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1602 case Intrinsic::uadd_with_overflow:
1603 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1605 case Intrinsic::ssub_with_overflow:
1606 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1608 case Intrinsic::usub_with_overflow:
1609 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1611 case Intrinsic::smul_with_overflow:
1612 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1614 case Intrinsic::umul_with_overflow:
1615 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1619 ConstantInt::get(F->getContext(), Res),
1620 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1622 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1624 case Intrinsic::cttz:
1625 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1626 return UndefValue::get(Ty);
1627 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1628 case Intrinsic::ctlz:
1629 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1630 return UndefValue::get(Ty);
1631 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());