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/SmallVector.h"
21 #include "llvm/ADT/StringMap.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/FEnv.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// DataLayout. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const DataLayout &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
55 // Handle a vector->integer cast.
56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
57 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
59 return ConstantExpr::getBitCast(C, DestTy);
61 unsigned NumSrcElts = CDV->getType()->getNumElements();
63 Type *SrcEltTy = CDV->getType()->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);
73 CDV = cast<ConstantDataVector>(C);
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
79 APInt Result(IT->getBitWidth(), 0);
80 for (unsigned i = 0; i != NumSrcElts; ++i) {
82 if (TD.isLittleEndian())
83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
85 Result |= CDV->getElementAsInteger(i);
88 return ConstantInt::get(IT, Result);
91 // The code below only handles casts to vectors currently.
92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
94 return ConstantExpr::getBitCast(C, DestTy);
96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
97 // vector so the code below can handle it uniformly.
98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
99 Constant *Ops = C; // don't take the address of C!
100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
103 // If this is a bitcast from constant vector -> vector, fold it.
104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
105 return ConstantExpr::getBitCast(C, DestTy);
107 // If the element types match, IR can fold it.
108 unsigned NumDstElt = DestVTy->getNumElements();
109 unsigned NumSrcElt = C->getType()->getVectorNumElements();
110 if (NumDstElt == NumSrcElt)
111 return ConstantExpr::getBitCast(C, DestTy);
113 Type *SrcEltTy = C->getType()->getVectorElementType();
114 Type *DstEltTy = DestVTy->getElementType();
116 // Otherwise, we're changing the number of elements in a vector, which
117 // requires endianness information to do the right thing. For example,
118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
119 // folds to (little endian):
120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
121 // and to (big endian):
122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
124 // First thing is first. We only want to think about integer here, so if
125 // we have something in FP form, recast it as integer.
126 if (DstEltTy->isFloatingPointTy()) {
127 // Fold to an vector of integers with same size as our FP type.
128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
131 // Recursively handle this integer conversion, if possible.
132 C = FoldBitCast(C, DestIVTy, TD);
134 // Finally, IR can handle this now that #elts line up.
135 return ConstantExpr::getBitCast(C, DestTy);
138 // Okay, we know the destination is integer, if the input is FP, convert
139 // it to integer first.
140 if (SrcEltTy->isFloatingPointTy()) {
141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
144 // Ask IR to do the conversion now that #elts line up.
145 C = ConstantExpr::getBitCast(C, SrcIVTy);
146 // If IR wasn't able to fold it, bail out.
147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
148 !isa<ConstantDataVector>(C))
152 // Now we know that the input and output vectors are both integer vectors
153 // of the same size, and that their #elements is not the same. Do the
154 // conversion here, which depends on whether the input or output has
156 bool isLittleEndian = TD.isLittleEndian();
158 SmallVector<Constant*, 32> Result;
159 if (NumDstElt < NumSrcElt) {
160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
161 Constant *Zero = Constant::getNullValue(DstEltTy);
162 unsigned Ratio = NumSrcElt/NumDstElt;
163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
165 for (unsigned i = 0; i != NumDstElt; ++i) {
166 // Build each element of the result.
167 Constant *Elt = Zero;
168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
169 for (unsigned j = 0; j != Ratio; ++j) {
170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
171 if (!Src) // Reject constantexpr elements.
172 return ConstantExpr::getBitCast(C, DestTy);
174 // Zero extend the element to the right size.
175 Src = ConstantExpr::getZExt(Src, Elt->getType());
177 // Shift it to the right place, depending on endianness.
178 Src = ConstantExpr::getShl(Src,
179 ConstantInt::get(Src->getType(), ShiftAmt));
180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
183 Elt = ConstantExpr::getOr(Elt, Src);
185 Result.push_back(Elt);
187 return ConstantVector::get(Result);
190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
191 unsigned Ratio = NumDstElt/NumSrcElt;
192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
194 // Loop over each source value, expanding into multiple results.
195 for (unsigned i = 0; i != NumSrcElt; ++i) {
196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
197 if (!Src) // Reject constantexpr elements.
198 return ConstantExpr::getBitCast(C, DestTy);
200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
201 for (unsigned j = 0; j != Ratio; ++j) {
202 // Shift the piece of the value into the right place, depending on
204 Constant *Elt = ConstantExpr::getLShr(Src,
205 ConstantInt::get(Src->getType(), ShiftAmt));
206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
208 // Truncate and remember this piece.
209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
213 return ConstantVector::get(Result);
217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
218 /// from a global, return the global and the constant. Because of
219 /// constantexprs, this function is recursive.
220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
221 APInt &Offset, const DataLayout &TD) {
222 // Trivial case, constant is the global.
223 if ((GV = dyn_cast<GlobalValue>(C))) {
224 Offset.clearAllBits();
228 // Otherwise, if this isn't a constant expr, bail out.
229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
230 if (!CE) return false;
232 // Look through ptr->int and ptr->ptr casts.
233 if (CE->getOpcode() == Instruction::PtrToInt ||
234 CE->getOpcode() == Instruction::BitCast)
235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
238 if (GEPOperator *GEP = dyn_cast<GEPOperator>(CE)) {
239 // If the base isn't a global+constant, we aren't either.
240 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
243 // Otherwise, add any offset that our operands provide.
244 return GEP->accumulateConstantOffset(TD, Offset);
250 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
251 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
252 /// pointer to copy results into and BytesLeft is the number of bytes left in
253 /// the CurPtr buffer. TD is the target data.
254 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
255 unsigned char *CurPtr, unsigned BytesLeft,
256 const DataLayout &TD) {
257 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
258 "Out of range access");
260 // If this element is zero or undefined, we can just return since *CurPtr is
262 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
265 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
266 if (CI->getBitWidth() > 64 ||
267 (CI->getBitWidth() & 7) != 0)
270 uint64_t Val = CI->getZExtValue();
271 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
273 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
275 if (!TD.isLittleEndian())
276 n = IntBytes - n - 1;
277 CurPtr[i] = (unsigned char)(Val >> (n * 8));
283 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
284 if (CFP->getType()->isDoubleTy()) {
285 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
286 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
288 if (CFP->getType()->isFloatTy()){
289 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
290 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
295 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
296 const StructLayout *SL = TD.getStructLayout(CS->getType());
297 unsigned Index = SL->getElementContainingOffset(ByteOffset);
298 uint64_t CurEltOffset = SL->getElementOffset(Index);
299 ByteOffset -= CurEltOffset;
302 // If the element access is to the element itself and not to tail padding,
303 // read the bytes from the element.
304 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
306 if (ByteOffset < EltSize &&
307 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
313 // Check to see if we read from the last struct element, if so we're done.
314 if (Index == CS->getType()->getNumElements())
317 // If we read all of the bytes we needed from this element we're done.
318 uint64_t NextEltOffset = SL->getElementOffset(Index);
320 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
323 // Move to the next element of the struct.
324 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
325 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
327 CurEltOffset = NextEltOffset;
332 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
333 isa<ConstantDataSequential>(C)) {
334 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
335 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
336 uint64_t Index = ByteOffset / EltSize;
337 uint64_t Offset = ByteOffset - Index * EltSize;
339 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
340 NumElts = AT->getNumElements();
342 NumElts = cast<VectorType>(C->getType())->getNumElements();
344 for (; Index != NumElts; ++Index) {
345 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
349 uint64_t BytesWritten = EltSize - Offset;
350 assert(BytesWritten <= EltSize && "Not indexing into this element?");
351 if (BytesWritten >= BytesLeft)
355 BytesLeft -= BytesWritten;
356 CurPtr += BytesWritten;
361 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
362 if (CE->getOpcode() == Instruction::IntToPtr &&
363 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
364 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
368 // Otherwise, unknown initializer type.
372 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
373 const DataLayout &TD) {
374 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
375 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
377 // If this isn't an integer load we can't fold it directly.
379 // If this is a float/double load, we can try folding it as an int32/64 load
380 // and then bitcast the result. This can be useful for union cases. Note
381 // that address spaces don't matter here since we're not going to result in
382 // an actual new load.
384 if (LoadTy->isFloatTy())
385 MapTy = Type::getInt32PtrTy(C->getContext());
386 else if (LoadTy->isDoubleTy())
387 MapTy = Type::getInt64PtrTy(C->getContext());
388 else if (LoadTy->isVectorTy()) {
389 MapTy = IntegerType::get(C->getContext(),
390 TD.getTypeAllocSizeInBits(LoadTy));
391 MapTy = PointerType::getUnqual(MapTy);
395 C = FoldBitCast(C, MapTy, TD);
396 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
397 return FoldBitCast(Res, LoadTy, TD);
401 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
402 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
405 APInt Offset(TD.getPointerSizeInBits(), 0);
406 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
409 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
410 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
411 !GV->getInitializer()->getType()->isSized())
414 // If we're loading off the beginning of the global, some bytes may be valid,
415 // but we don't try to handle this.
416 if (Offset.isNegative()) return 0;
418 // If we're not accessing anything in this constant, the result is undefined.
419 if (Offset.getZExtValue() >=
420 TD.getTypeAllocSize(GV->getInitializer()->getType()))
421 return UndefValue::get(IntType);
423 unsigned char RawBytes[32] = {0};
424 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
428 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
429 if (TD.isLittleEndian()) {
430 ResultVal = RawBytes[BytesLoaded - 1];
431 for (unsigned i = 1; i != BytesLoaded; ++i) {
433 ResultVal |= RawBytes[BytesLoaded-1-i];
436 ResultVal = RawBytes[0];
437 for (unsigned i = 1; i != BytesLoaded; ++i) {
439 ResultVal |= RawBytes[i];
443 return ConstantInt::get(IntType->getContext(), ResultVal);
446 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
447 /// produce if it is constant and determinable. If this is not determinable,
449 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
450 const DataLayout *TD) {
451 // First, try the easy cases:
452 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
453 if (GV->isConstant() && GV->hasDefinitiveInitializer())
454 return GV->getInitializer();
456 // If the loaded value isn't a constant expr, we can't handle it.
457 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
460 if (CE->getOpcode() == Instruction::GetElementPtr) {
461 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
462 if (GV->isConstant() && GV->hasDefinitiveInitializer())
464 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
468 // Instead of loading constant c string, use corresponding integer value
469 // directly if string length is small enough.
471 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
472 unsigned StrLen = Str.size();
473 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
474 unsigned NumBits = Ty->getPrimitiveSizeInBits();
475 // Replace load with immediate integer if the result is an integer or fp
477 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
478 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
479 APInt StrVal(NumBits, 0);
480 APInt SingleChar(NumBits, 0);
481 if (TD->isLittleEndian()) {
482 for (signed i = StrLen-1; i >= 0; i--) {
483 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
484 StrVal = (StrVal << 8) | SingleChar;
487 for (unsigned i = 0; i < StrLen; i++) {
488 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
489 StrVal = (StrVal << 8) | SingleChar;
491 // Append NULL at the end.
493 StrVal = (StrVal << 8) | SingleChar;
496 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
497 if (Ty->isFloatingPointTy())
498 Res = ConstantExpr::getBitCast(Res, Ty);
503 // If this load comes from anywhere in a constant global, and if the global
504 // is all undef or zero, we know what it loads.
505 if (GlobalVariable *GV =
506 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
507 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
508 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
509 if (GV->getInitializer()->isNullValue())
510 return Constant::getNullValue(ResTy);
511 if (isa<UndefValue>(GV->getInitializer()))
512 return UndefValue::get(ResTy);
516 // Try hard to fold loads from bitcasted strange and non-type-safe things.
518 return FoldReinterpretLoadFromConstPtr(CE, *TD);
522 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
523 if (LI->isVolatile()) return 0;
525 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
526 return ConstantFoldLoadFromConstPtr(C, TD);
531 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
532 /// Attempt to symbolically evaluate the result of a binary operator merging
533 /// these together. If target data info is available, it is provided as TD,
534 /// otherwise TD is null.
535 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
536 Constant *Op1, const DataLayout *TD){
539 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
540 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
544 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
545 // constant. This happens frequently when iterating over a global array.
546 if (Opc == Instruction::Sub && TD) {
547 GlobalValue *GV1, *GV2;
548 APInt Offs1(TD->getPointerSizeInBits(), 0),
549 Offs2(TD->getPointerSizeInBits(), 0);
551 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
552 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
554 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
555 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
562 /// CastGEPIndices - If array indices are not pointer-sized integers,
563 /// explicitly cast them so that they aren't implicitly casted by the
565 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
566 Type *ResultTy, const DataLayout *TD,
567 const TargetLibraryInfo *TLI) {
569 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
572 SmallVector<Constant*, 32> NewIdxs;
573 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
575 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
576 Ops.slice(1, i-1)))) &&
577 Ops[i]->getType() != IntPtrTy) {
579 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
585 NewIdxs.push_back(Ops[i]);
590 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
591 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
592 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
597 /// Strip the pointer casts, but preserve the address space information.
598 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
599 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
600 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
601 Ptr = cast<Constant>(Ptr->stripPointerCasts());
602 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
604 // Preserve the address space number of the pointer.
605 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
606 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
607 OldPtrTy->getAddressSpace());
608 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy);
613 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
614 /// constant expression, do so.
615 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
616 Type *ResultTy, const DataLayout *TD,
617 const TargetLibraryInfo *TLI) {
618 Constant *Ptr = Ops[0];
619 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
620 !Ptr->getType()->isPointerTy())
623 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
625 // If this is a constant expr gep that is effectively computing an
626 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
627 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
628 if (!isa<ConstantInt>(Ops[i])) {
630 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
631 // "inttoptr (sub (ptrtoint Ptr), V)"
632 if (Ops.size() == 2 &&
633 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
634 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
635 assert((CE == 0 || CE->getType() == IntPtrTy) &&
636 "CastGEPIndices didn't canonicalize index types!");
637 if (CE && CE->getOpcode() == Instruction::Sub &&
638 CE->getOperand(0)->isNullValue()) {
639 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
640 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
641 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
642 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
643 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
650 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
652 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
653 makeArrayRef((Value *const*)
656 Ptr = StripPtrCastKeepAS(Ptr);
658 // If this is a GEP of a GEP, fold it all into a single GEP.
659 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
660 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
662 // Do not try the incorporate the sub-GEP if some index is not a number.
663 bool AllConstantInt = true;
664 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
665 if (!isa<ConstantInt>(NestedOps[i])) {
666 AllConstantInt = false;
672 Ptr = cast<Constant>(GEP->getOperand(0));
673 Offset += APInt(BitWidth,
674 TD->getIndexedOffset(Ptr->getType(), NestedOps));
675 Ptr = StripPtrCastKeepAS(Ptr);
678 // If the base value for this address is a literal integer value, fold the
679 // getelementptr to the resulting integer value casted to the pointer type.
680 APInt BasePtr(BitWidth, 0);
681 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
682 if (CE->getOpcode() == Instruction::IntToPtr)
683 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
684 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
685 if (Ptr->isNullValue() || BasePtr != 0) {
686 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
687 return ConstantExpr::getIntToPtr(C, ResultTy);
690 // Otherwise form a regular getelementptr. Recompute the indices so that
691 // we eliminate over-indexing of the notional static type array bounds.
692 // This makes it easy to determine if the getelementptr is "inbounds".
693 // Also, this helps GlobalOpt do SROA on GlobalVariables.
694 Type *Ty = Ptr->getType();
695 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
696 SmallVector<Constant*, 32> NewIdxs;
698 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
699 if (ATy->isPointerTy()) {
700 // The only pointer indexing we'll do is on the first index of the GEP.
701 if (!NewIdxs.empty())
704 // Only handle pointers to sized types, not pointers to functions.
705 if (!ATy->getElementType()->isSized())
709 // Determine which element of the array the offset points into.
710 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
711 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
713 // The element size is 0. This may be [0 x Ty]*, so just use a zero
714 // index for this level and proceed to the next level to see if it can
715 // accommodate the offset.
716 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
718 // The element size is non-zero divide the offset by the element
719 // size (rounding down), to compute the index at this level.
720 APInt NewIdx = Offset.udiv(ElemSize);
721 Offset -= NewIdx * ElemSize;
722 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
724 Ty = ATy->getElementType();
725 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
726 // If we end up with an offset that isn't valid for this struct type, we
727 // can't re-form this GEP in a regular form, so bail out. The pointer
728 // operand likely went through casts that are necessary to make the GEP
730 const StructLayout &SL = *TD->getStructLayout(STy);
731 if (Offset.uge(SL.getSizeInBytes()))
734 // Determine which field of the struct the offset points into. The
735 // getZExtValue is fine as we've already ensured that the offset is
736 // within the range representable by the StructLayout API.
737 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
738 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
740 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
741 Ty = STy->getTypeAtIndex(ElIdx);
743 // We've reached some non-indexable type.
746 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
748 // If we haven't used up the entire offset by descending the static
749 // type, then the offset is pointing into the middle of an indivisible
750 // member, so we can't simplify it.
756 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
757 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
758 "Computed GetElementPtr has unexpected type!");
760 // If we ended up indexing a member with a type that doesn't match
761 // the type of what the original indices indexed, add a cast.
762 if (Ty != cast<PointerType>(ResultTy)->getElementType())
763 C = FoldBitCast(C, ResultTy, *TD);
770 //===----------------------------------------------------------------------===//
771 // Constant Folding public APIs
772 //===----------------------------------------------------------------------===//
774 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
775 /// If successful, the constant result is returned, if not, null is returned.
776 /// Note that this fails if not all of the operands are constant. Otherwise,
777 /// this function can only fail when attempting to fold instructions like loads
778 /// and stores, which have no constant expression form.
779 Constant *llvm::ConstantFoldInstruction(Instruction *I,
780 const DataLayout *TD,
781 const TargetLibraryInfo *TLI) {
782 // Handle PHI nodes quickly here...
783 if (PHINode *PN = dyn_cast<PHINode>(I)) {
784 Constant *CommonValue = 0;
786 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
787 Value *Incoming = PN->getIncomingValue(i);
788 // If the incoming value is undef then skip it. Note that while we could
789 // skip the value if it is equal to the phi node itself we choose not to
790 // because that would break the rule that constant folding only applies if
791 // all operands are constants.
792 if (isa<UndefValue>(Incoming))
794 // If the incoming value is not a constant, then give up.
795 Constant *C = dyn_cast<Constant>(Incoming);
798 // Fold the PHI's operands.
799 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
800 C = ConstantFoldConstantExpression(NewC, TD, TLI);
801 // If the incoming value is a different constant to
802 // the one we saw previously, then give up.
803 if (CommonValue && C != CommonValue)
809 // If we reach here, all incoming values are the same constant or undef.
810 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
813 // Scan the operand list, checking to see if they are all constants, if so,
814 // hand off to ConstantFoldInstOperands.
815 SmallVector<Constant*, 8> Ops;
816 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
817 Constant *Op = dyn_cast<Constant>(*i);
819 return 0; // All operands not constant!
821 // Fold the Instruction's operands.
822 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
823 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
828 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
829 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
832 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
833 return ConstantFoldLoadInst(LI, TD);
835 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
836 return ConstantExpr::getInsertValue(
837 cast<Constant>(IVI->getAggregateOperand()),
838 cast<Constant>(IVI->getInsertedValueOperand()),
841 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
842 return ConstantExpr::getExtractValue(
843 cast<Constant>(EVI->getAggregateOperand()),
846 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
849 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
850 /// using the specified DataLayout. If successful, the constant result is
851 /// result is returned, if not, null is returned.
852 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
853 const DataLayout *TD,
854 const TargetLibraryInfo *TLI) {
855 SmallVector<Constant*, 8> Ops;
856 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
858 Constant *NewC = cast<Constant>(*i);
859 // Recursively fold the ConstantExpr's operands.
860 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
861 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
866 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
868 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
871 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
872 /// specified opcode and operands. If successful, the constant result is
873 /// returned, if not, null is returned. Note that this function can fail when
874 /// attempting to fold instructions like loads and stores, which have no
875 /// constant expression form.
877 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
878 /// information, due to only being passed an opcode and operands. Constant
879 /// folding using this function strips this information.
881 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
882 ArrayRef<Constant *> Ops,
883 const DataLayout *TD,
884 const TargetLibraryInfo *TLI) {
885 // Handle easy binops first.
886 if (Instruction::isBinaryOp(Opcode)) {
887 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
888 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
891 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
896 case Instruction::ICmp:
897 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
898 case Instruction::Call:
899 if (Function *F = dyn_cast<Function>(Ops.back()))
900 if (canConstantFoldCallTo(F))
901 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
903 case Instruction::PtrToInt:
904 // If the input is a inttoptr, eliminate the pair. This requires knowing
905 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
906 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
907 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
908 Constant *Input = CE->getOperand(0);
909 unsigned InWidth = Input->getType()->getScalarSizeInBits();
910 if (TD->getPointerSizeInBits() < InWidth) {
912 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
913 TD->getPointerSizeInBits()));
914 Input = ConstantExpr::getAnd(Input, Mask);
916 // Do a zext or trunc to get to the dest size.
917 return ConstantExpr::getIntegerCast(Input, DestTy, false);
920 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
921 case Instruction::IntToPtr:
922 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
923 // the int size is >= the ptr size. This requires knowing the width of a
924 // pointer, so it can't be done in ConstantExpr::getCast.
925 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
927 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
928 CE->getOpcode() == Instruction::PtrToInt)
929 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
931 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
932 case Instruction::Trunc:
933 case Instruction::ZExt:
934 case Instruction::SExt:
935 case Instruction::FPTrunc:
936 case Instruction::FPExt:
937 case Instruction::UIToFP:
938 case Instruction::SIToFP:
939 case Instruction::FPToUI:
940 case Instruction::FPToSI:
941 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
942 case Instruction::BitCast:
944 return FoldBitCast(Ops[0], DestTy, *TD);
945 return ConstantExpr::getBitCast(Ops[0], DestTy);
946 case Instruction::Select:
947 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
948 case Instruction::ExtractElement:
949 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
950 case Instruction::InsertElement:
951 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
952 case Instruction::ShuffleVector:
953 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
954 case Instruction::GetElementPtr:
955 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
957 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
960 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
964 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
965 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
966 /// returns a constant expression of the specified operands.
968 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
969 Constant *Ops0, Constant *Ops1,
970 const DataLayout *TD,
971 const TargetLibraryInfo *TLI) {
972 // fold: icmp (inttoptr x), null -> icmp x, 0
973 // fold: icmp (ptrtoint x), 0 -> icmp x, null
974 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
975 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
977 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
978 // around to know if bit truncation is happening.
979 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
980 if (TD && Ops1->isNullValue()) {
981 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
982 if (CE0->getOpcode() == Instruction::IntToPtr) {
983 // Convert the integer value to the right size to ensure we get the
984 // proper extension or truncation.
985 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
987 Constant *Null = Constant::getNullValue(C->getType());
988 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
991 // Only do this transformation if the int is intptrty in size, otherwise
992 // there is a truncation or extension that we aren't modeling.
993 if (CE0->getOpcode() == Instruction::PtrToInt &&
994 CE0->getType() == IntPtrTy) {
995 Constant *C = CE0->getOperand(0);
996 Constant *Null = Constant::getNullValue(C->getType());
997 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1001 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1002 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1003 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1005 if (CE0->getOpcode() == Instruction::IntToPtr) {
1006 // Convert the integer value to the right size to ensure we get the
1007 // proper extension or truncation.
1008 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1010 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1012 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1015 // Only do this transformation if the int is intptrty in size, otherwise
1016 // there is a truncation or extension that we aren't modeling.
1017 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1018 CE0->getType() == IntPtrTy &&
1019 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1020 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1021 CE1->getOperand(0), TD, TLI);
1025 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1026 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1027 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1028 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1030 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1033 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1036 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1037 Constant *Ops[] = { LHS, RHS };
1038 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1042 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1046 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1047 /// getelementptr constantexpr, return the constant value being addressed by the
1048 /// constant expression, or null if something is funny and we can't decide.
1049 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1051 if (!CE->getOperand(1)->isNullValue())
1052 return 0; // Do not allow stepping over the value!
1054 // Loop over all of the operands, tracking down which value we are
1056 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1057 C = C->getAggregateElement(CE->getOperand(i));
1058 if (C == 0) return 0;
1063 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1064 /// indices (with an *implied* zero pointer index that is not in the list),
1065 /// return the constant value being addressed by a virtual load, or null if
1066 /// something is funny and we can't decide.
1067 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1068 ArrayRef<Constant*> Indices) {
1069 // Loop over all of the operands, tracking down which value we are
1071 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1072 C = C->getAggregateElement(Indices[i]);
1073 if (C == 0) return 0;
1079 //===----------------------------------------------------------------------===//
1080 // Constant Folding for Calls
1083 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1084 /// the specified function.
1086 llvm::canConstantFoldCallTo(const Function *F) {
1087 switch (F->getIntrinsicID()) {
1088 case Intrinsic::sqrt:
1089 case Intrinsic::pow:
1090 case Intrinsic::powi:
1091 case Intrinsic::bswap:
1092 case Intrinsic::ctpop:
1093 case Intrinsic::ctlz:
1094 case Intrinsic::cttz:
1095 case Intrinsic::sadd_with_overflow:
1096 case Intrinsic::uadd_with_overflow:
1097 case Intrinsic::ssub_with_overflow:
1098 case Intrinsic::usub_with_overflow:
1099 case Intrinsic::smul_with_overflow:
1100 case Intrinsic::umul_with_overflow:
1101 case Intrinsic::convert_from_fp16:
1102 case Intrinsic::convert_to_fp16:
1103 case Intrinsic::x86_sse_cvtss2si:
1104 case Intrinsic::x86_sse_cvtss2si64:
1105 case Intrinsic::x86_sse_cvttss2si:
1106 case Intrinsic::x86_sse_cvttss2si64:
1107 case Intrinsic::x86_sse2_cvtsd2si:
1108 case Intrinsic::x86_sse2_cvtsd2si64:
1109 case Intrinsic::x86_sse2_cvttsd2si:
1110 case Intrinsic::x86_sse2_cvttsd2si64:
1117 if (!F->hasName()) return false;
1118 StringRef Name = F->getName();
1120 // In these cases, the check of the length is required. We don't want to
1121 // return true for a name like "cos\0blah" which strcmp would return equal to
1122 // "cos", but has length 8.
1124 default: return false;
1126 return Name == "acos" || Name == "asin" ||
1127 Name == "atan" || Name == "atan2";
1129 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1131 return Name == "exp" || Name == "exp2";
1133 return Name == "fabs" || Name == "fmod" || Name == "floor";
1135 return Name == "log" || Name == "log10";
1137 return Name == "pow";
1139 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1140 Name == "sinf" || Name == "sqrtf";
1142 return Name == "tan" || Name == "tanh";
1146 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1148 sys::llvm_fenv_clearexcept();
1150 if (sys::llvm_fenv_testexcept()) {
1151 sys::llvm_fenv_clearexcept();
1155 if (Ty->isFloatTy())
1156 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1157 if (Ty->isDoubleTy())
1158 return ConstantFP::get(Ty->getContext(), APFloat(V));
1159 llvm_unreachable("Can only constant fold float/double");
1162 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1163 double V, double W, Type *Ty) {
1164 sys::llvm_fenv_clearexcept();
1166 if (sys::llvm_fenv_testexcept()) {
1167 sys::llvm_fenv_clearexcept();
1171 if (Ty->isFloatTy())
1172 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1173 if (Ty->isDoubleTy())
1174 return ConstantFP::get(Ty->getContext(), APFloat(V));
1175 llvm_unreachable("Can only constant fold float/double");
1178 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1179 /// conversion of a constant floating point. If roundTowardZero is false, the
1180 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1181 /// the behavior of the non-truncating SSE instructions in the default rounding
1182 /// mode. The desired integer type Ty is used to select how many bits are
1183 /// available for the result. Returns null if the conversion cannot be
1184 /// performed, otherwise returns the Constant value resulting from the
1186 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1187 bool roundTowardZero, Type *Ty) {
1188 // All of these conversion intrinsics form an integer of at most 64bits.
1189 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1190 assert(ResultWidth <= 64 &&
1191 "Can only constant fold conversions to 64 and 32 bit ints");
1194 bool isExact = false;
1195 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1196 : APFloat::rmNearestTiesToEven;
1197 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1198 /*isSigned=*/true, mode,
1200 if (status != APFloat::opOK && status != APFloat::opInexact)
1202 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1205 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1206 /// with the specified arguments, returning null if unsuccessful.
1208 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1209 const TargetLibraryInfo *TLI) {
1210 if (!F->hasName()) return 0;
1211 StringRef Name = F->getName();
1213 Type *Ty = F->getReturnType();
1214 if (Operands.size() == 1) {
1215 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1216 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1217 APFloat Val(Op->getValueAPF());
1220 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1222 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1227 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1230 /// We only fold functions with finite arguments. Folding NaN and inf is
1231 /// likely to be aborted with an exception anyway, and some host libms
1232 /// have known errors raising exceptions.
1233 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1236 /// Currently APFloat versions of these functions do not exist, so we use
1237 /// the host native double versions. Float versions are not called
1238 /// directly but for all these it is true (float)(f((double)arg)) ==
1239 /// f(arg). Long double not supported yet.
1240 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1241 Op->getValueAPF().convertToDouble();
1244 if (Name == "acos" && TLI->has(LibFunc::acos))
1245 return ConstantFoldFP(acos, V, Ty);
1246 else if (Name == "asin" && TLI->has(LibFunc::asin))
1247 return ConstantFoldFP(asin, V, Ty);
1248 else if (Name == "atan" && TLI->has(LibFunc::atan))
1249 return ConstantFoldFP(atan, V, Ty);
1252 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1253 return ConstantFoldFP(ceil, V, Ty);
1254 else if (Name == "cos" && TLI->has(LibFunc::cos))
1255 return ConstantFoldFP(cos, V, Ty);
1256 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1257 return ConstantFoldFP(cosh, V, Ty);
1258 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1259 return ConstantFoldFP(cos, V, Ty);
1262 if (Name == "exp" && TLI->has(LibFunc::exp))
1263 return ConstantFoldFP(exp, V, Ty);
1265 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1266 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1268 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1272 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1273 return ConstantFoldFP(fabs, V, Ty);
1274 else if (Name == "floor" && TLI->has(LibFunc::floor))
1275 return ConstantFoldFP(floor, V, Ty);
1278 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1279 return ConstantFoldFP(log, V, Ty);
1280 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1281 return ConstantFoldFP(log10, V, Ty);
1282 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1283 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1285 return ConstantFoldFP(sqrt, V, Ty);
1287 return Constant::getNullValue(Ty);
1291 if (Name == "sin" && TLI->has(LibFunc::sin))
1292 return ConstantFoldFP(sin, V, Ty);
1293 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1294 return ConstantFoldFP(sinh, V, Ty);
1295 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1296 return ConstantFoldFP(sqrt, V, Ty);
1297 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1298 return ConstantFoldFP(sqrt, V, Ty);
1299 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1300 return ConstantFoldFP(sin, V, Ty);
1303 if (Name == "tan" && TLI->has(LibFunc::tan))
1304 return ConstantFoldFP(tan, V, Ty);
1305 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1306 return ConstantFoldFP(tanh, V, Ty);
1314 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1315 switch (F->getIntrinsicID()) {
1316 case Intrinsic::bswap:
1317 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1318 case Intrinsic::ctpop:
1319 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1320 case Intrinsic::convert_from_fp16: {
1321 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1324 APFloat::opStatus status =
1325 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1327 // Conversion is always precise.
1329 assert(status == APFloat::opOK && !lost &&
1330 "Precision lost during fp16 constfolding");
1332 return ConstantFP::get(F->getContext(), Val);
1339 // Support ConstantVector in case we have an Undef in the top.
1340 if (isa<ConstantVector>(Operands[0]) ||
1341 isa<ConstantDataVector>(Operands[0])) {
1342 Constant *Op = cast<Constant>(Operands[0]);
1343 switch (F->getIntrinsicID()) {
1345 case Intrinsic::x86_sse_cvtss2si:
1346 case Intrinsic::x86_sse_cvtss2si64:
1347 case Intrinsic::x86_sse2_cvtsd2si:
1348 case Intrinsic::x86_sse2_cvtsd2si64:
1349 if (ConstantFP *FPOp =
1350 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1351 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1352 /*roundTowardZero=*/false, Ty);
1353 case Intrinsic::x86_sse_cvttss2si:
1354 case Intrinsic::x86_sse_cvttss2si64:
1355 case Intrinsic::x86_sse2_cvttsd2si:
1356 case Intrinsic::x86_sse2_cvttsd2si64:
1357 if (ConstantFP *FPOp =
1358 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1359 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1360 /*roundTowardZero=*/true, Ty);
1364 if (isa<UndefValue>(Operands[0])) {
1365 if (F->getIntrinsicID() == Intrinsic::bswap)
1373 if (Operands.size() == 2) {
1374 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1375 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1377 double Op1V = Ty->isFloatTy() ?
1378 (double)Op1->getValueAPF().convertToFloat() :
1379 Op1->getValueAPF().convertToDouble();
1380 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1381 if (Op2->getType() != Op1->getType())
1384 double Op2V = Ty->isFloatTy() ?
1385 (double)Op2->getValueAPF().convertToFloat():
1386 Op2->getValueAPF().convertToDouble();
1388 if (F->getIntrinsicID() == Intrinsic::pow) {
1389 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1393 if (Name == "pow" && TLI->has(LibFunc::pow))
1394 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1395 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1396 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1397 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1398 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1399 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1400 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1401 return ConstantFP::get(F->getContext(),
1402 APFloat((float)std::pow((float)Op1V,
1403 (int)Op2C->getZExtValue())));
1404 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1405 return ConstantFP::get(F->getContext(),
1406 APFloat((double)std::pow((double)Op1V,
1407 (int)Op2C->getZExtValue())));
1412 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1413 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1414 switch (F->getIntrinsicID()) {
1416 case Intrinsic::sadd_with_overflow:
1417 case Intrinsic::uadd_with_overflow:
1418 case Intrinsic::ssub_with_overflow:
1419 case Intrinsic::usub_with_overflow:
1420 case Intrinsic::smul_with_overflow:
1421 case Intrinsic::umul_with_overflow: {
1424 switch (F->getIntrinsicID()) {
1425 default: llvm_unreachable("Invalid case");
1426 case Intrinsic::sadd_with_overflow:
1427 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1429 case Intrinsic::uadd_with_overflow:
1430 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1432 case Intrinsic::ssub_with_overflow:
1433 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1435 case Intrinsic::usub_with_overflow:
1436 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1438 case Intrinsic::smul_with_overflow:
1439 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1441 case Intrinsic::umul_with_overflow:
1442 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1446 ConstantInt::get(F->getContext(), Res),
1447 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1449 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1451 case Intrinsic::cttz:
1452 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1453 return UndefValue::get(Ty);
1454 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1455 case Intrinsic::ctlz:
1456 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1457 return UndefValue::get(Ty);
1458 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());