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 VMCore ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // TargetData information. These functions cannot go in VMCore due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/LLVMContext.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/StringMap.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
38 //===----------------------------------------------------------------------===//
39 // Constant Folding internal helper functions
40 //===----------------------------------------------------------------------===//
42 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
43 /// TargetData. This always returns a non-null constant, but it may be a
44 /// ConstantExpr if unfoldable.
45 static Constant *FoldBitCast(Constant *C, const Type *DestTy,
46 const TargetData &TD) {
48 // This only handles casts to vectors currently.
49 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
51 return ConstantExpr::getBitCast(C, DestTy);
53 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
54 // vector so the code below can handle it uniformly.
55 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
56 Constant *Ops = C; // don't take the address of C!
57 return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD);
60 // If this is a bitcast from constant vector -> vector, fold it.
61 ConstantVector *CV = dyn_cast<ConstantVector>(C);
63 return ConstantExpr::getBitCast(C, DestTy);
65 // If the element types match, VMCore can fold it.
66 unsigned NumDstElt = DestVTy->getNumElements();
67 unsigned NumSrcElt = CV->getNumOperands();
68 if (NumDstElt == NumSrcElt)
69 return ConstantExpr::getBitCast(C, DestTy);
71 const Type *SrcEltTy = CV->getType()->getElementType();
72 const Type *DstEltTy = DestVTy->getElementType();
74 // Otherwise, we're changing the number of elements in a vector, which
75 // requires endianness information to do the right thing. For example,
76 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
77 // folds to (little endian):
78 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
79 // and to (big endian):
80 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
82 // First thing is first. We only want to think about integer here, so if
83 // we have something in FP form, recast it as integer.
84 if (DstEltTy->isFloatingPoint()) {
85 // Fold to an vector of integers with same size as our FP type.
86 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
87 const Type *DestIVTy =
88 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
89 // Recursively handle this integer conversion, if possible.
90 C = FoldBitCast(C, DestIVTy, TD);
91 if (!C) return ConstantExpr::getBitCast(C, DestTy);
93 // Finally, VMCore can handle this now that #elts line up.
94 return ConstantExpr::getBitCast(C, DestTy);
97 // Okay, we know the destination is integer, if the input is FP, convert
98 // it to integer first.
99 if (SrcEltTy->isFloatingPoint()) {
100 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
101 const Type *SrcIVTy =
102 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
103 // Ask VMCore to do the conversion now that #elts line up.
104 C = ConstantExpr::getBitCast(C, SrcIVTy);
105 CV = dyn_cast<ConstantVector>(C);
106 if (!CV) // If VMCore wasn't able to fold it, bail out.
110 // Now we know that the input and output vectors are both integer vectors
111 // of the same size, and that their #elements is not the same. Do the
112 // conversion here, which depends on whether the input or output has
114 bool isLittleEndian = TD.isLittleEndian();
116 SmallVector<Constant*, 32> Result;
117 if (NumDstElt < NumSrcElt) {
118 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
119 Constant *Zero = Constant::getNullValue(DstEltTy);
120 unsigned Ratio = NumSrcElt/NumDstElt;
121 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
123 for (unsigned i = 0; i != NumDstElt; ++i) {
124 // Build each element of the result.
125 Constant *Elt = Zero;
126 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
127 for (unsigned j = 0; j != Ratio; ++j) {
128 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
129 if (!Src) // Reject constantexpr elements.
130 return ConstantExpr::getBitCast(C, DestTy);
132 // Zero extend the element to the right size.
133 Src = ConstantExpr::getZExt(Src, Elt->getType());
135 // Shift it to the right place, depending on endianness.
136 Src = ConstantExpr::getShl(Src,
137 ConstantInt::get(Src->getType(), ShiftAmt));
138 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
141 Elt = ConstantExpr::getOr(Elt, Src);
143 Result.push_back(Elt);
146 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
147 unsigned Ratio = NumDstElt/NumSrcElt;
148 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
150 // Loop over each source value, expanding into multiple results.
151 for (unsigned i = 0; i != NumSrcElt; ++i) {
152 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
153 if (!Src) // Reject constantexpr elements.
154 return ConstantExpr::getBitCast(C, DestTy);
156 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
157 for (unsigned j = 0; j != Ratio; ++j) {
158 // Shift the piece of the value into the right place, depending on
160 Constant *Elt = ConstantExpr::getLShr(Src,
161 ConstantInt::get(Src->getType(), ShiftAmt));
162 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
164 // Truncate and remember this piece.
165 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
170 return ConstantVector::get(Result.data(), Result.size());
174 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
175 /// from a global, return the global and the constant. Because of
176 /// constantexprs, this function is recursive.
177 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
178 int64_t &Offset, const TargetData &TD) {
179 // Trivial case, constant is the global.
180 if ((GV = dyn_cast<GlobalValue>(C))) {
185 // Otherwise, if this isn't a constant expr, bail out.
186 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
187 if (!CE) return false;
189 // Look through ptr->int and ptr->ptr casts.
190 if (CE->getOpcode() == Instruction::PtrToInt ||
191 CE->getOpcode() == Instruction::BitCast)
192 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
194 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
195 if (CE->getOpcode() == Instruction::GetElementPtr) {
196 // Cannot compute this if the element type of the pointer is missing size
198 if (!cast<PointerType>(CE->getOperand(0)->getType())
199 ->getElementType()->isSized())
202 // If the base isn't a global+constant, we aren't either.
203 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
206 // Otherwise, add any offset that our operands provide.
207 gep_type_iterator GTI = gep_type_begin(CE);
208 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
209 i != e; ++i, ++GTI) {
210 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
211 if (!CI) return false; // Index isn't a simple constant?
212 if (CI->getZExtValue() == 0) continue; // Not adding anything.
214 if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
216 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
218 const SequentialType *SQT = cast<SequentialType>(*GTI);
219 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
228 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
229 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
230 /// pointer to copy results into and BytesLeft is the number of bytes left in
231 /// the CurPtr buffer. TD is the target data.
232 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
233 unsigned char *CurPtr, unsigned BytesLeft,
234 const TargetData &TD) {
235 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
236 "Out of range access");
238 // If this element is zero or undefined, we can just return since *CurPtr is
240 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
243 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
244 if (CI->getBitWidth() > 64 ||
245 (CI->getBitWidth() & 7) != 0)
248 uint64_t Val = CI->getZExtValue();
249 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
251 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
252 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
258 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
259 if (CFP->getType()->isDoubleTy()) {
260 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
261 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
263 if (CFP->getType()->isFloatTy()){
264 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
265 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
270 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
271 const StructLayout *SL = TD.getStructLayout(CS->getType());
272 unsigned Index = SL->getElementContainingOffset(ByteOffset);
273 uint64_t CurEltOffset = SL->getElementOffset(Index);
274 ByteOffset -= CurEltOffset;
277 // If the element access is to the element itself and not to tail padding,
278 // read the bytes from the element.
279 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
281 if (ByteOffset < EltSize &&
282 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
288 // Check to see if we read from the last struct element, if so we're done.
289 if (Index == CS->getType()->getNumElements())
292 // If we read all of the bytes we needed from this element we're done.
293 uint64_t NextEltOffset = SL->getElementOffset(Index);
295 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
298 // Move to the next element of the struct.
299 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
300 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
302 CurEltOffset = NextEltOffset;
307 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
308 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
309 uint64_t Index = ByteOffset / EltSize;
310 uint64_t Offset = ByteOffset - Index * EltSize;
311 for (; Index != CA->getType()->getNumElements(); ++Index) {
312 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
315 if (EltSize >= BytesLeft)
319 BytesLeft -= EltSize;
325 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
326 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
327 uint64_t Index = ByteOffset / EltSize;
328 uint64_t Offset = ByteOffset - Index * EltSize;
329 for (; Index != CV->getType()->getNumElements(); ++Index) {
330 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
333 if (EltSize >= BytesLeft)
337 BytesLeft -= EltSize;
343 // Otherwise, unknown initializer type.
347 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
348 const TargetData &TD) {
349 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
350 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
352 // If this isn't an integer load we can't fold it directly.
354 // If this is a float/double load, we can try folding it as an int32/64 load
355 // and then bitcast the result. This can be useful for union cases. Note
356 // that address spaces don't matter here since we're not going to result in
357 // an actual new load.
359 if (LoadTy->isFloatTy())
360 MapTy = Type::getInt32PtrTy(C->getContext());
361 else if (LoadTy->isDoubleTy())
362 MapTy = Type::getInt64PtrTy(C->getContext());
363 else if (isa<VectorType>(LoadTy)) {
364 MapTy = IntegerType::get(C->getContext(),
365 TD.getTypeAllocSizeInBits(LoadTy));
366 MapTy = PointerType::getUnqual(MapTy);
370 C = FoldBitCast(C, MapTy, TD);
371 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
372 return FoldBitCast(Res, LoadTy, TD);
376 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
377 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
381 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
384 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
385 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
386 !GV->getInitializer()->getType()->isSized())
389 // If we're loading off the beginning of the global, some bytes may be valid,
390 // but we don't try to handle this.
391 if (Offset < 0) return 0;
393 // If we're not accessing anything in this constant, the result is undefined.
394 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
395 return UndefValue::get(IntType);
397 unsigned char RawBytes[32] = {0};
398 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
402 APInt ResultVal(IntType->getBitWidth(), 0);
403 for (unsigned i = 0; i != BytesLoaded; ++i) {
405 ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]);
408 return ConstantInt::get(IntType->getContext(), ResultVal);
411 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
412 /// produce if it is constant and determinable. If this is not determinable,
414 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
415 const TargetData *TD) {
416 // First, try the easy cases:
417 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
418 if (GV->isConstant() && GV->hasDefinitiveInitializer())
419 return GV->getInitializer();
421 // If the loaded value isn't a constant expr, we can't handle it.
422 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
425 if (CE->getOpcode() == Instruction::GetElementPtr) {
426 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
427 if (GV->isConstant() && GV->hasDefinitiveInitializer())
429 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
433 // Instead of loading constant c string, use corresponding integer value
434 // directly if string length is small enough.
436 if (TD && GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
437 unsigned StrLen = Str.length();
438 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
439 unsigned NumBits = Ty->getPrimitiveSizeInBits();
440 // Replace LI with immediate integer store.
441 if ((NumBits >> 3) == StrLen + 1) {
442 APInt StrVal(NumBits, 0);
443 APInt SingleChar(NumBits, 0);
444 if (TD->isLittleEndian()) {
445 for (signed i = StrLen-1; i >= 0; i--) {
446 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
447 StrVal = (StrVal << 8) | SingleChar;
450 for (unsigned i = 0; i < StrLen; i++) {
451 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
452 StrVal = (StrVal << 8) | SingleChar;
454 // Append NULL at the end.
456 StrVal = (StrVal << 8) | SingleChar;
458 return ConstantInt::get(CE->getContext(), StrVal);
462 // If this load comes from anywhere in a constant global, and if the global
463 // is all undef or zero, we know what it loads.
464 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){
465 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
466 const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
467 if (GV->getInitializer()->isNullValue())
468 return Constant::getNullValue(ResTy);
469 if (isa<UndefValue>(GV->getInitializer()))
470 return UndefValue::get(ResTy);
474 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
475 // currently don't do any of this for big endian systems. It can be
476 // generalized in the future if someone is interested.
477 if (TD && TD->isLittleEndian())
478 return FoldReinterpretLoadFromConstPtr(CE, *TD);
482 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
483 if (LI->isVolatile()) return 0;
485 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
486 return ConstantFoldLoadFromConstPtr(C, TD);
491 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
492 /// Attempt to symbolically evaluate the result of a binary operator merging
493 /// these together. If target data info is available, it is provided as TD,
494 /// otherwise TD is null.
495 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
496 Constant *Op1, const TargetData *TD,
497 LLVMContext &Context){
500 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
501 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
505 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
506 // constant. This happens frequently when iterating over a global array.
507 if (Opc == Instruction::Sub && TD) {
508 GlobalValue *GV1, *GV2;
509 int64_t Offs1, Offs2;
511 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
512 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
514 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
515 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
522 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
523 /// constant expression, do so.
524 static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps,
525 const Type *ResultTy,
526 LLVMContext &Context,
527 const TargetData *TD) {
528 Constant *Ptr = Ops[0];
529 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
532 unsigned BitWidth = TD->getTypeSizeInBits(TD->getIntPtrType(Context));
533 APInt BasePtr(BitWidth, 0);
534 bool BaseIsInt = true;
535 if (!Ptr->isNullValue()) {
536 // If this is a inttoptr from a constant int, we can fold this as the base,
537 // otherwise we can't.
538 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
539 if (CE->getOpcode() == Instruction::IntToPtr)
540 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
541 BasePtr = Base->getValue();
542 BasePtr.zextOrTrunc(BitWidth);
549 // If this is a constant expr gep that is effectively computing an
550 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
551 for (unsigned i = 1; i != NumOps; ++i)
552 if (!isa<ConstantInt>(Ops[i]))
555 APInt Offset = APInt(BitWidth,
556 TD->getIndexedOffset(Ptr->getType(),
557 (Value**)Ops+1, NumOps-1));
558 // If the base value for this address is a literal integer value, fold the
559 // getelementptr to the resulting integer value casted to the pointer type.
561 Constant *C = ConstantInt::get(Context, Offset+BasePtr);
562 return ConstantExpr::getIntToPtr(C, ResultTy);
565 // Otherwise form a regular getelementptr. Recompute the indices so that
566 // we eliminate over-indexing of the notional static type array bounds.
567 // This makes it easy to determine if the getelementptr is "inbounds".
568 // Also, this helps GlobalOpt do SROA on GlobalVariables.
569 const Type *Ty = Ptr->getType();
570 SmallVector<Constant*, 32> NewIdxs;
572 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
573 // The only pointer indexing we'll do is on the first index of the GEP.
574 if (isa<PointerType>(ATy) && !NewIdxs.empty())
576 // Determine which element of the array the offset points into.
577 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
580 APInt NewIdx = Offset.udiv(ElemSize);
581 Offset -= NewIdx * ElemSize;
582 NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Context), NewIdx));
583 Ty = ATy->getElementType();
584 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
585 // Determine which field of the struct the offset points into. The
586 // getZExtValue is at least as safe as the StructLayout API because we
587 // know the offset is within the struct at this point.
588 const StructLayout &SL = *TD->getStructLayout(STy);
589 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
590 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Context), ElIdx));
591 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
592 Ty = STy->getTypeAtIndex(ElIdx);
594 // We've reached some non-indexable type.
597 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
599 // If we haven't used up the entire offset by descending the static
600 // type, then the offset is pointing into the middle of an indivisible
601 // member, so we can't simplify it.
607 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
608 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
609 "Computed GetElementPtr has unexpected type!");
611 // If we ended up indexing a member with a type that doesn't match
612 // the type of what the original indices indexed, add a cast.
613 if (Ty != cast<PointerType>(ResultTy)->getElementType())
614 C = FoldBitCast(C, ResultTy, *TD);
621 //===----------------------------------------------------------------------===//
622 // Constant Folding public APIs
623 //===----------------------------------------------------------------------===//
626 /// ConstantFoldInstruction - Attempt to constant fold the specified
627 /// instruction. If successful, the constant result is returned, if not, null
628 /// is returned. Note that this function can only fail when attempting to fold
629 /// instructions like loads and stores, which have no constant expression form.
631 Constant *llvm::ConstantFoldInstruction(Instruction *I, LLVMContext &Context,
632 const TargetData *TD) {
633 if (PHINode *PN = dyn_cast<PHINode>(I)) {
634 if (PN->getNumIncomingValues() == 0)
635 return UndefValue::get(PN->getType());
637 Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
638 if (Result == 0) return 0;
640 // Handle PHI nodes specially here...
641 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
642 if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
643 return 0; // Not all the same incoming constants...
645 // If we reach here, all incoming values are the same constant.
649 // Scan the operand list, checking to see if they are all constants, if so,
650 // hand off to ConstantFoldInstOperands.
651 SmallVector<Constant*, 8> Ops;
652 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
653 if (Constant *Op = dyn_cast<Constant>(*i))
656 return 0; // All operands not constant!
658 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
659 return ConstantFoldCompareInstOperands(CI->getPredicate(),
660 Ops.data(), Ops.size(),
663 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
664 return ConstantFoldLoadInst(LI, TD);
666 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
667 Ops.data(), Ops.size(), Context, TD);
670 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
671 /// using the specified TargetData. If successful, the constant result is
672 /// result is returned, if not, null is returned.
673 Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE,
674 LLVMContext &Context,
675 const TargetData *TD) {
676 SmallVector<Constant*, 8> Ops;
677 for (User::op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i)
678 Ops.push_back(cast<Constant>(*i));
681 return ConstantFoldCompareInstOperands(CE->getPredicate(),
682 Ops.data(), Ops.size(),
684 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
685 Ops.data(), Ops.size(), Context, TD);
688 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
689 /// specified opcode and operands. If successful, the constant result is
690 /// returned, if not, null is returned. Note that this function can fail when
691 /// attempting to fold instructions like loads and stores, which have no
692 /// constant expression form.
694 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
695 Constant* const* Ops, unsigned NumOps,
696 LLVMContext &Context,
697 const TargetData *TD) {
698 // Handle easy binops first.
699 if (Instruction::isBinaryOp(Opcode)) {
700 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
701 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD,
705 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
710 case Instruction::Call:
711 if (Function *F = dyn_cast<Function>(Ops[0]))
712 if (canConstantFoldCallTo(F))
713 return ConstantFoldCall(F, Ops+1, NumOps-1);
715 case Instruction::ICmp:
716 case Instruction::FCmp:
717 llvm_unreachable("This function is invalid for compares: no predicate specified");
718 case Instruction::PtrToInt:
719 // If the input is a inttoptr, eliminate the pair. This requires knowing
720 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
721 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
722 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
723 Constant *Input = CE->getOperand(0);
724 unsigned InWidth = Input->getType()->getScalarSizeInBits();
725 if (TD->getPointerSizeInBits() < InWidth) {
727 ConstantInt::get(Context, APInt::getLowBitsSet(InWidth,
728 TD->getPointerSizeInBits()));
729 Input = ConstantExpr::getAnd(Input, Mask);
731 // Do a zext or trunc to get to the dest size.
732 return ConstantExpr::getIntegerCast(Input, DestTy, false);
735 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
736 case Instruction::IntToPtr:
737 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
738 // the int size is >= the ptr size. This requires knowing the width of a
739 // pointer, so it can't be done in ConstantExpr::getCast.
740 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
742 TD->getPointerSizeInBits() <=
743 CE->getType()->getScalarSizeInBits()) {
744 if (CE->getOpcode() == Instruction::PtrToInt)
745 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
747 // If there's a constant offset added to the integer value before
748 // it is casted back to a pointer, see if the expression can be
749 // converted into a GEP.
750 if (CE->getOpcode() == Instruction::Add)
751 if (ConstantInt *L = dyn_cast<ConstantInt>(CE->getOperand(0)))
752 if (ConstantExpr *R = dyn_cast<ConstantExpr>(CE->getOperand(1)))
753 if (R->getOpcode() == Instruction::PtrToInt)
754 if (GlobalVariable *GV =
755 dyn_cast<GlobalVariable>(R->getOperand(0))) {
756 const PointerType *GVTy = cast<PointerType>(GV->getType());
757 if (const ArrayType *AT =
758 dyn_cast<ArrayType>(GVTy->getElementType())) {
759 const Type *ElTy = AT->getElementType();
760 uint64_t AllocSize = TD->getTypeAllocSize(ElTy);
761 APInt PSA(L->getValue().getBitWidth(), AllocSize);
762 if (ElTy == cast<PointerType>(DestTy)->getElementType() &&
763 L->getValue().urem(PSA) == 0) {
764 APInt ElemIdx = L->getValue().udiv(PSA);
765 if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(),
766 AT->getNumElements()))) {
767 Constant *Index[] = {
768 Constant::getNullValue(CE->getType()),
769 ConstantInt::get(Context, ElemIdx)
772 ConstantExpr::getGetElementPtr(GV, &Index[0], 2);
779 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
780 case Instruction::Trunc:
781 case Instruction::ZExt:
782 case Instruction::SExt:
783 case Instruction::FPTrunc:
784 case Instruction::FPExt:
785 case Instruction::UIToFP:
786 case Instruction::SIToFP:
787 case Instruction::FPToUI:
788 case Instruction::FPToSI:
789 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
790 case Instruction::BitCast:
792 return FoldBitCast(Ops[0], DestTy, *TD);
793 return ConstantExpr::getBitCast(Ops[0], DestTy);
794 case Instruction::Select:
795 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
796 case Instruction::ExtractElement:
797 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
798 case Instruction::InsertElement:
799 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
800 case Instruction::ShuffleVector:
801 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
802 case Instruction::GetElementPtr:
803 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, Context, TD))
806 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
810 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
811 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
812 /// returns a constant expression of the specified operands.
814 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
815 Constant*const * Ops,
817 LLVMContext &Context,
818 const TargetData *TD) {
819 // fold: icmp (inttoptr x), null -> icmp x, 0
820 // fold: icmp (ptrtoint x), 0 -> icmp x, null
821 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
822 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
824 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
825 // around to know if bit truncation is happening.
826 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops[0])) {
827 if (TD && Ops[1]->isNullValue()) {
828 const Type *IntPtrTy = TD->getIntPtrType(Context);
829 if (CE0->getOpcode() == Instruction::IntToPtr) {
830 // Convert the integer value to the right size to ensure we get the
831 // proper extension or truncation.
832 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
834 Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
835 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
839 // Only do this transformation if the int is intptrty in size, otherwise
840 // there is a truncation or extension that we aren't modeling.
841 if (CE0->getOpcode() == Instruction::PtrToInt &&
842 CE0->getType() == IntPtrTy) {
843 Constant *C = CE0->getOperand(0);
844 Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
846 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
851 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops[1])) {
852 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
853 const Type *IntPtrTy = TD->getIntPtrType(Context);
855 if (CE0->getOpcode() == Instruction::IntToPtr) {
856 // Convert the integer value to the right size to ensure we get the
857 // proper extension or truncation.
858 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
860 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
862 Constant *NewOps[] = { C0, C1 };
863 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
867 // Only do this transformation if the int is intptrty in size, otherwise
868 // there is a truncation or extension that we aren't modeling.
869 if ((CE0->getOpcode() == Instruction::PtrToInt &&
870 CE0->getType() == IntPtrTy &&
871 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) {
872 Constant *NewOps[] = {
873 CE0->getOperand(0), CE1->getOperand(0)
875 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
881 return ConstantExpr::getCompare(Predicate, Ops[0], Ops[1]);
885 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
886 /// getelementptr constantexpr, return the constant value being addressed by the
887 /// constant expression, or null if something is funny and we can't decide.
888 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
890 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
891 return 0; // Do not allow stepping over the value!
893 // Loop over all of the operands, tracking down which value we are
895 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
896 for (++I; I != E; ++I)
897 if (const StructType *STy = dyn_cast<StructType>(*I)) {
898 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
899 assert(CU->getZExtValue() < STy->getNumElements() &&
900 "Struct index out of range!");
901 unsigned El = (unsigned)CU->getZExtValue();
902 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
903 C = CS->getOperand(El);
904 } else if (isa<ConstantAggregateZero>(C)) {
905 C = Constant::getNullValue(STy->getElementType(El));
906 } else if (isa<UndefValue>(C)) {
907 C = UndefValue::get(STy->getElementType(El));
911 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
912 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
913 if (CI->getZExtValue() >= ATy->getNumElements())
915 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
916 C = CA->getOperand(CI->getZExtValue());
917 else if (isa<ConstantAggregateZero>(C))
918 C = Constant::getNullValue(ATy->getElementType());
919 else if (isa<UndefValue>(C))
920 C = UndefValue::get(ATy->getElementType());
923 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
924 if (CI->getZExtValue() >= VTy->getNumElements())
926 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
927 C = CP->getOperand(CI->getZExtValue());
928 else if (isa<ConstantAggregateZero>(C))
929 C = Constant::getNullValue(VTy->getElementType());
930 else if (isa<UndefValue>(C))
931 C = UndefValue::get(VTy->getElementType());
944 //===----------------------------------------------------------------------===//
945 // Constant Folding for Calls
948 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
949 /// the specified function.
951 llvm::canConstantFoldCallTo(const Function *F) {
952 switch (F->getIntrinsicID()) {
953 case Intrinsic::sqrt:
954 case Intrinsic::powi:
955 case Intrinsic::bswap:
956 case Intrinsic::ctpop:
957 case Intrinsic::ctlz:
958 case Intrinsic::cttz:
959 case Intrinsic::uadd_with_overflow:
960 case Intrinsic::usub_with_overflow:
961 case Intrinsic::sadd_with_overflow:
962 case Intrinsic::ssub_with_overflow:
969 if (!F->hasName()) return false;
970 StringRef Name = F->getName();
972 // In these cases, the check of the length is required. We don't want to
973 // return true for a name like "cos\0blah" which strcmp would return equal to
974 // "cos", but has length 8.
976 default: return false;
978 return Name == "acos" || Name == "asin" ||
979 Name == "atan" || Name == "atan2";
981 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
983 return Name == "exp";
985 return Name == "fabs" || Name == "fmod" || Name == "floor";
987 return Name == "log" || Name == "log10";
989 return Name == "pow";
991 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
992 Name == "sinf" || Name == "sqrtf";
994 return Name == "tan" || Name == "tanh";
998 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
999 const Type *Ty, LLVMContext &Context) {
1007 if (Ty->isFloatTy())
1008 return ConstantFP::get(Context, APFloat((float)V));
1009 if (Ty->isDoubleTy())
1010 return ConstantFP::get(Context, APFloat(V));
1011 llvm_unreachable("Can only constant fold float/double");
1012 return 0; // dummy return to suppress warning
1015 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1018 LLVMContext &Context) {
1026 if (Ty->isFloatTy())
1027 return ConstantFP::get(Context, APFloat((float)V));
1028 if (Ty->isDoubleTy())
1029 return ConstantFP::get(Context, APFloat(V));
1030 llvm_unreachable("Can only constant fold float/double");
1031 return 0; // dummy return to suppress warning
1034 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1035 /// with the specified arguments, returning null if unsuccessful.
1037 llvm::ConstantFoldCall(Function *F,
1038 Constant *const *Operands, unsigned NumOperands) {
1039 if (!F->hasName()) return 0;
1040 LLVMContext &Context = F->getContext();
1041 StringRef Name = F->getName();
1043 const Type *Ty = F->getReturnType();
1044 if (NumOperands == 1) {
1045 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1046 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1048 /// Currently APFloat versions of these functions do not exist, so we use
1049 /// the host native double versions. Float versions are not called
1050 /// directly but for all these it is true (float)(f((double)arg)) ==
1051 /// f(arg). Long double not supported yet.
1052 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1053 Op->getValueAPF().convertToDouble();
1057 return ConstantFoldFP(acos, V, Ty, Context);
1058 else if (Name == "asin")
1059 return ConstantFoldFP(asin, V, Ty, Context);
1060 else if (Name == "atan")
1061 return ConstantFoldFP(atan, V, Ty, Context);
1065 return ConstantFoldFP(ceil, V, Ty, Context);
1066 else if (Name == "cos")
1067 return ConstantFoldFP(cos, V, Ty, Context);
1068 else if (Name == "cosh")
1069 return ConstantFoldFP(cosh, V, Ty, Context);
1070 else if (Name == "cosf")
1071 return ConstantFoldFP(cos, V, Ty, Context);
1075 return ConstantFoldFP(exp, V, Ty, Context);
1079 return ConstantFoldFP(fabs, V, Ty, Context);
1080 else if (Name == "floor")
1081 return ConstantFoldFP(floor, V, Ty, Context);
1084 if (Name == "log" && V > 0)
1085 return ConstantFoldFP(log, V, Ty, Context);
1086 else if (Name == "log10" && V > 0)
1087 return ConstantFoldFP(log10, V, Ty, Context);
1088 else if (Name == "llvm.sqrt.f32" ||
1089 Name == "llvm.sqrt.f64") {
1091 return ConstantFoldFP(sqrt, V, Ty, Context);
1093 return Constant::getNullValue(Ty);
1098 return ConstantFoldFP(sin, V, Ty, Context);
1099 else if (Name == "sinh")
1100 return ConstantFoldFP(sinh, V, Ty, Context);
1101 else if (Name == "sqrt" && V >= 0)
1102 return ConstantFoldFP(sqrt, V, Ty, Context);
1103 else if (Name == "sqrtf" && V >= 0)
1104 return ConstantFoldFP(sqrt, V, Ty, Context);
1105 else if (Name == "sinf")
1106 return ConstantFoldFP(sin, V, Ty, Context);
1110 return ConstantFoldFP(tan, V, Ty, Context);
1111 else if (Name == "tanh")
1112 return ConstantFoldFP(tanh, V, Ty, Context);
1121 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1122 if (Name.startswith("llvm.bswap"))
1123 return ConstantInt::get(Context, Op->getValue().byteSwap());
1124 else if (Name.startswith("llvm.ctpop"))
1125 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1126 else if (Name.startswith("llvm.cttz"))
1127 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1128 else if (Name.startswith("llvm.ctlz"))
1129 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1136 if (NumOperands == 2) {
1137 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1138 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1140 double Op1V = Ty->isFloatTy() ?
1141 (double)Op1->getValueAPF().convertToFloat() :
1142 Op1->getValueAPF().convertToDouble();
1143 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1144 if (Op2->getType() != Op1->getType())
1147 double Op2V = Ty->isFloatTy() ?
1148 (double)Op2->getValueAPF().convertToFloat():
1149 Op2->getValueAPF().convertToDouble();
1152 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty, Context);
1154 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty, Context);
1155 if (Name == "atan2")
1156 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty, Context);
1157 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1158 if (Name == "llvm.powi.f32")
1159 return ConstantFP::get(Context, APFloat((float)std::pow((float)Op1V,
1160 (int)Op2C->getZExtValue())));
1161 if (Name == "llvm.powi.f64")
1162 return ConstantFP::get(Context, APFloat((double)std::pow((double)Op1V,
1163 (int)Op2C->getZExtValue())));
1169 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1170 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1171 switch (F->getIntrinsicID()) {
1173 case Intrinsic::uadd_with_overflow: {
1174 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1176 Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow.
1178 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1180 case Intrinsic::usub_with_overflow: {
1181 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1183 Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow.
1185 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1187 case Intrinsic::sadd_with_overflow: {
1188 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1189 Constant *Overflow = ConstantExpr::getSelect(
1190 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1191 ConstantInt::get(Op1->getType(), 0), Op1),
1192 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2),
1193 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow.
1195 Constant *Ops[] = { Res, Overflow };
1196 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1198 case Intrinsic::ssub_with_overflow: {
1199 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1200 Constant *Overflow = ConstantExpr::getSelect(
1201 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1202 ConstantInt::get(Op2->getType(), 0), Op2),
1203 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1),
1204 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow.
1206 Constant *Ops[] = { Res, Overflow };
1207 return ConstantStruct::get(F->getContext(), Ops, 2, false);