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/Analysis/ValueTracking.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringMap.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/MathExtras.h"
37 //===----------------------------------------------------------------------===//
38 // Constant Folding internal helper functions
39 //===----------------------------------------------------------------------===//
41 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
42 /// TargetData. This always returns a non-null constant, but it may be a
43 /// ConstantExpr if unfoldable.
44 static Constant *FoldBitCast(Constant *C, const Type *DestTy,
45 const TargetData &TD) {
47 // This only handles casts to vectors currently.
48 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
50 return ConstantExpr::getBitCast(C, DestTy);
52 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
53 // vector so the code below can handle it uniformly.
54 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
55 Constant *Ops = C; // don't take the address of C!
56 return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD);
59 // If this is a bitcast from constant vector -> vector, fold it.
60 ConstantVector *CV = dyn_cast<ConstantVector>(C);
62 return ConstantExpr::getBitCast(C, DestTy);
64 // If the element types match, VMCore can fold it.
65 unsigned NumDstElt = DestVTy->getNumElements();
66 unsigned NumSrcElt = CV->getNumOperands();
67 if (NumDstElt == NumSrcElt)
68 return ConstantExpr::getBitCast(C, DestTy);
70 const Type *SrcEltTy = CV->getType()->getElementType();
71 const Type *DstEltTy = DestVTy->getElementType();
73 // Otherwise, we're changing the number of elements in a vector, which
74 // requires endianness information to do the right thing. For example,
75 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
76 // folds to (little endian):
77 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
78 // and to (big endian):
79 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
81 // First thing is first. We only want to think about integer here, so if
82 // we have something in FP form, recast it as integer.
83 if (DstEltTy->isFloatingPoint()) {
84 // Fold to an vector of integers with same size as our FP type.
85 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
86 const Type *DestIVTy =
87 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
88 // Recursively handle this integer conversion, if possible.
89 C = FoldBitCast(C, DestIVTy, TD);
90 if (!C) return ConstantExpr::getBitCast(C, DestTy);
92 // Finally, VMCore can handle this now that #elts line up.
93 return ConstantExpr::getBitCast(C, DestTy);
96 // Okay, we know the destination is integer, if the input is FP, convert
97 // it to integer first.
98 if (SrcEltTy->isFloatingPoint()) {
99 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
100 const Type *SrcIVTy =
101 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
102 // Ask VMCore to do the conversion now that #elts line up.
103 C = ConstantExpr::getBitCast(C, SrcIVTy);
104 CV = dyn_cast<ConstantVector>(C);
105 if (!CV) // If VMCore wasn't able to fold it, bail out.
109 // Now we know that the input and output vectors are both integer vectors
110 // of the same size, and that their #elements is not the same. Do the
111 // conversion here, which depends on whether the input or output has
113 bool isLittleEndian = TD.isLittleEndian();
115 SmallVector<Constant*, 32> Result;
116 if (NumDstElt < NumSrcElt) {
117 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
118 Constant *Zero = Constant::getNullValue(DstEltTy);
119 unsigned Ratio = NumSrcElt/NumDstElt;
120 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
122 for (unsigned i = 0; i != NumDstElt; ++i) {
123 // Build each element of the result.
124 Constant *Elt = Zero;
125 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
126 for (unsigned j = 0; j != Ratio; ++j) {
127 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
128 if (!Src) // Reject constantexpr elements.
129 return ConstantExpr::getBitCast(C, DestTy);
131 // Zero extend the element to the right size.
132 Src = ConstantExpr::getZExt(Src, Elt->getType());
134 // Shift it to the right place, depending on endianness.
135 Src = ConstantExpr::getShl(Src,
136 ConstantInt::get(Src->getType(), ShiftAmt));
137 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
140 Elt = ConstantExpr::getOr(Elt, Src);
142 Result.push_back(Elt);
145 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
146 unsigned Ratio = NumDstElt/NumSrcElt;
147 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
149 // Loop over each source value, expanding into multiple results.
150 for (unsigned i = 0; i != NumSrcElt; ++i) {
151 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
152 if (!Src) // Reject constantexpr elements.
153 return ConstantExpr::getBitCast(C, DestTy);
155 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
156 for (unsigned j = 0; j != Ratio; ++j) {
157 // Shift the piece of the value into the right place, depending on
159 Constant *Elt = ConstantExpr::getLShr(Src,
160 ConstantInt::get(Src->getType(), ShiftAmt));
161 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
163 // Truncate and remember this piece.
164 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
169 return ConstantVector::get(Result.data(), Result.size());
173 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
174 /// from a global, return the global and the constant. Because of
175 /// constantexprs, this function is recursive.
176 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
177 int64_t &Offset, const TargetData &TD) {
178 // Trivial case, constant is the global.
179 if ((GV = dyn_cast<GlobalValue>(C))) {
184 // Otherwise, if this isn't a constant expr, bail out.
185 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
186 if (!CE) return false;
188 // Look through ptr->int and ptr->ptr casts.
189 if (CE->getOpcode() == Instruction::PtrToInt ||
190 CE->getOpcode() == Instruction::BitCast)
191 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
193 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
194 if (CE->getOpcode() == Instruction::GetElementPtr) {
195 // Cannot compute this if the element type of the pointer is missing size
197 if (!cast<PointerType>(CE->getOperand(0)->getType())
198 ->getElementType()->isSized())
201 // If the base isn't a global+constant, we aren't either.
202 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
205 // Otherwise, add any offset that our operands provide.
206 gep_type_iterator GTI = gep_type_begin(CE);
207 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
208 i != e; ++i, ++GTI) {
209 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
210 if (!CI) return false; // Index isn't a simple constant?
211 if (CI->getZExtValue() == 0) continue; // Not adding anything.
213 if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
215 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
217 const SequentialType *SQT = cast<SequentialType>(*GTI);
218 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
227 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
228 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
229 /// pointer to copy results into and BytesLeft is the number of bytes left in
230 /// the CurPtr buffer. TD is the target data.
231 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
232 unsigned char *CurPtr, unsigned BytesLeft,
233 const TargetData &TD) {
234 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
235 "Out of range access");
237 // If this element is zero or undefined, we can just return since *CurPtr is
239 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
242 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
243 if (CI->getBitWidth() > 64 ||
244 (CI->getBitWidth() & 7) != 0)
247 uint64_t Val = CI->getZExtValue();
248 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
250 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
251 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
257 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
258 if (CFP->getType()->isDoubleTy()) {
259 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
260 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
262 if (CFP->getType()->isFloatTy()){
263 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
264 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
269 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
270 const StructLayout *SL = TD.getStructLayout(CS->getType());
271 unsigned Index = SL->getElementContainingOffset(ByteOffset);
272 uint64_t CurEltOffset = SL->getElementOffset(Index);
273 ByteOffset -= CurEltOffset;
276 // If the element access is to the element itself and not to tail padding,
277 // read the bytes from the element.
278 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
280 if (ByteOffset < EltSize &&
281 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
287 // Check to see if we read from the last struct element, if so we're done.
288 if (Index == CS->getType()->getNumElements())
291 // If we read all of the bytes we needed from this element we're done.
292 uint64_t NextEltOffset = SL->getElementOffset(Index);
294 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
297 // Move to the next element of the struct.
298 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
299 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
301 CurEltOffset = NextEltOffset;
306 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
307 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
308 uint64_t Index = ByteOffset / EltSize;
309 uint64_t Offset = ByteOffset - Index * EltSize;
310 for (; Index != CA->getType()->getNumElements(); ++Index) {
311 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
314 if (EltSize >= BytesLeft)
318 BytesLeft -= EltSize;
324 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
325 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
326 uint64_t Index = ByteOffset / EltSize;
327 uint64_t Offset = ByteOffset - Index * EltSize;
328 for (; Index != CV->getType()->getNumElements(); ++Index) {
329 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
332 if (EltSize >= BytesLeft)
336 BytesLeft -= EltSize;
342 // Otherwise, unknown initializer type.
346 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
347 const TargetData &TD) {
348 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
349 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
351 // If this isn't an integer load we can't fold it directly.
353 // If this is a float/double load, we can try folding it as an int32/64 load
354 // and then bitcast the result. This can be useful for union cases. Note
355 // that address spaces don't matter here since we're not going to result in
356 // an actual new load.
358 if (LoadTy->isFloatTy())
359 MapTy = Type::getInt32PtrTy(C->getContext());
360 else if (LoadTy->isDoubleTy())
361 MapTy = Type::getInt64PtrTy(C->getContext());
362 else if (isa<VectorType>(LoadTy)) {
363 MapTy = IntegerType::get(C->getContext(),
364 TD.getTypeAllocSizeInBits(LoadTy));
365 MapTy = PointerType::getUnqual(MapTy);
369 C = FoldBitCast(C, MapTy, TD);
370 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
371 return FoldBitCast(Res, LoadTy, TD);
375 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
376 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
380 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
383 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
384 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
385 !GV->getInitializer()->getType()->isSized())
388 // If we're loading off the beginning of the global, some bytes may be valid,
389 // but we don't try to handle this.
390 if (Offset < 0) return 0;
392 // If we're not accessing anything in this constant, the result is undefined.
393 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
394 return UndefValue::get(IntType);
396 unsigned char RawBytes[32] = {0};
397 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
401 APInt ResultVal(IntType->getBitWidth(), 0);
402 for (unsigned i = 0; i != BytesLoaded; ++i) {
404 ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]);
407 return ConstantInt::get(IntType->getContext(), ResultVal);
410 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
411 /// produce if it is constant and determinable. If this is not determinable,
413 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
414 const TargetData *TD) {
415 // First, try the easy cases:
416 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
417 if (GV->isConstant() && GV->hasDefinitiveInitializer())
418 return GV->getInitializer();
420 // If the loaded value isn't a constant expr, we can't handle it.
421 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
424 if (CE->getOpcode() == Instruction::GetElementPtr) {
425 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
426 if (GV->isConstant() && GV->hasDefinitiveInitializer())
428 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
432 // Instead of loading constant c string, use corresponding integer value
433 // directly if string length is small enough.
435 if (TD && GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
436 unsigned StrLen = Str.length();
437 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
438 unsigned NumBits = Ty->getPrimitiveSizeInBits();
439 // Replace LI with immediate integer store.
440 if ((NumBits >> 3) == StrLen + 1) {
441 APInt StrVal(NumBits, 0);
442 APInt SingleChar(NumBits, 0);
443 if (TD->isLittleEndian()) {
444 for (signed i = StrLen-1; i >= 0; i--) {
445 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
446 StrVal = (StrVal << 8) | SingleChar;
449 for (unsigned i = 0; i < StrLen; i++) {
450 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
451 StrVal = (StrVal << 8) | SingleChar;
453 // Append NULL at the end.
455 StrVal = (StrVal << 8) | SingleChar;
457 return ConstantInt::get(CE->getContext(), StrVal);
461 // If this load comes from anywhere in a constant global, and if the global
462 // is all undef or zero, we know what it loads.
463 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){
464 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
465 const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
466 if (GV->getInitializer()->isNullValue())
467 return Constant::getNullValue(ResTy);
468 if (isa<UndefValue>(GV->getInitializer()))
469 return UndefValue::get(ResTy);
473 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
474 // currently don't do any of this for big endian systems. It can be
475 // generalized in the future if someone is interested.
476 if (TD && TD->isLittleEndian())
477 return FoldReinterpretLoadFromConstPtr(CE, *TD);
481 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
482 if (LI->isVolatile()) return 0;
484 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
485 return ConstantFoldLoadFromConstPtr(C, TD);
490 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
491 /// Attempt to symbolically evaluate the result of a binary operator merging
492 /// these together. If target data info is available, it is provided as TD,
493 /// otherwise TD is null.
494 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
495 Constant *Op1, const TargetData *TD){
498 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
499 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
503 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
504 // constant. This happens frequently when iterating over a global array.
505 if (Opc == Instruction::Sub && TD) {
506 GlobalValue *GV1, *GV2;
507 int64_t Offs1, Offs2;
509 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
510 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
512 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
513 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
520 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
521 /// constant expression, do so.
522 static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps,
523 const Type *ResultTy,
524 const TargetData *TD) {
525 Constant *Ptr = Ops[0];
526 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
530 TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext()));
531 APInt BasePtr(BitWidth, 0);
532 bool BaseIsInt = true;
533 if (!Ptr->isNullValue()) {
534 // If this is a inttoptr from a constant int, we can fold this as the base,
535 // otherwise we can't.
536 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
537 if (CE->getOpcode() == Instruction::IntToPtr)
538 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
539 BasePtr = Base->getValue();
540 BasePtr.zextOrTrunc(BitWidth);
547 // If this is a constant expr gep that is effectively computing an
548 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
549 for (unsigned i = 1; i != NumOps; ++i)
550 if (!isa<ConstantInt>(Ops[i]))
553 APInt Offset = APInt(BitWidth,
554 TD->getIndexedOffset(Ptr->getType(),
555 (Value**)Ops+1, NumOps-1));
556 // If the base value for this address is a literal integer value, fold the
557 // getelementptr to the resulting integer value casted to the pointer type.
559 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
560 return ConstantExpr::getIntToPtr(C, ResultTy);
563 // Otherwise form a regular getelementptr. Recompute the indices so that
564 // we eliminate over-indexing of the notional static type array bounds.
565 // This makes it easy to determine if the getelementptr is "inbounds".
566 // Also, this helps GlobalOpt do SROA on GlobalVariables.
567 const Type *Ty = Ptr->getType();
568 SmallVector<Constant*, 32> NewIdxs;
570 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
571 // The only pointer indexing we'll do is on the first index of the GEP.
572 if (isa<PointerType>(ATy) && !NewIdxs.empty())
574 // Determine which element of the array the offset points into.
575 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
578 APInt NewIdx = Offset.udiv(ElemSize);
579 Offset -= NewIdx * ElemSize;
580 NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Ty->getContext()),
582 Ty = ATy->getElementType();
583 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
584 // Determine which field of the struct the offset points into. The
585 // getZExtValue is at least as safe as the StructLayout API because we
586 // know the offset is within the struct at this point.
587 const StructLayout &SL = *TD->getStructLayout(STy);
588 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
589 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
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, const TargetData *TD) {
632 if (PHINode *PN = dyn_cast<PHINode>(I)) {
633 if (PN->getNumIncomingValues() == 0)
634 return UndefValue::get(PN->getType());
636 Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
637 if (Result == 0) return 0;
639 // Handle PHI nodes specially here...
640 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
641 if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
642 return 0; // Not all the same incoming constants...
644 // If we reach here, all incoming values are the same constant.
648 // Scan the operand list, checking to see if they are all constants, if so,
649 // hand off to ConstantFoldInstOperands.
650 SmallVector<Constant*, 8> Ops;
651 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
652 if (Constant *Op = dyn_cast<Constant>(*i))
655 return 0; // All operands not constant!
657 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
658 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
661 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
662 return ConstantFoldLoadInst(LI, TD);
664 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
665 Ops.data(), Ops.size(), TD);
668 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
669 /// using the specified TargetData. If successful, the constant result is
670 /// result is returned, if not, null is returned.
671 Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE,
672 const TargetData *TD) {
673 SmallVector<Constant*, 8> Ops;
674 for (User::op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i) {
675 Constant *NewC = cast<Constant>(*i);
676 // Recursively fold the ConstantExpr's operands.
677 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
678 NewC = ConstantFoldConstantExpression(NewCE, TD);
683 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
685 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
686 Ops.data(), Ops.size(), TD);
689 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
690 /// specified opcode and operands. If successful, the constant result is
691 /// returned, if not, null is returned. Note that this function can fail when
692 /// attempting to fold instructions like loads and stores, which have no
693 /// constant expression form.
695 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
696 /// information, due to only being passed an opcode and operands. Constant
697 /// folding using this function strips this information.
699 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
700 Constant* const* Ops, unsigned NumOps,
701 const TargetData *TD) {
702 // Handle easy binops first.
703 if (Instruction::isBinaryOp(Opcode)) {
704 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
705 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
708 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
713 case Instruction::Call:
714 if (Function *F = dyn_cast<Function>(Ops[0]))
715 if (canConstantFoldCallTo(F))
716 return ConstantFoldCall(F, Ops+1, NumOps-1);
718 case Instruction::ICmp:
719 case Instruction::FCmp:
720 llvm_unreachable("This function is invalid for compares: no predicate specified");
721 case Instruction::PtrToInt:
722 // If the input is a inttoptr, eliminate the pair. This requires knowing
723 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
724 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
725 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
726 Constant *Input = CE->getOperand(0);
727 unsigned InWidth = Input->getType()->getScalarSizeInBits();
728 if (TD->getPointerSizeInBits() < InWidth) {
730 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
731 TD->getPointerSizeInBits()));
732 Input = ConstantExpr::getAnd(Input, Mask);
734 // Do a zext or trunc to get to the dest size.
735 return ConstantExpr::getIntegerCast(Input, DestTy, false);
738 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
739 case Instruction::IntToPtr:
740 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
741 // the int size is >= the ptr size. This requires knowing the width of a
742 // pointer, so it can't be done in ConstantExpr::getCast.
743 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
745 TD->getPointerSizeInBits() <=
746 CE->getType()->getScalarSizeInBits()) {
747 if (CE->getOpcode() == Instruction::PtrToInt)
748 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
750 // If there's a constant offset added to the integer value before
751 // it is casted back to a pointer, see if the expression can be
752 // converted into a GEP.
753 if (CE->getOpcode() == Instruction::Add)
754 if (ConstantInt *L = dyn_cast<ConstantInt>(CE->getOperand(0)))
755 if (ConstantExpr *R = dyn_cast<ConstantExpr>(CE->getOperand(1)))
756 if (R->getOpcode() == Instruction::PtrToInt)
757 if (GlobalVariable *GV =
758 dyn_cast<GlobalVariable>(R->getOperand(0))) {
759 const PointerType *GVTy = cast<PointerType>(GV->getType());
760 if (const ArrayType *AT =
761 dyn_cast<ArrayType>(GVTy->getElementType())) {
762 const Type *ElTy = AT->getElementType();
763 uint64_t AllocSize = TD->getTypeAllocSize(ElTy);
764 APInt PSA(L->getValue().getBitWidth(), AllocSize);
765 if (ElTy == cast<PointerType>(DestTy)->getElementType() &&
766 L->getValue().urem(PSA) == 0) {
767 APInt ElemIdx = L->getValue().udiv(PSA);
768 if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(),
769 AT->getNumElements()))) {
770 Constant *Index[] = {
771 Constant::getNullValue(CE->getType()),
772 ConstantInt::get(ElTy->getContext(), ElemIdx)
775 ConstantExpr::getGetElementPtr(GV, &Index[0], 2);
782 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
783 case Instruction::Trunc:
784 case Instruction::ZExt:
785 case Instruction::SExt:
786 case Instruction::FPTrunc:
787 case Instruction::FPExt:
788 case Instruction::UIToFP:
789 case Instruction::SIToFP:
790 case Instruction::FPToUI:
791 case Instruction::FPToSI:
792 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
793 case Instruction::BitCast:
795 return FoldBitCast(Ops[0], DestTy, *TD);
796 return ConstantExpr::getBitCast(Ops[0], DestTy);
797 case Instruction::Select:
798 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
799 case Instruction::ExtractElement:
800 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
801 case Instruction::InsertElement:
802 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
803 case Instruction::ShuffleVector:
804 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
805 case Instruction::GetElementPtr:
806 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD))
809 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
813 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
814 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
815 /// returns a constant expression of the specified operands.
817 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
818 Constant *Ops0, Constant *Ops1,
819 const TargetData *TD) {
820 // fold: icmp (inttoptr x), null -> icmp x, 0
821 // fold: icmp (ptrtoint x), 0 -> icmp x, null
822 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
823 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
825 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
826 // around to know if bit truncation is happening.
827 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
828 if (TD && Ops1->isNullValue()) {
829 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
830 if (CE0->getOpcode() == Instruction::IntToPtr) {
831 // Convert the integer value to the right size to ensure we get the
832 // proper extension or truncation.
833 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
835 Constant *Null = Constant::getNullValue(C->getType());
836 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
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 *Null = Constant::getNullValue(C->getType());
845 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
849 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
850 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
851 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
853 if (CE0->getOpcode() == Instruction::IntToPtr) {
854 // Convert the integer value to the right size to ensure we get the
855 // proper extension or truncation.
856 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
858 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
860 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
863 // Only do this transformation if the int is intptrty in size, otherwise
864 // there is a truncation or extension that we aren't modeling.
865 if ((CE0->getOpcode() == Instruction::PtrToInt &&
866 CE0->getType() == IntPtrTy &&
867 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
868 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
869 CE1->getOperand(0), TD);
874 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
878 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
879 /// getelementptr constantexpr, return the constant value being addressed by the
880 /// constant expression, or null if something is funny and we can't decide.
881 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
883 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
884 return 0; // Do not allow stepping over the value!
886 // Loop over all of the operands, tracking down which value we are
888 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
889 for (++I; I != E; ++I)
890 if (const StructType *STy = dyn_cast<StructType>(*I)) {
891 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
892 assert(CU->getZExtValue() < STy->getNumElements() &&
893 "Struct index out of range!");
894 unsigned El = (unsigned)CU->getZExtValue();
895 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
896 C = CS->getOperand(El);
897 } else if (isa<ConstantAggregateZero>(C)) {
898 C = Constant::getNullValue(STy->getElementType(El));
899 } else if (isa<UndefValue>(C)) {
900 C = UndefValue::get(STy->getElementType(El));
904 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
905 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
906 if (CI->getZExtValue() >= ATy->getNumElements())
908 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
909 C = CA->getOperand(CI->getZExtValue());
910 else if (isa<ConstantAggregateZero>(C))
911 C = Constant::getNullValue(ATy->getElementType());
912 else if (isa<UndefValue>(C))
913 C = UndefValue::get(ATy->getElementType());
916 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
917 if (CI->getZExtValue() >= VTy->getNumElements())
919 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
920 C = CP->getOperand(CI->getZExtValue());
921 else if (isa<ConstantAggregateZero>(C))
922 C = Constant::getNullValue(VTy->getElementType());
923 else if (isa<UndefValue>(C))
924 C = UndefValue::get(VTy->getElementType());
937 //===----------------------------------------------------------------------===//
938 // Constant Folding for Calls
941 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
942 /// the specified function.
944 llvm::canConstantFoldCallTo(const Function *F) {
945 switch (F->getIntrinsicID()) {
946 case Intrinsic::sqrt:
947 case Intrinsic::powi:
948 case Intrinsic::bswap:
949 case Intrinsic::ctpop:
950 case Intrinsic::ctlz:
951 case Intrinsic::cttz:
952 case Intrinsic::uadd_with_overflow:
953 case Intrinsic::usub_with_overflow:
954 case Intrinsic::sadd_with_overflow:
955 case Intrinsic::ssub_with_overflow:
962 if (!F->hasName()) return false;
963 StringRef Name = F->getName();
965 // In these cases, the check of the length is required. We don't want to
966 // return true for a name like "cos\0blah" which strcmp would return equal to
967 // "cos", but has length 8.
969 default: return false;
971 return Name == "acos" || Name == "asin" ||
972 Name == "atan" || Name == "atan2";
974 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
976 return Name == "exp";
978 return Name == "fabs" || Name == "fmod" || Name == "floor";
980 return Name == "log" || Name == "log10";
982 return Name == "pow";
984 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
985 Name == "sinf" || Name == "sqrtf";
987 return Name == "tan" || Name == "tanh";
991 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1000 if (Ty->isFloatTy())
1001 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1002 if (Ty->isDoubleTy())
1003 return ConstantFP::get(Ty->getContext(), APFloat(V));
1004 llvm_unreachable("Can only constant fold float/double");
1005 return 0; // dummy return to suppress warning
1008 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1009 double V, double W, const Type *Ty) {
1017 if (Ty->isFloatTy())
1018 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1019 if (Ty->isDoubleTy())
1020 return ConstantFP::get(Ty->getContext(), APFloat(V));
1021 llvm_unreachable("Can only constant fold float/double");
1022 return 0; // dummy return to suppress warning
1025 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1026 /// with the specified arguments, returning null if unsuccessful.
1028 llvm::ConstantFoldCall(Function *F,
1029 Constant *const *Operands, unsigned NumOperands) {
1030 if (!F->hasName()) return 0;
1031 StringRef Name = F->getName();
1033 const Type *Ty = F->getReturnType();
1034 if (NumOperands == 1) {
1035 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1036 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1038 /// Currently APFloat versions of these functions do not exist, so we use
1039 /// the host native double versions. Float versions are not called
1040 /// directly but for all these it is true (float)(f((double)arg)) ==
1041 /// f(arg). Long double not supported yet.
1042 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1043 Op->getValueAPF().convertToDouble();
1047 return ConstantFoldFP(acos, V, Ty);
1048 else if (Name == "asin")
1049 return ConstantFoldFP(asin, V, Ty);
1050 else if (Name == "atan")
1051 return ConstantFoldFP(atan, V, Ty);
1055 return ConstantFoldFP(ceil, V, Ty);
1056 else if (Name == "cos")
1057 return ConstantFoldFP(cos, V, Ty);
1058 else if (Name == "cosh")
1059 return ConstantFoldFP(cosh, V, Ty);
1060 else if (Name == "cosf")
1061 return ConstantFoldFP(cos, V, Ty);
1065 return ConstantFoldFP(exp, V, Ty);
1069 return ConstantFoldFP(fabs, V, Ty);
1070 else if (Name == "floor")
1071 return ConstantFoldFP(floor, V, Ty);
1074 if (Name == "log" && V > 0)
1075 return ConstantFoldFP(log, V, Ty);
1076 else if (Name == "log10" && V > 0)
1077 return ConstantFoldFP(log10, V, Ty);
1078 else if (Name == "llvm.sqrt.f32" ||
1079 Name == "llvm.sqrt.f64") {
1081 return ConstantFoldFP(sqrt, V, Ty);
1083 return Constant::getNullValue(Ty);
1088 return ConstantFoldFP(sin, V, Ty);
1089 else if (Name == "sinh")
1090 return ConstantFoldFP(sinh, V, Ty);
1091 else if (Name == "sqrt" && V >= 0)
1092 return ConstantFoldFP(sqrt, V, Ty);
1093 else if (Name == "sqrtf" && V >= 0)
1094 return ConstantFoldFP(sqrt, V, Ty);
1095 else if (Name == "sinf")
1096 return ConstantFoldFP(sin, V, Ty);
1100 return ConstantFoldFP(tan, V, Ty);
1101 else if (Name == "tanh")
1102 return ConstantFoldFP(tanh, V, Ty);
1111 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1112 if (Name.startswith("llvm.bswap"))
1113 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1114 else if (Name.startswith("llvm.ctpop"))
1115 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1116 else if (Name.startswith("llvm.cttz"))
1117 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1118 else if (Name.startswith("llvm.ctlz"))
1119 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1126 if (NumOperands == 2) {
1127 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1128 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1130 double Op1V = Ty->isFloatTy() ?
1131 (double)Op1->getValueAPF().convertToFloat() :
1132 Op1->getValueAPF().convertToDouble();
1133 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1134 if (Op2->getType() != Op1->getType())
1137 double Op2V = Ty->isFloatTy() ?
1138 (double)Op2->getValueAPF().convertToFloat():
1139 Op2->getValueAPF().convertToDouble();
1142 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1144 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1145 if (Name == "atan2")
1146 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1147 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1148 if (Name == "llvm.powi.f32")
1149 return ConstantFP::get(F->getContext(),
1150 APFloat((float)std::pow((float)Op1V,
1151 (int)Op2C->getZExtValue())));
1152 if (Name == "llvm.powi.f64")
1153 return ConstantFP::get(F->getContext(),
1154 APFloat((double)std::pow((double)Op1V,
1155 (int)Op2C->getZExtValue())));
1161 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1162 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1163 switch (F->getIntrinsicID()) {
1165 case Intrinsic::uadd_with_overflow: {
1166 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1168 Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow.
1170 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1172 case Intrinsic::usub_with_overflow: {
1173 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1175 Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow.
1177 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1179 case Intrinsic::sadd_with_overflow: {
1180 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1181 Constant *Overflow = ConstantExpr::getSelect(
1182 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1183 ConstantInt::get(Op1->getType(), 0), Op1),
1184 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2),
1185 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow.
1187 Constant *Ops[] = { Res, Overflow };
1188 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1190 case Intrinsic::ssub_with_overflow: {
1191 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1192 Constant *Overflow = ConstantExpr::getSelect(
1193 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1194 ConstantInt::get(Op2->getType(), 0), Op2),
1195 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1),
1196 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow.
1198 Constant *Ops[] = { Res, Overflow };
1199 return ConstantStruct::get(F->getContext(), Ops, 2, false);