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 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
43 /// from a global, return the global and the constant. Because of
44 /// constantexprs, this function is recursive.
45 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
46 int64_t &Offset, const TargetData &TD) {
47 // Trivial case, constant is the global.
48 if ((GV = dyn_cast<GlobalValue>(C))) {
53 // Otherwise, if this isn't a constant expr, bail out.
54 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
55 if (!CE) return false;
57 // Look through ptr->int and ptr->ptr casts.
58 if (CE->getOpcode() == Instruction::PtrToInt ||
59 CE->getOpcode() == Instruction::BitCast)
60 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
62 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
63 if (CE->getOpcode() == Instruction::GetElementPtr) {
64 // Cannot compute this if the element type of the pointer is missing size
66 if (!cast<PointerType>(CE->getOperand(0)->getType())
67 ->getElementType()->isSized())
70 // If the base isn't a global+constant, we aren't either.
71 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
74 // Otherwise, add any offset that our operands provide.
75 gep_type_iterator GTI = gep_type_begin(CE);
76 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
78 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
79 if (!CI) return false; // Index isn't a simple constant?
80 if (CI->getZExtValue() == 0) continue; // Not adding anything.
82 if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
84 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
86 const SequentialType *SQT = cast<SequentialType>(*GTI);
87 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
96 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
97 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
98 /// pointer to copy results into and BytesLeft is the number of bytes left in
99 /// the CurPtr buffer. TD is the target data.
100 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
101 unsigned char *CurPtr, unsigned BytesLeft,
102 const TargetData &TD) {
103 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
104 "Out of range access");
106 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
109 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
110 if (CI->getBitWidth() > 64 ||
111 (CI->getBitWidth() & 7) != 0)
114 uint64_t Val = CI->getZExtValue();
115 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
117 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
118 CurPtr[i] = (unsigned char)(Val >> ByteOffset * 8);
124 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
125 if (CFP->getType()->isDoubleTy()) {
126 C = ConstantExpr::getBitCast(C, Type::getInt64Ty(C->getContext()));
127 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
129 if (CFP->getType()->isFloatTy()){
130 C = ConstantExpr::getBitCast(C, Type::getInt32Ty(C->getContext()));
131 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
135 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
136 const StructLayout *SL = TD.getStructLayout(CS->getType());
137 unsigned Index = SL->getElementContainingOffset(ByteOffset);
138 uint64_t CurEltOffset = SL->getElementOffset(Index);
139 ByteOffset -= CurEltOffset;
142 // If the element access is to the element itself and not to tail padding,
143 // read the bytes from the element.
144 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
146 if (ByteOffset < EltSize &&
147 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
153 // Check to see if we read from the last struct element, if so we're done.
154 if (Index == CS->getType()->getNumElements())
157 // If we read all of the bytes we needed from this element we're done.
158 uint64_t NextEltOffset = SL->getElementOffset(Index);
160 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
163 // Move to the next element of the struct.
164 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
166 CurEltOffset = NextEltOffset;
171 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
172 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
173 uint64_t Index = ByteOffset / EltSize;
174 uint64_t Offset = ByteOffset - Index * EltSize;
175 for (; Index != CA->getType()->getNumElements(); ++Index) {
176 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
179 if (EltSize >= BytesLeft)
183 BytesLeft -= EltSize;
189 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
190 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
191 uint64_t Index = ByteOffset / EltSize;
192 uint64_t Offset = ByteOffset - Index * EltSize;
193 for (; Index != CV->getType()->getNumElements(); ++Index) {
194 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
197 if (EltSize >= BytesLeft)
201 BytesLeft -= EltSize;
207 // Otherwise, unknown initializer type.
211 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
212 const TargetData &TD) {
213 const Type *InitializerTy = cast<PointerType>(C->getType())->getElementType();
214 const IntegerType *IntType = dyn_cast<IntegerType>(InitializerTy);
216 // If this isn't an integer load we can't fold it directly.
218 // If this is a float/double load, we can try folding it as an int32/64 load
219 // and then bitcast the result. This can be useful for union cases.
221 if (InitializerTy->isFloatTy())
222 MapTy = Type::getInt32PtrTy(C->getContext());
223 else if (InitializerTy->isDoubleTy())
224 MapTy = Type::getInt64PtrTy(C->getContext());
228 C = ConstantExpr::getBitCast(C, MapTy);
229 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
230 return ConstantExpr::getBitCast(Res, InitializerTy);
234 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
235 if (BytesLoaded > 8 || BytesLoaded == 0) return 0;
239 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
242 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
243 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
244 !GV->hasDefinitiveInitializer() ||
245 !GV->getInitializer()->getType()->isSized())
248 // If we're loading off the beginning of the global, some bytes may be valid,
249 // but we don't try to handle this.
250 if (Offset < 0) return 0;
252 // If we're not accessing anything in this constant, the result is undefined.
253 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
254 return UndefValue::get(IntType);
256 unsigned char RawBytes[8] = {0};
257 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
261 uint64_t ResultVal = 0;
262 for (unsigned i = 0; i != BytesLoaded; ++i)
263 ResultVal |= (uint64_t)RawBytes[i] << (i * 8);
265 return ConstantInt::get(IntType, ResultVal);
268 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
269 /// produce if it is constant and determinable. If this is not determinable,
271 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
272 const TargetData *TD) {
273 // First, try the easy cases:
274 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
275 if (GV->isConstant() && GV->hasDefinitiveInitializer())
276 return GV->getInitializer();
278 // If the loaded value isn't a constant expr, we can't handle it.
279 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
282 if (CE->getOpcode() == Instruction::GetElementPtr) {
283 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
284 if (GV->isConstant() && GV->hasDefinitiveInitializer())
286 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
290 // Instead of loading constant c string, use corresponding integer value
291 // directly if string length is small enough.
293 if (TD && GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
294 unsigned StrLen = Str.length();
295 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
296 unsigned NumBits = Ty->getPrimitiveSizeInBits();
297 // Replace LI with immediate integer store.
298 if ((NumBits >> 3) == StrLen + 1) {
299 APInt StrVal(NumBits, 0);
300 APInt SingleChar(NumBits, 0);
301 if (TD->isLittleEndian()) {
302 for (signed i = StrLen-1; i >= 0; i--) {
303 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
304 StrVal = (StrVal << 8) | SingleChar;
307 for (unsigned i = 0; i < StrLen; i++) {
308 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
309 StrVal = (StrVal << 8) | SingleChar;
311 // Append NULL at the end.
313 StrVal = (StrVal << 8) | SingleChar;
315 return ConstantInt::get(CE->getContext(), StrVal);
319 // If this load comes from anywhere in a constant global, and if the global
320 // is all undef or zero, we know what it loads.
321 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){
322 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
323 const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
324 if (GV->getInitializer()->isNullValue())
325 return Constant::getNullValue(ResTy);
326 if (isa<UndefValue>(GV->getInitializer()))
327 return UndefValue::get(ResTy);
331 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
332 // currently don't do any of this for big endian systems. It can be
333 // generalized in the future if someone is interested.
334 if (TD && TD->isLittleEndian())
335 return FoldReinterpretLoadFromConstPtr(CE, *TD);
339 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
340 if (LI->isVolatile()) return 0;
342 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
343 return ConstantFoldLoadFromConstPtr(C, TD);
348 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
349 /// Attempt to symbolically evaluate the result of a binary operator merging
350 /// these together. If target data info is available, it is provided as TD,
351 /// otherwise TD is null.
352 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
353 Constant *Op1, const TargetData *TD,
354 LLVMContext &Context){
357 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
358 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
362 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
363 // constant. This happens frequently when iterating over a global array.
364 if (Opc == Instruction::Sub && TD) {
365 GlobalValue *GV1, *GV2;
366 int64_t Offs1, Offs2;
368 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
369 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
371 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
372 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
379 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
380 /// constant expression, do so.
381 static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps,
382 const Type *ResultTy,
383 LLVMContext &Context,
384 const TargetData *TD) {
385 Constant *Ptr = Ops[0];
386 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
389 unsigned BitWidth = TD->getTypeSizeInBits(TD->getIntPtrType(Context));
390 APInt BasePtr(BitWidth, 0);
391 bool BaseIsInt = true;
392 if (!Ptr->isNullValue()) {
393 // If this is a inttoptr from a constant int, we can fold this as the base,
394 // otherwise we can't.
395 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
396 if (CE->getOpcode() == Instruction::IntToPtr)
397 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
398 BasePtr = Base->getValue();
399 BasePtr.zextOrTrunc(BitWidth);
406 // If this is a constant expr gep that is effectively computing an
407 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
408 for (unsigned i = 1; i != NumOps; ++i)
409 if (!isa<ConstantInt>(Ops[i]))
412 APInt Offset = APInt(BitWidth,
413 TD->getIndexedOffset(Ptr->getType(),
414 (Value**)Ops+1, NumOps-1));
415 // If the base value for this address is a literal integer value, fold the
416 // getelementptr to the resulting integer value casted to the pointer type.
418 Constant *C = ConstantInt::get(Context, Offset+BasePtr);
419 return ConstantExpr::getIntToPtr(C, ResultTy);
422 // Otherwise form a regular getelementptr. Recompute the indices so that
423 // we eliminate over-indexing of the notional static type array bounds.
424 // This makes it easy to determine if the getelementptr is "inbounds".
425 // Also, this helps GlobalOpt do SROA on GlobalVariables.
426 const Type *Ty = Ptr->getType();
427 SmallVector<Constant*, 32> NewIdxs;
429 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
430 // The only pointer indexing we'll do is on the first index of the GEP.
431 if (isa<PointerType>(ATy) && !NewIdxs.empty())
433 // Determine which element of the array the offset points into.
434 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
437 APInt NewIdx = Offset.udiv(ElemSize);
438 Offset -= NewIdx * ElemSize;
439 NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Context), NewIdx));
440 Ty = ATy->getElementType();
441 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
442 // Determine which field of the struct the offset points into. The
443 // getZExtValue is at least as safe as the StructLayout API because we
444 // know the offset is within the struct at this point.
445 const StructLayout &SL = *TD->getStructLayout(STy);
446 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
447 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Context), ElIdx));
448 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
449 Ty = STy->getTypeAtIndex(ElIdx);
451 // We've reached some non-indexable type.
454 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
456 // If we haven't used up the entire offset by descending the static
457 // type, then the offset is pointing into the middle of an indivisible
458 // member, so we can't simplify it.
464 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
465 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
466 "Computed GetElementPtr has unexpected type!");
468 // If we ended up indexing a member with a type that doesn't match
469 // the type of what the original indices indexed, add a cast.
470 if (Ty != cast<PointerType>(ResultTy)->getElementType())
471 C = ConstantExpr::getBitCast(C, ResultTy);
476 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
477 /// targetdata. Return 0 if unfoldable.
478 static Constant *FoldBitCast(Constant *C, const Type *DestTy,
479 const TargetData &TD, LLVMContext &Context) {
480 // If this is a bitcast from constant vector -> vector, fold it.
481 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
482 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
483 // If the element types match, VMCore can fold it.
484 unsigned NumDstElt = DestVTy->getNumElements();
485 unsigned NumSrcElt = CV->getNumOperands();
486 if (NumDstElt == NumSrcElt)
489 const Type *SrcEltTy = CV->getType()->getElementType();
490 const Type *DstEltTy = DestVTy->getElementType();
492 // Otherwise, we're changing the number of elements in a vector, which
493 // requires endianness information to do the right thing. For example,
494 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
495 // folds to (little endian):
496 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
497 // and to (big endian):
498 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
500 // First thing is first. We only want to think about integer here, so if
501 // we have something in FP form, recast it as integer.
502 if (DstEltTy->isFloatingPoint()) {
503 // Fold to an vector of integers with same size as our FP type.
504 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
505 const Type *DestIVTy = VectorType::get(
506 IntegerType::get(Context, FPWidth), NumDstElt);
507 // Recursively handle this integer conversion, if possible.
508 C = FoldBitCast(C, DestIVTy, TD, Context);
511 // Finally, VMCore can handle this now that #elts line up.
512 return ConstantExpr::getBitCast(C, DestTy);
515 // Okay, we know the destination is integer, if the input is FP, convert
516 // it to integer first.
517 if (SrcEltTy->isFloatingPoint()) {
518 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
519 const Type *SrcIVTy = VectorType::get(
520 IntegerType::get(Context, FPWidth), NumSrcElt);
521 // Ask VMCore to do the conversion now that #elts line up.
522 C = ConstantExpr::getBitCast(C, SrcIVTy);
523 CV = dyn_cast<ConstantVector>(C);
524 if (!CV) return 0; // If VMCore wasn't able to fold it, bail out.
527 // Now we know that the input and output vectors are both integer vectors
528 // of the same size, and that their #elements is not the same. Do the
529 // conversion here, which depends on whether the input or output has
531 bool isLittleEndian = TD.isLittleEndian();
533 SmallVector<Constant*, 32> Result;
534 if (NumDstElt < NumSrcElt) {
535 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
536 Constant *Zero = Constant::getNullValue(DstEltTy);
537 unsigned Ratio = NumSrcElt/NumDstElt;
538 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
540 for (unsigned i = 0; i != NumDstElt; ++i) {
541 // Build each element of the result.
542 Constant *Elt = Zero;
543 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
544 for (unsigned j = 0; j != Ratio; ++j) {
545 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
546 if (!Src) return 0; // Reject constantexpr elements.
548 // Zero extend the element to the right size.
549 Src = ConstantExpr::getZExt(Src, Elt->getType());
551 // Shift it to the right place, depending on endianness.
552 Src = ConstantExpr::getShl(Src,
553 ConstantInt::get(Src->getType(), ShiftAmt));
554 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
557 Elt = ConstantExpr::getOr(Elt, Src);
559 Result.push_back(Elt);
562 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
563 unsigned Ratio = NumDstElt/NumSrcElt;
564 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
566 // Loop over each source value, expanding into multiple results.
567 for (unsigned i = 0; i != NumSrcElt; ++i) {
568 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
569 if (!Src) return 0; // Reject constantexpr elements.
571 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
572 for (unsigned j = 0; j != Ratio; ++j) {
573 // Shift the piece of the value into the right place, depending on
575 Constant *Elt = ConstantExpr::getLShr(Src,
576 ConstantInt::get(Src->getType(), ShiftAmt));
577 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
579 // Truncate and remember this piece.
580 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
585 return ConstantVector::get(Result.data(), Result.size());
593 //===----------------------------------------------------------------------===//
594 // Constant Folding public APIs
595 //===----------------------------------------------------------------------===//
598 /// ConstantFoldInstruction - Attempt to constant fold the specified
599 /// instruction. If successful, the constant result is returned, if not, null
600 /// is returned. Note that this function can only fail when attempting to fold
601 /// instructions like loads and stores, which have no constant expression form.
603 Constant *llvm::ConstantFoldInstruction(Instruction *I, LLVMContext &Context,
604 const TargetData *TD) {
605 if (PHINode *PN = dyn_cast<PHINode>(I)) {
606 if (PN->getNumIncomingValues() == 0)
607 return UndefValue::get(PN->getType());
609 Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
610 if (Result == 0) return 0;
612 // Handle PHI nodes specially here...
613 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
614 if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
615 return 0; // Not all the same incoming constants...
617 // If we reach here, all incoming values are the same constant.
621 // Scan the operand list, checking to see if they are all constants, if so,
622 // hand off to ConstantFoldInstOperands.
623 SmallVector<Constant*, 8> Ops;
624 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
625 if (Constant *Op = dyn_cast<Constant>(*i))
628 return 0; // All operands not constant!
630 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
631 return ConstantFoldCompareInstOperands(CI->getPredicate(),
632 Ops.data(), Ops.size(),
635 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
636 return ConstantFoldLoadInst(LI, TD);
638 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
639 Ops.data(), Ops.size(), Context, TD);
642 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
643 /// using the specified TargetData. If successful, the constant result is
644 /// result is returned, if not, null is returned.
645 Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE,
646 LLVMContext &Context,
647 const TargetData *TD) {
648 SmallVector<Constant*, 8> Ops;
649 for (User::op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i)
650 Ops.push_back(cast<Constant>(*i));
653 return ConstantFoldCompareInstOperands(CE->getPredicate(),
654 Ops.data(), Ops.size(),
656 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
657 Ops.data(), Ops.size(), Context, TD);
660 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
661 /// specified opcode and operands. If successful, the constant result is
662 /// returned, if not, null is returned. Note that this function can fail when
663 /// attempting to fold instructions like loads and stores, which have no
664 /// constant expression form.
666 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
667 Constant* const* Ops, unsigned NumOps,
668 LLVMContext &Context,
669 const TargetData *TD) {
670 // Handle easy binops first.
671 if (Instruction::isBinaryOp(Opcode)) {
672 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
673 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD,
677 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
682 case Instruction::Call:
683 if (Function *F = dyn_cast<Function>(Ops[0]))
684 if (canConstantFoldCallTo(F))
685 return ConstantFoldCall(F, Ops+1, NumOps-1);
687 case Instruction::ICmp:
688 case Instruction::FCmp:
689 llvm_unreachable("This function is invalid for compares: no predicate specified");
690 case Instruction::PtrToInt:
691 // If the input is a inttoptr, eliminate the pair. This requires knowing
692 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
693 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
694 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
695 Constant *Input = CE->getOperand(0);
696 unsigned InWidth = Input->getType()->getScalarSizeInBits();
697 if (TD->getPointerSizeInBits() < InWidth) {
699 ConstantInt::get(Context, APInt::getLowBitsSet(InWidth,
700 TD->getPointerSizeInBits()));
701 Input = ConstantExpr::getAnd(Input, Mask);
703 // Do a zext or trunc to get to the dest size.
704 return ConstantExpr::getIntegerCast(Input, DestTy, false);
707 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
708 case Instruction::IntToPtr:
709 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
710 // the int size is >= the ptr size. This requires knowing the width of a
711 // pointer, so it can't be done in ConstantExpr::getCast.
712 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
714 TD->getPointerSizeInBits() <=
715 CE->getType()->getScalarSizeInBits()) {
716 if (CE->getOpcode() == Instruction::PtrToInt) {
717 Constant *Input = CE->getOperand(0);
718 Constant *C = FoldBitCast(Input, DestTy, *TD, Context);
719 return C ? C : ConstantExpr::getBitCast(Input, DestTy);
721 // If there's a constant offset added to the integer value before
722 // it is casted back to a pointer, see if the expression can be
723 // converted into a GEP.
724 if (CE->getOpcode() == Instruction::Add)
725 if (ConstantInt *L = dyn_cast<ConstantInt>(CE->getOperand(0)))
726 if (ConstantExpr *R = dyn_cast<ConstantExpr>(CE->getOperand(1)))
727 if (R->getOpcode() == Instruction::PtrToInt)
728 if (GlobalVariable *GV =
729 dyn_cast<GlobalVariable>(R->getOperand(0))) {
730 const PointerType *GVTy = cast<PointerType>(GV->getType());
731 if (const ArrayType *AT =
732 dyn_cast<ArrayType>(GVTy->getElementType())) {
733 const Type *ElTy = AT->getElementType();
734 uint64_t AllocSize = TD->getTypeAllocSize(ElTy);
735 APInt PSA(L->getValue().getBitWidth(), AllocSize);
736 if (ElTy == cast<PointerType>(DestTy)->getElementType() &&
737 L->getValue().urem(PSA) == 0) {
738 APInt ElemIdx = L->getValue().udiv(PSA);
739 if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(),
740 AT->getNumElements()))) {
741 Constant *Index[] = {
742 Constant::getNullValue(CE->getType()),
743 ConstantInt::get(Context, ElemIdx)
746 ConstantExpr::getGetElementPtr(GV, &Index[0], 2);
753 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
754 case Instruction::Trunc:
755 case Instruction::ZExt:
756 case Instruction::SExt:
757 case Instruction::FPTrunc:
758 case Instruction::FPExt:
759 case Instruction::UIToFP:
760 case Instruction::SIToFP:
761 case Instruction::FPToUI:
762 case Instruction::FPToSI:
763 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
764 case Instruction::BitCast:
766 if (Constant *C = FoldBitCast(Ops[0], DestTy, *TD, Context))
768 return ConstantExpr::getBitCast(Ops[0], DestTy);
769 case Instruction::Select:
770 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
771 case Instruction::ExtractElement:
772 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
773 case Instruction::InsertElement:
774 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
775 case Instruction::ShuffleVector:
776 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
777 case Instruction::GetElementPtr:
778 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, Context, TD))
781 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
785 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
786 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
787 /// returns a constant expression of the specified operands.
789 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
790 Constant*const * Ops,
792 LLVMContext &Context,
793 const TargetData *TD) {
794 // fold: icmp (inttoptr x), null -> icmp x, 0
795 // fold: icmp (ptrtoint x), 0 -> icmp x, null
796 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
797 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
799 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
800 // around to know if bit truncation is happening.
801 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops[0])) {
802 if (TD && Ops[1]->isNullValue()) {
803 const Type *IntPtrTy = TD->getIntPtrType(Context);
804 if (CE0->getOpcode() == Instruction::IntToPtr) {
805 // Convert the integer value to the right size to ensure we get the
806 // proper extension or truncation.
807 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
809 Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
810 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
814 // Only do this transformation if the int is intptrty in size, otherwise
815 // there is a truncation or extension that we aren't modeling.
816 if (CE0->getOpcode() == Instruction::PtrToInt &&
817 CE0->getType() == IntPtrTy) {
818 Constant *C = CE0->getOperand(0);
819 Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
821 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
826 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops[1])) {
827 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
828 const Type *IntPtrTy = TD->getIntPtrType(Context);
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 *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
835 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
837 Constant *NewOps[] = { C0, C1 };
838 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
842 // Only do this transformation if the int is intptrty in size, otherwise
843 // there is a truncation or extension that we aren't modeling.
844 if ((CE0->getOpcode() == Instruction::PtrToInt &&
845 CE0->getType() == IntPtrTy &&
846 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) {
847 Constant *NewOps[] = {
848 CE0->getOperand(0), CE1->getOperand(0)
850 return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
856 return ConstantExpr::getCompare(Predicate, Ops[0], Ops[1]);
860 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
861 /// getelementptr constantexpr, return the constant value being addressed by the
862 /// constant expression, or null if something is funny and we can't decide.
863 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
865 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
866 return 0; // Do not allow stepping over the value!
868 // Loop over all of the operands, tracking down which value we are
870 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
871 for (++I; I != E; ++I)
872 if (const StructType *STy = dyn_cast<StructType>(*I)) {
873 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
874 assert(CU->getZExtValue() < STy->getNumElements() &&
875 "Struct index out of range!");
876 unsigned El = (unsigned)CU->getZExtValue();
877 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
878 C = CS->getOperand(El);
879 } else if (isa<ConstantAggregateZero>(C)) {
880 C = Constant::getNullValue(STy->getElementType(El));
881 } else if (isa<UndefValue>(C)) {
882 C = UndefValue::get(STy->getElementType(El));
886 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
887 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
888 if (CI->getZExtValue() >= ATy->getNumElements())
890 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
891 C = CA->getOperand(CI->getZExtValue());
892 else if (isa<ConstantAggregateZero>(C))
893 C = Constant::getNullValue(ATy->getElementType());
894 else if (isa<UndefValue>(C))
895 C = UndefValue::get(ATy->getElementType());
898 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
899 if (CI->getZExtValue() >= VTy->getNumElements())
901 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
902 C = CP->getOperand(CI->getZExtValue());
903 else if (isa<ConstantAggregateZero>(C))
904 C = Constant::getNullValue(VTy->getElementType());
905 else if (isa<UndefValue>(C))
906 C = UndefValue::get(VTy->getElementType());
919 //===----------------------------------------------------------------------===//
920 // Constant Folding for Calls
923 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
924 /// the specified function.
926 llvm::canConstantFoldCallTo(const Function *F) {
927 switch (F->getIntrinsicID()) {
928 case Intrinsic::sqrt:
929 case Intrinsic::powi:
930 case Intrinsic::bswap:
931 case Intrinsic::ctpop:
932 case Intrinsic::ctlz:
933 case Intrinsic::cttz:
934 case Intrinsic::uadd_with_overflow:
935 case Intrinsic::usub_with_overflow:
936 case Intrinsic::sadd_with_overflow:
937 case Intrinsic::ssub_with_overflow:
944 if (!F->hasName()) return false;
945 StringRef Name = F->getName();
947 // In these cases, the check of the length is required. We don't want to
948 // return true for a name like "cos\0blah" which strcmp would return equal to
949 // "cos", but has length 8.
951 default: return false;
953 return Name == "acos" || Name == "asin" ||
954 Name == "atan" || Name == "atan2";
956 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
958 return Name == "exp";
960 return Name == "fabs" || Name == "fmod" || Name == "floor";
962 return Name == "log" || Name == "log10";
964 return Name == "pow";
966 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
967 Name == "sinf" || Name == "sqrtf";
969 return Name == "tan" || Name == "tanh";
973 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
974 const Type *Ty, LLVMContext &Context) {
983 return ConstantFP::get(Context, APFloat((float)V));
984 if (Ty->isDoubleTy())
985 return ConstantFP::get(Context, APFloat(V));
986 llvm_unreachable("Can only constant fold float/double");
987 return 0; // dummy return to suppress warning
990 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
993 LLVMContext &Context) {
1001 if (Ty->isFloatTy())
1002 return ConstantFP::get(Context, APFloat((float)V));
1003 if (Ty->isDoubleTy())
1004 return ConstantFP::get(Context, APFloat(V));
1005 llvm_unreachable("Can only constant fold float/double");
1006 return 0; // dummy return to suppress warning
1009 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1010 /// with the specified arguments, returning null if unsuccessful.
1012 llvm::ConstantFoldCall(Function *F,
1013 Constant *const *Operands, unsigned NumOperands) {
1014 if (!F->hasName()) return 0;
1015 LLVMContext &Context = F->getContext();
1016 StringRef Name = F->getName();
1018 const Type *Ty = F->getReturnType();
1019 if (NumOperands == 1) {
1020 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1021 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1023 /// Currently APFloat versions of these functions do not exist, so we use
1024 /// the host native double versions. Float versions are not called
1025 /// directly but for all these it is true (float)(f((double)arg)) ==
1026 /// f(arg). Long double not supported yet.
1027 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1028 Op->getValueAPF().convertToDouble();
1032 return ConstantFoldFP(acos, V, Ty, Context);
1033 else if (Name == "asin")
1034 return ConstantFoldFP(asin, V, Ty, Context);
1035 else if (Name == "atan")
1036 return ConstantFoldFP(atan, V, Ty, Context);
1040 return ConstantFoldFP(ceil, V, Ty, Context);
1041 else if (Name == "cos")
1042 return ConstantFoldFP(cos, V, Ty, Context);
1043 else if (Name == "cosh")
1044 return ConstantFoldFP(cosh, V, Ty, Context);
1045 else if (Name == "cosf")
1046 return ConstantFoldFP(cos, V, Ty, Context);
1050 return ConstantFoldFP(exp, V, Ty, Context);
1054 return ConstantFoldFP(fabs, V, Ty, Context);
1055 else if (Name == "floor")
1056 return ConstantFoldFP(floor, V, Ty, Context);
1059 if (Name == "log" && V > 0)
1060 return ConstantFoldFP(log, V, Ty, Context);
1061 else if (Name == "log10" && V > 0)
1062 return ConstantFoldFP(log10, V, Ty, Context);
1063 else if (Name == "llvm.sqrt.f32" ||
1064 Name == "llvm.sqrt.f64") {
1066 return ConstantFoldFP(sqrt, V, Ty, Context);
1068 return Constant::getNullValue(Ty);
1073 return ConstantFoldFP(sin, V, Ty, Context);
1074 else if (Name == "sinh")
1075 return ConstantFoldFP(sinh, V, Ty, Context);
1076 else if (Name == "sqrt" && V >= 0)
1077 return ConstantFoldFP(sqrt, V, Ty, Context);
1078 else if (Name == "sqrtf" && V >= 0)
1079 return ConstantFoldFP(sqrt, V, Ty, Context);
1080 else if (Name == "sinf")
1081 return ConstantFoldFP(sin, V, Ty, Context);
1085 return ConstantFoldFP(tan, V, Ty, Context);
1086 else if (Name == "tanh")
1087 return ConstantFoldFP(tanh, V, Ty, Context);
1096 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1097 if (Name.startswith("llvm.bswap"))
1098 return ConstantInt::get(Context, Op->getValue().byteSwap());
1099 else if (Name.startswith("llvm.ctpop"))
1100 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1101 else if (Name.startswith("llvm.cttz"))
1102 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1103 else if (Name.startswith("llvm.ctlz"))
1104 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1111 if (NumOperands == 2) {
1112 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1113 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1115 double Op1V = Ty->isFloatTy() ?
1116 (double)Op1->getValueAPF().convertToFloat() :
1117 Op1->getValueAPF().convertToDouble();
1118 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1119 if (Op2->getType() != Op1->getType())
1122 double Op2V = Ty->isFloatTy() ?
1123 (double)Op2->getValueAPF().convertToFloat():
1124 Op2->getValueAPF().convertToDouble();
1127 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty, Context);
1129 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty, Context);
1130 if (Name == "atan2")
1131 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty, Context);
1132 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1133 if (Name == "llvm.powi.f32")
1134 return ConstantFP::get(Context, APFloat((float)std::pow((float)Op1V,
1135 (int)Op2C->getZExtValue())));
1136 if (Name == "llvm.powi.f64")
1137 return ConstantFP::get(Context, APFloat((double)std::pow((double)Op1V,
1138 (int)Op2C->getZExtValue())));
1144 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1145 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1146 switch (F->getIntrinsicID()) {
1148 case Intrinsic::uadd_with_overflow: {
1149 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1151 Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow.
1153 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1155 case Intrinsic::usub_with_overflow: {
1156 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1158 Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow.
1160 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1162 case Intrinsic::sadd_with_overflow: {
1163 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
1164 Constant *Overflow = ConstantExpr::getSelect(
1165 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1166 ConstantInt::get(Op1->getType(), 0), Op1),
1167 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2),
1168 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow.
1170 Constant *Ops[] = { Res, Overflow };
1171 return ConstantStruct::get(F->getContext(), Ops, 2, false);
1173 case Intrinsic::ssub_with_overflow: {
1174 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
1175 Constant *Overflow = ConstantExpr::getSelect(
1176 ConstantExpr::getICmp(CmpInst::ICMP_SGT,
1177 ConstantInt::get(Op2->getType(), 0), Op2),
1178 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1),
1179 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow.
1181 Constant *Ops[] = { Res, Overflow };
1182 return ConstantStruct::get(F->getContext(), Ops, 2, false);