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/Operator.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLibraryInfo.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/StringMap.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/FEnv.h"
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// TargetData. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const TargetData &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
55 // The code below only handles casts to vectors currently.
56 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
58 return ConstantExpr::getBitCast(C, DestTy);
60 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
61 // vector so the code below can handle it uniformly.
62 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
63 Constant *Ops = C; // don't take the address of C!
64 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
67 // If this is a bitcast from constant vector -> vector, fold it.
68 ConstantVector *CV = dyn_cast<ConstantVector>(C);
70 return ConstantExpr::getBitCast(C, DestTy);
72 // If the element types match, VMCore can fold it.
73 unsigned NumDstElt = DestVTy->getNumElements();
74 unsigned NumSrcElt = CV->getNumOperands();
75 if (NumDstElt == NumSrcElt)
76 return ConstantExpr::getBitCast(C, DestTy);
78 Type *SrcEltTy = CV->getType()->getElementType();
79 Type *DstEltTy = DestVTy->getElementType();
81 // Otherwise, we're changing the number of elements in a vector, which
82 // requires endianness information to do the right thing. For example,
83 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
84 // folds to (little endian):
85 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
86 // and to (big endian):
87 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
89 // First thing is first. We only want to think about integer here, so if
90 // we have something in FP form, recast it as integer.
91 if (DstEltTy->isFloatingPointTy()) {
92 // Fold to an vector of integers with same size as our FP type.
93 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
95 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
96 // Recursively handle this integer conversion, if possible.
97 C = FoldBitCast(C, DestIVTy, TD);
98 if (!C) return ConstantExpr::getBitCast(C, DestTy);
100 // Finally, VMCore can handle this now that #elts line up.
101 return ConstantExpr::getBitCast(C, DestTy);
104 // Okay, we know the destination is integer, if the input is FP, convert
105 // it to integer first.
106 if (SrcEltTy->isFloatingPointTy()) {
107 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
109 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
110 // Ask VMCore to do the conversion now that #elts line up.
111 C = ConstantExpr::getBitCast(C, SrcIVTy);
112 CV = dyn_cast<ConstantVector>(C);
113 if (!CV) // If VMCore wasn't able to fold it, bail out.
117 // Now we know that the input and output vectors are both integer vectors
118 // of the same size, and that their #elements is not the same. Do the
119 // conversion here, which depends on whether the input or output has
121 bool isLittleEndian = TD.isLittleEndian();
123 SmallVector<Constant*, 32> Result;
124 if (NumDstElt < NumSrcElt) {
125 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
126 Constant *Zero = Constant::getNullValue(DstEltTy);
127 unsigned Ratio = NumSrcElt/NumDstElt;
128 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
130 for (unsigned i = 0; i != NumDstElt; ++i) {
131 // Build each element of the result.
132 Constant *Elt = Zero;
133 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
134 for (unsigned j = 0; j != Ratio; ++j) {
135 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
136 if (!Src) // Reject constantexpr elements.
137 return ConstantExpr::getBitCast(C, DestTy);
139 // Zero extend the element to the right size.
140 Src = ConstantExpr::getZExt(Src, Elt->getType());
142 // Shift it to the right place, depending on endianness.
143 Src = ConstantExpr::getShl(Src,
144 ConstantInt::get(Src->getType(), ShiftAmt));
145 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
148 Elt = ConstantExpr::getOr(Elt, Src);
150 Result.push_back(Elt);
153 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
154 unsigned Ratio = NumDstElt/NumSrcElt;
155 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
157 // Loop over each source value, expanding into multiple results.
158 for (unsigned i = 0; i != NumSrcElt; ++i) {
159 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
160 if (!Src) // Reject constantexpr elements.
161 return ConstantExpr::getBitCast(C, DestTy);
163 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
164 for (unsigned j = 0; j != Ratio; ++j) {
165 // Shift the piece of the value into the right place, depending on
167 Constant *Elt = ConstantExpr::getLShr(Src,
168 ConstantInt::get(Src->getType(), ShiftAmt));
169 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
171 // Truncate and remember this piece.
172 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
177 return ConstantVector::get(Result);
181 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
182 /// from a global, return the global and the constant. Because of
183 /// constantexprs, this function is recursive.
184 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
185 int64_t &Offset, const TargetData &TD) {
186 // Trivial case, constant is the global.
187 if ((GV = dyn_cast<GlobalValue>(C))) {
192 // Otherwise, if this isn't a constant expr, bail out.
193 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
194 if (!CE) return false;
196 // Look through ptr->int and ptr->ptr casts.
197 if (CE->getOpcode() == Instruction::PtrToInt ||
198 CE->getOpcode() == Instruction::BitCast)
199 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
201 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
202 if (CE->getOpcode() == Instruction::GetElementPtr) {
203 // Cannot compute this if the element type of the pointer is missing size
205 if (!cast<PointerType>(CE->getOperand(0)->getType())
206 ->getElementType()->isSized())
209 // If the base isn't a global+constant, we aren't either.
210 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
213 // Otherwise, add any offset that our operands provide.
214 gep_type_iterator GTI = gep_type_begin(CE);
215 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
216 i != e; ++i, ++GTI) {
217 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
218 if (!CI) return false; // Index isn't a simple constant?
219 if (CI->isZero()) continue; // Not adding anything.
221 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
223 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
225 SequentialType *SQT = cast<SequentialType>(*GTI);
226 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
235 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
236 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
237 /// pointer to copy results into and BytesLeft is the number of bytes left in
238 /// the CurPtr buffer. TD is the target data.
239 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
240 unsigned char *CurPtr, unsigned BytesLeft,
241 const TargetData &TD) {
242 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
243 "Out of range access");
245 // If this element is zero or undefined, we can just return since *CurPtr is
247 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
250 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
251 if (CI->getBitWidth() > 64 ||
252 (CI->getBitWidth() & 7) != 0)
255 uint64_t Val = CI->getZExtValue();
256 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
258 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
259 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
265 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
266 if (CFP->getType()->isDoubleTy()) {
267 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
268 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
270 if (CFP->getType()->isFloatTy()){
271 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
272 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
277 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
278 const StructLayout *SL = TD.getStructLayout(CS->getType());
279 unsigned Index = SL->getElementContainingOffset(ByteOffset);
280 uint64_t CurEltOffset = SL->getElementOffset(Index);
281 ByteOffset -= CurEltOffset;
284 // If the element access is to the element itself and not to tail padding,
285 // read the bytes from the element.
286 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
288 if (ByteOffset < EltSize &&
289 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
295 // Check to see if we read from the last struct element, if so we're done.
296 if (Index == CS->getType()->getNumElements())
299 // If we read all of the bytes we needed from this element we're done.
300 uint64_t NextEltOffset = SL->getElementOffset(Index);
302 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
305 // Move to the next element of the struct.
306 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
307 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
309 CurEltOffset = NextEltOffset;
314 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
315 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
316 uint64_t Index = ByteOffset / EltSize;
317 uint64_t Offset = ByteOffset - Index * EltSize;
318 for (; Index != CA->getType()->getNumElements(); ++Index) {
319 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
322 if (EltSize >= BytesLeft)
326 BytesLeft -= EltSize;
332 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
333 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
334 uint64_t Index = ByteOffset / EltSize;
335 uint64_t Offset = ByteOffset - Index * EltSize;
336 for (; Index != CV->getType()->getNumElements(); ++Index) {
337 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
340 if (EltSize >= BytesLeft)
344 BytesLeft -= EltSize;
350 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
351 if (CE->getOpcode() == Instruction::IntToPtr &&
352 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
353 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
357 // Otherwise, unknown initializer type.
361 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
362 const TargetData &TD) {
363 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
364 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
366 // If this isn't an integer load we can't fold it directly.
368 // If this is a float/double load, we can try folding it as an int32/64 load
369 // and then bitcast the result. This can be useful for union cases. Note
370 // that address spaces don't matter here since we're not going to result in
371 // an actual new load.
373 if (LoadTy->isFloatTy())
374 MapTy = Type::getInt32PtrTy(C->getContext());
375 else if (LoadTy->isDoubleTy())
376 MapTy = Type::getInt64PtrTy(C->getContext());
377 else if (LoadTy->isVectorTy()) {
378 MapTy = IntegerType::get(C->getContext(),
379 TD.getTypeAllocSizeInBits(LoadTy));
380 MapTy = PointerType::getUnqual(MapTy);
384 C = FoldBitCast(C, MapTy, TD);
385 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
386 return FoldBitCast(Res, LoadTy, TD);
390 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
391 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
395 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
398 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
399 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
400 !GV->getInitializer()->getType()->isSized())
403 // If we're loading off the beginning of the global, some bytes may be valid,
404 // but we don't try to handle this.
405 if (Offset < 0) return 0;
407 // If we're not accessing anything in this constant, the result is undefined.
408 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
409 return UndefValue::get(IntType);
411 unsigned char RawBytes[32] = {0};
412 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
416 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
417 for (unsigned i = 1; i != BytesLoaded; ++i) {
419 ResultVal |= RawBytes[BytesLoaded-1-i];
422 return ConstantInt::get(IntType->getContext(), ResultVal);
425 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
426 /// produce if it is constant and determinable. If this is not determinable,
428 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
429 const TargetData *TD) {
430 // First, try the easy cases:
431 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
432 if (GV->isConstant() && GV->hasDefinitiveInitializer())
433 return GV->getInitializer();
435 // If the loaded value isn't a constant expr, we can't handle it.
436 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
439 if (CE->getOpcode() == Instruction::GetElementPtr) {
440 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
441 if (GV->isConstant() && GV->hasDefinitiveInitializer())
443 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
447 // Instead of loading constant c string, use corresponding integer value
448 // directly if string length is small enough.
450 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
451 unsigned StrLen = Str.length();
452 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
453 unsigned NumBits = Ty->getPrimitiveSizeInBits();
454 // Replace load with immediate integer if the result is an integer or fp
456 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
457 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
458 APInt StrVal(NumBits, 0);
459 APInt SingleChar(NumBits, 0);
460 if (TD->isLittleEndian()) {
461 for (signed i = StrLen-1; i >= 0; i--) {
462 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
463 StrVal = (StrVal << 8) | SingleChar;
466 for (unsigned i = 0; i < StrLen; i++) {
467 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
468 StrVal = (StrVal << 8) | SingleChar;
470 // Append NULL at the end.
472 StrVal = (StrVal << 8) | SingleChar;
475 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
476 if (Ty->isFloatingPointTy())
477 Res = ConstantExpr::getBitCast(Res, Ty);
482 // If this load comes from anywhere in a constant global, and if the global
483 // is all undef or zero, we know what it loads.
484 if (GlobalVariable *GV =
485 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
486 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
487 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
488 if (GV->getInitializer()->isNullValue())
489 return Constant::getNullValue(ResTy);
490 if (isa<UndefValue>(GV->getInitializer()))
491 return UndefValue::get(ResTy);
495 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
496 // currently don't do any of this for big endian systems. It can be
497 // generalized in the future if someone is interested.
498 if (TD && TD->isLittleEndian())
499 return FoldReinterpretLoadFromConstPtr(CE, *TD);
503 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
504 if (LI->isVolatile()) return 0;
506 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
507 return ConstantFoldLoadFromConstPtr(C, TD);
512 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
513 /// Attempt to symbolically evaluate the result of a binary operator merging
514 /// these together. If target data info is available, it is provided as TD,
515 /// otherwise TD is null.
516 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
517 Constant *Op1, const TargetData *TD){
520 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
521 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
525 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
526 // constant. This happens frequently when iterating over a global array.
527 if (Opc == Instruction::Sub && TD) {
528 GlobalValue *GV1, *GV2;
529 int64_t Offs1, Offs2;
531 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
532 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
534 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
535 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
542 /// CastGEPIndices - If array indices are not pointer-sized integers,
543 /// explicitly cast them so that they aren't implicitly casted by the
545 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
546 Type *ResultTy, const TargetData *TD,
547 const TargetLibraryInfo *TLI) {
549 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
552 SmallVector<Constant*, 32> NewIdxs;
553 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
555 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
556 Ops.slice(1, i-1)))) &&
557 Ops[i]->getType() != IntPtrTy) {
559 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
565 NewIdxs.push_back(Ops[i]);
570 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
571 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
572 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
577 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
578 /// constant expression, do so.
579 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
580 Type *ResultTy, const TargetData *TD,
581 const TargetLibraryInfo *TLI) {
582 Constant *Ptr = Ops[0];
583 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
584 !Ptr->getType()->isPointerTy())
587 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
589 // If this is a constant expr gep that is effectively computing an
590 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
591 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
592 if (!isa<ConstantInt>(Ops[i])) {
594 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
595 // "inttoptr (sub (ptrtoint Ptr), V)"
596 if (Ops.size() == 2 &&
597 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
598 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
599 assert((CE == 0 || CE->getType() == IntPtrTy) &&
600 "CastGEPIndices didn't canonicalize index types!");
601 if (CE && CE->getOpcode() == Instruction::Sub &&
602 CE->getOperand(0)->isNullValue()) {
603 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
604 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
605 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
606 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
607 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
614 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
616 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
617 makeArrayRef((Value **)Ops.data() + 1,
619 Ptr = cast<Constant>(Ptr->stripPointerCasts());
621 // If this is a GEP of a GEP, fold it all into a single GEP.
622 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
623 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
625 // Do not try the incorporate the sub-GEP if some index is not a number.
626 bool AllConstantInt = true;
627 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
628 if (!isa<ConstantInt>(NestedOps[i])) {
629 AllConstantInt = false;
635 Ptr = cast<Constant>(GEP->getOperand(0));
636 Offset += APInt(BitWidth,
637 TD->getIndexedOffset(Ptr->getType(), NestedOps));
638 Ptr = cast<Constant>(Ptr->stripPointerCasts());
641 // If the base value for this address is a literal integer value, fold the
642 // getelementptr to the resulting integer value casted to the pointer type.
643 APInt BasePtr(BitWidth, 0);
644 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
645 if (CE->getOpcode() == Instruction::IntToPtr)
646 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
647 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
648 if (Ptr->isNullValue() || BasePtr != 0) {
649 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
650 return ConstantExpr::getIntToPtr(C, ResultTy);
653 // Otherwise form a regular getelementptr. Recompute the indices so that
654 // we eliminate over-indexing of the notional static type array bounds.
655 // This makes it easy to determine if the getelementptr is "inbounds".
656 // Also, this helps GlobalOpt do SROA on GlobalVariables.
657 Type *Ty = Ptr->getType();
658 SmallVector<Constant*, 32> NewIdxs;
660 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
661 if (ATy->isPointerTy()) {
662 // The only pointer indexing we'll do is on the first index of the GEP.
663 if (!NewIdxs.empty())
666 // Only handle pointers to sized types, not pointers to functions.
667 if (!ATy->getElementType()->isSized())
671 // Determine which element of the array the offset points into.
672 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
673 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
675 // The element size is 0. This may be [0 x Ty]*, so just use a zero
676 // index for this level and proceed to the next level to see if it can
677 // accommodate the offset.
678 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
680 // The element size is non-zero divide the offset by the element
681 // size (rounding down), to compute the index at this level.
682 APInt NewIdx = Offset.udiv(ElemSize);
683 Offset -= NewIdx * ElemSize;
684 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
686 Ty = ATy->getElementType();
687 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
688 // Determine which field of the struct the offset points into. The
689 // getZExtValue is at least as safe as the StructLayout API because we
690 // know the offset is within the struct at this point.
691 const StructLayout &SL = *TD->getStructLayout(STy);
692 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
693 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
695 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
696 Ty = STy->getTypeAtIndex(ElIdx);
698 // We've reached some non-indexable type.
701 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
703 // If we haven't used up the entire offset by descending the static
704 // type, then the offset is pointing into the middle of an indivisible
705 // member, so we can't simplify it.
711 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
712 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
713 "Computed GetElementPtr has unexpected type!");
715 // If we ended up indexing a member with a type that doesn't match
716 // the type of what the original indices indexed, add a cast.
717 if (Ty != cast<PointerType>(ResultTy)->getElementType())
718 C = FoldBitCast(C, ResultTy, *TD);
725 //===----------------------------------------------------------------------===//
726 // Constant Folding public APIs
727 //===----------------------------------------------------------------------===//
729 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
730 /// If successful, the constant result is returned, if not, null is returned.
731 /// Note that this fails if not all of the operands are constant. Otherwise,
732 /// this function can only fail when attempting to fold instructions like loads
733 /// and stores, which have no constant expression form.
734 Constant *llvm::ConstantFoldInstruction(Instruction *I,
735 const TargetData *TD,
736 const TargetLibraryInfo *TLI) {
737 // Handle PHI nodes quickly here...
738 if (PHINode *PN = dyn_cast<PHINode>(I)) {
739 Constant *CommonValue = 0;
741 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
742 Value *Incoming = PN->getIncomingValue(i);
743 // If the incoming value is undef then skip it. Note that while we could
744 // skip the value if it is equal to the phi node itself we choose not to
745 // because that would break the rule that constant folding only applies if
746 // all operands are constants.
747 if (isa<UndefValue>(Incoming))
749 // If the incoming value is not a constant, or is a different constant to
750 // the one we saw previously, then give up.
751 Constant *C = dyn_cast<Constant>(Incoming);
752 if (!C || (CommonValue && C != CommonValue))
757 // If we reach here, all incoming values are the same constant or undef.
758 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
761 // Scan the operand list, checking to see if they are all constants, if so,
762 // hand off to ConstantFoldInstOperands.
763 SmallVector<Constant*, 8> Ops;
764 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
765 if (Constant *Op = dyn_cast<Constant>(*i))
768 return 0; // All operands not constant!
770 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
771 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
774 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
775 return ConstantFoldLoadInst(LI, TD);
777 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
778 return ConstantExpr::getInsertValue(
779 cast<Constant>(IVI->getAggregateOperand()),
780 cast<Constant>(IVI->getInsertedValueOperand()),
783 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
784 return ConstantExpr::getExtractValue(
785 cast<Constant>(EVI->getAggregateOperand()),
788 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
791 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
792 /// using the specified TargetData. If successful, the constant result is
793 /// result is returned, if not, null is returned.
794 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
795 const TargetData *TD,
796 const TargetLibraryInfo *TLI) {
797 SmallVector<Constant*, 8> Ops;
798 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
800 Constant *NewC = cast<Constant>(*i);
801 // Recursively fold the ConstantExpr's operands.
802 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
803 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
808 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
810 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
813 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
814 /// specified opcode and operands. If successful, the constant result is
815 /// returned, if not, null is returned. Note that this function can fail when
816 /// attempting to fold instructions like loads and stores, which have no
817 /// constant expression form.
819 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
820 /// information, due to only being passed an opcode and operands. Constant
821 /// folding using this function strips this information.
823 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
824 ArrayRef<Constant *> Ops,
825 const TargetData *TD,
826 const TargetLibraryInfo *TLI) {
827 // Handle easy binops first.
828 if (Instruction::isBinaryOp(Opcode)) {
829 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
830 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
833 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
838 case Instruction::ICmp:
839 case Instruction::FCmp: assert(0 && "Invalid for compares");
840 case Instruction::Call:
841 if (Function *F = dyn_cast<Function>(Ops.back()))
842 if (canConstantFoldCallTo(F))
843 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
845 case Instruction::PtrToInt:
846 // If the input is a inttoptr, eliminate the pair. This requires knowing
847 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
848 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
849 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
850 Constant *Input = CE->getOperand(0);
851 unsigned InWidth = Input->getType()->getScalarSizeInBits();
852 if (TD->getPointerSizeInBits() < InWidth) {
854 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
855 TD->getPointerSizeInBits()));
856 Input = ConstantExpr::getAnd(Input, Mask);
858 // Do a zext or trunc to get to the dest size.
859 return ConstantExpr::getIntegerCast(Input, DestTy, false);
862 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
863 case Instruction::IntToPtr:
864 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
865 // the int size is >= the ptr size. This requires knowing the width of a
866 // pointer, so it can't be done in ConstantExpr::getCast.
867 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
869 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
870 CE->getOpcode() == Instruction::PtrToInt)
871 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
873 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
874 case Instruction::Trunc:
875 case Instruction::ZExt:
876 case Instruction::SExt:
877 case Instruction::FPTrunc:
878 case Instruction::FPExt:
879 case Instruction::UIToFP:
880 case Instruction::SIToFP:
881 case Instruction::FPToUI:
882 case Instruction::FPToSI:
883 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
884 case Instruction::BitCast:
886 return FoldBitCast(Ops[0], DestTy, *TD);
887 return ConstantExpr::getBitCast(Ops[0], DestTy);
888 case Instruction::Select:
889 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
890 case Instruction::ExtractElement:
891 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
892 case Instruction::InsertElement:
893 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
894 case Instruction::ShuffleVector:
895 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
896 case Instruction::GetElementPtr:
897 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
899 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
902 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
906 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
907 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
908 /// returns a constant expression of the specified operands.
910 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
911 Constant *Ops0, Constant *Ops1,
912 const TargetData *TD,
913 const TargetLibraryInfo *TLI) {
914 // fold: icmp (inttoptr x), null -> icmp x, 0
915 // fold: icmp (ptrtoint x), 0 -> icmp x, null
916 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
917 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
919 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
920 // around to know if bit truncation is happening.
921 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
922 if (TD && Ops1->isNullValue()) {
923 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
924 if (CE0->getOpcode() == Instruction::IntToPtr) {
925 // Convert the integer value to the right size to ensure we get the
926 // proper extension or truncation.
927 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
929 Constant *Null = Constant::getNullValue(C->getType());
930 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
933 // Only do this transformation if the int is intptrty in size, otherwise
934 // there is a truncation or extension that we aren't modeling.
935 if (CE0->getOpcode() == Instruction::PtrToInt &&
936 CE0->getType() == IntPtrTy) {
937 Constant *C = CE0->getOperand(0);
938 Constant *Null = Constant::getNullValue(C->getType());
939 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
943 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
944 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
945 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
947 if (CE0->getOpcode() == Instruction::IntToPtr) {
948 // Convert the integer value to the right size to ensure we get the
949 // proper extension or truncation.
950 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
952 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
954 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
957 // Only do this transformation if the int is intptrty in size, otherwise
958 // there is a truncation or extension that we aren't modeling.
959 if ((CE0->getOpcode() == Instruction::PtrToInt &&
960 CE0->getType() == IntPtrTy &&
961 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
962 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
963 CE1->getOperand(0), TD, TLI);
967 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
968 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
969 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
970 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
972 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
975 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
978 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
979 Constant *Ops[] = { LHS, RHS };
980 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
984 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
988 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
989 /// getelementptr constantexpr, return the constant value being addressed by the
990 /// constant expression, or null if something is funny and we can't decide.
991 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
993 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
994 return 0; // Do not allow stepping over the value!
996 // Loop over all of the operands, tracking down which value we are
998 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
999 for (++I; I != E; ++I)
1000 if (StructType *STy = dyn_cast<StructType>(*I)) {
1001 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
1002 assert(CU->getZExtValue() < STy->getNumElements() &&
1003 "Struct index out of range!");
1004 unsigned El = (unsigned)CU->getZExtValue();
1005 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1006 C = CS->getOperand(El);
1007 } else if (isa<ConstantAggregateZero>(C)) {
1008 C = Constant::getNullValue(STy->getElementType(El));
1009 } else if (isa<UndefValue>(C)) {
1010 C = UndefValue::get(STy->getElementType(El));
1014 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
1015 if (ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
1016 if (CI->getZExtValue() >= ATy->getNumElements())
1018 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
1019 C = CA->getOperand(CI->getZExtValue());
1020 else if (isa<ConstantAggregateZero>(C))
1021 C = Constant::getNullValue(ATy->getElementType());
1022 else if (isa<UndefValue>(C))
1023 C = UndefValue::get(ATy->getElementType());
1026 } else if (VectorType *VTy = dyn_cast<VectorType>(*I)) {
1027 if (CI->getZExtValue() >= VTy->getNumElements())
1029 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
1030 C = CP->getOperand(CI->getZExtValue());
1031 else if (isa<ConstantAggregateZero>(C))
1032 C = Constant::getNullValue(VTy->getElementType());
1033 else if (isa<UndefValue>(C))
1034 C = UndefValue::get(VTy->getElementType());
1047 //===----------------------------------------------------------------------===//
1048 // Constant Folding for Calls
1051 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1052 /// the specified function.
1054 llvm::canConstantFoldCallTo(const Function *F) {
1055 switch (F->getIntrinsicID()) {
1056 case Intrinsic::sqrt:
1057 case Intrinsic::pow:
1058 case Intrinsic::powi:
1059 case Intrinsic::bswap:
1060 case Intrinsic::ctpop:
1061 case Intrinsic::ctlz:
1062 case Intrinsic::cttz:
1063 case Intrinsic::sadd_with_overflow:
1064 case Intrinsic::uadd_with_overflow:
1065 case Intrinsic::ssub_with_overflow:
1066 case Intrinsic::usub_with_overflow:
1067 case Intrinsic::smul_with_overflow:
1068 case Intrinsic::umul_with_overflow:
1069 case Intrinsic::convert_from_fp16:
1070 case Intrinsic::convert_to_fp16:
1071 case Intrinsic::x86_sse_cvtss2si:
1072 case Intrinsic::x86_sse_cvtss2si64:
1073 case Intrinsic::x86_sse_cvttss2si:
1074 case Intrinsic::x86_sse_cvttss2si64:
1075 case Intrinsic::x86_sse2_cvtsd2si:
1076 case Intrinsic::x86_sse2_cvtsd2si64:
1077 case Intrinsic::x86_sse2_cvttsd2si:
1078 case Intrinsic::x86_sse2_cvttsd2si64:
1085 if (!F->hasName()) return false;
1086 StringRef Name = F->getName();
1088 // In these cases, the check of the length is required. We don't want to
1089 // return true for a name like "cos\0blah" which strcmp would return equal to
1090 // "cos", but has length 8.
1092 default: return false;
1094 return Name == "acos" || Name == "asin" ||
1095 Name == "atan" || Name == "atan2";
1097 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1099 return Name == "exp" || Name == "exp2";
1101 return Name == "fabs" || Name == "fmod" || Name == "floor";
1103 return Name == "log" || Name == "log10";
1105 return Name == "pow";
1107 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1108 Name == "sinf" || Name == "sqrtf";
1110 return Name == "tan" || Name == "tanh";
1114 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1116 sys::llvm_fenv_clearexcept();
1118 if (sys::llvm_fenv_testexcept()) {
1119 sys::llvm_fenv_clearexcept();
1123 if (Ty->isFloatTy())
1124 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1125 if (Ty->isDoubleTy())
1126 return ConstantFP::get(Ty->getContext(), APFloat(V));
1127 llvm_unreachable("Can only constant fold float/double");
1128 return 0; // dummy return to suppress warning
1131 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1132 double V, double W, Type *Ty) {
1133 sys::llvm_fenv_clearexcept();
1135 if (sys::llvm_fenv_testexcept()) {
1136 sys::llvm_fenv_clearexcept();
1140 if (Ty->isFloatTy())
1141 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1142 if (Ty->isDoubleTy())
1143 return ConstantFP::get(Ty->getContext(), APFloat(V));
1144 llvm_unreachable("Can only constant fold float/double");
1145 return 0; // dummy return to suppress warning
1148 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1149 /// conversion of a constant floating point. If roundTowardZero is false, the
1150 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1151 /// the behavior of the non-truncating SSE instructions in the default rounding
1152 /// mode. The desired integer type Ty is used to select how many bits are
1153 /// available for the result. Returns null if the conversion cannot be
1154 /// performed, otherwise returns the Constant value resulting from the
1156 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1158 assert(Op && "Called with NULL operand");
1159 APFloat Val(Op->getValueAPF());
1161 // All of these conversion intrinsics form an integer of at most 64bits.
1162 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1163 assert(ResultWidth <= 64 &&
1164 "Can only constant fold conversions to 64 and 32 bit ints");
1167 bool isExact = false;
1168 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1169 : APFloat::rmNearestTiesToEven;
1170 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1171 /*isSigned=*/true, mode,
1173 if (status != APFloat::opOK && status != APFloat::opInexact)
1175 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1178 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1179 /// with the specified arguments, returning null if unsuccessful.
1181 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1182 const TargetLibraryInfo *TLI) {
1183 if (!F->hasName()) return 0;
1184 StringRef Name = F->getName();
1186 Type *Ty = F->getReturnType();
1187 if (Operands.size() == 1) {
1188 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1189 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1190 APFloat Val(Op->getValueAPF());
1193 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1195 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1200 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1203 /// We only fold functions with finite arguments. Folding NaN and inf is
1204 /// likely to be aborted with an exception anyway, and some host libms
1205 /// have known errors raising exceptions.
1206 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1209 /// Currently APFloat versions of these functions do not exist, so we use
1210 /// the host native double versions. Float versions are not called
1211 /// directly but for all these it is true (float)(f((double)arg)) ==
1212 /// f(arg). Long double not supported yet.
1213 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1214 Op->getValueAPF().convertToDouble();
1217 if (Name == "acos" && TLI->has(LibFunc::acos))
1218 return ConstantFoldFP(acos, V, Ty);
1219 else if (Name == "asin" && TLI->has(LibFunc::asin))
1220 return ConstantFoldFP(asin, V, Ty);
1221 else if (Name == "atan" && TLI->has(LibFunc::atan))
1222 return ConstantFoldFP(atan, V, Ty);
1225 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1226 return ConstantFoldFP(ceil, V, Ty);
1227 else if (Name == "cos" && TLI->has(LibFunc::cos))
1228 return ConstantFoldFP(cos, V, Ty);
1229 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1230 return ConstantFoldFP(cosh, V, Ty);
1231 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1232 return ConstantFoldFP(cos, V, Ty);
1235 if (Name == "exp" && TLI->has(LibFunc::exp))
1236 return ConstantFoldFP(exp, V, Ty);
1238 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1239 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1241 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1245 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1246 return ConstantFoldFP(fabs, V, Ty);
1247 else if (Name == "floor" && TLI->has(LibFunc::floor))
1248 return ConstantFoldFP(floor, V, Ty);
1251 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1252 return ConstantFoldFP(log, V, Ty);
1253 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1254 return ConstantFoldFP(log10, V, Ty);
1255 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1256 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1258 return ConstantFoldFP(sqrt, V, Ty);
1260 return Constant::getNullValue(Ty);
1264 if (Name == "sin" && TLI->has(LibFunc::sin))
1265 return ConstantFoldFP(sin, V, Ty);
1266 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1267 return ConstantFoldFP(sinh, V, Ty);
1268 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1269 return ConstantFoldFP(sqrt, V, Ty);
1270 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1271 return ConstantFoldFP(sqrt, V, Ty);
1272 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1273 return ConstantFoldFP(sin, V, Ty);
1276 if (Name == "tan" && TLI->has(LibFunc::tan))
1277 return ConstantFoldFP(tan, V, Ty);
1278 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1279 return ConstantFoldFP(tanh, V, Ty);
1287 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1288 switch (F->getIntrinsicID()) {
1289 case Intrinsic::bswap:
1290 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1291 case Intrinsic::ctpop:
1292 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1293 case Intrinsic::cttz:
1294 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1295 case Intrinsic::ctlz:
1296 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1297 case Intrinsic::convert_from_fp16: {
1298 APFloat Val(Op->getValue());
1301 APFloat::opStatus status =
1302 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1304 // Conversion is always precise.
1306 assert(status == APFloat::opOK && !lost &&
1307 "Precision lost during fp16 constfolding");
1309 return ConstantFP::get(F->getContext(), Val);
1316 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1317 switch (F->getIntrinsicID()) {
1319 case Intrinsic::x86_sse_cvtss2si:
1320 case Intrinsic::x86_sse_cvtss2si64:
1321 case Intrinsic::x86_sse2_cvtsd2si:
1322 case Intrinsic::x86_sse2_cvtsd2si64:
1323 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1324 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1325 case Intrinsic::x86_sse_cvttss2si:
1326 case Intrinsic::x86_sse_cvttss2si64:
1327 case Intrinsic::x86_sse2_cvttsd2si:
1328 case Intrinsic::x86_sse2_cvttsd2si64:
1329 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1330 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1334 if (isa<UndefValue>(Operands[0])) {
1335 if (F->getIntrinsicID() == Intrinsic::bswap)
1343 if (Operands.size() == 2) {
1344 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1345 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1347 double Op1V = Ty->isFloatTy() ?
1348 (double)Op1->getValueAPF().convertToFloat() :
1349 Op1->getValueAPF().convertToDouble();
1350 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1351 if (Op2->getType() != Op1->getType())
1354 double Op2V = Ty->isFloatTy() ?
1355 (double)Op2->getValueAPF().convertToFloat():
1356 Op2->getValueAPF().convertToDouble();
1358 if (F->getIntrinsicID() == Intrinsic::pow) {
1359 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1363 if (Name == "pow" && TLI->has(LibFunc::pow))
1364 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1365 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1366 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1367 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1368 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1369 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1370 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1371 return ConstantFP::get(F->getContext(),
1372 APFloat((float)std::pow((float)Op1V,
1373 (int)Op2C->getZExtValue())));
1374 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1375 return ConstantFP::get(F->getContext(),
1376 APFloat((double)std::pow((double)Op1V,
1377 (int)Op2C->getZExtValue())));
1382 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1383 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1384 switch (F->getIntrinsicID()) {
1386 case Intrinsic::sadd_with_overflow:
1387 case Intrinsic::uadd_with_overflow:
1388 case Intrinsic::ssub_with_overflow:
1389 case Intrinsic::usub_with_overflow:
1390 case Intrinsic::smul_with_overflow:
1391 case Intrinsic::umul_with_overflow: {
1394 switch (F->getIntrinsicID()) {
1395 default: assert(0 && "Invalid case");
1396 case Intrinsic::sadd_with_overflow:
1397 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1399 case Intrinsic::uadd_with_overflow:
1400 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1402 case Intrinsic::ssub_with_overflow:
1403 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1405 case Intrinsic::usub_with_overflow:
1406 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1408 case Intrinsic::smul_with_overflow:
1409 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1411 case Intrinsic::umul_with_overflow:
1412 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1416 ConstantInt::get(F->getContext(), Res),
1417 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1419 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);