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"
33 #include "llvm/Support/FEnv.h"
38 //===----------------------------------------------------------------------===//
39 // Constant Folding internal helper functions
40 //===----------------------------------------------------------------------===//
42 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
43 /// TargetData. This always returns a non-null constant, but it may be a
44 /// ConstantExpr if unfoldable.
45 static Constant *FoldBitCast(Constant *C, const Type *DestTy,
46 const TargetData &TD) {
48 // This only handles casts to vectors currently.
49 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
51 return ConstantExpr::getBitCast(C, DestTy);
53 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
54 // vector so the code below can handle it uniformly.
55 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
56 Constant *Ops = C; // don't take the address of C!
57 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
60 // If this is a bitcast from constant vector -> vector, fold it.
61 ConstantVector *CV = dyn_cast<ConstantVector>(C);
63 return ConstantExpr::getBitCast(C, DestTy);
65 // If the element types match, VMCore can fold it.
66 unsigned NumDstElt = DestVTy->getNumElements();
67 unsigned NumSrcElt = CV->getNumOperands();
68 if (NumDstElt == NumSrcElt)
69 return ConstantExpr::getBitCast(C, DestTy);
71 const Type *SrcEltTy = CV->getType()->getElementType();
72 const Type *DstEltTy = DestVTy->getElementType();
74 // Otherwise, we're changing the number of elements in a vector, which
75 // requires endianness information to do the right thing. For example,
76 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
77 // folds to (little endian):
78 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
79 // and to (big endian):
80 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
82 // First thing is first. We only want to think about integer here, so if
83 // we have something in FP form, recast it as integer.
84 if (DstEltTy->isFloatingPointTy()) {
85 // Fold to an vector of integers with same size as our FP type.
86 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
87 const Type *DestIVTy =
88 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
89 // Recursively handle this integer conversion, if possible.
90 C = FoldBitCast(C, DestIVTy, TD);
91 if (!C) return ConstantExpr::getBitCast(C, DestTy);
93 // Finally, VMCore can handle this now that #elts line up.
94 return ConstantExpr::getBitCast(C, DestTy);
97 // Okay, we know the destination is integer, if the input is FP, convert
98 // it to integer first.
99 if (SrcEltTy->isFloatingPointTy()) {
100 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
101 const Type *SrcIVTy =
102 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
103 // Ask VMCore to do the conversion now that #elts line up.
104 C = ConstantExpr::getBitCast(C, SrcIVTy);
105 CV = dyn_cast<ConstantVector>(C);
106 if (!CV) // If VMCore wasn't able to fold it, bail out.
110 // Now we know that the input and output vectors are both integer vectors
111 // of the same size, and that their #elements is not the same. Do the
112 // conversion here, which depends on whether the input or output has
114 bool isLittleEndian = TD.isLittleEndian();
116 SmallVector<Constant*, 32> Result;
117 if (NumDstElt < NumSrcElt) {
118 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
119 Constant *Zero = Constant::getNullValue(DstEltTy);
120 unsigned Ratio = NumSrcElt/NumDstElt;
121 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
123 for (unsigned i = 0; i != NumDstElt; ++i) {
124 // Build each element of the result.
125 Constant *Elt = Zero;
126 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
127 for (unsigned j = 0; j != Ratio; ++j) {
128 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
129 if (!Src) // Reject constantexpr elements.
130 return ConstantExpr::getBitCast(C, DestTy);
132 // Zero extend the element to the right size.
133 Src = ConstantExpr::getZExt(Src, Elt->getType());
135 // Shift it to the right place, depending on endianness.
136 Src = ConstantExpr::getShl(Src,
137 ConstantInt::get(Src->getType(), ShiftAmt));
138 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
141 Elt = ConstantExpr::getOr(Elt, Src);
143 Result.push_back(Elt);
146 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
147 unsigned Ratio = NumDstElt/NumSrcElt;
148 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
150 // Loop over each source value, expanding into multiple results.
151 for (unsigned i = 0; i != NumSrcElt; ++i) {
152 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
153 if (!Src) // Reject constantexpr elements.
154 return ConstantExpr::getBitCast(C, DestTy);
156 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
157 for (unsigned j = 0; j != Ratio; ++j) {
158 // Shift the piece of the value into the right place, depending on
160 Constant *Elt = ConstantExpr::getLShr(Src,
161 ConstantInt::get(Src->getType(), ShiftAmt));
162 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
164 // Truncate and remember this piece.
165 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
170 return ConstantVector::get(Result);
174 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
175 /// from a global, return the global and the constant. Because of
176 /// constantexprs, this function is recursive.
177 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
178 int64_t &Offset, const TargetData &TD) {
179 // Trivial case, constant is the global.
180 if ((GV = dyn_cast<GlobalValue>(C))) {
185 // Otherwise, if this isn't a constant expr, bail out.
186 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
187 if (!CE) return false;
189 // Look through ptr->int and ptr->ptr casts.
190 if (CE->getOpcode() == Instruction::PtrToInt ||
191 CE->getOpcode() == Instruction::BitCast)
192 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
194 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
195 if (CE->getOpcode() == Instruction::GetElementPtr) {
196 // Cannot compute this if the element type of the pointer is missing size
198 if (!cast<PointerType>(CE->getOperand(0)->getType())
199 ->getElementType()->isSized())
202 // If the base isn't a global+constant, we aren't either.
203 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
206 // Otherwise, add any offset that our operands provide.
207 gep_type_iterator GTI = gep_type_begin(CE);
208 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
209 i != e; ++i, ++GTI) {
210 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
211 if (!CI) return false; // Index isn't a simple constant?
212 if (CI->isZero()) continue; // Not adding anything.
214 if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
216 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
218 const SequentialType *SQT = cast<SequentialType>(*GTI);
219 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
228 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
229 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
230 /// pointer to copy results into and BytesLeft is the number of bytes left in
231 /// the CurPtr buffer. TD is the target data.
232 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
233 unsigned char *CurPtr, unsigned BytesLeft,
234 const TargetData &TD) {
235 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
236 "Out of range access");
238 // If this element is zero or undefined, we can just return since *CurPtr is
240 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
243 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
244 if (CI->getBitWidth() > 64 ||
245 (CI->getBitWidth() & 7) != 0)
248 uint64_t Val = CI->getZExtValue();
249 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
251 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
252 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
258 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
259 if (CFP->getType()->isDoubleTy()) {
260 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
261 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
263 if (CFP->getType()->isFloatTy()){
264 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
265 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
270 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
271 const StructLayout *SL = TD.getStructLayout(CS->getType());
272 unsigned Index = SL->getElementContainingOffset(ByteOffset);
273 uint64_t CurEltOffset = SL->getElementOffset(Index);
274 ByteOffset -= CurEltOffset;
277 // If the element access is to the element itself and not to tail padding,
278 // read the bytes from the element.
279 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
281 if (ByteOffset < EltSize &&
282 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
288 // Check to see if we read from the last struct element, if so we're done.
289 if (Index == CS->getType()->getNumElements())
292 // If we read all of the bytes we needed from this element we're done.
293 uint64_t NextEltOffset = SL->getElementOffset(Index);
295 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
298 // Move to the next element of the struct.
299 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
300 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
302 CurEltOffset = NextEltOffset;
307 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
308 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
309 uint64_t Index = ByteOffset / EltSize;
310 uint64_t Offset = ByteOffset - Index * EltSize;
311 for (; Index != CA->getType()->getNumElements(); ++Index) {
312 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
315 if (EltSize >= BytesLeft)
319 BytesLeft -= EltSize;
325 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
326 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
327 uint64_t Index = ByteOffset / EltSize;
328 uint64_t Offset = ByteOffset - Index * EltSize;
329 for (; Index != CV->getType()->getNumElements(); ++Index) {
330 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
333 if (EltSize >= BytesLeft)
337 BytesLeft -= EltSize;
343 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
344 if (CE->getOpcode() == Instruction::IntToPtr &&
345 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
346 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
350 // Otherwise, unknown initializer type.
354 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
355 const TargetData &TD) {
356 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
357 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
359 // If this isn't an integer load we can't fold it directly.
361 // If this is a float/double load, we can try folding it as an int32/64 load
362 // and then bitcast the result. This can be useful for union cases. Note
363 // that address spaces don't matter here since we're not going to result in
364 // an actual new load.
366 if (LoadTy->isFloatTy())
367 MapTy = Type::getInt32PtrTy(C->getContext());
368 else if (LoadTy->isDoubleTy())
369 MapTy = Type::getInt64PtrTy(C->getContext());
370 else if (LoadTy->isVectorTy()) {
371 MapTy = IntegerType::get(C->getContext(),
372 TD.getTypeAllocSizeInBits(LoadTy));
373 MapTy = PointerType::getUnqual(MapTy);
377 C = FoldBitCast(C, MapTy, TD);
378 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
379 return FoldBitCast(Res, LoadTy, TD);
383 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
384 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
388 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
391 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
392 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
393 !GV->getInitializer()->getType()->isSized())
396 // If we're loading off the beginning of the global, some bytes may be valid,
397 // but we don't try to handle this.
398 if (Offset < 0) return 0;
400 // If we're not accessing anything in this constant, the result is undefined.
401 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
402 return UndefValue::get(IntType);
404 unsigned char RawBytes[32] = {0};
405 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
409 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
410 for (unsigned i = 1; i != BytesLoaded; ++i) {
412 ResultVal |= RawBytes[BytesLoaded-1-i];
415 return ConstantInt::get(IntType->getContext(), ResultVal);
418 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
419 /// produce if it is constant and determinable. If this is not determinable,
421 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
422 const TargetData *TD) {
423 // First, try the easy cases:
424 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
425 if (GV->isConstant() && GV->hasDefinitiveInitializer())
426 return GV->getInitializer();
428 // If the loaded value isn't a constant expr, we can't handle it.
429 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
432 if (CE->getOpcode() == Instruction::GetElementPtr) {
433 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
434 if (GV->isConstant() && GV->hasDefinitiveInitializer())
436 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
440 // Instead of loading constant c string, use corresponding integer value
441 // directly if string length is small enough.
443 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
444 unsigned StrLen = Str.length();
445 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
446 unsigned NumBits = Ty->getPrimitiveSizeInBits();
447 // Replace load with immediate integer if the result is an integer or fp
449 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
450 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
451 APInt StrVal(NumBits, 0);
452 APInt SingleChar(NumBits, 0);
453 if (TD->isLittleEndian()) {
454 for (signed i = StrLen-1; i >= 0; i--) {
455 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
456 StrVal = (StrVal << 8) | SingleChar;
459 for (unsigned i = 0; i < StrLen; i++) {
460 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
461 StrVal = (StrVal << 8) | SingleChar;
463 // Append NULL at the end.
465 StrVal = (StrVal << 8) | SingleChar;
468 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
469 if (Ty->isFloatingPointTy())
470 Res = ConstantExpr::getBitCast(Res, Ty);
475 // If this load comes from anywhere in a constant global, and if the global
476 // is all undef or zero, we know what it loads.
477 if (GlobalVariable *GV =
478 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
479 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
480 const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
481 if (GV->getInitializer()->isNullValue())
482 return Constant::getNullValue(ResTy);
483 if (isa<UndefValue>(GV->getInitializer()))
484 return UndefValue::get(ResTy);
488 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
489 // currently don't do any of this for big endian systems. It can be
490 // generalized in the future if someone is interested.
491 if (TD && TD->isLittleEndian())
492 return FoldReinterpretLoadFromConstPtr(CE, *TD);
496 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
497 if (LI->isVolatile()) return 0;
499 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
500 return ConstantFoldLoadFromConstPtr(C, TD);
505 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
506 /// Attempt to symbolically evaluate the result of a binary operator merging
507 /// these together. If target data info is available, it is provided as TD,
508 /// otherwise TD is null.
509 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
510 Constant *Op1, const TargetData *TD){
513 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
514 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
518 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
519 // constant. This happens frequently when iterating over a global array.
520 if (Opc == Instruction::Sub && TD) {
521 GlobalValue *GV1, *GV2;
522 int64_t Offs1, Offs2;
524 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
525 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
527 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
528 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
535 /// CastGEPIndices - If array indices are not pointer-sized integers,
536 /// explicitly cast them so that they aren't implicitly casted by the
538 static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps,
539 const Type *ResultTy,
540 const TargetData *TD) {
542 const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
545 SmallVector<Constant*, 32> NewIdxs;
546 for (unsigned i = 1; i != NumOps; ++i) {
548 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
549 reinterpret_cast<Value *const *>(Ops+1),
551 Ops[i]->getType() != IntPtrTy) {
553 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
559 NewIdxs.push_back(Ops[i]);
564 ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size());
565 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
566 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
571 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
572 /// constant expression, do so.
573 static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps,
574 const Type *ResultTy,
575 const TargetData *TD) {
576 Constant *Ptr = Ops[0];
577 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
580 const Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
582 // If this is a constant expr gep that is effectively computing an
583 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
584 for (unsigned i = 1; i != NumOps; ++i)
585 if (!isa<ConstantInt>(Ops[i])) {
587 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
588 // "inttoptr (sub (ptrtoint Ptr), V)"
590 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
591 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
592 assert((CE == 0 || CE->getType() == IntPtrTy) &&
593 "CastGEPIndices didn't canonicalize index types!");
594 if (CE && CE->getOpcode() == Instruction::Sub &&
595 CE->getOperand(0)->isNullValue()) {
596 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
597 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
598 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
599 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
600 Res = ConstantFoldConstantExpression(ResCE, TD);
607 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
608 APInt Offset = APInt(BitWidth,
609 TD->getIndexedOffset(Ptr->getType(),
610 (Value**)Ops+1, NumOps-1));
611 Ptr = cast<Constant>(Ptr->stripPointerCasts());
613 // If this is a GEP of a GEP, fold it all into a single GEP.
614 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
615 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
617 // Do not try the incorporate the sub-GEP if some index is not a number.
618 bool AllConstantInt = true;
619 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
620 if (!isa<ConstantInt>(NestedOps[i])) {
621 AllConstantInt = false;
627 Ptr = cast<Constant>(GEP->getOperand(0));
628 Offset += APInt(BitWidth,
629 TD->getIndexedOffset(Ptr->getType(),
630 (Value**)NestedOps.data(),
632 Ptr = cast<Constant>(Ptr->stripPointerCasts());
635 // If the base value for this address is a literal integer value, fold the
636 // getelementptr to the resulting integer value casted to the pointer type.
637 APInt BasePtr(BitWidth, 0);
638 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
639 if (CE->getOpcode() == Instruction::IntToPtr)
640 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
641 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
642 if (Ptr->isNullValue() || BasePtr != 0) {
643 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
644 return ConstantExpr::getIntToPtr(C, ResultTy);
647 // Otherwise form a regular getelementptr. Recompute the indices so that
648 // we eliminate over-indexing of the notional static type array bounds.
649 // This makes it easy to determine if the getelementptr is "inbounds".
650 // Also, this helps GlobalOpt do SROA on GlobalVariables.
651 const Type *Ty = Ptr->getType();
652 SmallVector<Constant*, 32> NewIdxs;
654 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
655 if (ATy->isPointerTy()) {
656 // The only pointer indexing we'll do is on the first index of the GEP.
657 if (!NewIdxs.empty())
660 // Only handle pointers to sized types, not pointers to functions.
661 if (!ATy->getElementType()->isSized())
665 // Determine which element of the array the offset points into.
666 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
667 const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
669 // The element size is 0. This may be [0 x Ty]*, so just use a zero
670 // index for this level and proceed to the next level to see if it can
671 // accommodate the offset.
672 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
674 // The element size is non-zero divide the offset by the element
675 // size (rounding down), to compute the index at this level.
676 APInt NewIdx = Offset.udiv(ElemSize);
677 Offset -= NewIdx * ElemSize;
678 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
680 Ty = ATy->getElementType();
681 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
682 // Determine which field of the struct the offset points into. The
683 // getZExtValue is at least as safe as the StructLayout API because we
684 // know the offset is within the struct at this point.
685 const StructLayout &SL = *TD->getStructLayout(STy);
686 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
687 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
689 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
690 Ty = STy->getTypeAtIndex(ElIdx);
692 // We've reached some non-indexable type.
695 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
697 // If we haven't used up the entire offset by descending the static
698 // type, then the offset is pointing into the middle of an indivisible
699 // member, so we can't simplify it.
705 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
706 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
707 "Computed GetElementPtr has unexpected type!");
709 // If we ended up indexing a member with a type that doesn't match
710 // the type of what the original indices indexed, add a cast.
711 if (Ty != cast<PointerType>(ResultTy)->getElementType())
712 C = FoldBitCast(C, ResultTy, *TD);
719 //===----------------------------------------------------------------------===//
720 // Constant Folding public APIs
721 //===----------------------------------------------------------------------===//
723 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
724 /// If successful, the constant result is returned, if not, null is returned.
725 /// Note that this fails if not all of the operands are constant. Otherwise,
726 /// this function can only fail when attempting to fold instructions like loads
727 /// and stores, which have no constant expression form.
728 Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
729 // Handle PHI nodes quickly here...
730 if (PHINode *PN = dyn_cast<PHINode>(I)) {
731 Constant *CommonValue = 0;
733 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
734 Value *Incoming = PN->getIncomingValue(i);
735 // If the incoming value is undef then skip it. Note that while we could
736 // skip the value if it is equal to the phi node itself we choose not to
737 // because that would break the rule that constant folding only applies if
738 // all operands are constants.
739 if (isa<UndefValue>(Incoming))
741 // If the incoming value is not a constant, or is a different constant to
742 // the one we saw previously, then give up.
743 Constant *C = dyn_cast<Constant>(Incoming);
744 if (!C || (CommonValue && C != CommonValue))
749 // If we reach here, all incoming values are the same constant or undef.
750 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
753 // Scan the operand list, checking to see if they are all constants, if so,
754 // hand off to ConstantFoldInstOperands.
755 SmallVector<Constant*, 8> Ops;
756 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
757 if (Constant *Op = dyn_cast<Constant>(*i))
760 return 0; // All operands not constant!
762 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
763 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
766 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
767 return ConstantFoldLoadInst(LI, TD);
769 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
770 return ConstantExpr::getInsertValue(
771 cast<Constant>(IVI->getAggregateOperand()),
772 cast<Constant>(IVI->getInsertedValueOperand()),
773 IVI->idx_begin(), IVI->getNumIndices());
775 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
776 return ConstantExpr::getExtractValue(
777 cast<Constant>(EVI->getAggregateOperand()),
778 EVI->idx_begin(), EVI->getNumIndices());
780 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
781 Ops.data(), Ops.size(), TD);
784 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
785 /// using the specified TargetData. If successful, the constant result is
786 /// result is returned, if not, null is returned.
787 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
788 const TargetData *TD) {
789 SmallVector<Constant*, 8> Ops;
790 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
792 Constant *NewC = cast<Constant>(*i);
793 // Recursively fold the ConstantExpr's operands.
794 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
795 NewC = ConstantFoldConstantExpression(NewCE, TD);
800 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
802 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
803 Ops.data(), Ops.size(), TD);
806 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
807 /// specified opcode and operands. If successful, the constant result is
808 /// returned, if not, null is returned. Note that this function can fail when
809 /// attempting to fold instructions like loads and stores, which have no
810 /// constant expression form.
812 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
813 /// information, due to only being passed an opcode and operands. Constant
814 /// folding using this function strips this information.
816 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
817 Constant* const* Ops, unsigned NumOps,
818 const TargetData *TD) {
819 // Handle easy binops first.
820 if (Instruction::isBinaryOp(Opcode)) {
821 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
822 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
825 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
830 case Instruction::ICmp:
831 case Instruction::FCmp: assert(0 && "Invalid for compares");
832 case Instruction::Call:
833 if (Function *F = dyn_cast<Function>(Ops[NumOps - 1]))
834 if (canConstantFoldCallTo(F))
835 return ConstantFoldCall(F, Ops, NumOps - 1);
837 case Instruction::PtrToInt:
838 // If the input is a inttoptr, eliminate the pair. This requires knowing
839 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
840 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
841 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
842 Constant *Input = CE->getOperand(0);
843 unsigned InWidth = Input->getType()->getScalarSizeInBits();
844 if (TD->getPointerSizeInBits() < InWidth) {
846 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
847 TD->getPointerSizeInBits()));
848 Input = ConstantExpr::getAnd(Input, Mask);
850 // Do a zext or trunc to get to the dest size.
851 return ConstantExpr::getIntegerCast(Input, DestTy, false);
854 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
855 case Instruction::IntToPtr:
856 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
857 // the int size is >= the ptr size. This requires knowing the width of a
858 // pointer, so it can't be done in ConstantExpr::getCast.
859 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
861 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
862 CE->getOpcode() == Instruction::PtrToInt)
863 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
865 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
866 case Instruction::Trunc:
867 case Instruction::ZExt:
868 case Instruction::SExt:
869 case Instruction::FPTrunc:
870 case Instruction::FPExt:
871 case Instruction::UIToFP:
872 case Instruction::SIToFP:
873 case Instruction::FPToUI:
874 case Instruction::FPToSI:
875 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
876 case Instruction::BitCast:
878 return FoldBitCast(Ops[0], DestTy, *TD);
879 return ConstantExpr::getBitCast(Ops[0], DestTy);
880 case Instruction::Select:
881 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
882 case Instruction::ExtractElement:
883 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
884 case Instruction::InsertElement:
885 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
886 case Instruction::ShuffleVector:
887 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
888 case Instruction::GetElementPtr:
889 if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD))
891 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD))
894 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
898 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
899 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
900 /// returns a constant expression of the specified operands.
902 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
903 Constant *Ops0, Constant *Ops1,
904 const TargetData *TD) {
905 // fold: icmp (inttoptr x), null -> icmp x, 0
906 // fold: icmp (ptrtoint x), 0 -> icmp x, null
907 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
908 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
910 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
911 // around to know if bit truncation is happening.
912 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
913 if (TD && Ops1->isNullValue()) {
914 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
915 if (CE0->getOpcode() == Instruction::IntToPtr) {
916 // Convert the integer value to the right size to ensure we get the
917 // proper extension or truncation.
918 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
920 Constant *Null = Constant::getNullValue(C->getType());
921 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
924 // Only do this transformation if the int is intptrty in size, otherwise
925 // there is a truncation or extension that we aren't modeling.
926 if (CE0->getOpcode() == Instruction::PtrToInt &&
927 CE0->getType() == IntPtrTy) {
928 Constant *C = CE0->getOperand(0);
929 Constant *Null = Constant::getNullValue(C->getType());
930 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
934 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
935 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
936 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
938 if (CE0->getOpcode() == Instruction::IntToPtr) {
939 // Convert the integer value to the right size to ensure we get the
940 // proper extension or truncation.
941 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
943 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
945 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
948 // Only do this transformation if the int is intptrty in size, otherwise
949 // there is a truncation or extension that we aren't modeling.
950 if ((CE0->getOpcode() == Instruction::PtrToInt &&
951 CE0->getType() == IntPtrTy &&
952 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
953 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
954 CE1->getOperand(0), TD);
958 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
959 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
960 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
961 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
963 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
965 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
967 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
968 Constant *Ops[] = { LHS, RHS };
969 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD);
973 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
977 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
978 /// getelementptr constantexpr, return the constant value being addressed by the
979 /// constant expression, or null if something is funny and we can't decide.
980 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
982 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
983 return 0; // Do not allow stepping over the value!
985 // Loop over all of the operands, tracking down which value we are
987 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
988 for (++I; I != E; ++I)
989 if (const StructType *STy = dyn_cast<StructType>(*I)) {
990 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
991 assert(CU->getZExtValue() < STy->getNumElements() &&
992 "Struct index out of range!");
993 unsigned El = (unsigned)CU->getZExtValue();
994 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
995 C = CS->getOperand(El);
996 } else if (isa<ConstantAggregateZero>(C)) {
997 C = Constant::getNullValue(STy->getElementType(El));
998 } else if (isa<UndefValue>(C)) {
999 C = UndefValue::get(STy->getElementType(El));
1003 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
1004 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
1005 if (CI->getZExtValue() >= ATy->getNumElements())
1007 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
1008 C = CA->getOperand(CI->getZExtValue());
1009 else if (isa<ConstantAggregateZero>(C))
1010 C = Constant::getNullValue(ATy->getElementType());
1011 else if (isa<UndefValue>(C))
1012 C = UndefValue::get(ATy->getElementType());
1015 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
1016 if (CI->getZExtValue() >= VTy->getNumElements())
1018 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
1019 C = CP->getOperand(CI->getZExtValue());
1020 else if (isa<ConstantAggregateZero>(C))
1021 C = Constant::getNullValue(VTy->getElementType());
1022 else if (isa<UndefValue>(C))
1023 C = UndefValue::get(VTy->getElementType());
1036 //===----------------------------------------------------------------------===//
1037 // Constant Folding for Calls
1040 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1041 /// the specified function.
1043 llvm::canConstantFoldCallTo(const Function *F) {
1044 switch (F->getIntrinsicID()) {
1045 case Intrinsic::sqrt:
1046 case Intrinsic::powi:
1047 case Intrinsic::bswap:
1048 case Intrinsic::ctpop:
1049 case Intrinsic::ctlz:
1050 case Intrinsic::cttz:
1051 case Intrinsic::uadd_with_overflow:
1052 case Intrinsic::usub_with_overflow:
1053 case Intrinsic::sadd_with_overflow:
1054 case Intrinsic::ssub_with_overflow:
1055 case Intrinsic::smul_with_overflow:
1056 case Intrinsic::convert_from_fp16:
1057 case Intrinsic::convert_to_fp16:
1058 case Intrinsic::x86_sse_cvtss2si:
1059 case Intrinsic::x86_sse_cvtss2si64:
1060 case Intrinsic::x86_sse_cvttss2si:
1061 case Intrinsic::x86_sse_cvttss2si64:
1062 case Intrinsic::x86_sse2_cvtsd2si:
1063 case Intrinsic::x86_sse2_cvtsd2si64:
1064 case Intrinsic::x86_sse2_cvttsd2si:
1065 case Intrinsic::x86_sse2_cvttsd2si64:
1072 if (!F->hasName()) return false;
1073 StringRef Name = F->getName();
1075 // In these cases, the check of the length is required. We don't want to
1076 // return true for a name like "cos\0blah" which strcmp would return equal to
1077 // "cos", but has length 8.
1079 default: return false;
1081 return Name == "acos" || Name == "asin" ||
1082 Name == "atan" || Name == "atan2";
1084 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1086 return Name == "exp";
1088 return Name == "fabs" || Name == "fmod" || Name == "floor";
1090 return Name == "log" || Name == "log10";
1092 return Name == "pow";
1094 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1095 Name == "sinf" || Name == "sqrtf";
1097 return Name == "tan" || Name == "tanh";
1101 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1103 sys::llvm_fenv_clearexcept();
1105 if (sys::llvm_fenv_testexcept()) {
1106 sys::llvm_fenv_clearexcept();
1110 if (Ty->isFloatTy())
1111 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1112 if (Ty->isDoubleTy())
1113 return ConstantFP::get(Ty->getContext(), APFloat(V));
1114 llvm_unreachable("Can only constant fold float/double");
1115 return 0; // dummy return to suppress warning
1118 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1119 double V, double W, const Type *Ty) {
1120 sys::llvm_fenv_clearexcept();
1122 if (sys::llvm_fenv_testexcept()) {
1123 sys::llvm_fenv_clearexcept();
1127 if (Ty->isFloatTy())
1128 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1129 if (Ty->isDoubleTy())
1130 return ConstantFP::get(Ty->getContext(), APFloat(V));
1131 llvm_unreachable("Can only constant fold float/double");
1132 return 0; // dummy return to suppress warning
1135 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1136 /// conversion of a constant floating point. If roundTowardZero is false, the
1137 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1138 /// the behavior of the non-truncating SSE instructions in the default rounding
1139 /// mode. The desired integer type Ty is used to select how many bits are
1140 /// available for the result. Returns null if the conversion cannot be
1141 /// performed, otherwise returns the Constant value resulting from the
1143 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1145 assert(Op && "Called with NULL operand");
1146 APFloat Val(Op->getValueAPF());
1148 // All of these conversion intrinsics form an integer of at most 64bits.
1149 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1150 assert(ResultWidth <= 64 &&
1151 "Can only constant fold conversions to 64 and 32 bit ints");
1154 bool isExact = false;
1155 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1156 : APFloat::rmNearestTiesToEven;
1157 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1158 /*isSigned=*/true, mode,
1160 if (status != APFloat::opOK && status != APFloat::opInexact)
1162 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1165 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1166 /// with the specified arguments, returning null if unsuccessful.
1168 llvm::ConstantFoldCall(Function *F,
1169 Constant *const *Operands, unsigned NumOperands) {
1170 if (!F->hasName()) return 0;
1171 StringRef Name = F->getName();
1173 const Type *Ty = F->getReturnType();
1174 if (NumOperands == 1) {
1175 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1176 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1177 APFloat Val(Op->getValueAPF());
1180 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1182 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1185 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1188 /// We only fold functions with finite arguments. Folding NaN and inf is
1189 /// likely to be aborted with an exception anyway, and some host libms
1190 /// have known errors raising exceptions.
1191 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1194 /// Currently APFloat versions of these functions do not exist, so we use
1195 /// the host native double versions. Float versions are not called
1196 /// directly but for all these it is true (float)(f((double)arg)) ==
1197 /// f(arg). Long double not supported yet.
1198 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1199 Op->getValueAPF().convertToDouble();
1203 return ConstantFoldFP(acos, V, Ty);
1204 else if (Name == "asin")
1205 return ConstantFoldFP(asin, V, Ty);
1206 else if (Name == "atan")
1207 return ConstantFoldFP(atan, V, Ty);
1211 return ConstantFoldFP(ceil, V, Ty);
1212 else if (Name == "cos")
1213 return ConstantFoldFP(cos, V, Ty);
1214 else if (Name == "cosh")
1215 return ConstantFoldFP(cosh, V, Ty);
1216 else if (Name == "cosf")
1217 return ConstantFoldFP(cos, V, Ty);
1221 return ConstantFoldFP(exp, V, Ty);
1225 return ConstantFoldFP(fabs, V, Ty);
1226 else if (Name == "floor")
1227 return ConstantFoldFP(floor, V, Ty);
1230 if (Name == "log" && V > 0)
1231 return ConstantFoldFP(log, V, Ty);
1232 else if (Name == "log10" && V > 0)
1233 return ConstantFoldFP(log10, V, Ty);
1234 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1235 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1237 return ConstantFoldFP(sqrt, V, Ty);
1239 return Constant::getNullValue(Ty);
1244 return ConstantFoldFP(sin, V, Ty);
1245 else if (Name == "sinh")
1246 return ConstantFoldFP(sinh, V, Ty);
1247 else if (Name == "sqrt" && V >= 0)
1248 return ConstantFoldFP(sqrt, V, Ty);
1249 else if (Name == "sqrtf" && V >= 0)
1250 return ConstantFoldFP(sqrt, V, Ty);
1251 else if (Name == "sinf")
1252 return ConstantFoldFP(sin, V, Ty);
1256 return ConstantFoldFP(tan, V, Ty);
1257 else if (Name == "tanh")
1258 return ConstantFoldFP(tanh, V, Ty);
1266 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1267 switch (F->getIntrinsicID()) {
1268 case Intrinsic::bswap:
1269 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1270 case Intrinsic::ctpop:
1271 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1272 case Intrinsic::cttz:
1273 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1274 case Intrinsic::ctlz:
1275 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1276 case Intrinsic::convert_from_fp16: {
1277 APFloat Val(Op->getValue());
1280 APFloat::opStatus status =
1281 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1283 // Conversion is always precise.
1285 assert(status == APFloat::opOK && !lost &&
1286 "Precision lost during fp16 constfolding");
1288 return ConstantFP::get(F->getContext(), Val);
1295 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1296 switch (F->getIntrinsicID()) {
1298 case Intrinsic::x86_sse_cvtss2si:
1299 case Intrinsic::x86_sse_cvtss2si64:
1300 case Intrinsic::x86_sse2_cvtsd2si:
1301 case Intrinsic::x86_sse2_cvtsd2si64:
1302 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1303 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1304 case Intrinsic::x86_sse_cvttss2si:
1305 case Intrinsic::x86_sse_cvttss2si64:
1306 case Intrinsic::x86_sse2_cvttsd2si:
1307 case Intrinsic::x86_sse2_cvttsd2si64:
1308 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1309 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1313 if (isa<UndefValue>(Operands[0])) {
1314 if (F->getIntrinsicID() == Intrinsic::bswap)
1322 if (NumOperands == 2) {
1323 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1324 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1326 double Op1V = Ty->isFloatTy() ?
1327 (double)Op1->getValueAPF().convertToFloat() :
1328 Op1->getValueAPF().convertToDouble();
1329 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1330 if (Op2->getType() != Op1->getType())
1333 double Op2V = Ty->isFloatTy() ?
1334 (double)Op2->getValueAPF().convertToFloat():
1335 Op2->getValueAPF().convertToDouble();
1338 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1340 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1341 if (Name == "atan2")
1342 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1343 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1344 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1345 return ConstantFP::get(F->getContext(),
1346 APFloat((float)std::pow((float)Op1V,
1347 (int)Op2C->getZExtValue())));
1348 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1349 return ConstantFP::get(F->getContext(),
1350 APFloat((double)std::pow((double)Op1V,
1351 (int)Op2C->getZExtValue())));
1357 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1358 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1359 switch (F->getIntrinsicID()) {
1361 case Intrinsic::sadd_with_overflow:
1362 case Intrinsic::uadd_with_overflow:
1363 case Intrinsic::ssub_with_overflow:
1364 case Intrinsic::usub_with_overflow:
1365 case Intrinsic::smul_with_overflow: {
1368 switch (F->getIntrinsicID()) {
1369 default: assert(0 && "Invalid case");
1370 case Intrinsic::sadd_with_overflow:
1371 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1373 case Intrinsic::uadd_with_overflow:
1374 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1376 case Intrinsic::ssub_with_overflow:
1377 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1379 case Intrinsic::usub_with_overflow:
1380 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1382 case Intrinsic::smul_with_overflow:
1383 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1387 ConstantInt::get(F->getContext(), Res),
1388 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1390 return ConstantStruct::get(F->getContext(), Ops, 2, false);