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/ADT/SmallVector.h"
30 #include "llvm/ADT/StringMap.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/FEnv.h"
39 //===----------------------------------------------------------------------===//
40 // Constant Folding internal helper functions
41 //===----------------------------------------------------------------------===//
43 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
44 /// TargetData. This always returns a non-null constant, but it may be a
45 /// ConstantExpr if unfoldable.
46 static Constant *FoldBitCast(Constant *C, Type *DestTy,
47 const TargetData &TD) {
48 // Catch the obvious splat cases.
49 if (C->isNullValue() && !DestTy->isX86_MMXTy())
50 return Constant::getNullValue(DestTy);
51 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
52 return Constant::getAllOnesValue(DestTy);
54 // The code below only handles casts to vectors currently.
55 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
57 return ConstantExpr::getBitCast(C, DestTy);
59 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
60 // vector so the code below can handle it uniformly.
61 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
62 Constant *Ops = C; // don't take the address of C!
63 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
66 // If this is a bitcast from constant vector -> vector, fold it.
67 ConstantVector *CV = dyn_cast<ConstantVector>(C);
69 return ConstantExpr::getBitCast(C, DestTy);
71 // If the element types match, VMCore can fold it.
72 unsigned NumDstElt = DestVTy->getNumElements();
73 unsigned NumSrcElt = CV->getNumOperands();
74 if (NumDstElt == NumSrcElt)
75 return ConstantExpr::getBitCast(C, DestTy);
77 Type *SrcEltTy = CV->getType()->getElementType();
78 Type *DstEltTy = DestVTy->getElementType();
80 // Otherwise, we're changing the number of elements in a vector, which
81 // requires endianness information to do the right thing. For example,
82 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
83 // folds to (little endian):
84 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
85 // and to (big endian):
86 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
88 // First thing is first. We only want to think about integer here, so if
89 // we have something in FP form, recast it as integer.
90 if (DstEltTy->isFloatingPointTy()) {
91 // Fold to an vector of integers with same size as our FP type.
92 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
94 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
95 // Recursively handle this integer conversion, if possible.
96 C = FoldBitCast(C, DestIVTy, TD);
97 if (!C) return ConstantExpr::getBitCast(C, DestTy);
99 // Finally, VMCore can handle this now that #elts line up.
100 return ConstantExpr::getBitCast(C, DestTy);
103 // Okay, we know the destination is integer, if the input is FP, convert
104 // it to integer first.
105 if (SrcEltTy->isFloatingPointTy()) {
106 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
108 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
109 // Ask VMCore to do the conversion now that #elts line up.
110 C = ConstantExpr::getBitCast(C, SrcIVTy);
111 CV = dyn_cast<ConstantVector>(C);
112 if (!CV) // If VMCore wasn't able to fold it, bail out.
116 // Now we know that the input and output vectors are both integer vectors
117 // of the same size, and that their #elements is not the same. Do the
118 // conversion here, which depends on whether the input or output has
120 bool isLittleEndian = TD.isLittleEndian();
122 SmallVector<Constant*, 32> Result;
123 if (NumDstElt < NumSrcElt) {
124 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
125 Constant *Zero = Constant::getNullValue(DstEltTy);
126 unsigned Ratio = NumSrcElt/NumDstElt;
127 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
129 for (unsigned i = 0; i != NumDstElt; ++i) {
130 // Build each element of the result.
131 Constant *Elt = Zero;
132 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
133 for (unsigned j = 0; j != Ratio; ++j) {
134 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
135 if (!Src) // Reject constantexpr elements.
136 return ConstantExpr::getBitCast(C, DestTy);
138 // Zero extend the element to the right size.
139 Src = ConstantExpr::getZExt(Src, Elt->getType());
141 // Shift it to the right place, depending on endianness.
142 Src = ConstantExpr::getShl(Src,
143 ConstantInt::get(Src->getType(), ShiftAmt));
144 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
147 Elt = ConstantExpr::getOr(Elt, Src);
149 Result.push_back(Elt);
152 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
153 unsigned Ratio = NumDstElt/NumSrcElt;
154 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
156 // Loop over each source value, expanding into multiple results.
157 for (unsigned i = 0; i != NumSrcElt; ++i) {
158 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
159 if (!Src) // Reject constantexpr elements.
160 return ConstantExpr::getBitCast(C, DestTy);
162 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
163 for (unsigned j = 0; j != Ratio; ++j) {
164 // Shift the piece of the value into the right place, depending on
166 Constant *Elt = ConstantExpr::getLShr(Src,
167 ConstantInt::get(Src->getType(), ShiftAmt));
168 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
170 // Truncate and remember this piece.
171 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
176 return ConstantVector::get(Result);
180 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
181 /// from a global, return the global and the constant. Because of
182 /// constantexprs, this function is recursive.
183 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
184 int64_t &Offset, const TargetData &TD) {
185 // Trivial case, constant is the global.
186 if ((GV = dyn_cast<GlobalValue>(C))) {
191 // Otherwise, if this isn't a constant expr, bail out.
192 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
193 if (!CE) return false;
195 // Look through ptr->int and ptr->ptr casts.
196 if (CE->getOpcode() == Instruction::PtrToInt ||
197 CE->getOpcode() == Instruction::BitCast)
198 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
200 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
201 if (CE->getOpcode() == Instruction::GetElementPtr) {
202 // Cannot compute this if the element type of the pointer is missing size
204 if (!cast<PointerType>(CE->getOperand(0)->getType())
205 ->getElementType()->isSized())
208 // If the base isn't a global+constant, we aren't either.
209 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
212 // Otherwise, add any offset that our operands provide.
213 gep_type_iterator GTI = gep_type_begin(CE);
214 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
215 i != e; ++i, ++GTI) {
216 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
217 if (!CI) return false; // Index isn't a simple constant?
218 if (CI->isZero()) continue; // Not adding anything.
220 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
222 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
224 SequentialType *SQT = cast<SequentialType>(*GTI);
225 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
234 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
235 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
236 /// pointer to copy results into and BytesLeft is the number of bytes left in
237 /// the CurPtr buffer. TD is the target data.
238 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
239 unsigned char *CurPtr, unsigned BytesLeft,
240 const TargetData &TD) {
241 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
242 "Out of range access");
244 // If this element is zero or undefined, we can just return since *CurPtr is
246 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
249 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
250 if (CI->getBitWidth() > 64 ||
251 (CI->getBitWidth() & 7) != 0)
254 uint64_t Val = CI->getZExtValue();
255 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
257 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
258 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
264 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
265 if (CFP->getType()->isDoubleTy()) {
266 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
267 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
269 if (CFP->getType()->isFloatTy()){
270 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
271 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
276 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
277 const StructLayout *SL = TD.getStructLayout(CS->getType());
278 unsigned Index = SL->getElementContainingOffset(ByteOffset);
279 uint64_t CurEltOffset = SL->getElementOffset(Index);
280 ByteOffset -= CurEltOffset;
283 // If the element access is to the element itself and not to tail padding,
284 // read the bytes from the element.
285 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
287 if (ByteOffset < EltSize &&
288 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
294 // Check to see if we read from the last struct element, if so we're done.
295 if (Index == CS->getType()->getNumElements())
298 // If we read all of the bytes we needed from this element we're done.
299 uint64_t NextEltOffset = SL->getElementOffset(Index);
301 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
304 // Move to the next element of the struct.
305 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
306 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
308 CurEltOffset = NextEltOffset;
313 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
314 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
315 uint64_t Index = ByteOffset / EltSize;
316 uint64_t Offset = ByteOffset - Index * EltSize;
317 for (; Index != CA->getType()->getNumElements(); ++Index) {
318 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
321 if (EltSize >= BytesLeft)
325 BytesLeft -= EltSize;
331 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
332 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
333 uint64_t Index = ByteOffset / EltSize;
334 uint64_t Offset = ByteOffset - Index * EltSize;
335 for (; Index != CV->getType()->getNumElements(); ++Index) {
336 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
339 if (EltSize >= BytesLeft)
343 BytesLeft -= EltSize;
349 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
350 if (CE->getOpcode() == Instruction::IntToPtr &&
351 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
352 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
356 // Otherwise, unknown initializer type.
360 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
361 const TargetData &TD) {
362 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
363 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
365 // If this isn't an integer load we can't fold it directly.
367 // If this is a float/double load, we can try folding it as an int32/64 load
368 // and then bitcast the result. This can be useful for union cases. Note
369 // that address spaces don't matter here since we're not going to result in
370 // an actual new load.
372 if (LoadTy->isFloatTy())
373 MapTy = Type::getInt32PtrTy(C->getContext());
374 else if (LoadTy->isDoubleTy())
375 MapTy = Type::getInt64PtrTy(C->getContext());
376 else if (LoadTy->isVectorTy()) {
377 MapTy = IntegerType::get(C->getContext(),
378 TD.getTypeAllocSizeInBits(LoadTy));
379 MapTy = PointerType::getUnqual(MapTy);
383 C = FoldBitCast(C, MapTy, TD);
384 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
385 return FoldBitCast(Res, LoadTy, TD);
389 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
390 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
394 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
397 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
398 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
399 !GV->getInitializer()->getType()->isSized())
402 // If we're loading off the beginning of the global, some bytes may be valid,
403 // but we don't try to handle this.
404 if (Offset < 0) return 0;
406 // If we're not accessing anything in this constant, the result is undefined.
407 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
408 return UndefValue::get(IntType);
410 unsigned char RawBytes[32] = {0};
411 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
415 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
416 for (unsigned i = 1; i != BytesLoaded; ++i) {
418 ResultVal |= RawBytes[BytesLoaded-1-i];
421 return ConstantInt::get(IntType->getContext(), ResultVal);
424 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
425 /// produce if it is constant and determinable. If this is not determinable,
427 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
428 const TargetData *TD) {
429 // First, try the easy cases:
430 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
431 if (GV->isConstant() && GV->hasDefinitiveInitializer())
432 return GV->getInitializer();
434 // If the loaded value isn't a constant expr, we can't handle it.
435 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
438 if (CE->getOpcode() == Instruction::GetElementPtr) {
439 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
440 if (GV->isConstant() && GV->hasDefinitiveInitializer())
442 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
446 // Instead of loading constant c string, use corresponding integer value
447 // directly if string length is small enough.
449 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
450 unsigned StrLen = Str.length();
451 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
452 unsigned NumBits = Ty->getPrimitiveSizeInBits();
453 // Replace load with immediate integer if the result is an integer or fp
455 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
456 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
457 APInt StrVal(NumBits, 0);
458 APInt SingleChar(NumBits, 0);
459 if (TD->isLittleEndian()) {
460 for (signed i = StrLen-1; i >= 0; i--) {
461 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
462 StrVal = (StrVal << 8) | SingleChar;
465 for (unsigned i = 0; i < StrLen; i++) {
466 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
467 StrVal = (StrVal << 8) | SingleChar;
469 // Append NULL at the end.
471 StrVal = (StrVal << 8) | SingleChar;
474 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
475 if (Ty->isFloatingPointTy())
476 Res = ConstantExpr::getBitCast(Res, Ty);
481 // If this load comes from anywhere in a constant global, and if the global
482 // is all undef or zero, we know what it loads.
483 if (GlobalVariable *GV =
484 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
485 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
486 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
487 if (GV->getInitializer()->isNullValue())
488 return Constant::getNullValue(ResTy);
489 if (isa<UndefValue>(GV->getInitializer()))
490 return UndefValue::get(ResTy);
494 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
495 // currently don't do any of this for big endian systems. It can be
496 // generalized in the future if someone is interested.
497 if (TD && TD->isLittleEndian())
498 return FoldReinterpretLoadFromConstPtr(CE, *TD);
502 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
503 if (LI->isVolatile()) return 0;
505 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
506 return ConstantFoldLoadFromConstPtr(C, TD);
511 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
512 /// Attempt to symbolically evaluate the result of a binary operator merging
513 /// these together. If target data info is available, it is provided as TD,
514 /// otherwise TD is null.
515 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
516 Constant *Op1, const TargetData *TD){
519 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
520 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
524 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
525 // constant. This happens frequently when iterating over a global array.
526 if (Opc == Instruction::Sub && TD) {
527 GlobalValue *GV1, *GV2;
528 int64_t Offs1, Offs2;
530 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
531 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
533 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
534 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
541 /// CastGEPIndices - If array indices are not pointer-sized integers,
542 /// explicitly cast them so that they aren't implicitly casted by the
544 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
546 const TargetData *TD) {
548 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
551 SmallVector<Constant*, 32> NewIdxs;
552 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
554 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
555 Ops.slice(1, i-1)))) &&
556 Ops[i]->getType() != IntPtrTy) {
558 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
564 NewIdxs.push_back(Ops[i]);
569 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
570 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
571 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
576 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
577 /// constant expression, do so.
578 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
580 const TargetData *TD) {
581 Constant *Ptr = Ops[0];
582 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
585 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
587 // If this is a constant expr gep that is effectively computing an
588 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
589 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
590 if (!isa<ConstantInt>(Ops[i])) {
592 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
593 // "inttoptr (sub (ptrtoint Ptr), V)"
594 if (Ops.size() == 2 &&
595 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
596 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
597 assert((CE == 0 || CE->getType() == IntPtrTy) &&
598 "CastGEPIndices didn't canonicalize index types!");
599 if (CE && CE->getOpcode() == Instruction::Sub &&
600 CE->getOperand(0)->isNullValue()) {
601 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
602 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
603 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
604 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
605 Res = ConstantFoldConstantExpression(ResCE, TD);
612 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
614 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
615 makeArrayRef((Value **)Ops.data() + 1,
617 Ptr = cast<Constant>(Ptr->stripPointerCasts());
619 // If this is a GEP of a GEP, fold it all into a single GEP.
620 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
621 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
623 // Do not try the incorporate the sub-GEP if some index is not a number.
624 bool AllConstantInt = true;
625 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
626 if (!isa<ConstantInt>(NestedOps[i])) {
627 AllConstantInt = false;
633 Ptr = cast<Constant>(GEP->getOperand(0));
634 Offset += APInt(BitWidth,
635 TD->getIndexedOffset(Ptr->getType(), NestedOps));
636 Ptr = cast<Constant>(Ptr->stripPointerCasts());
639 // If the base value for this address is a literal integer value, fold the
640 // getelementptr to the resulting integer value casted to the pointer type.
641 APInt BasePtr(BitWidth, 0);
642 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
643 if (CE->getOpcode() == Instruction::IntToPtr)
644 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
645 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
646 if (Ptr->isNullValue() || BasePtr != 0) {
647 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
648 return ConstantExpr::getIntToPtr(C, ResultTy);
651 // Otherwise form a regular getelementptr. Recompute the indices so that
652 // we eliminate over-indexing of the notional static type array bounds.
653 // This makes it easy to determine if the getelementptr is "inbounds".
654 // Also, this helps GlobalOpt do SROA on GlobalVariables.
655 Type *Ty = Ptr->getType();
656 SmallVector<Constant*, 32> NewIdxs;
658 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
659 if (ATy->isPointerTy()) {
660 // The only pointer indexing we'll do is on the first index of the GEP.
661 if (!NewIdxs.empty())
664 // Only handle pointers to sized types, not pointers to functions.
665 if (!ATy->getElementType()->isSized())
669 // Determine which element of the array the offset points into.
670 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
671 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
673 // The element size is 0. This may be [0 x Ty]*, so just use a zero
674 // index for this level and proceed to the next level to see if it can
675 // accommodate the offset.
676 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
678 // The element size is non-zero divide the offset by the element
679 // size (rounding down), to compute the index at this level.
680 APInt NewIdx = Offset.udiv(ElemSize);
681 Offset -= NewIdx * ElemSize;
682 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
684 Ty = ATy->getElementType();
685 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
686 // Determine which field of the struct the offset points into. The
687 // getZExtValue is at least as safe as the StructLayout API because we
688 // know the offset is within the struct at this point.
689 const StructLayout &SL = *TD->getStructLayout(STy);
690 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
691 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
693 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
694 Ty = STy->getTypeAtIndex(ElIdx);
696 // We've reached some non-indexable type.
699 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
701 // If we haven't used up the entire offset by descending the static
702 // type, then the offset is pointing into the middle of an indivisible
703 // member, so we can't simplify it.
709 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
710 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
711 "Computed GetElementPtr has unexpected type!");
713 // If we ended up indexing a member with a type that doesn't match
714 // the type of what the original indices indexed, add a cast.
715 if (Ty != cast<PointerType>(ResultTy)->getElementType())
716 C = FoldBitCast(C, ResultTy, *TD);
723 //===----------------------------------------------------------------------===//
724 // Constant Folding public APIs
725 //===----------------------------------------------------------------------===//
727 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
728 /// If successful, the constant result is returned, if not, null is returned.
729 /// Note that this fails if not all of the operands are constant. Otherwise,
730 /// this function can only fail when attempting to fold instructions like loads
731 /// and stores, which have no constant expression form.
732 Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
733 // Handle PHI nodes quickly here...
734 if (PHINode *PN = dyn_cast<PHINode>(I)) {
735 Constant *CommonValue = 0;
737 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
738 Value *Incoming = PN->getIncomingValue(i);
739 // If the incoming value is undef then skip it. Note that while we could
740 // skip the value if it is equal to the phi node itself we choose not to
741 // because that would break the rule that constant folding only applies if
742 // all operands are constants.
743 if (isa<UndefValue>(Incoming))
745 // If the incoming value is not a constant, or is a different constant to
746 // the one we saw previously, then give up.
747 Constant *C = dyn_cast<Constant>(Incoming);
748 if (!C || (CommonValue && C != CommonValue))
753 // If we reach here, all incoming values are the same constant or undef.
754 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
757 // Scan the operand list, checking to see if they are all constants, if so,
758 // hand off to ConstantFoldInstOperands.
759 SmallVector<Constant*, 8> Ops;
760 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
761 if (Constant *Op = dyn_cast<Constant>(*i))
764 return 0; // All operands not constant!
766 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
767 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
770 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
771 return ConstantFoldLoadInst(LI, TD);
773 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
774 return ConstantExpr::getInsertValue(
775 cast<Constant>(IVI->getAggregateOperand()),
776 cast<Constant>(IVI->getInsertedValueOperand()),
779 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
780 return ConstantExpr::getExtractValue(
781 cast<Constant>(EVI->getAggregateOperand()),
784 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD);
787 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
788 /// using the specified TargetData. If successful, the constant result is
789 /// result is returned, if not, null is returned.
790 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
791 const TargetData *TD) {
792 SmallVector<Constant*, 8> Ops;
793 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
795 Constant *NewC = cast<Constant>(*i);
796 // Recursively fold the ConstantExpr's operands.
797 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
798 NewC = ConstantFoldConstantExpression(NewCE, TD);
803 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
805 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD);
808 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
809 /// specified opcode and operands. If successful, the constant result is
810 /// returned, if not, null is returned. Note that this function can fail when
811 /// attempting to fold instructions like loads and stores, which have no
812 /// constant expression form.
814 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
815 /// information, due to only being passed an opcode and operands. Constant
816 /// folding using this function strips this information.
818 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
819 ArrayRef<Constant *> Ops,
820 const TargetData *TD) {
821 // Handle easy binops first.
822 if (Instruction::isBinaryOp(Opcode)) {
823 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
824 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
827 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
832 case Instruction::ICmp:
833 case Instruction::FCmp: assert(0 && "Invalid for compares");
834 case Instruction::Call:
835 if (Function *F = dyn_cast<Function>(Ops.back()))
836 if (canConstantFoldCallTo(F))
837 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1));
839 case Instruction::PtrToInt:
840 // If the input is a inttoptr, eliminate the pair. This requires knowing
841 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
842 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
843 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
844 Constant *Input = CE->getOperand(0);
845 unsigned InWidth = Input->getType()->getScalarSizeInBits();
846 if (TD->getPointerSizeInBits() < InWidth) {
848 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
849 TD->getPointerSizeInBits()));
850 Input = ConstantExpr::getAnd(Input, Mask);
852 // Do a zext or trunc to get to the dest size.
853 return ConstantExpr::getIntegerCast(Input, DestTy, false);
856 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
857 case Instruction::IntToPtr:
858 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
859 // the int size is >= the ptr size. This requires knowing the width of a
860 // pointer, so it can't be done in ConstantExpr::getCast.
861 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
863 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
864 CE->getOpcode() == Instruction::PtrToInt)
865 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
867 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
868 case Instruction::Trunc:
869 case Instruction::ZExt:
870 case Instruction::SExt:
871 case Instruction::FPTrunc:
872 case Instruction::FPExt:
873 case Instruction::UIToFP:
874 case Instruction::SIToFP:
875 case Instruction::FPToUI:
876 case Instruction::FPToSI:
877 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
878 case Instruction::BitCast:
880 return FoldBitCast(Ops[0], DestTy, *TD);
881 return ConstantExpr::getBitCast(Ops[0], DestTy);
882 case Instruction::Select:
883 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
884 case Instruction::ExtractElement:
885 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
886 case Instruction::InsertElement:
887 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
888 case Instruction::ShuffleVector:
889 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
890 case Instruction::GetElementPtr:
891 if (Constant *C = CastGEPIndices(Ops, DestTy, TD))
893 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD))
896 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
900 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
901 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
902 /// returns a constant expression of the specified operands.
904 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
905 Constant *Ops0, Constant *Ops1,
906 const TargetData *TD) {
907 // fold: icmp (inttoptr x), null -> icmp x, 0
908 // fold: icmp (ptrtoint x), 0 -> icmp x, null
909 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
910 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
912 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
913 // around to know if bit truncation is happening.
914 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
915 if (TD && Ops1->isNullValue()) {
916 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
917 if (CE0->getOpcode() == Instruction::IntToPtr) {
918 // Convert the integer value to the right size to ensure we get the
919 // proper extension or truncation.
920 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
922 Constant *Null = Constant::getNullValue(C->getType());
923 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
926 // Only do this transformation if the int is intptrty in size, otherwise
927 // there is a truncation or extension that we aren't modeling.
928 if (CE0->getOpcode() == Instruction::PtrToInt &&
929 CE0->getType() == IntPtrTy) {
930 Constant *C = CE0->getOperand(0);
931 Constant *Null = Constant::getNullValue(C->getType());
932 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
936 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
937 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
938 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
940 if (CE0->getOpcode() == Instruction::IntToPtr) {
941 // Convert the integer value to the right size to ensure we get the
942 // proper extension or truncation.
943 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
945 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
947 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
950 // Only do this transformation if the int is intptrty in size, otherwise
951 // there is a truncation or extension that we aren't modeling.
952 if ((CE0->getOpcode() == Instruction::PtrToInt &&
953 CE0->getType() == IntPtrTy &&
954 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
955 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
956 CE1->getOperand(0), TD);
960 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
961 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
962 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
963 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
965 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
967 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
969 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
970 Constant *Ops[] = { LHS, RHS };
971 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD);
975 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
979 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
980 /// getelementptr constantexpr, return the constant value being addressed by the
981 /// constant expression, or null if something is funny and we can't decide.
982 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
984 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
985 return 0; // Do not allow stepping over the value!
987 // Loop over all of the operands, tracking down which value we are
989 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
990 for (++I; I != E; ++I)
991 if (StructType *STy = dyn_cast<StructType>(*I)) {
992 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
993 assert(CU->getZExtValue() < STy->getNumElements() &&
994 "Struct index out of range!");
995 unsigned El = (unsigned)CU->getZExtValue();
996 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
997 C = CS->getOperand(El);
998 } else if (isa<ConstantAggregateZero>(C)) {
999 C = Constant::getNullValue(STy->getElementType(El));
1000 } else if (isa<UndefValue>(C)) {
1001 C = UndefValue::get(STy->getElementType(El));
1005 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
1006 if (ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
1007 if (CI->getZExtValue() >= ATy->getNumElements())
1009 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
1010 C = CA->getOperand(CI->getZExtValue());
1011 else if (isa<ConstantAggregateZero>(C))
1012 C = Constant::getNullValue(ATy->getElementType());
1013 else if (isa<UndefValue>(C))
1014 C = UndefValue::get(ATy->getElementType());
1017 } else if (VectorType *VTy = dyn_cast<VectorType>(*I)) {
1018 if (CI->getZExtValue() >= VTy->getNumElements())
1020 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
1021 C = CP->getOperand(CI->getZExtValue());
1022 else if (isa<ConstantAggregateZero>(C))
1023 C = Constant::getNullValue(VTy->getElementType());
1024 else if (isa<UndefValue>(C))
1025 C = UndefValue::get(VTy->getElementType());
1038 //===----------------------------------------------------------------------===//
1039 // Constant Folding for Calls
1042 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1043 /// the specified function.
1045 llvm::canConstantFoldCallTo(const Function *F) {
1046 switch (F->getIntrinsicID()) {
1047 case Intrinsic::sqrt:
1048 case Intrinsic::powi:
1049 case Intrinsic::bswap:
1050 case Intrinsic::ctpop:
1051 case Intrinsic::ctlz:
1052 case Intrinsic::cttz:
1053 case Intrinsic::sadd_with_overflow:
1054 case Intrinsic::uadd_with_overflow:
1055 case Intrinsic::ssub_with_overflow:
1056 case Intrinsic::usub_with_overflow:
1057 case Intrinsic::smul_with_overflow:
1058 case Intrinsic::umul_with_overflow:
1059 case Intrinsic::convert_from_fp16:
1060 case Intrinsic::convert_to_fp16:
1061 case Intrinsic::x86_sse_cvtss2si:
1062 case Intrinsic::x86_sse_cvtss2si64:
1063 case Intrinsic::x86_sse_cvttss2si:
1064 case Intrinsic::x86_sse_cvttss2si64:
1065 case Intrinsic::x86_sse2_cvtsd2si:
1066 case Intrinsic::x86_sse2_cvtsd2si64:
1067 case Intrinsic::x86_sse2_cvttsd2si:
1068 case Intrinsic::x86_sse2_cvttsd2si64:
1075 if (!F->hasName()) return false;
1076 StringRef Name = F->getName();
1078 // In these cases, the check of the length is required. We don't want to
1079 // return true for a name like "cos\0blah" which strcmp would return equal to
1080 // "cos", but has length 8.
1082 default: return false;
1084 return Name == "acos" || Name == "asin" ||
1085 Name == "atan" || Name == "atan2";
1087 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1089 return Name == "exp" || Name == "exp2";
1091 return Name == "fabs" || Name == "fmod" || Name == "floor";
1093 return Name == "log" || Name == "log10";
1095 return Name == "pow";
1097 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1098 Name == "sinf" || Name == "sqrtf";
1100 return Name == "tan" || Name == "tanh";
1104 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1106 sys::llvm_fenv_clearexcept();
1108 if (sys::llvm_fenv_testexcept()) {
1109 sys::llvm_fenv_clearexcept();
1113 if (Ty->isFloatTy())
1114 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1115 if (Ty->isDoubleTy())
1116 return ConstantFP::get(Ty->getContext(), APFloat(V));
1117 llvm_unreachable("Can only constant fold float/double");
1118 return 0; // dummy return to suppress warning
1121 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1122 double V, double W, Type *Ty) {
1123 sys::llvm_fenv_clearexcept();
1125 if (sys::llvm_fenv_testexcept()) {
1126 sys::llvm_fenv_clearexcept();
1130 if (Ty->isFloatTy())
1131 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1132 if (Ty->isDoubleTy())
1133 return ConstantFP::get(Ty->getContext(), APFloat(V));
1134 llvm_unreachable("Can only constant fold float/double");
1135 return 0; // dummy return to suppress warning
1138 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1139 /// conversion of a constant floating point. If roundTowardZero is false, the
1140 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1141 /// the behavior of the non-truncating SSE instructions in the default rounding
1142 /// mode. The desired integer type Ty is used to select how many bits are
1143 /// available for the result. Returns null if the conversion cannot be
1144 /// performed, otherwise returns the Constant value resulting from the
1146 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1148 assert(Op && "Called with NULL operand");
1149 APFloat Val(Op->getValueAPF());
1151 // All of these conversion intrinsics form an integer of at most 64bits.
1152 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1153 assert(ResultWidth <= 64 &&
1154 "Can only constant fold conversions to 64 and 32 bit ints");
1157 bool isExact = false;
1158 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1159 : APFloat::rmNearestTiesToEven;
1160 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1161 /*isSigned=*/true, mode,
1163 if (status != APFloat::opOK && status != APFloat::opInexact)
1165 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1168 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1169 /// with the specified arguments, returning null if unsuccessful.
1171 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands) {
1172 if (!F->hasName()) return 0;
1173 StringRef Name = F->getName();
1175 Type *Ty = F->getReturnType();
1176 if (Operands.size() == 1) {
1177 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1178 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1179 APFloat Val(Op->getValueAPF());
1182 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1184 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1187 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1190 /// We only fold functions with finite arguments. Folding NaN and inf is
1191 /// likely to be aborted with an exception anyway, and some host libms
1192 /// have known errors raising exceptions.
1193 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1196 /// Currently APFloat versions of these functions do not exist, so we use
1197 /// the host native double versions. Float versions are not called
1198 /// directly but for all these it is true (float)(f((double)arg)) ==
1199 /// f(arg). Long double not supported yet.
1200 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1201 Op->getValueAPF().convertToDouble();
1205 return ConstantFoldFP(acos, V, Ty);
1206 else if (Name == "asin")
1207 return ConstantFoldFP(asin, V, Ty);
1208 else if (Name == "atan")
1209 return ConstantFoldFP(atan, V, Ty);
1213 return ConstantFoldFP(ceil, V, Ty);
1214 else if (Name == "cos")
1215 return ConstantFoldFP(cos, V, Ty);
1216 else if (Name == "cosh")
1217 return ConstantFoldFP(cosh, V, Ty);
1218 else if (Name == "cosf")
1219 return ConstantFoldFP(cos, V, Ty);
1223 return ConstantFoldFP(exp, V, Ty);
1225 if (Name == "exp2") {
1226 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1228 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1233 return ConstantFoldFP(fabs, V, Ty);
1234 else if (Name == "floor")
1235 return ConstantFoldFP(floor, V, Ty);
1238 if (Name == "log" && V > 0)
1239 return ConstantFoldFP(log, V, Ty);
1240 else if (Name == "log10" && V > 0)
1241 return ConstantFoldFP(log10, V, Ty);
1242 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1243 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1245 return ConstantFoldFP(sqrt, V, Ty);
1247 return Constant::getNullValue(Ty);
1252 return ConstantFoldFP(sin, V, Ty);
1253 else if (Name == "sinh")
1254 return ConstantFoldFP(sinh, V, Ty);
1255 else if (Name == "sqrt" && V >= 0)
1256 return ConstantFoldFP(sqrt, V, Ty);
1257 else if (Name == "sqrtf" && V >= 0)
1258 return ConstantFoldFP(sqrt, V, Ty);
1259 else if (Name == "sinf")
1260 return ConstantFoldFP(sin, V, Ty);
1264 return ConstantFoldFP(tan, V, Ty);
1265 else if (Name == "tanh")
1266 return ConstantFoldFP(tanh, V, Ty);
1274 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1275 switch (F->getIntrinsicID()) {
1276 case Intrinsic::bswap:
1277 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1278 case Intrinsic::ctpop:
1279 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1280 case Intrinsic::cttz:
1281 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1282 case Intrinsic::ctlz:
1283 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1284 case Intrinsic::convert_from_fp16: {
1285 APFloat Val(Op->getValue());
1288 APFloat::opStatus status =
1289 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1291 // Conversion is always precise.
1293 assert(status == APFloat::opOK && !lost &&
1294 "Precision lost during fp16 constfolding");
1296 return ConstantFP::get(F->getContext(), Val);
1303 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1304 switch (F->getIntrinsicID()) {
1306 case Intrinsic::x86_sse_cvtss2si:
1307 case Intrinsic::x86_sse_cvtss2si64:
1308 case Intrinsic::x86_sse2_cvtsd2si:
1309 case Intrinsic::x86_sse2_cvtsd2si64:
1310 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1311 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1312 case Intrinsic::x86_sse_cvttss2si:
1313 case Intrinsic::x86_sse_cvttss2si64:
1314 case Intrinsic::x86_sse2_cvttsd2si:
1315 case Intrinsic::x86_sse2_cvttsd2si64:
1316 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1317 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1321 if (isa<UndefValue>(Operands[0])) {
1322 if (F->getIntrinsicID() == Intrinsic::bswap)
1330 if (Operands.size() == 2) {
1331 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1332 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1334 double Op1V = Ty->isFloatTy() ?
1335 (double)Op1->getValueAPF().convertToFloat() :
1336 Op1->getValueAPF().convertToDouble();
1337 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1338 if (Op2->getType() != Op1->getType())
1341 double Op2V = Ty->isFloatTy() ?
1342 (double)Op2->getValueAPF().convertToFloat():
1343 Op2->getValueAPF().convertToDouble();
1346 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1348 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1349 if (Name == "atan2")
1350 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1351 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1352 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1353 return ConstantFP::get(F->getContext(),
1354 APFloat((float)std::pow((float)Op1V,
1355 (int)Op2C->getZExtValue())));
1356 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1357 return ConstantFP::get(F->getContext(),
1358 APFloat((double)std::pow((double)Op1V,
1359 (int)Op2C->getZExtValue())));
1365 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1366 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1367 switch (F->getIntrinsicID()) {
1369 case Intrinsic::sadd_with_overflow:
1370 case Intrinsic::uadd_with_overflow:
1371 case Intrinsic::ssub_with_overflow:
1372 case Intrinsic::usub_with_overflow:
1373 case Intrinsic::smul_with_overflow:
1374 case Intrinsic::umul_with_overflow: {
1377 switch (F->getIntrinsicID()) {
1378 default: assert(0 && "Invalid case");
1379 case Intrinsic::sadd_with_overflow:
1380 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1382 case Intrinsic::uadd_with_overflow:
1383 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1385 case Intrinsic::ssub_with_overflow:
1386 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1388 case Intrinsic::usub_with_overflow:
1389 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1391 case Intrinsic::smul_with_overflow:
1392 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1394 case Intrinsic::umul_with_overflow:
1395 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1399 ConstantInt::get(F->getContext(), Res),
1400 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1402 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);