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) {
49 ConstantVector *CV = dyn_cast<ConstantVector>(C);
50 IntegerType *IntVTy = dyn_cast<IntegerType>(DestTy);
51 // When casting vectors to scalar integers, catch the
52 // obvious splat cases.
54 if (CV->isNullValue()) return ConstantInt::getNullValue(IntVTy);
55 if (CV->isAllOnesValue()) return ConstantInt::getAllOnesValue(IntVTy);
58 // The code below only handles casts to vectors currently.
59 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
61 return ConstantExpr::getBitCast(C, DestTy);
63 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
64 // vector so the code below can handle it uniformly.
65 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
66 Constant *Ops = C; // don't take the address of C!
67 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
70 // If this is a bitcast from constant vector -> vector, fold it.
72 return ConstantExpr::getBitCast(C, DestTy);
74 // If the element types match, VMCore can fold it.
75 unsigned NumDstElt = DestVTy->getNumElements();
76 unsigned NumSrcElt = CV->getNumOperands();
77 if (NumDstElt == NumSrcElt)
78 return ConstantExpr::getBitCast(C, DestTy);
80 Type *SrcEltTy = CV->getType()->getElementType();
81 Type *DstEltTy = DestVTy->getElementType();
83 // Otherwise, we're changing the number of elements in a vector, which
84 // requires endianness information to do the right thing. For example,
85 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
86 // folds to (little endian):
87 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
88 // and to (big endian):
89 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
91 // First thing is first. We only want to think about integer here, so if
92 // we have something in FP form, recast it as integer.
93 if (DstEltTy->isFloatingPointTy()) {
94 // Fold to an vector of integers with same size as our FP type.
95 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
97 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
98 // Recursively handle this integer conversion, if possible.
99 C = FoldBitCast(C, DestIVTy, TD);
100 if (!C) return ConstantExpr::getBitCast(C, DestTy);
102 // Finally, VMCore can handle this now that #elts line up.
103 return ConstantExpr::getBitCast(C, DestTy);
106 // Okay, we know the destination is integer, if the input is FP, convert
107 // it to integer first.
108 if (SrcEltTy->isFloatingPointTy()) {
109 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
111 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
112 // Ask VMCore to do the conversion now that #elts line up.
113 C = ConstantExpr::getBitCast(C, SrcIVTy);
114 CV = dyn_cast<ConstantVector>(C);
115 if (!CV) // If VMCore wasn't able to fold it, bail out.
119 // Now we know that the input and output vectors are both integer vectors
120 // of the same size, and that their #elements is not the same. Do the
121 // conversion here, which depends on whether the input or output has
123 bool isLittleEndian = TD.isLittleEndian();
125 SmallVector<Constant*, 32> Result;
126 if (NumDstElt < NumSrcElt) {
127 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
128 Constant *Zero = Constant::getNullValue(DstEltTy);
129 unsigned Ratio = NumSrcElt/NumDstElt;
130 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
132 for (unsigned i = 0; i != NumDstElt; ++i) {
133 // Build each element of the result.
134 Constant *Elt = Zero;
135 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
136 for (unsigned j = 0; j != Ratio; ++j) {
137 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
138 if (!Src) // Reject constantexpr elements.
139 return ConstantExpr::getBitCast(C, DestTy);
141 // Zero extend the element to the right size.
142 Src = ConstantExpr::getZExt(Src, Elt->getType());
144 // Shift it to the right place, depending on endianness.
145 Src = ConstantExpr::getShl(Src,
146 ConstantInt::get(Src->getType(), ShiftAmt));
147 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
150 Elt = ConstantExpr::getOr(Elt, Src);
152 Result.push_back(Elt);
155 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
156 unsigned Ratio = NumDstElt/NumSrcElt;
157 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
159 // Loop over each source value, expanding into multiple results.
160 for (unsigned i = 0; i != NumSrcElt; ++i) {
161 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
162 if (!Src) // Reject constantexpr elements.
163 return ConstantExpr::getBitCast(C, DestTy);
165 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
166 for (unsigned j = 0; j != Ratio; ++j) {
167 // Shift the piece of the value into the right place, depending on
169 Constant *Elt = ConstantExpr::getLShr(Src,
170 ConstantInt::get(Src->getType(), ShiftAmt));
171 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
173 // Truncate and remember this piece.
174 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
179 return ConstantVector::get(Result);
183 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
184 /// from a global, return the global and the constant. Because of
185 /// constantexprs, this function is recursive.
186 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
187 int64_t &Offset, const TargetData &TD) {
188 // Trivial case, constant is the global.
189 if ((GV = dyn_cast<GlobalValue>(C))) {
194 // Otherwise, if this isn't a constant expr, bail out.
195 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
196 if (!CE) return false;
198 // Look through ptr->int and ptr->ptr casts.
199 if (CE->getOpcode() == Instruction::PtrToInt ||
200 CE->getOpcode() == Instruction::BitCast)
201 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
203 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
204 if (CE->getOpcode() == Instruction::GetElementPtr) {
205 // Cannot compute this if the element type of the pointer is missing size
207 if (!cast<PointerType>(CE->getOperand(0)->getType())
208 ->getElementType()->isSized())
211 // If the base isn't a global+constant, we aren't either.
212 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
215 // Otherwise, add any offset that our operands provide.
216 gep_type_iterator GTI = gep_type_begin(CE);
217 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
218 i != e; ++i, ++GTI) {
219 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
220 if (!CI) return false; // Index isn't a simple constant?
221 if (CI->isZero()) continue; // Not adding anything.
223 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
225 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
227 SequentialType *SQT = cast<SequentialType>(*GTI);
228 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
237 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
238 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
239 /// pointer to copy results into and BytesLeft is the number of bytes left in
240 /// the CurPtr buffer. TD is the target data.
241 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
242 unsigned char *CurPtr, unsigned BytesLeft,
243 const TargetData &TD) {
244 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
245 "Out of range access");
247 // If this element is zero or undefined, we can just return since *CurPtr is
249 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
252 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
253 if (CI->getBitWidth() > 64 ||
254 (CI->getBitWidth() & 7) != 0)
257 uint64_t Val = CI->getZExtValue();
258 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
260 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
261 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
267 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
268 if (CFP->getType()->isDoubleTy()) {
269 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
270 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
272 if (CFP->getType()->isFloatTy()){
273 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
274 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
279 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
280 const StructLayout *SL = TD.getStructLayout(CS->getType());
281 unsigned Index = SL->getElementContainingOffset(ByteOffset);
282 uint64_t CurEltOffset = SL->getElementOffset(Index);
283 ByteOffset -= CurEltOffset;
286 // If the element access is to the element itself and not to tail padding,
287 // read the bytes from the element.
288 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
290 if (ByteOffset < EltSize &&
291 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
297 // Check to see if we read from the last struct element, if so we're done.
298 if (Index == CS->getType()->getNumElements())
301 // If we read all of the bytes we needed from this element we're done.
302 uint64_t NextEltOffset = SL->getElementOffset(Index);
304 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
307 // Move to the next element of the struct.
308 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
309 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
311 CurEltOffset = NextEltOffset;
316 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
317 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
318 uint64_t Index = ByteOffset / EltSize;
319 uint64_t Offset = ByteOffset - Index * EltSize;
320 for (; Index != CA->getType()->getNumElements(); ++Index) {
321 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
324 if (EltSize >= BytesLeft)
328 BytesLeft -= EltSize;
334 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
335 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
336 uint64_t Index = ByteOffset / EltSize;
337 uint64_t Offset = ByteOffset - Index * EltSize;
338 for (; Index != CV->getType()->getNumElements(); ++Index) {
339 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
342 if (EltSize >= BytesLeft)
346 BytesLeft -= EltSize;
352 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
353 if (CE->getOpcode() == Instruction::IntToPtr &&
354 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
355 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
359 // Otherwise, unknown initializer type.
363 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
364 const TargetData &TD) {
365 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
366 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
368 // If this isn't an integer load we can't fold it directly.
370 // If this is a float/double load, we can try folding it as an int32/64 load
371 // and then bitcast the result. This can be useful for union cases. Note
372 // that address spaces don't matter here since we're not going to result in
373 // an actual new load.
375 if (LoadTy->isFloatTy())
376 MapTy = Type::getInt32PtrTy(C->getContext());
377 else if (LoadTy->isDoubleTy())
378 MapTy = Type::getInt64PtrTy(C->getContext());
379 else if (LoadTy->isVectorTy()) {
380 MapTy = IntegerType::get(C->getContext(),
381 TD.getTypeAllocSizeInBits(LoadTy));
382 MapTy = PointerType::getUnqual(MapTy);
386 C = FoldBitCast(C, MapTy, TD);
387 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
388 return FoldBitCast(Res, LoadTy, TD);
392 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
393 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
397 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
400 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
401 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
402 !GV->getInitializer()->getType()->isSized())
405 // If we're loading off the beginning of the global, some bytes may be valid,
406 // but we don't try to handle this.
407 if (Offset < 0) return 0;
409 // If we're not accessing anything in this constant, the result is undefined.
410 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
411 return UndefValue::get(IntType);
413 unsigned char RawBytes[32] = {0};
414 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
418 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
419 for (unsigned i = 1; i != BytesLoaded; ++i) {
421 ResultVal |= RawBytes[BytesLoaded-1-i];
424 return ConstantInt::get(IntType->getContext(), ResultVal);
427 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
428 /// produce if it is constant and determinable. If this is not determinable,
430 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
431 const TargetData *TD) {
432 // First, try the easy cases:
433 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
434 if (GV->isConstant() && GV->hasDefinitiveInitializer())
435 return GV->getInitializer();
437 // If the loaded value isn't a constant expr, we can't handle it.
438 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
441 if (CE->getOpcode() == Instruction::GetElementPtr) {
442 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
443 if (GV->isConstant() && GV->hasDefinitiveInitializer())
445 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
449 // Instead of loading constant c string, use corresponding integer value
450 // directly if string length is small enough.
452 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
453 unsigned StrLen = Str.length();
454 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
455 unsigned NumBits = Ty->getPrimitiveSizeInBits();
456 // Replace load with immediate integer if the result is an integer or fp
458 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
459 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
460 APInt StrVal(NumBits, 0);
461 APInt SingleChar(NumBits, 0);
462 if (TD->isLittleEndian()) {
463 for (signed i = StrLen-1; i >= 0; i--) {
464 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
465 StrVal = (StrVal << 8) | SingleChar;
468 for (unsigned i = 0; i < StrLen; i++) {
469 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
470 StrVal = (StrVal << 8) | SingleChar;
472 // Append NULL at the end.
474 StrVal = (StrVal << 8) | SingleChar;
477 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
478 if (Ty->isFloatingPointTy())
479 Res = ConstantExpr::getBitCast(Res, Ty);
484 // If this load comes from anywhere in a constant global, and if the global
485 // is all undef or zero, we know what it loads.
486 if (GlobalVariable *GV =
487 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
488 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
489 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
490 if (GV->getInitializer()->isNullValue())
491 return Constant::getNullValue(ResTy);
492 if (isa<UndefValue>(GV->getInitializer()))
493 return UndefValue::get(ResTy);
497 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
498 // currently don't do any of this for big endian systems. It can be
499 // generalized in the future if someone is interested.
500 if (TD && TD->isLittleEndian())
501 return FoldReinterpretLoadFromConstPtr(CE, *TD);
505 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
506 if (LI->isVolatile()) return 0;
508 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
509 return ConstantFoldLoadFromConstPtr(C, TD);
514 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
515 /// Attempt to symbolically evaluate the result of a binary operator merging
516 /// these together. If target data info is available, it is provided as TD,
517 /// otherwise TD is null.
518 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
519 Constant *Op1, const TargetData *TD){
522 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
523 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
527 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
528 // constant. This happens frequently when iterating over a global array.
529 if (Opc == Instruction::Sub && TD) {
530 GlobalValue *GV1, *GV2;
531 int64_t Offs1, Offs2;
533 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
534 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
536 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
537 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
544 /// CastGEPIndices - If array indices are not pointer-sized integers,
545 /// explicitly cast them so that they aren't implicitly casted by the
547 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
549 const TargetData *TD) {
551 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
554 SmallVector<Constant*, 32> NewIdxs;
555 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
557 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
558 Ops.slice(1, i-1)))) &&
559 Ops[i]->getType() != IntPtrTy) {
561 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
567 NewIdxs.push_back(Ops[i]);
572 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
573 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
574 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
579 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
580 /// constant expression, do so.
581 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
583 const TargetData *TD) {
584 Constant *Ptr = Ops[0];
585 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
588 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
590 // If this is a constant expr gep that is effectively computing an
591 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
592 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
593 if (!isa<ConstantInt>(Ops[i])) {
595 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
596 // "inttoptr (sub (ptrtoint Ptr), V)"
597 if (Ops.size() == 2 &&
598 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
599 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
600 assert((CE == 0 || CE->getType() == IntPtrTy) &&
601 "CastGEPIndices didn't canonicalize index types!");
602 if (CE && CE->getOpcode() == Instruction::Sub &&
603 CE->getOperand(0)->isNullValue()) {
604 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
605 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
606 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
607 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
608 Res = ConstantFoldConstantExpression(ResCE, TD);
615 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
617 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
618 makeArrayRef((Value **)Ops.data() + 1,
620 Ptr = cast<Constant>(Ptr->stripPointerCasts());
622 // If this is a GEP of a GEP, fold it all into a single GEP.
623 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
624 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
626 // Do not try the incorporate the sub-GEP if some index is not a number.
627 bool AllConstantInt = true;
628 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
629 if (!isa<ConstantInt>(NestedOps[i])) {
630 AllConstantInt = false;
636 Ptr = cast<Constant>(GEP->getOperand(0));
637 Offset += APInt(BitWidth,
638 TD->getIndexedOffset(Ptr->getType(), NestedOps));
639 Ptr = cast<Constant>(Ptr->stripPointerCasts());
642 // If the base value for this address is a literal integer value, fold the
643 // getelementptr to the resulting integer value casted to the pointer type.
644 APInt BasePtr(BitWidth, 0);
645 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
646 if (CE->getOpcode() == Instruction::IntToPtr)
647 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
648 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
649 if (Ptr->isNullValue() || BasePtr != 0) {
650 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
651 return ConstantExpr::getIntToPtr(C, ResultTy);
654 // Otherwise form a regular getelementptr. Recompute the indices so that
655 // we eliminate over-indexing of the notional static type array bounds.
656 // This makes it easy to determine if the getelementptr is "inbounds".
657 // Also, this helps GlobalOpt do SROA on GlobalVariables.
658 Type *Ty = Ptr->getType();
659 SmallVector<Constant*, 32> NewIdxs;
661 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
662 if (ATy->isPointerTy()) {
663 // The only pointer indexing we'll do is on the first index of the GEP.
664 if (!NewIdxs.empty())
667 // Only handle pointers to sized types, not pointers to functions.
668 if (!ATy->getElementType()->isSized())
672 // Determine which element of the array the offset points into.
673 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
674 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
676 // The element size is 0. This may be [0 x Ty]*, so just use a zero
677 // index for this level and proceed to the next level to see if it can
678 // accommodate the offset.
679 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
681 // The element size is non-zero divide the offset by the element
682 // size (rounding down), to compute the index at this level.
683 APInt NewIdx = Offset.udiv(ElemSize);
684 Offset -= NewIdx * ElemSize;
685 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
687 Ty = ATy->getElementType();
688 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
689 // Determine which field of the struct the offset points into. The
690 // getZExtValue is at least as safe as the StructLayout API because we
691 // know the offset is within the struct at this point.
692 const StructLayout &SL = *TD->getStructLayout(STy);
693 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
694 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
696 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
697 Ty = STy->getTypeAtIndex(ElIdx);
699 // We've reached some non-indexable type.
702 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
704 // If we haven't used up the entire offset by descending the static
705 // type, then the offset is pointing into the middle of an indivisible
706 // member, so we can't simplify it.
712 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
713 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
714 "Computed GetElementPtr has unexpected type!");
716 // If we ended up indexing a member with a type that doesn't match
717 // the type of what the original indices indexed, add a cast.
718 if (Ty != cast<PointerType>(ResultTy)->getElementType())
719 C = FoldBitCast(C, ResultTy, *TD);
726 //===----------------------------------------------------------------------===//
727 // Constant Folding public APIs
728 //===----------------------------------------------------------------------===//
730 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
731 /// If successful, the constant result is returned, if not, null is returned.
732 /// Note that this fails if not all of the operands are constant. Otherwise,
733 /// this function can only fail when attempting to fold instructions like loads
734 /// and stores, which have no constant expression form.
735 Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
736 // Handle PHI nodes quickly here...
737 if (PHINode *PN = dyn_cast<PHINode>(I)) {
738 Constant *CommonValue = 0;
740 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
741 Value *Incoming = PN->getIncomingValue(i);
742 // If the incoming value is undef then skip it. Note that while we could
743 // skip the value if it is equal to the phi node itself we choose not to
744 // because that would break the rule that constant folding only applies if
745 // all operands are constants.
746 if (isa<UndefValue>(Incoming))
748 // If the incoming value is not a constant, or is a different constant to
749 // the one we saw previously, then give up.
750 Constant *C = dyn_cast<Constant>(Incoming);
751 if (!C || (CommonValue && C != CommonValue))
756 // If we reach here, all incoming values are the same constant or undef.
757 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
760 // Scan the operand list, checking to see if they are all constants, if so,
761 // hand off to ConstantFoldInstOperands.
762 SmallVector<Constant*, 8> Ops;
763 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
764 if (Constant *Op = dyn_cast<Constant>(*i))
767 return 0; // All operands not constant!
769 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
770 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
773 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
774 return ConstantFoldLoadInst(LI, TD);
776 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
777 return ConstantExpr::getInsertValue(
778 cast<Constant>(IVI->getAggregateOperand()),
779 cast<Constant>(IVI->getInsertedValueOperand()),
782 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
783 return ConstantExpr::getExtractValue(
784 cast<Constant>(EVI->getAggregateOperand()),
787 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD);
790 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
791 /// using the specified TargetData. If successful, the constant result is
792 /// result is returned, if not, null is returned.
793 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
794 const TargetData *TD) {
795 SmallVector<Constant*, 8> Ops;
796 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
798 Constant *NewC = cast<Constant>(*i);
799 // Recursively fold the ConstantExpr's operands.
800 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
801 NewC = ConstantFoldConstantExpression(NewCE, TD);
806 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
808 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD);
811 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
812 /// specified opcode and operands. If successful, the constant result is
813 /// returned, if not, null is returned. Note that this function can fail when
814 /// attempting to fold instructions like loads and stores, which have no
815 /// constant expression form.
817 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
818 /// information, due to only being passed an opcode and operands. Constant
819 /// folding using this function strips this information.
821 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
822 ArrayRef<Constant *> Ops,
823 const TargetData *TD) {
824 // Handle easy binops first.
825 if (Instruction::isBinaryOp(Opcode)) {
826 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
827 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
830 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
835 case Instruction::ICmp:
836 case Instruction::FCmp: assert(0 && "Invalid for compares");
837 case Instruction::Call:
838 if (Function *F = dyn_cast<Function>(Ops.back()))
839 if (canConstantFoldCallTo(F))
840 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1));
842 case Instruction::PtrToInt:
843 // If the input is a inttoptr, eliminate the pair. This requires knowing
844 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
845 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
846 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
847 Constant *Input = CE->getOperand(0);
848 unsigned InWidth = Input->getType()->getScalarSizeInBits();
849 if (TD->getPointerSizeInBits() < InWidth) {
851 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
852 TD->getPointerSizeInBits()));
853 Input = ConstantExpr::getAnd(Input, Mask);
855 // Do a zext or trunc to get to the dest size.
856 return ConstantExpr::getIntegerCast(Input, DestTy, false);
859 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
860 case Instruction::IntToPtr:
861 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
862 // the int size is >= the ptr size. This requires knowing the width of a
863 // pointer, so it can't be done in ConstantExpr::getCast.
864 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
866 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
867 CE->getOpcode() == Instruction::PtrToInt)
868 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
870 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
871 case Instruction::Trunc:
872 case Instruction::ZExt:
873 case Instruction::SExt:
874 case Instruction::FPTrunc:
875 case Instruction::FPExt:
876 case Instruction::UIToFP:
877 case Instruction::SIToFP:
878 case Instruction::FPToUI:
879 case Instruction::FPToSI:
880 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
881 case Instruction::BitCast:
883 return FoldBitCast(Ops[0], DestTy, *TD);
884 return ConstantExpr::getBitCast(Ops[0], DestTy);
885 case Instruction::Select:
886 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
887 case Instruction::ExtractElement:
888 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
889 case Instruction::InsertElement:
890 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
891 case Instruction::ShuffleVector:
892 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
893 case Instruction::GetElementPtr:
894 if (Constant *C = CastGEPIndices(Ops, DestTy, TD))
896 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD))
899 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
903 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
904 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
905 /// returns a constant expression of the specified operands.
907 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
908 Constant *Ops0, Constant *Ops1,
909 const TargetData *TD) {
910 // fold: icmp (inttoptr x), null -> icmp x, 0
911 // fold: icmp (ptrtoint x), 0 -> icmp x, null
912 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
913 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
915 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
916 // around to know if bit truncation is happening.
917 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
918 if (TD && Ops1->isNullValue()) {
919 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
920 if (CE0->getOpcode() == Instruction::IntToPtr) {
921 // Convert the integer value to the right size to ensure we get the
922 // proper extension or truncation.
923 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
925 Constant *Null = Constant::getNullValue(C->getType());
926 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
929 // Only do this transformation if the int is intptrty in size, otherwise
930 // there is a truncation or extension that we aren't modeling.
931 if (CE0->getOpcode() == Instruction::PtrToInt &&
932 CE0->getType() == IntPtrTy) {
933 Constant *C = CE0->getOperand(0);
934 Constant *Null = Constant::getNullValue(C->getType());
935 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
939 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
940 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
941 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
943 if (CE0->getOpcode() == Instruction::IntToPtr) {
944 // Convert the integer value to the right size to ensure we get the
945 // proper extension or truncation.
946 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
948 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
950 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
953 // Only do this transformation if the int is intptrty in size, otherwise
954 // there is a truncation or extension that we aren't modeling.
955 if ((CE0->getOpcode() == Instruction::PtrToInt &&
956 CE0->getType() == IntPtrTy &&
957 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
958 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
959 CE1->getOperand(0), TD);
963 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
964 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
965 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
966 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
968 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
970 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
972 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
973 Constant *Ops[] = { LHS, RHS };
974 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD);
978 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
982 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
983 /// getelementptr constantexpr, return the constant value being addressed by the
984 /// constant expression, or null if something is funny and we can't decide.
985 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
987 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
988 return 0; // Do not allow stepping over the value!
990 // Loop over all of the operands, tracking down which value we are
992 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
993 for (++I; I != E; ++I)
994 if (StructType *STy = dyn_cast<StructType>(*I)) {
995 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
996 assert(CU->getZExtValue() < STy->getNumElements() &&
997 "Struct index out of range!");
998 unsigned El = (unsigned)CU->getZExtValue();
999 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1000 C = CS->getOperand(El);
1001 } else if (isa<ConstantAggregateZero>(C)) {
1002 C = Constant::getNullValue(STy->getElementType(El));
1003 } else if (isa<UndefValue>(C)) {
1004 C = UndefValue::get(STy->getElementType(El));
1008 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
1009 if (ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
1010 if (CI->getZExtValue() >= ATy->getNumElements())
1012 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
1013 C = CA->getOperand(CI->getZExtValue());
1014 else if (isa<ConstantAggregateZero>(C))
1015 C = Constant::getNullValue(ATy->getElementType());
1016 else if (isa<UndefValue>(C))
1017 C = UndefValue::get(ATy->getElementType());
1020 } else if (VectorType *VTy = dyn_cast<VectorType>(*I)) {
1021 if (CI->getZExtValue() >= VTy->getNumElements())
1023 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
1024 C = CP->getOperand(CI->getZExtValue());
1025 else if (isa<ConstantAggregateZero>(C))
1026 C = Constant::getNullValue(VTy->getElementType());
1027 else if (isa<UndefValue>(C))
1028 C = UndefValue::get(VTy->getElementType());
1041 //===----------------------------------------------------------------------===//
1042 // Constant Folding for Calls
1045 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1046 /// the specified function.
1048 llvm::canConstantFoldCallTo(const Function *F) {
1049 switch (F->getIntrinsicID()) {
1050 case Intrinsic::sqrt:
1051 case Intrinsic::powi:
1052 case Intrinsic::bswap:
1053 case Intrinsic::ctpop:
1054 case Intrinsic::ctlz:
1055 case Intrinsic::cttz:
1056 case Intrinsic::sadd_with_overflow:
1057 case Intrinsic::uadd_with_overflow:
1058 case Intrinsic::ssub_with_overflow:
1059 case Intrinsic::usub_with_overflow:
1060 case Intrinsic::smul_with_overflow:
1061 case Intrinsic::umul_with_overflow:
1062 case Intrinsic::convert_from_fp16:
1063 case Intrinsic::convert_to_fp16:
1064 case Intrinsic::x86_sse_cvtss2si:
1065 case Intrinsic::x86_sse_cvtss2si64:
1066 case Intrinsic::x86_sse_cvttss2si:
1067 case Intrinsic::x86_sse_cvttss2si64:
1068 case Intrinsic::x86_sse2_cvtsd2si:
1069 case Intrinsic::x86_sse2_cvtsd2si64:
1070 case Intrinsic::x86_sse2_cvttsd2si:
1071 case Intrinsic::x86_sse2_cvttsd2si64:
1078 if (!F->hasName()) return false;
1079 StringRef Name = F->getName();
1081 // In these cases, the check of the length is required. We don't want to
1082 // return true for a name like "cos\0blah" which strcmp would return equal to
1083 // "cos", but has length 8.
1085 default: return false;
1087 return Name == "acos" || Name == "asin" ||
1088 Name == "atan" || Name == "atan2";
1090 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1092 return Name == "exp" || Name == "exp2";
1094 return Name == "fabs" || Name == "fmod" || Name == "floor";
1096 return Name == "log" || Name == "log10";
1098 return Name == "pow";
1100 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1101 Name == "sinf" || Name == "sqrtf";
1103 return Name == "tan" || Name == "tanh";
1107 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1109 sys::llvm_fenv_clearexcept();
1111 if (sys::llvm_fenv_testexcept()) {
1112 sys::llvm_fenv_clearexcept();
1116 if (Ty->isFloatTy())
1117 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1118 if (Ty->isDoubleTy())
1119 return ConstantFP::get(Ty->getContext(), APFloat(V));
1120 llvm_unreachable("Can only constant fold float/double");
1121 return 0; // dummy return to suppress warning
1124 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1125 double V, double W, Type *Ty) {
1126 sys::llvm_fenv_clearexcept();
1128 if (sys::llvm_fenv_testexcept()) {
1129 sys::llvm_fenv_clearexcept();
1133 if (Ty->isFloatTy())
1134 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1135 if (Ty->isDoubleTy())
1136 return ConstantFP::get(Ty->getContext(), APFloat(V));
1137 llvm_unreachable("Can only constant fold float/double");
1138 return 0; // dummy return to suppress warning
1141 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1142 /// conversion of a constant floating point. If roundTowardZero is false, the
1143 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1144 /// the behavior of the non-truncating SSE instructions in the default rounding
1145 /// mode. The desired integer type Ty is used to select how many bits are
1146 /// available for the result. Returns null if the conversion cannot be
1147 /// performed, otherwise returns the Constant value resulting from the
1149 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1151 assert(Op && "Called with NULL operand");
1152 APFloat Val(Op->getValueAPF());
1154 // All of these conversion intrinsics form an integer of at most 64bits.
1155 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1156 assert(ResultWidth <= 64 &&
1157 "Can only constant fold conversions to 64 and 32 bit ints");
1160 bool isExact = false;
1161 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1162 : APFloat::rmNearestTiesToEven;
1163 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1164 /*isSigned=*/true, mode,
1166 if (status != APFloat::opOK && status != APFloat::opInexact)
1168 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1171 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1172 /// with the specified arguments, returning null if unsuccessful.
1174 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands) {
1175 if (!F->hasName()) return 0;
1176 StringRef Name = F->getName();
1178 Type *Ty = F->getReturnType();
1179 if (Operands.size() == 1) {
1180 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1181 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1182 APFloat Val(Op->getValueAPF());
1185 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1187 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1190 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1193 /// We only fold functions with finite arguments. Folding NaN and inf is
1194 /// likely to be aborted with an exception anyway, and some host libms
1195 /// have known errors raising exceptions.
1196 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1199 /// Currently APFloat versions of these functions do not exist, so we use
1200 /// the host native double versions. Float versions are not called
1201 /// directly but for all these it is true (float)(f((double)arg)) ==
1202 /// f(arg). Long double not supported yet.
1203 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1204 Op->getValueAPF().convertToDouble();
1208 return ConstantFoldFP(acos, V, Ty);
1209 else if (Name == "asin")
1210 return ConstantFoldFP(asin, V, Ty);
1211 else if (Name == "atan")
1212 return ConstantFoldFP(atan, V, Ty);
1216 return ConstantFoldFP(ceil, V, Ty);
1217 else if (Name == "cos")
1218 return ConstantFoldFP(cos, V, Ty);
1219 else if (Name == "cosh")
1220 return ConstantFoldFP(cosh, V, Ty);
1221 else if (Name == "cosf")
1222 return ConstantFoldFP(cos, V, Ty);
1226 return ConstantFoldFP(exp, V, Ty);
1228 if (Name == "exp2") {
1229 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1231 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1236 return ConstantFoldFP(fabs, V, Ty);
1237 else if (Name == "floor")
1238 return ConstantFoldFP(floor, V, Ty);
1241 if (Name == "log" && V > 0)
1242 return ConstantFoldFP(log, V, Ty);
1243 else if (Name == "log10" && V > 0)
1244 return ConstantFoldFP(log10, V, Ty);
1245 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1246 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1248 return ConstantFoldFP(sqrt, V, Ty);
1250 return Constant::getNullValue(Ty);
1255 return ConstantFoldFP(sin, V, Ty);
1256 else if (Name == "sinh")
1257 return ConstantFoldFP(sinh, V, Ty);
1258 else if (Name == "sqrt" && V >= 0)
1259 return ConstantFoldFP(sqrt, V, Ty);
1260 else if (Name == "sqrtf" && V >= 0)
1261 return ConstantFoldFP(sqrt, V, Ty);
1262 else if (Name == "sinf")
1263 return ConstantFoldFP(sin, V, Ty);
1267 return ConstantFoldFP(tan, V, Ty);
1268 else if (Name == "tanh")
1269 return ConstantFoldFP(tanh, V, Ty);
1277 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1278 switch (F->getIntrinsicID()) {
1279 case Intrinsic::bswap:
1280 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1281 case Intrinsic::ctpop:
1282 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1283 case Intrinsic::cttz:
1284 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1285 case Intrinsic::ctlz:
1286 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1287 case Intrinsic::convert_from_fp16: {
1288 APFloat Val(Op->getValue());
1291 APFloat::opStatus status =
1292 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1294 // Conversion is always precise.
1296 assert(status == APFloat::opOK && !lost &&
1297 "Precision lost during fp16 constfolding");
1299 return ConstantFP::get(F->getContext(), Val);
1306 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1307 switch (F->getIntrinsicID()) {
1309 case Intrinsic::x86_sse_cvtss2si:
1310 case Intrinsic::x86_sse_cvtss2si64:
1311 case Intrinsic::x86_sse2_cvtsd2si:
1312 case Intrinsic::x86_sse2_cvtsd2si64:
1313 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1314 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1315 case Intrinsic::x86_sse_cvttss2si:
1316 case Intrinsic::x86_sse_cvttss2si64:
1317 case Intrinsic::x86_sse2_cvttsd2si:
1318 case Intrinsic::x86_sse2_cvttsd2si64:
1319 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1320 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1324 if (isa<UndefValue>(Operands[0])) {
1325 if (F->getIntrinsicID() == Intrinsic::bswap)
1333 if (Operands.size() == 2) {
1334 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1335 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1337 double Op1V = Ty->isFloatTy() ?
1338 (double)Op1->getValueAPF().convertToFloat() :
1339 Op1->getValueAPF().convertToDouble();
1340 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1341 if (Op2->getType() != Op1->getType())
1344 double Op2V = Ty->isFloatTy() ?
1345 (double)Op2->getValueAPF().convertToFloat():
1346 Op2->getValueAPF().convertToDouble();
1349 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1351 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1352 if (Name == "atan2")
1353 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1354 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1355 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1356 return ConstantFP::get(F->getContext(),
1357 APFloat((float)std::pow((float)Op1V,
1358 (int)Op2C->getZExtValue())));
1359 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1360 return ConstantFP::get(F->getContext(),
1361 APFloat((double)std::pow((double)Op1V,
1362 (int)Op2C->getZExtValue())));
1368 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1369 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1370 switch (F->getIntrinsicID()) {
1372 case Intrinsic::sadd_with_overflow:
1373 case Intrinsic::uadd_with_overflow:
1374 case Intrinsic::ssub_with_overflow:
1375 case Intrinsic::usub_with_overflow:
1376 case Intrinsic::smul_with_overflow:
1377 case Intrinsic::umul_with_overflow: {
1380 switch (F->getIntrinsicID()) {
1381 default: assert(0 && "Invalid case");
1382 case Intrinsic::sadd_with_overflow:
1383 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1385 case Intrinsic::uadd_with_overflow:
1386 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1388 case Intrinsic::ssub_with_overflow:
1389 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1391 case Intrinsic::usub_with_overflow:
1392 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1394 case Intrinsic::smul_with_overflow:
1395 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1397 case Intrinsic::umul_with_overflow:
1398 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1402 ConstantInt::get(F->getContext(), Res),
1403 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1405 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);