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 IR ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // DataLayout information. These functions cannot go in IR due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/StringMap.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/FEnv.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// DataLayout. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const DataLayout &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
55 // Handle a vector->integer cast.
56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
57 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
59 return ConstantExpr::getBitCast(C, DestTy);
61 unsigned NumSrcElts = CDV->getType()->getNumElements();
63 Type *SrcEltTy = CDV->getType()->getElementType();
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (SrcEltTy->isFloatingPointTy()) {
68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
71 // Ask IR to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
73 CDV = cast<ConstantDataVector>(C);
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
79 APInt Result(IT->getBitWidth(), 0);
80 for (unsigned i = 0; i != NumSrcElts; ++i) {
82 if (TD.isLittleEndian())
83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
85 Result |= CDV->getElementAsInteger(i);
88 return ConstantInt::get(IT, Result);
91 // The code below only handles casts to vectors currently.
92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
94 return ConstantExpr::getBitCast(C, DestTy);
96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
97 // vector so the code below can handle it uniformly.
98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
99 Constant *Ops = C; // don't take the address of C!
100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
103 // If this is a bitcast from constant vector -> vector, fold it.
104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
105 return ConstantExpr::getBitCast(C, DestTy);
107 // If the element types match, IR can fold it.
108 unsigned NumDstElt = DestVTy->getNumElements();
109 unsigned NumSrcElt = C->getType()->getVectorNumElements();
110 if (NumDstElt == NumSrcElt)
111 return ConstantExpr::getBitCast(C, DestTy);
113 Type *SrcEltTy = C->getType()->getVectorElementType();
114 Type *DstEltTy = DestVTy->getElementType();
116 // Otherwise, we're changing the number of elements in a vector, which
117 // requires endianness information to do the right thing. For example,
118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
119 // folds to (little endian):
120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
121 // and to (big endian):
122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
124 // First thing is first. We only want to think about integer here, so if
125 // we have something in FP form, recast it as integer.
126 if (DstEltTy->isFloatingPointTy()) {
127 // Fold to an vector of integers with same size as our FP type.
128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
131 // Recursively handle this integer conversion, if possible.
132 C = FoldBitCast(C, DestIVTy, TD);
134 // Finally, IR can handle this now that #elts line up.
135 return ConstantExpr::getBitCast(C, DestTy);
138 // Okay, we know the destination is integer, if the input is FP, convert
139 // it to integer first.
140 if (SrcEltTy->isFloatingPointTy()) {
141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
144 // Ask IR to do the conversion now that #elts line up.
145 C = ConstantExpr::getBitCast(C, SrcIVTy);
146 // If IR wasn't able to fold it, bail out.
147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
148 !isa<ConstantDataVector>(C))
152 // Now we know that the input and output vectors are both integer vectors
153 // of the same size, and that their #elements is not the same. Do the
154 // conversion here, which depends on whether the input or output has
156 bool isLittleEndian = TD.isLittleEndian();
158 SmallVector<Constant*, 32> Result;
159 if (NumDstElt < NumSrcElt) {
160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
161 Constant *Zero = Constant::getNullValue(DstEltTy);
162 unsigned Ratio = NumSrcElt/NumDstElt;
163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
165 for (unsigned i = 0; i != NumDstElt; ++i) {
166 // Build each element of the result.
167 Constant *Elt = Zero;
168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
169 for (unsigned j = 0; j != Ratio; ++j) {
170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
171 if (!Src) // Reject constantexpr elements.
172 return ConstantExpr::getBitCast(C, DestTy);
174 // Zero extend the element to the right size.
175 Src = ConstantExpr::getZExt(Src, Elt->getType());
177 // Shift it to the right place, depending on endianness.
178 Src = ConstantExpr::getShl(Src,
179 ConstantInt::get(Src->getType(), ShiftAmt));
180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
183 Elt = ConstantExpr::getOr(Elt, Src);
185 Result.push_back(Elt);
187 return ConstantVector::get(Result);
190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
191 unsigned Ratio = NumDstElt/NumSrcElt;
192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
194 // Loop over each source value, expanding into multiple results.
195 for (unsigned i = 0; i != NumSrcElt; ++i) {
196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
197 if (!Src) // Reject constantexpr elements.
198 return ConstantExpr::getBitCast(C, DestTy);
200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
201 for (unsigned j = 0; j != Ratio; ++j) {
202 // Shift the piece of the value into the right place, depending on
204 Constant *Elt = ConstantExpr::getLShr(Src,
205 ConstantInt::get(Src->getType(), ShiftAmt));
206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
208 // Truncate and remember this piece.
209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
213 return ConstantVector::get(Result);
217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
218 /// from a global, return the global and the constant. Because of
219 /// constantexprs, this function is recursive.
220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
221 APInt &Offset, const DataLayout &TD) {
222 // Trivial case, constant is the global.
223 if ((GV = dyn_cast<GlobalValue>(C))) {
224 Offset.clearAllBits();
228 // Otherwise, if this isn't a constant expr, bail out.
229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
230 if (!CE) return false;
232 // Look through ptr->int and ptr->ptr casts.
233 if (CE->getOpcode() == Instruction::PtrToInt ||
234 CE->getOpcode() == Instruction::BitCast)
235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
238 if (GEPOperator *GEP = dyn_cast<GEPOperator>(CE)) {
239 // If the base isn't a global+constant, we aren't either.
240 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
243 // Otherwise, add any offset that our operands provide.
244 return GEP->accumulateConstantOffset(TD, Offset);
250 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
251 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
252 /// pointer to copy results into and BytesLeft is the number of bytes left in
253 /// the CurPtr buffer. TD is the target data.
254 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
255 unsigned char *CurPtr, unsigned BytesLeft,
256 const DataLayout &TD) {
257 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
258 "Out of range access");
260 // If this element is zero or undefined, we can just return since *CurPtr is
262 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
265 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
266 if (CI->getBitWidth() > 64 ||
267 (CI->getBitWidth() & 7) != 0)
270 uint64_t Val = CI->getZExtValue();
271 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
273 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
275 if (!TD.isLittleEndian())
276 n = IntBytes - n - 1;
277 CurPtr[i] = (unsigned char)(Val >> (n * 8));
283 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
284 if (CFP->getType()->isDoubleTy()) {
285 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
286 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
288 if (CFP->getType()->isFloatTy()){
289 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
290 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
292 if (CFP->getType()->isHalfTy()){
293 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), TD);
294 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
299 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
300 const StructLayout *SL = TD.getStructLayout(CS->getType());
301 unsigned Index = SL->getElementContainingOffset(ByteOffset);
302 uint64_t CurEltOffset = SL->getElementOffset(Index);
303 ByteOffset -= CurEltOffset;
306 // If the element access is to the element itself and not to tail padding,
307 // read the bytes from the element.
308 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
310 if (ByteOffset < EltSize &&
311 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
317 // Check to see if we read from the last struct element, if so we're done.
318 if (Index == CS->getType()->getNumElements())
321 // If we read all of the bytes we needed from this element we're done.
322 uint64_t NextEltOffset = SL->getElementOffset(Index);
324 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
327 // Move to the next element of the struct.
328 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
329 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
331 CurEltOffset = NextEltOffset;
336 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
337 isa<ConstantDataSequential>(C)) {
338 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
339 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
340 uint64_t Index = ByteOffset / EltSize;
341 uint64_t Offset = ByteOffset - Index * EltSize;
343 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
344 NumElts = AT->getNumElements();
346 NumElts = cast<VectorType>(C->getType())->getNumElements();
348 for (; Index != NumElts; ++Index) {
349 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
353 uint64_t BytesWritten = EltSize - Offset;
354 assert(BytesWritten <= EltSize && "Not indexing into this element?");
355 if (BytesWritten >= BytesLeft)
359 BytesLeft -= BytesWritten;
360 CurPtr += BytesWritten;
365 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
366 if (CE->getOpcode() == Instruction::IntToPtr &&
367 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
368 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
372 // Otherwise, unknown initializer type.
376 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
377 const DataLayout &TD) {
378 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
379 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
381 // If this isn't an integer load we can't fold it directly.
383 // If this is a float/double load, we can try folding it as an int32/64 load
384 // and then bitcast the result. This can be useful for union cases. Note
385 // that address spaces don't matter here since we're not going to result in
386 // an actual new load.
388 if (LoadTy->isHalfTy())
389 MapTy = Type::getInt16PtrTy(C->getContext());
390 else if (LoadTy->isFloatTy())
391 MapTy = Type::getInt32PtrTy(C->getContext());
392 else if (LoadTy->isDoubleTy())
393 MapTy = Type::getInt64PtrTy(C->getContext());
394 else if (LoadTy->isVectorTy()) {
395 MapTy = IntegerType::get(C->getContext(),
396 TD.getTypeAllocSizeInBits(LoadTy));
397 MapTy = PointerType::getUnqual(MapTy);
401 C = FoldBitCast(C, MapTy, TD);
402 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
403 return FoldBitCast(Res, LoadTy, TD);
407 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
408 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
411 APInt Offset(TD.getPointerSizeInBits(), 0);
412 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
415 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
416 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
417 !GV->getInitializer()->getType()->isSized())
420 // If we're loading off the beginning of the global, some bytes may be valid,
421 // but we don't try to handle this.
422 if (Offset.isNegative()) return 0;
424 // If we're not accessing anything in this constant, the result is undefined.
425 if (Offset.getZExtValue() >=
426 TD.getTypeAllocSize(GV->getInitializer()->getType()))
427 return UndefValue::get(IntType);
429 unsigned char RawBytes[32] = {0};
430 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
434 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
435 if (TD.isLittleEndian()) {
436 ResultVal = RawBytes[BytesLoaded - 1];
437 for (unsigned i = 1; i != BytesLoaded; ++i) {
439 ResultVal |= RawBytes[BytesLoaded-1-i];
442 ResultVal = RawBytes[0];
443 for (unsigned i = 1; i != BytesLoaded; ++i) {
445 ResultVal |= RawBytes[i];
449 return ConstantInt::get(IntType->getContext(), ResultVal);
452 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
453 /// produce if it is constant and determinable. If this is not determinable,
455 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
456 const DataLayout *TD) {
457 // First, try the easy cases:
458 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
459 if (GV->isConstant() && GV->hasDefinitiveInitializer())
460 return GV->getInitializer();
462 // If the loaded value isn't a constant expr, we can't handle it.
463 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
466 if (CE->getOpcode() == Instruction::GetElementPtr) {
467 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
468 if (GV->isConstant() && GV->hasDefinitiveInitializer())
470 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
474 // Instead of loading constant c string, use corresponding integer value
475 // directly if string length is small enough.
477 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
478 unsigned StrLen = Str.size();
479 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
480 unsigned NumBits = Ty->getPrimitiveSizeInBits();
481 // Replace load with immediate integer if the result is an integer or fp
483 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
484 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
485 APInt StrVal(NumBits, 0);
486 APInt SingleChar(NumBits, 0);
487 if (TD->isLittleEndian()) {
488 for (signed i = StrLen-1; i >= 0; i--) {
489 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
490 StrVal = (StrVal << 8) | SingleChar;
493 for (unsigned i = 0; i < StrLen; i++) {
494 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
495 StrVal = (StrVal << 8) | SingleChar;
497 // Append NULL at the end.
499 StrVal = (StrVal << 8) | SingleChar;
502 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
503 if (Ty->isFloatingPointTy())
504 Res = ConstantExpr::getBitCast(Res, Ty);
509 // If this load comes from anywhere in a constant global, and if the global
510 // is all undef or zero, we know what it loads.
511 if (GlobalVariable *GV =
512 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
513 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
514 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
515 if (GV->getInitializer()->isNullValue())
516 return Constant::getNullValue(ResTy);
517 if (isa<UndefValue>(GV->getInitializer()))
518 return UndefValue::get(ResTy);
522 // Try hard to fold loads from bitcasted strange and non-type-safe things.
524 return FoldReinterpretLoadFromConstPtr(CE, *TD);
528 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
529 if (LI->isVolatile()) return 0;
531 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
532 return ConstantFoldLoadFromConstPtr(C, TD);
537 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
538 /// Attempt to symbolically evaluate the result of a binary operator merging
539 /// these together. If target data info is available, it is provided as TD,
540 /// otherwise TD is null.
541 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
542 Constant *Op1, const DataLayout *TD){
545 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
546 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
550 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
551 // constant. This happens frequently when iterating over a global array.
552 if (Opc == Instruction::Sub && TD) {
553 GlobalValue *GV1, *GV2;
554 unsigned PtrSize = TD->getPointerSizeInBits();
555 unsigned OpSize = TD->getTypeSizeInBits(Op0->getType());
556 APInt Offs1(PtrSize, 0), Offs2(PtrSize, 0);
558 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
559 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
561 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
562 // PtrToInt may change the bitwidth so we have convert to the right size
564 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
565 Offs2.zextOrTrunc(OpSize));
572 /// CastGEPIndices - If array indices are not pointer-sized integers,
573 /// explicitly cast them so that they aren't implicitly casted by the
575 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
576 Type *ResultTy, const DataLayout *TD,
577 const TargetLibraryInfo *TLI) {
579 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
582 SmallVector<Constant*, 32> NewIdxs;
583 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
585 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
586 Ops.slice(1, i-1)))) &&
587 Ops[i]->getType() != IntPtrTy) {
589 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
595 NewIdxs.push_back(Ops[i]);
600 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
601 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
602 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
607 /// Strip the pointer casts, but preserve the address space information.
608 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
609 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
610 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
611 Ptr = cast<Constant>(Ptr->stripPointerCasts());
612 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
614 // Preserve the address space number of the pointer.
615 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
616 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
617 OldPtrTy->getAddressSpace());
618 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy);
623 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
624 /// constant expression, do so.
625 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
626 Type *ResultTy, const DataLayout *TD,
627 const TargetLibraryInfo *TLI) {
628 Constant *Ptr = Ops[0];
629 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
630 !Ptr->getType()->isPointerTy())
633 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
635 // If this is a constant expr gep that is effectively computing an
636 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
637 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
638 if (!isa<ConstantInt>(Ops[i])) {
640 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
641 // "inttoptr (sub (ptrtoint Ptr), V)"
642 if (Ops.size() == 2 &&
643 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
644 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
645 assert((CE == 0 || CE->getType() == IntPtrTy) &&
646 "CastGEPIndices didn't canonicalize index types!");
647 if (CE && CE->getOpcode() == Instruction::Sub &&
648 CE->getOperand(0)->isNullValue()) {
649 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
650 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
651 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
652 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
653 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
660 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
662 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
663 makeArrayRef((Value *const*)
666 Ptr = StripPtrCastKeepAS(Ptr);
668 // If this is a GEP of a GEP, fold it all into a single GEP.
669 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
670 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
672 // Do not try the incorporate the sub-GEP if some index is not a number.
673 bool AllConstantInt = true;
674 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
675 if (!isa<ConstantInt>(NestedOps[i])) {
676 AllConstantInt = false;
682 Ptr = cast<Constant>(GEP->getOperand(0));
683 Offset += APInt(BitWidth,
684 TD->getIndexedOffset(Ptr->getType(), NestedOps));
685 Ptr = StripPtrCastKeepAS(Ptr);
688 // If the base value for this address is a literal integer value, fold the
689 // getelementptr to the resulting integer value casted to the pointer type.
690 APInt BasePtr(BitWidth, 0);
691 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
692 if (CE->getOpcode() == Instruction::IntToPtr)
693 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
694 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
695 if (Ptr->isNullValue() || BasePtr != 0) {
696 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
697 return ConstantExpr::getIntToPtr(C, ResultTy);
700 // Otherwise form a regular getelementptr. Recompute the indices so that
701 // we eliminate over-indexing of the notional static type array bounds.
702 // This makes it easy to determine if the getelementptr is "inbounds".
703 // Also, this helps GlobalOpt do SROA on GlobalVariables.
704 Type *Ty = Ptr->getType();
705 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
706 SmallVector<Constant*, 32> NewIdxs;
708 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
709 if (ATy->isPointerTy()) {
710 // The only pointer indexing we'll do is on the first index of the GEP.
711 if (!NewIdxs.empty())
714 // Only handle pointers to sized types, not pointers to functions.
715 if (!ATy->getElementType()->isSized())
719 // Determine which element of the array the offset points into.
720 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
721 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
723 // The element size is 0. This may be [0 x Ty]*, so just use a zero
724 // index for this level and proceed to the next level to see if it can
725 // accommodate the offset.
726 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
728 // The element size is non-zero divide the offset by the element
729 // size (rounding down), to compute the index at this level.
730 APInt NewIdx = Offset.udiv(ElemSize);
731 Offset -= NewIdx * ElemSize;
732 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
734 Ty = ATy->getElementType();
735 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
736 // If we end up with an offset that isn't valid for this struct type, we
737 // can't re-form this GEP in a regular form, so bail out. The pointer
738 // operand likely went through casts that are necessary to make the GEP
740 const StructLayout &SL = *TD->getStructLayout(STy);
741 if (Offset.uge(SL.getSizeInBytes()))
744 // Determine which field of the struct the offset points into. The
745 // getZExtValue is fine as we've already ensured that the offset is
746 // within the range representable by the StructLayout API.
747 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
748 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
750 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
751 Ty = STy->getTypeAtIndex(ElIdx);
753 // We've reached some non-indexable type.
756 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
758 // If we haven't used up the entire offset by descending the static
759 // type, then the offset is pointing into the middle of an indivisible
760 // member, so we can't simplify it.
766 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
767 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
768 "Computed GetElementPtr has unexpected type!");
770 // If we ended up indexing a member with a type that doesn't match
771 // the type of what the original indices indexed, add a cast.
772 if (Ty != cast<PointerType>(ResultTy)->getElementType())
773 C = FoldBitCast(C, ResultTy, *TD);
780 //===----------------------------------------------------------------------===//
781 // Constant Folding public APIs
782 //===----------------------------------------------------------------------===//
784 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
785 /// If successful, the constant result is returned, if not, null is returned.
786 /// Note that this fails if not all of the operands are constant. Otherwise,
787 /// this function can only fail when attempting to fold instructions like loads
788 /// and stores, which have no constant expression form.
789 Constant *llvm::ConstantFoldInstruction(Instruction *I,
790 const DataLayout *TD,
791 const TargetLibraryInfo *TLI) {
792 // Handle PHI nodes quickly here...
793 if (PHINode *PN = dyn_cast<PHINode>(I)) {
794 Constant *CommonValue = 0;
796 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
797 Value *Incoming = PN->getIncomingValue(i);
798 // If the incoming value is undef then skip it. Note that while we could
799 // skip the value if it is equal to the phi node itself we choose not to
800 // because that would break the rule that constant folding only applies if
801 // all operands are constants.
802 if (isa<UndefValue>(Incoming))
804 // If the incoming value is not a constant, then give up.
805 Constant *C = dyn_cast<Constant>(Incoming);
808 // Fold the PHI's operands.
809 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
810 C = ConstantFoldConstantExpression(NewC, TD, TLI);
811 // If the incoming value is a different constant to
812 // the one we saw previously, then give up.
813 if (CommonValue && C != CommonValue)
819 // If we reach here, all incoming values are the same constant or undef.
820 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
823 // Scan the operand list, checking to see if they are all constants, if so,
824 // hand off to ConstantFoldInstOperands.
825 SmallVector<Constant*, 8> Ops;
826 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
827 Constant *Op = dyn_cast<Constant>(*i);
829 return 0; // All operands not constant!
831 // Fold the Instruction's operands.
832 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
833 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
838 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
839 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
842 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
843 return ConstantFoldLoadInst(LI, TD);
845 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
846 return ConstantExpr::getInsertValue(
847 cast<Constant>(IVI->getAggregateOperand()),
848 cast<Constant>(IVI->getInsertedValueOperand()),
851 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
852 return ConstantExpr::getExtractValue(
853 cast<Constant>(EVI->getAggregateOperand()),
856 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
859 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
860 /// using the specified DataLayout. If successful, the constant result is
861 /// result is returned, if not, null is returned.
862 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
863 const DataLayout *TD,
864 const TargetLibraryInfo *TLI) {
865 SmallVector<Constant*, 8> Ops;
866 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
868 Constant *NewC = cast<Constant>(*i);
869 // Recursively fold the ConstantExpr's operands.
870 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
871 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
876 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
878 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
881 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
882 /// specified opcode and operands. If successful, the constant result is
883 /// returned, if not, null is returned. Note that this function can fail when
884 /// attempting to fold instructions like loads and stores, which have no
885 /// constant expression form.
887 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
888 /// information, due to only being passed an opcode and operands. Constant
889 /// folding using this function strips this information.
891 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
892 ArrayRef<Constant *> Ops,
893 const DataLayout *TD,
894 const TargetLibraryInfo *TLI) {
895 // Handle easy binops first.
896 if (Instruction::isBinaryOp(Opcode)) {
897 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
898 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
901 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
906 case Instruction::ICmp:
907 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
908 case Instruction::Call:
909 if (Function *F = dyn_cast<Function>(Ops.back()))
910 if (canConstantFoldCallTo(F))
911 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
913 case Instruction::PtrToInt:
914 // If the input is a inttoptr, eliminate the pair. This requires knowing
915 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
916 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
917 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
918 Constant *Input = CE->getOperand(0);
919 unsigned InWidth = Input->getType()->getScalarSizeInBits();
920 if (TD->getPointerSizeInBits() < InWidth) {
922 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
923 TD->getPointerSizeInBits()));
924 Input = ConstantExpr::getAnd(Input, Mask);
926 // Do a zext or trunc to get to the dest size.
927 return ConstantExpr::getIntegerCast(Input, DestTy, false);
930 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
931 case Instruction::IntToPtr:
932 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
933 // the int size is >= the ptr size. This requires knowing the width of a
934 // pointer, so it can't be done in ConstantExpr::getCast.
935 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
937 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
938 CE->getOpcode() == Instruction::PtrToInt)
939 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
941 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
942 case Instruction::Trunc:
943 case Instruction::ZExt:
944 case Instruction::SExt:
945 case Instruction::FPTrunc:
946 case Instruction::FPExt:
947 case Instruction::UIToFP:
948 case Instruction::SIToFP:
949 case Instruction::FPToUI:
950 case Instruction::FPToSI:
951 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
952 case Instruction::BitCast:
954 return FoldBitCast(Ops[0], DestTy, *TD);
955 return ConstantExpr::getBitCast(Ops[0], DestTy);
956 case Instruction::Select:
957 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
958 case Instruction::ExtractElement:
959 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
960 case Instruction::InsertElement:
961 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
962 case Instruction::ShuffleVector:
963 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
964 case Instruction::GetElementPtr:
965 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
967 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
970 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
974 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
975 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
976 /// returns a constant expression of the specified operands.
978 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
979 Constant *Ops0, Constant *Ops1,
980 const DataLayout *TD,
981 const TargetLibraryInfo *TLI) {
982 // fold: icmp (inttoptr x), null -> icmp x, 0
983 // fold: icmp (ptrtoint x), 0 -> icmp x, null
984 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
985 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
987 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
988 // around to know if bit truncation is happening.
989 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
990 if (TD && Ops1->isNullValue()) {
991 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
992 if (CE0->getOpcode() == Instruction::IntToPtr) {
993 // Convert the integer value to the right size to ensure we get the
994 // proper extension or truncation.
995 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
997 Constant *Null = Constant::getNullValue(C->getType());
998 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1001 // Only do this transformation if the int is intptrty in size, otherwise
1002 // there is a truncation or extension that we aren't modeling.
1003 if (CE0->getOpcode() == Instruction::PtrToInt &&
1004 CE0->getType() == IntPtrTy) {
1005 Constant *C = CE0->getOperand(0);
1006 Constant *Null = Constant::getNullValue(C->getType());
1007 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1011 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1012 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1013 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1015 if (CE0->getOpcode() == Instruction::IntToPtr) {
1016 // Convert the integer value to the right size to ensure we get the
1017 // proper extension or truncation.
1018 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1020 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1022 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1025 // Only do this transformation if the int is intptrty in size, otherwise
1026 // there is a truncation or extension that we aren't modeling.
1027 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1028 CE0->getType() == IntPtrTy &&
1029 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1030 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1031 CE1->getOperand(0), TD, TLI);
1035 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1036 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1037 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1038 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1040 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1043 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1046 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1047 Constant *Ops[] = { LHS, RHS };
1048 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1052 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1056 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1057 /// getelementptr constantexpr, return the constant value being addressed by the
1058 /// constant expression, or null if something is funny and we can't decide.
1059 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1061 if (!CE->getOperand(1)->isNullValue())
1062 return 0; // Do not allow stepping over the value!
1064 // Loop over all of the operands, tracking down which value we are
1066 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1067 C = C->getAggregateElement(CE->getOperand(i));
1068 if (C == 0) return 0;
1073 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1074 /// indices (with an *implied* zero pointer index that is not in the list),
1075 /// return the constant value being addressed by a virtual load, or null if
1076 /// something is funny and we can't decide.
1077 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1078 ArrayRef<Constant*> Indices) {
1079 // Loop over all of the operands, tracking down which value we are
1081 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1082 C = C->getAggregateElement(Indices[i]);
1083 if (C == 0) return 0;
1089 //===----------------------------------------------------------------------===//
1090 // Constant Folding for Calls
1093 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1094 /// the specified function.
1096 llvm::canConstantFoldCallTo(const Function *F) {
1097 switch (F->getIntrinsicID()) {
1098 case Intrinsic::fabs:
1099 case Intrinsic::log:
1100 case Intrinsic::log2:
1101 case Intrinsic::log10:
1102 case Intrinsic::exp:
1103 case Intrinsic::exp2:
1104 case Intrinsic::floor:
1105 case Intrinsic::sqrt:
1106 case Intrinsic::pow:
1107 case Intrinsic::powi:
1108 case Intrinsic::bswap:
1109 case Intrinsic::ctpop:
1110 case Intrinsic::ctlz:
1111 case Intrinsic::cttz:
1112 case Intrinsic::sadd_with_overflow:
1113 case Intrinsic::uadd_with_overflow:
1114 case Intrinsic::ssub_with_overflow:
1115 case Intrinsic::usub_with_overflow:
1116 case Intrinsic::smul_with_overflow:
1117 case Intrinsic::umul_with_overflow:
1118 case Intrinsic::convert_from_fp16:
1119 case Intrinsic::convert_to_fp16:
1120 case Intrinsic::x86_sse_cvtss2si:
1121 case Intrinsic::x86_sse_cvtss2si64:
1122 case Intrinsic::x86_sse_cvttss2si:
1123 case Intrinsic::x86_sse_cvttss2si64:
1124 case Intrinsic::x86_sse2_cvtsd2si:
1125 case Intrinsic::x86_sse2_cvtsd2si64:
1126 case Intrinsic::x86_sse2_cvttsd2si:
1127 case Intrinsic::x86_sse2_cvttsd2si64:
1134 if (!F->hasName()) return false;
1135 StringRef Name = F->getName();
1137 // In these cases, the check of the length is required. We don't want to
1138 // return true for a name like "cos\0blah" which strcmp would return equal to
1139 // "cos", but has length 8.
1141 default: return false;
1143 return Name == "acos" || Name == "asin" ||
1144 Name == "atan" || Name == "atan2";
1146 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1148 return Name == "exp" || Name == "exp2";
1150 return Name == "fabs" || Name == "fmod" || Name == "floor";
1152 return Name == "log" || Name == "log10";
1154 return Name == "pow";
1156 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1157 Name == "sinf" || Name == "sqrtf";
1159 return Name == "tan" || Name == "tanh";
1163 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1165 sys::llvm_fenv_clearexcept();
1167 if (sys::llvm_fenv_testexcept()) {
1168 sys::llvm_fenv_clearexcept();
1172 if (Ty->isHalfTy()) {
1175 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1176 return ConstantFP::get(Ty->getContext(), APF);
1178 if (Ty->isFloatTy())
1179 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1180 if (Ty->isDoubleTy())
1181 return ConstantFP::get(Ty->getContext(), APFloat(V));
1182 llvm_unreachable("Can only constant fold half/float/double");
1185 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1186 double V, double W, Type *Ty) {
1187 sys::llvm_fenv_clearexcept();
1189 if (sys::llvm_fenv_testexcept()) {
1190 sys::llvm_fenv_clearexcept();
1194 if (Ty->isHalfTy()) {
1197 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1198 return ConstantFP::get(Ty->getContext(), APF);
1200 if (Ty->isFloatTy())
1201 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1202 if (Ty->isDoubleTy())
1203 return ConstantFP::get(Ty->getContext(), APFloat(V));
1204 llvm_unreachable("Can only constant fold half/float/double");
1207 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1208 /// conversion of a constant floating point. If roundTowardZero is false, the
1209 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1210 /// the behavior of the non-truncating SSE instructions in the default rounding
1211 /// mode. The desired integer type Ty is used to select how many bits are
1212 /// available for the result. Returns null if the conversion cannot be
1213 /// performed, otherwise returns the Constant value resulting from the
1215 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1216 bool roundTowardZero, Type *Ty) {
1217 // All of these conversion intrinsics form an integer of at most 64bits.
1218 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1219 assert(ResultWidth <= 64 &&
1220 "Can only constant fold conversions to 64 and 32 bit ints");
1223 bool isExact = false;
1224 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1225 : APFloat::rmNearestTiesToEven;
1226 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1227 /*isSigned=*/true, mode,
1229 if (status != APFloat::opOK && status != APFloat::opInexact)
1231 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1234 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1235 /// with the specified arguments, returning null if unsuccessful.
1237 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1238 const TargetLibraryInfo *TLI) {
1239 if (!F->hasName()) return 0;
1240 StringRef Name = F->getName();
1242 Type *Ty = F->getReturnType();
1243 if (Operands.size() == 1) {
1244 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1245 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1246 APFloat Val(Op->getValueAPF());
1249 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1251 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1256 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1259 /// We only fold functions with finite arguments. Folding NaN and inf is
1260 /// likely to be aborted with an exception anyway, and some host libms
1261 /// have known errors raising exceptions.
1262 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1265 /// Currently APFloat versions of these functions do not exist, so we use
1266 /// the host native double versions. Float versions are not called
1267 /// directly but for all these it is true (float)(f((double)arg)) ==
1268 /// f(arg). Long double not supported yet.
1270 if (Ty->isFloatTy())
1271 V = Op->getValueAPF().convertToFloat();
1272 else if (Ty->isDoubleTy())
1273 V = Op->getValueAPF().convertToDouble();
1276 APFloat APF = Op->getValueAPF();
1277 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1278 V = APF.convertToDouble();
1281 switch (F->getIntrinsicID()) {
1283 case Intrinsic::fabs:
1284 return ConstantFoldFP(fabs, V, Ty);
1286 case Intrinsic::log2:
1287 return ConstantFoldFP(log2, V, Ty);
1290 case Intrinsic::log:
1291 return ConstantFoldFP(log, V, Ty);
1294 case Intrinsic::log10:
1295 return ConstantFoldFP(log10, V, Ty);
1298 case Intrinsic::exp:
1299 return ConstantFoldFP(exp, V, Ty);
1302 case Intrinsic::exp2:
1303 return ConstantFoldFP(exp2, V, Ty);
1305 case Intrinsic::floor:
1306 return ConstantFoldFP(floor, V, Ty);
1311 if (Name == "acos" && TLI->has(LibFunc::acos))
1312 return ConstantFoldFP(acos, V, Ty);
1313 else if (Name == "asin" && TLI->has(LibFunc::asin))
1314 return ConstantFoldFP(asin, V, Ty);
1315 else if (Name == "atan" && TLI->has(LibFunc::atan))
1316 return ConstantFoldFP(atan, V, Ty);
1319 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1320 return ConstantFoldFP(ceil, V, Ty);
1321 else if (Name == "cos" && TLI->has(LibFunc::cos))
1322 return ConstantFoldFP(cos, V, Ty);
1323 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1324 return ConstantFoldFP(cosh, V, Ty);
1325 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1326 return ConstantFoldFP(cos, V, Ty);
1329 if (Name == "exp" && TLI->has(LibFunc::exp))
1330 return ConstantFoldFP(exp, V, Ty);
1332 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1333 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1335 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1339 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1340 return ConstantFoldFP(fabs, V, Ty);
1341 else if (Name == "floor" && TLI->has(LibFunc::floor))
1342 return ConstantFoldFP(floor, V, Ty);
1345 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1346 return ConstantFoldFP(log, V, Ty);
1347 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1348 return ConstantFoldFP(log10, V, Ty);
1349 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1350 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1352 return ConstantFoldFP(sqrt, V, Ty);
1354 return Constant::getNullValue(Ty);
1358 if (Name == "sin" && TLI->has(LibFunc::sin))
1359 return ConstantFoldFP(sin, V, Ty);
1360 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1361 return ConstantFoldFP(sinh, V, Ty);
1362 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1363 return ConstantFoldFP(sqrt, V, Ty);
1364 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1365 return ConstantFoldFP(sqrt, V, Ty);
1366 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1367 return ConstantFoldFP(sin, V, Ty);
1370 if (Name == "tan" && TLI->has(LibFunc::tan))
1371 return ConstantFoldFP(tan, V, Ty);
1372 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1373 return ConstantFoldFP(tanh, V, Ty);
1381 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1382 switch (F->getIntrinsicID()) {
1383 case Intrinsic::bswap:
1384 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1385 case Intrinsic::ctpop:
1386 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1387 case Intrinsic::convert_from_fp16: {
1388 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1391 APFloat::opStatus status =
1392 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1394 // Conversion is always precise.
1396 assert(status == APFloat::opOK && !lost &&
1397 "Precision lost during fp16 constfolding");
1399 return ConstantFP::get(F->getContext(), Val);
1406 // Support ConstantVector in case we have an Undef in the top.
1407 if (isa<ConstantVector>(Operands[0]) ||
1408 isa<ConstantDataVector>(Operands[0])) {
1409 Constant *Op = cast<Constant>(Operands[0]);
1410 switch (F->getIntrinsicID()) {
1412 case Intrinsic::x86_sse_cvtss2si:
1413 case Intrinsic::x86_sse_cvtss2si64:
1414 case Intrinsic::x86_sse2_cvtsd2si:
1415 case Intrinsic::x86_sse2_cvtsd2si64:
1416 if (ConstantFP *FPOp =
1417 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1418 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1419 /*roundTowardZero=*/false, Ty);
1420 case Intrinsic::x86_sse_cvttss2si:
1421 case Intrinsic::x86_sse_cvttss2si64:
1422 case Intrinsic::x86_sse2_cvttsd2si:
1423 case Intrinsic::x86_sse2_cvttsd2si64:
1424 if (ConstantFP *FPOp =
1425 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1426 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1427 /*roundTowardZero=*/true, Ty);
1431 if (isa<UndefValue>(Operands[0])) {
1432 if (F->getIntrinsicID() == Intrinsic::bswap)
1440 if (Operands.size() == 2) {
1441 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1442 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1445 if (Ty->isFloatTy())
1446 Op1V = Op1->getValueAPF().convertToFloat();
1447 else if (Ty->isDoubleTy())
1448 Op1V = Op1->getValueAPF().convertToDouble();
1451 APFloat APF = Op1->getValueAPF();
1452 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1453 Op1V = APF.convertToDouble();
1456 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1457 if (Op2->getType() != Op1->getType())
1461 if (Ty->isFloatTy())
1462 Op2V = Op2->getValueAPF().convertToFloat();
1463 else if (Ty->isDoubleTy())
1464 Op2V = Op2->getValueAPF().convertToDouble();
1467 APFloat APF = Op2->getValueAPF();
1468 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1469 Op2V = APF.convertToDouble();
1472 if (F->getIntrinsicID() == Intrinsic::pow) {
1473 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1477 if (Name == "pow" && TLI->has(LibFunc::pow))
1478 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1479 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1480 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1481 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1482 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1483 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1484 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isHalfTy())
1485 return ConstantFP::get(F->getContext(),
1486 APFloat((float)std::pow((float)Op1V,
1487 (int)Op2C->getZExtValue())));
1488 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1489 return ConstantFP::get(F->getContext(),
1490 APFloat((float)std::pow((float)Op1V,
1491 (int)Op2C->getZExtValue())));
1492 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1493 return ConstantFP::get(F->getContext(),
1494 APFloat((double)std::pow((double)Op1V,
1495 (int)Op2C->getZExtValue())));
1500 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1501 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1502 switch (F->getIntrinsicID()) {
1504 case Intrinsic::sadd_with_overflow:
1505 case Intrinsic::uadd_with_overflow:
1506 case Intrinsic::ssub_with_overflow:
1507 case Intrinsic::usub_with_overflow:
1508 case Intrinsic::smul_with_overflow:
1509 case Intrinsic::umul_with_overflow: {
1512 switch (F->getIntrinsicID()) {
1513 default: llvm_unreachable("Invalid case");
1514 case Intrinsic::sadd_with_overflow:
1515 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1517 case Intrinsic::uadd_with_overflow:
1518 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1520 case Intrinsic::ssub_with_overflow:
1521 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1523 case Intrinsic::usub_with_overflow:
1524 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1526 case Intrinsic::smul_with_overflow:
1527 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1529 case Intrinsic::umul_with_overflow:
1530 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1534 ConstantInt::get(F->getContext(), Res),
1535 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1537 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1539 case Intrinsic::cttz:
1540 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1541 return UndefValue::get(Ty);
1542 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1543 case Intrinsic::ctlz:
1544 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1545 return UndefValue::get(Ty);
1546 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());