1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines routines for folding instructions into constants.
12 // Also, to supplement the basic VMCore ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // TargetData information. These functions cannot go in VMCore due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLibraryInfo.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/StringMap.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/FEnv.h"
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// TargetData. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const TargetData &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
55 // Handle a vector->integer cast.
56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
57 // FIXME: Remove ConstantVector support.
58 if ((!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C)) ||
59 // TODO: Handle big endian someday.
61 return ConstantExpr::getBitCast(C, DestTy);
63 unsigned NumSrcElts = C->getType()->getVectorNumElements();
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (C->getType()->getVectorElementType()->isFloatingPointTy()) {
69 C->getType()->getVectorElementType()->getPrimitiveSizeInBits();
71 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
72 // Ask VMCore to do the conversion now that #elts line up.
73 C = ConstantExpr::getBitCast(C, SrcIVTy);
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
79 TD.getTypeAllocSizeInBits(C->getType()->getVectorElementType());
80 APInt Result(IT->getBitWidth(), 0);
81 for (unsigned i = 0; i != NumSrcElts; ++i) {
82 // FIXME: Rework when we have ConstantDataVector.
83 ConstantInt *Elt=dyn_cast_or_null<ConstantInt>(C->getAggregateElement(i));
84 if (Elt == 0) // Elt must be a constant expr or something.
85 return ConstantExpr::getBitCast(C, DestTy);
87 Result |= Elt->getValue().zext(IT->getBitWidth()) << i*BitShift;
90 return ConstantInt::get(IT, Result);
93 // The code below only handles casts to vectors currently.
94 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
96 return ConstantExpr::getBitCast(C, DestTy);
98 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
99 // vector so the code below can handle it uniformly.
100 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
101 Constant *Ops = C; // don't take the address of C!
102 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
105 // If this is a bitcast from constant vector -> vector, fold it.
106 // FIXME: Remove ConstantVector support.
107 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
108 return ConstantExpr::getBitCast(C, DestTy);
110 // If the element types match, VMCore can fold it.
111 unsigned NumDstElt = DestVTy->getNumElements();
112 unsigned NumSrcElt = C->getType()->getVectorNumElements();
113 if (NumDstElt == NumSrcElt)
114 return ConstantExpr::getBitCast(C, DestTy);
116 Type *SrcEltTy = C->getType()->getVectorElementType();
117 Type *DstEltTy = DestVTy->getElementType();
119 // Otherwise, we're changing the number of elements in a vector, which
120 // requires endianness information to do the right thing. For example,
121 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
122 // folds to (little endian):
123 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
124 // and to (big endian):
125 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
127 // First thing is first. We only want to think about integer here, so if
128 // we have something in FP form, recast it as integer.
129 if (DstEltTy->isFloatingPointTy()) {
130 // Fold to an vector of integers with same size as our FP type.
131 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
133 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
134 // Recursively handle this integer conversion, if possible.
135 C = FoldBitCast(C, DestIVTy, TD);
137 // Finally, VMCore can handle this now that #elts line up.
138 return ConstantExpr::getBitCast(C, DestTy);
141 // Okay, we know the destination is integer, if the input is FP, convert
142 // it to integer first.
143 if (SrcEltTy->isFloatingPointTy()) {
144 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
146 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
147 // Ask VMCore to do the conversion now that #elts line up.
148 C = ConstantExpr::getBitCast(C, SrcIVTy);
149 // If VMCore wasn't able to fold it, bail out.
150 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
151 !isa<ConstantDataVector>(C))
155 // Now we know that the input and output vectors are both integer vectors
156 // of the same size, and that their #elements is not the same. Do the
157 // conversion here, which depends on whether the input or output has
159 bool isLittleEndian = TD.isLittleEndian();
161 SmallVector<Constant*, 32> Result;
162 if (NumDstElt < NumSrcElt) {
163 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
164 Constant *Zero = Constant::getNullValue(DstEltTy);
165 unsigned Ratio = NumSrcElt/NumDstElt;
166 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
168 for (unsigned i = 0; i != NumDstElt; ++i) {
169 // Build each element of the result.
170 Constant *Elt = Zero;
171 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
172 for (unsigned j = 0; j != Ratio; ++j) {
173 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
174 if (!Src) // Reject constantexpr elements.
175 return ConstantExpr::getBitCast(C, DestTy);
177 // Zero extend the element to the right size.
178 Src = ConstantExpr::getZExt(Src, Elt->getType());
180 // Shift it to the right place, depending on endianness.
181 Src = ConstantExpr::getShl(Src,
182 ConstantInt::get(Src->getType(), ShiftAmt));
183 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
186 Elt = ConstantExpr::getOr(Elt, Src);
188 Result.push_back(Elt);
190 return ConstantVector::get(Result);
193 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
194 unsigned Ratio = NumDstElt/NumSrcElt;
195 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
197 // Loop over each source value, expanding into multiple results.
198 for (unsigned i = 0; i != NumSrcElt; ++i) {
199 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
200 if (!Src) // Reject constantexpr elements.
201 return ConstantExpr::getBitCast(C, DestTy);
203 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
204 for (unsigned j = 0; j != Ratio; ++j) {
205 // Shift the piece of the value into the right place, depending on
207 Constant *Elt = ConstantExpr::getLShr(Src,
208 ConstantInt::get(Src->getType(), ShiftAmt));
209 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
211 // Truncate and remember this piece.
212 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
216 return ConstantVector::get(Result);
220 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
221 /// from a global, return the global and the constant. Because of
222 /// constantexprs, this function is recursive.
223 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
224 int64_t &Offset, const TargetData &TD) {
225 // Trivial case, constant is the global.
226 if ((GV = dyn_cast<GlobalValue>(C))) {
231 // Otherwise, if this isn't a constant expr, bail out.
232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
233 if (!CE) return false;
235 // Look through ptr->int and ptr->ptr casts.
236 if (CE->getOpcode() == Instruction::PtrToInt ||
237 CE->getOpcode() == Instruction::BitCast)
238 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
240 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
241 if (CE->getOpcode() == Instruction::GetElementPtr) {
242 // Cannot compute this if the element type of the pointer is missing size
244 if (!cast<PointerType>(CE->getOperand(0)->getType())
245 ->getElementType()->isSized())
248 // If the base isn't a global+constant, we aren't either.
249 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
252 // Otherwise, add any offset that our operands provide.
253 gep_type_iterator GTI = gep_type_begin(CE);
254 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
255 i != e; ++i, ++GTI) {
256 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
257 if (!CI) return false; // Index isn't a simple constant?
258 if (CI->isZero()) continue; // Not adding anything.
260 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
262 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
264 SequentialType *SQT = cast<SequentialType>(*GTI);
265 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
274 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
275 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
276 /// pointer to copy results into and BytesLeft is the number of bytes left in
277 /// the CurPtr buffer. TD is the target data.
278 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
279 unsigned char *CurPtr, unsigned BytesLeft,
280 const TargetData &TD) {
281 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
282 "Out of range access");
284 // If this element is zero or undefined, we can just return since *CurPtr is
286 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
289 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
290 if (CI->getBitWidth() > 64 ||
291 (CI->getBitWidth() & 7) != 0)
294 uint64_t Val = CI->getZExtValue();
295 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
297 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
298 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
304 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
305 if (CFP->getType()->isDoubleTy()) {
306 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
307 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
309 if (CFP->getType()->isFloatTy()){
310 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
311 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
316 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
317 const StructLayout *SL = TD.getStructLayout(CS->getType());
318 unsigned Index = SL->getElementContainingOffset(ByteOffset);
319 uint64_t CurEltOffset = SL->getElementOffset(Index);
320 ByteOffset -= CurEltOffset;
323 // If the element access is to the element itself and not to tail padding,
324 // read the bytes from the element.
325 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
327 if (ByteOffset < EltSize &&
328 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
334 // Check to see if we read from the last struct element, if so we're done.
335 if (Index == CS->getType()->getNumElements())
338 // If we read all of the bytes we needed from this element we're done.
339 uint64_t NextEltOffset = SL->getElementOffset(Index);
341 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
344 // Move to the next element of the struct.
345 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
346 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
348 CurEltOffset = NextEltOffset;
353 // FIXME: Remove ConstantVector
354 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
355 isa<ConstantDataSequential>(C)) {
356 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
357 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
358 uint64_t Index = ByteOffset / EltSize;
359 uint64_t Offset = ByteOffset - Index * EltSize;
361 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
362 NumElts = AT->getNumElements();
364 NumElts = cast<VectorType>(C->getType())->getNumElements();
366 for (; Index != NumElts; ++Index) {
367 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
370 if (EltSize >= BytesLeft)
374 BytesLeft -= EltSize;
380 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
381 if (CE->getOpcode() == Instruction::IntToPtr &&
382 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
383 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
387 // Otherwise, unknown initializer type.
391 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
392 const TargetData &TD) {
393 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
394 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
396 // If this isn't an integer load we can't fold it directly.
398 // If this is a float/double load, we can try folding it as an int32/64 load
399 // and then bitcast the result. This can be useful for union cases. Note
400 // that address spaces don't matter here since we're not going to result in
401 // an actual new load.
403 if (LoadTy->isFloatTy())
404 MapTy = Type::getInt32PtrTy(C->getContext());
405 else if (LoadTy->isDoubleTy())
406 MapTy = Type::getInt64PtrTy(C->getContext());
407 else if (LoadTy->isVectorTy()) {
408 MapTy = IntegerType::get(C->getContext(),
409 TD.getTypeAllocSizeInBits(LoadTy));
410 MapTy = PointerType::getUnqual(MapTy);
414 C = FoldBitCast(C, MapTy, TD);
415 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
416 return FoldBitCast(Res, LoadTy, TD);
420 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
421 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
425 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
428 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
429 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
430 !GV->getInitializer()->getType()->isSized())
433 // If we're loading off the beginning of the global, some bytes may be valid,
434 // but we don't try to handle this.
435 if (Offset < 0) return 0;
437 // If we're not accessing anything in this constant, the result is undefined.
438 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
439 return UndefValue::get(IntType);
441 unsigned char RawBytes[32] = {0};
442 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
446 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
447 for (unsigned i = 1; i != BytesLoaded; ++i) {
449 ResultVal |= RawBytes[BytesLoaded-1-i];
452 return ConstantInt::get(IntType->getContext(), ResultVal);
455 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
456 /// produce if it is constant and determinable. If this is not determinable,
458 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
459 const TargetData *TD) {
460 // First, try the easy cases:
461 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
462 if (GV->isConstant() && GV->hasDefinitiveInitializer())
463 return GV->getInitializer();
465 // If the loaded value isn't a constant expr, we can't handle it.
466 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
469 if (CE->getOpcode() == Instruction::GetElementPtr) {
470 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
471 if (GV->isConstant() && GV->hasDefinitiveInitializer())
473 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
477 // Instead of loading constant c string, use corresponding integer value
478 // directly if string length is small enough.
480 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
481 unsigned StrLen = Str.length();
482 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
483 unsigned NumBits = Ty->getPrimitiveSizeInBits();
484 // Replace load with immediate integer if the result is an integer or fp
486 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
487 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
488 APInt StrVal(NumBits, 0);
489 APInt SingleChar(NumBits, 0);
490 if (TD->isLittleEndian()) {
491 for (signed i = StrLen-1; i >= 0; i--) {
492 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
493 StrVal = (StrVal << 8) | SingleChar;
496 for (unsigned i = 0; i < StrLen; i++) {
497 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
498 StrVal = (StrVal << 8) | SingleChar;
500 // Append NULL at the end.
502 StrVal = (StrVal << 8) | SingleChar;
505 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
506 if (Ty->isFloatingPointTy())
507 Res = ConstantExpr::getBitCast(Res, Ty);
512 // If this load comes from anywhere in a constant global, and if the global
513 // is all undef or zero, we know what it loads.
514 if (GlobalVariable *GV =
515 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
516 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
517 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
518 if (GV->getInitializer()->isNullValue())
519 return Constant::getNullValue(ResTy);
520 if (isa<UndefValue>(GV->getInitializer()))
521 return UndefValue::get(ResTy);
525 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
526 // currently don't do any of this for big endian systems. It can be
527 // generalized in the future if someone is interested.
528 if (TD && TD->isLittleEndian())
529 return FoldReinterpretLoadFromConstPtr(CE, *TD);
533 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
534 if (LI->isVolatile()) return 0;
536 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
537 return ConstantFoldLoadFromConstPtr(C, TD);
542 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
543 /// Attempt to symbolically evaluate the result of a binary operator merging
544 /// these together. If target data info is available, it is provided as TD,
545 /// otherwise TD is null.
546 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
547 Constant *Op1, const TargetData *TD){
550 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
551 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
555 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
556 // constant. This happens frequently when iterating over a global array.
557 if (Opc == Instruction::Sub && TD) {
558 GlobalValue *GV1, *GV2;
559 int64_t Offs1, Offs2;
561 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
562 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
564 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
565 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
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 TargetData *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 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
608 /// constant expression, do so.
609 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
610 Type *ResultTy, const TargetData *TD,
611 const TargetLibraryInfo *TLI) {
612 Constant *Ptr = Ops[0];
613 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
614 !Ptr->getType()->isPointerTy())
617 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
619 // If this is a constant expr gep that is effectively computing an
620 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
621 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
622 if (!isa<ConstantInt>(Ops[i])) {
624 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
625 // "inttoptr (sub (ptrtoint Ptr), V)"
626 if (Ops.size() == 2 &&
627 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
628 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
629 assert((CE == 0 || CE->getType() == IntPtrTy) &&
630 "CastGEPIndices didn't canonicalize index types!");
631 if (CE && CE->getOpcode() == Instruction::Sub &&
632 CE->getOperand(0)->isNullValue()) {
633 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
634 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
635 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
636 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
637 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
644 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
646 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
647 makeArrayRef((Value **)Ops.data() + 1,
649 Ptr = cast<Constant>(Ptr->stripPointerCasts());
651 // If this is a GEP of a GEP, fold it all into a single GEP.
652 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
653 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
655 // Do not try the incorporate the sub-GEP if some index is not a number.
656 bool AllConstantInt = true;
657 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
658 if (!isa<ConstantInt>(NestedOps[i])) {
659 AllConstantInt = false;
665 Ptr = cast<Constant>(GEP->getOperand(0));
666 Offset += APInt(BitWidth,
667 TD->getIndexedOffset(Ptr->getType(), NestedOps));
668 Ptr = cast<Constant>(Ptr->stripPointerCasts());
671 // If the base value for this address is a literal integer value, fold the
672 // getelementptr to the resulting integer value casted to the pointer type.
673 APInt BasePtr(BitWidth, 0);
674 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
675 if (CE->getOpcode() == Instruction::IntToPtr)
676 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
677 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
678 if (Ptr->isNullValue() || BasePtr != 0) {
679 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
680 return ConstantExpr::getIntToPtr(C, ResultTy);
683 // Otherwise form a regular getelementptr. Recompute the indices so that
684 // we eliminate over-indexing of the notional static type array bounds.
685 // This makes it easy to determine if the getelementptr is "inbounds".
686 // Also, this helps GlobalOpt do SROA on GlobalVariables.
687 Type *Ty = Ptr->getType();
688 SmallVector<Constant*, 32> NewIdxs;
690 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
691 if (ATy->isPointerTy()) {
692 // The only pointer indexing we'll do is on the first index of the GEP.
693 if (!NewIdxs.empty())
696 // Only handle pointers to sized types, not pointers to functions.
697 if (!ATy->getElementType()->isSized())
701 // Determine which element of the array the offset points into.
702 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
703 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
705 // The element size is 0. This may be [0 x Ty]*, so just use a zero
706 // index for this level and proceed to the next level to see if it can
707 // accommodate the offset.
708 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
710 // The element size is non-zero divide the offset by the element
711 // size (rounding down), to compute the index at this level.
712 APInt NewIdx = Offset.udiv(ElemSize);
713 Offset -= NewIdx * ElemSize;
714 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
716 Ty = ATy->getElementType();
717 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
718 // Determine which field of the struct the offset points into. The
719 // getZExtValue is at least as safe as the StructLayout API because we
720 // know the offset is within the struct at this point.
721 const StructLayout &SL = *TD->getStructLayout(STy);
722 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
723 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
725 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
726 Ty = STy->getTypeAtIndex(ElIdx);
728 // We've reached some non-indexable type.
731 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
733 // If we haven't used up the entire offset by descending the static
734 // type, then the offset is pointing into the middle of an indivisible
735 // member, so we can't simplify it.
741 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
742 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
743 "Computed GetElementPtr has unexpected type!");
745 // If we ended up indexing a member with a type that doesn't match
746 // the type of what the original indices indexed, add a cast.
747 if (Ty != cast<PointerType>(ResultTy)->getElementType())
748 C = FoldBitCast(C, ResultTy, *TD);
755 //===----------------------------------------------------------------------===//
756 // Constant Folding public APIs
757 //===----------------------------------------------------------------------===//
759 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
760 /// If successful, the constant result is returned, if not, null is returned.
761 /// Note that this fails if not all of the operands are constant. Otherwise,
762 /// this function can only fail when attempting to fold instructions like loads
763 /// and stores, which have no constant expression form.
764 Constant *llvm::ConstantFoldInstruction(Instruction *I,
765 const TargetData *TD,
766 const TargetLibraryInfo *TLI) {
767 // Handle PHI nodes quickly here...
768 if (PHINode *PN = dyn_cast<PHINode>(I)) {
769 Constant *CommonValue = 0;
771 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
772 Value *Incoming = PN->getIncomingValue(i);
773 // If the incoming value is undef then skip it. Note that while we could
774 // skip the value if it is equal to the phi node itself we choose not to
775 // because that would break the rule that constant folding only applies if
776 // all operands are constants.
777 if (isa<UndefValue>(Incoming))
779 // If the incoming value is not a constant, or is a different constant to
780 // the one we saw previously, then give up.
781 Constant *C = dyn_cast<Constant>(Incoming);
782 if (!C || (CommonValue && C != CommonValue))
787 // If we reach here, all incoming values are the same constant or undef.
788 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
791 // Scan the operand list, checking to see if they are all constants, if so,
792 // hand off to ConstantFoldInstOperands.
793 SmallVector<Constant*, 8> Ops;
794 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
795 if (Constant *Op = dyn_cast<Constant>(*i))
798 return 0; // All operands not constant!
800 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
801 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
804 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
805 return ConstantFoldLoadInst(LI, TD);
807 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
808 return ConstantExpr::getInsertValue(
809 cast<Constant>(IVI->getAggregateOperand()),
810 cast<Constant>(IVI->getInsertedValueOperand()),
813 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
814 return ConstantExpr::getExtractValue(
815 cast<Constant>(EVI->getAggregateOperand()),
818 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
821 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
822 /// using the specified TargetData. If successful, the constant result is
823 /// result is returned, if not, null is returned.
824 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
825 const TargetData *TD,
826 const TargetLibraryInfo *TLI) {
827 SmallVector<Constant*, 8> Ops;
828 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
830 Constant *NewC = cast<Constant>(*i);
831 // Recursively fold the ConstantExpr's operands.
832 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
833 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
838 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
840 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
843 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
844 /// specified opcode and operands. If successful, the constant result is
845 /// returned, if not, null is returned. Note that this function can fail when
846 /// attempting to fold instructions like loads and stores, which have no
847 /// constant expression form.
849 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
850 /// information, due to only being passed an opcode and operands. Constant
851 /// folding using this function strips this information.
853 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
854 ArrayRef<Constant *> Ops,
855 const TargetData *TD,
856 const TargetLibraryInfo *TLI) {
857 // Handle easy binops first.
858 if (Instruction::isBinaryOp(Opcode)) {
859 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
860 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
863 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
868 case Instruction::ICmp:
869 case Instruction::FCmp: assert(0 && "Invalid for compares");
870 case Instruction::Call:
871 if (Function *F = dyn_cast<Function>(Ops.back()))
872 if (canConstantFoldCallTo(F))
873 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
875 case Instruction::PtrToInt:
876 // If the input is a inttoptr, eliminate the pair. This requires knowing
877 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
878 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
879 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
880 Constant *Input = CE->getOperand(0);
881 unsigned InWidth = Input->getType()->getScalarSizeInBits();
882 if (TD->getPointerSizeInBits() < InWidth) {
884 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
885 TD->getPointerSizeInBits()));
886 Input = ConstantExpr::getAnd(Input, Mask);
888 // Do a zext or trunc to get to the dest size.
889 return ConstantExpr::getIntegerCast(Input, DestTy, false);
892 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
893 case Instruction::IntToPtr:
894 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
895 // the int size is >= the ptr size. This requires knowing the width of a
896 // pointer, so it can't be done in ConstantExpr::getCast.
897 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
899 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
900 CE->getOpcode() == Instruction::PtrToInt)
901 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
903 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
904 case Instruction::Trunc:
905 case Instruction::ZExt:
906 case Instruction::SExt:
907 case Instruction::FPTrunc:
908 case Instruction::FPExt:
909 case Instruction::UIToFP:
910 case Instruction::SIToFP:
911 case Instruction::FPToUI:
912 case Instruction::FPToSI:
913 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
914 case Instruction::BitCast:
916 return FoldBitCast(Ops[0], DestTy, *TD);
917 return ConstantExpr::getBitCast(Ops[0], DestTy);
918 case Instruction::Select:
919 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
920 case Instruction::ExtractElement:
921 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
922 case Instruction::InsertElement:
923 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
924 case Instruction::ShuffleVector:
925 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
926 case Instruction::GetElementPtr:
927 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
929 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
932 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
936 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
937 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
938 /// returns a constant expression of the specified operands.
940 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
941 Constant *Ops0, Constant *Ops1,
942 const TargetData *TD,
943 const TargetLibraryInfo *TLI) {
944 // fold: icmp (inttoptr x), null -> icmp x, 0
945 // fold: icmp (ptrtoint x), 0 -> icmp x, null
946 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
947 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
949 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
950 // around to know if bit truncation is happening.
951 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
952 if (TD && Ops1->isNullValue()) {
953 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
954 if (CE0->getOpcode() == Instruction::IntToPtr) {
955 // Convert the integer value to the right size to ensure we get the
956 // proper extension or truncation.
957 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
959 Constant *Null = Constant::getNullValue(C->getType());
960 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
963 // Only do this transformation if the int is intptrty in size, otherwise
964 // there is a truncation or extension that we aren't modeling.
965 if (CE0->getOpcode() == Instruction::PtrToInt &&
966 CE0->getType() == IntPtrTy) {
967 Constant *C = CE0->getOperand(0);
968 Constant *Null = Constant::getNullValue(C->getType());
969 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
973 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
974 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
975 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
977 if (CE0->getOpcode() == Instruction::IntToPtr) {
978 // Convert the integer value to the right size to ensure we get the
979 // proper extension or truncation.
980 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
982 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
984 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
987 // Only do this transformation if the int is intptrty in size, otherwise
988 // there is a truncation or extension that we aren't modeling.
989 if ((CE0->getOpcode() == Instruction::PtrToInt &&
990 CE0->getType() == IntPtrTy &&
991 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
992 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
993 CE1->getOperand(0), TD, TLI);
997 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
998 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
999 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1000 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1002 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1005 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1008 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1009 Constant *Ops[] = { LHS, RHS };
1010 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1014 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1018 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1019 /// getelementptr constantexpr, return the constant value being addressed by the
1020 /// constant expression, or null if something is funny and we can't decide.
1021 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1023 if (!CE->getOperand(1)->isNullValue())
1024 return 0; // Do not allow stepping over the value!
1026 // Loop over all of the operands, tracking down which value we are
1028 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1029 C = C->getAggregateElement(CE->getOperand(i));
1030 if (C == 0) return 0;
1035 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1036 /// indices (with an *implied* zero pointer index that is not in the list),
1037 /// return the constant value being addressed by a virtual load, or null if
1038 /// something is funny and we can't decide.
1039 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1040 ArrayRef<Constant*> Indices) {
1041 // Loop over all of the operands, tracking down which value we are
1043 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1044 C = C->getAggregateElement(Indices[i]);
1045 if (C == 0) return 0;
1051 //===----------------------------------------------------------------------===//
1052 // Constant Folding for Calls
1055 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1056 /// the specified function.
1058 llvm::canConstantFoldCallTo(const Function *F) {
1059 switch (F->getIntrinsicID()) {
1060 case Intrinsic::sqrt:
1061 case Intrinsic::pow:
1062 case Intrinsic::powi:
1063 case Intrinsic::bswap:
1064 case Intrinsic::ctpop:
1065 case Intrinsic::ctlz:
1066 case Intrinsic::cttz:
1067 case Intrinsic::sadd_with_overflow:
1068 case Intrinsic::uadd_with_overflow:
1069 case Intrinsic::ssub_with_overflow:
1070 case Intrinsic::usub_with_overflow:
1071 case Intrinsic::smul_with_overflow:
1072 case Intrinsic::umul_with_overflow:
1073 case Intrinsic::convert_from_fp16:
1074 case Intrinsic::convert_to_fp16:
1075 case Intrinsic::x86_sse_cvtss2si:
1076 case Intrinsic::x86_sse_cvtss2si64:
1077 case Intrinsic::x86_sse_cvttss2si:
1078 case Intrinsic::x86_sse_cvttss2si64:
1079 case Intrinsic::x86_sse2_cvtsd2si:
1080 case Intrinsic::x86_sse2_cvtsd2si64:
1081 case Intrinsic::x86_sse2_cvttsd2si:
1082 case Intrinsic::x86_sse2_cvttsd2si64:
1089 if (!F->hasName()) return false;
1090 StringRef Name = F->getName();
1092 // In these cases, the check of the length is required. We don't want to
1093 // return true for a name like "cos\0blah" which strcmp would return equal to
1094 // "cos", but has length 8.
1096 default: return false;
1098 return Name == "acos" || Name == "asin" ||
1099 Name == "atan" || Name == "atan2";
1101 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1103 return Name == "exp" || Name == "exp2";
1105 return Name == "fabs" || Name == "fmod" || Name == "floor";
1107 return Name == "log" || Name == "log10";
1109 return Name == "pow";
1111 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1112 Name == "sinf" || Name == "sqrtf";
1114 return Name == "tan" || Name == "tanh";
1118 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1120 sys::llvm_fenv_clearexcept();
1122 if (sys::llvm_fenv_testexcept()) {
1123 sys::llvm_fenv_clearexcept();
1127 if (Ty->isFloatTy())
1128 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1129 if (Ty->isDoubleTy())
1130 return ConstantFP::get(Ty->getContext(), APFloat(V));
1131 llvm_unreachable("Can only constant fold float/double");
1134 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1135 double V, double W, Type *Ty) {
1136 sys::llvm_fenv_clearexcept();
1138 if (sys::llvm_fenv_testexcept()) {
1139 sys::llvm_fenv_clearexcept();
1143 if (Ty->isFloatTy())
1144 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1145 if (Ty->isDoubleTy())
1146 return ConstantFP::get(Ty->getContext(), APFloat(V));
1147 llvm_unreachable("Can only constant fold float/double");
1150 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1151 /// conversion of a constant floating point. If roundTowardZero is false, the
1152 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1153 /// the behavior of the non-truncating SSE instructions in the default rounding
1154 /// mode. The desired integer type Ty is used to select how many bits are
1155 /// available for the result. Returns null if the conversion cannot be
1156 /// performed, otherwise returns the Constant value resulting from the
1158 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1159 bool roundTowardZero, Type *Ty) {
1160 // All of these conversion intrinsics form an integer of at most 64bits.
1161 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1162 assert(ResultWidth <= 64 &&
1163 "Can only constant fold conversions to 64 and 32 bit ints");
1166 bool isExact = false;
1167 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1168 : APFloat::rmNearestTiesToEven;
1169 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1170 /*isSigned=*/true, mode,
1172 if (status != APFloat::opOK && status != APFloat::opInexact)
1174 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1177 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1178 /// with the specified arguments, returning null if unsuccessful.
1180 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1181 const TargetLibraryInfo *TLI) {
1182 if (!F->hasName()) return 0;
1183 StringRef Name = F->getName();
1185 Type *Ty = F->getReturnType();
1186 if (Operands.size() == 1) {
1187 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1188 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1189 APFloat Val(Op->getValueAPF());
1192 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1194 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1199 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1202 /// We only fold functions with finite arguments. Folding NaN and inf is
1203 /// likely to be aborted with an exception anyway, and some host libms
1204 /// have known errors raising exceptions.
1205 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1208 /// Currently APFloat versions of these functions do not exist, so we use
1209 /// the host native double versions. Float versions are not called
1210 /// directly but for all these it is true (float)(f((double)arg)) ==
1211 /// f(arg). Long double not supported yet.
1212 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1213 Op->getValueAPF().convertToDouble();
1216 if (Name == "acos" && TLI->has(LibFunc::acos))
1217 return ConstantFoldFP(acos, V, Ty);
1218 else if (Name == "asin" && TLI->has(LibFunc::asin))
1219 return ConstantFoldFP(asin, V, Ty);
1220 else if (Name == "atan" && TLI->has(LibFunc::atan))
1221 return ConstantFoldFP(atan, V, Ty);
1224 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1225 return ConstantFoldFP(ceil, V, Ty);
1226 else if (Name == "cos" && TLI->has(LibFunc::cos))
1227 return ConstantFoldFP(cos, V, Ty);
1228 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1229 return ConstantFoldFP(cosh, V, Ty);
1230 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1231 return ConstantFoldFP(cos, V, Ty);
1234 if (Name == "exp" && TLI->has(LibFunc::exp))
1235 return ConstantFoldFP(exp, V, Ty);
1237 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1238 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1240 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1244 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1245 return ConstantFoldFP(fabs, V, Ty);
1246 else if (Name == "floor" && TLI->has(LibFunc::floor))
1247 return ConstantFoldFP(floor, V, Ty);
1250 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1251 return ConstantFoldFP(log, V, Ty);
1252 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1253 return ConstantFoldFP(log10, V, Ty);
1254 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1255 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1257 return ConstantFoldFP(sqrt, V, Ty);
1259 return Constant::getNullValue(Ty);
1263 if (Name == "sin" && TLI->has(LibFunc::sin))
1264 return ConstantFoldFP(sin, V, Ty);
1265 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1266 return ConstantFoldFP(sinh, V, Ty);
1267 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1268 return ConstantFoldFP(sqrt, V, Ty);
1269 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1270 return ConstantFoldFP(sqrt, V, Ty);
1271 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1272 return ConstantFoldFP(sin, V, Ty);
1275 if (Name == "tan" && TLI->has(LibFunc::tan))
1276 return ConstantFoldFP(tan, V, Ty);
1277 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1278 return ConstantFoldFP(tanh, V, Ty);
1286 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1287 switch (F->getIntrinsicID()) {
1288 case Intrinsic::bswap:
1289 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1290 case Intrinsic::ctpop:
1291 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1292 case Intrinsic::convert_from_fp16: {
1293 APFloat Val(Op->getValue());
1296 APFloat::opStatus status =
1297 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1299 // Conversion is always precise.
1301 assert(status == APFloat::opOK && !lost &&
1302 "Precision lost during fp16 constfolding");
1304 return ConstantFP::get(F->getContext(), Val);
1311 // Support ConstantVector in case we have an Undef in the top.
1312 if (isa<ConstantVector>(Operands[0]) ||
1313 isa<ConstantDataVector>(Operands[0])) {
1314 Constant *Op = cast<Constant>(Operands[0]);
1315 switch (F->getIntrinsicID()) {
1317 case Intrinsic::x86_sse_cvtss2si:
1318 case Intrinsic::x86_sse_cvtss2si64:
1319 case Intrinsic::x86_sse2_cvtsd2si:
1320 case Intrinsic::x86_sse2_cvtsd2si64:
1321 if (ConstantFP *FPOp =
1322 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1323 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1324 /*roundTowardZero=*/false, Ty);
1325 case Intrinsic::x86_sse_cvttss2si:
1326 case Intrinsic::x86_sse_cvttss2si64:
1327 case Intrinsic::x86_sse2_cvttsd2si:
1328 case Intrinsic::x86_sse2_cvttsd2si64:
1329 if (ConstantFP *FPOp =
1330 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1331 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1332 /*roundTowardZero=*/true, Ty);
1336 if (isa<UndefValue>(Operands[0])) {
1337 if (F->getIntrinsicID() == Intrinsic::bswap)
1345 if (Operands.size() == 2) {
1346 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1347 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1349 double Op1V = Ty->isFloatTy() ?
1350 (double)Op1->getValueAPF().convertToFloat() :
1351 Op1->getValueAPF().convertToDouble();
1352 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1353 if (Op2->getType() != Op1->getType())
1356 double Op2V = Ty->isFloatTy() ?
1357 (double)Op2->getValueAPF().convertToFloat():
1358 Op2->getValueAPF().convertToDouble();
1360 if (F->getIntrinsicID() == Intrinsic::pow) {
1361 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1365 if (Name == "pow" && TLI->has(LibFunc::pow))
1366 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1367 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1368 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1369 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1370 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1371 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1372 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1373 return ConstantFP::get(F->getContext(),
1374 APFloat((float)std::pow((float)Op1V,
1375 (int)Op2C->getZExtValue())));
1376 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1377 return ConstantFP::get(F->getContext(),
1378 APFloat((double)std::pow((double)Op1V,
1379 (int)Op2C->getZExtValue())));
1384 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1385 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1386 switch (F->getIntrinsicID()) {
1388 case Intrinsic::sadd_with_overflow:
1389 case Intrinsic::uadd_with_overflow:
1390 case Intrinsic::ssub_with_overflow:
1391 case Intrinsic::usub_with_overflow:
1392 case Intrinsic::smul_with_overflow:
1393 case Intrinsic::umul_with_overflow: {
1396 switch (F->getIntrinsicID()) {
1397 default: assert(0 && "Invalid case");
1398 case Intrinsic::sadd_with_overflow:
1399 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1401 case Intrinsic::uadd_with_overflow:
1402 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1404 case Intrinsic::ssub_with_overflow:
1405 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1407 case Intrinsic::usub_with_overflow:
1408 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1410 case Intrinsic::smul_with_overflow:
1411 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1413 case Intrinsic::umul_with_overflow:
1414 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1418 ConstantInt::get(F->getContext(), Res),
1419 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1421 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1423 case Intrinsic::cttz:
1424 // FIXME: This should check for Op2 == 1, and become unreachable if
1426 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1427 case Intrinsic::ctlz:
1428 // FIXME: This should check for Op2 == 1, and become unreachable if
1430 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());