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 // The code below only handles casts to vectors currently.
56 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
58 return ConstantExpr::getBitCast(C, DestTy);
60 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
61 // vector so the code below can handle it uniformly.
62 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
63 Constant *Ops = C; // don't take the address of C!
64 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
67 // If this is a bitcast from constant vector -> vector, fold it.
68 ConstantVector *CV = dyn_cast<ConstantVector>(C);
70 return ConstantExpr::getBitCast(C, DestTy);
72 // If the element types match, VMCore can fold it.
73 unsigned NumDstElt = DestVTy->getNumElements();
74 unsigned NumSrcElt = CV->getNumOperands();
75 if (NumDstElt == NumSrcElt)
76 return ConstantExpr::getBitCast(C, DestTy);
78 Type *SrcEltTy = CV->getType()->getElementType();
79 Type *DstEltTy = DestVTy->getElementType();
81 // Otherwise, we're changing the number of elements in a vector, which
82 // requires endianness information to do the right thing. For example,
83 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
84 // folds to (little endian):
85 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
86 // and to (big endian):
87 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
89 // First thing is first. We only want to think about integer here, so if
90 // we have something in FP form, recast it as integer.
91 if (DstEltTy->isFloatingPointTy()) {
92 // Fold to an vector of integers with same size as our FP type.
93 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
95 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
96 // Recursively handle this integer conversion, if possible.
97 C = FoldBitCast(C, DestIVTy, TD);
98 if (!C) return ConstantExpr::getBitCast(C, DestTy);
100 // Finally, VMCore can handle this now that #elts line up.
101 return ConstantExpr::getBitCast(C, DestTy);
104 // Okay, we know the destination is integer, if the input is FP, convert
105 // it to integer first.
106 if (SrcEltTy->isFloatingPointTy()) {
107 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
109 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
110 // Ask VMCore to do the conversion now that #elts line up.
111 C = ConstantExpr::getBitCast(C, SrcIVTy);
112 CV = dyn_cast<ConstantVector>(C);
113 if (!CV) // If VMCore wasn't able to fold it, bail out.
117 // Now we know that the input and output vectors are both integer vectors
118 // of the same size, and that their #elements is not the same. Do the
119 // conversion here, which depends on whether the input or output has
121 bool isLittleEndian = TD.isLittleEndian();
123 SmallVector<Constant*, 32> Result;
124 if (NumDstElt < NumSrcElt) {
125 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
126 Constant *Zero = Constant::getNullValue(DstEltTy);
127 unsigned Ratio = NumSrcElt/NumDstElt;
128 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
130 for (unsigned i = 0; i != NumDstElt; ++i) {
131 // Build each element of the result.
132 Constant *Elt = Zero;
133 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
134 for (unsigned j = 0; j != Ratio; ++j) {
135 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
136 if (!Src) // Reject constantexpr elements.
137 return ConstantExpr::getBitCast(C, DestTy);
139 // Zero extend the element to the right size.
140 Src = ConstantExpr::getZExt(Src, Elt->getType());
142 // Shift it to the right place, depending on endianness.
143 Src = ConstantExpr::getShl(Src,
144 ConstantInt::get(Src->getType(), ShiftAmt));
145 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
148 Elt = ConstantExpr::getOr(Elt, Src);
150 Result.push_back(Elt);
153 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
154 unsigned Ratio = NumDstElt/NumSrcElt;
155 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
157 // Loop over each source value, expanding into multiple results.
158 for (unsigned i = 0; i != NumSrcElt; ++i) {
159 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
160 if (!Src) // Reject constantexpr elements.
161 return ConstantExpr::getBitCast(C, DestTy);
163 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
164 for (unsigned j = 0; j != Ratio; ++j) {
165 // Shift the piece of the value into the right place, depending on
167 Constant *Elt = ConstantExpr::getLShr(Src,
168 ConstantInt::get(Src->getType(), ShiftAmt));
169 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
171 // Truncate and remember this piece.
172 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
177 return ConstantVector::get(Result);
181 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
182 /// from a global, return the global and the constant. Because of
183 /// constantexprs, this function is recursive.
184 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
185 int64_t &Offset, const TargetData &TD) {
186 // Trivial case, constant is the global.
187 if ((GV = dyn_cast<GlobalValue>(C))) {
192 // Otherwise, if this isn't a constant expr, bail out.
193 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
194 if (!CE) return false;
196 // Look through ptr->int and ptr->ptr casts.
197 if (CE->getOpcode() == Instruction::PtrToInt ||
198 CE->getOpcode() == Instruction::BitCast)
199 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
201 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
202 if (CE->getOpcode() == Instruction::GetElementPtr) {
203 // Cannot compute this if the element type of the pointer is missing size
205 if (!cast<PointerType>(CE->getOperand(0)->getType())
206 ->getElementType()->isSized())
209 // If the base isn't a global+constant, we aren't either.
210 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
213 // Otherwise, add any offset that our operands provide.
214 gep_type_iterator GTI = gep_type_begin(CE);
215 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
216 i != e; ++i, ++GTI) {
217 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
218 if (!CI) return false; // Index isn't a simple constant?
219 if (CI->isZero()) continue; // Not adding anything.
221 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
223 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
225 SequentialType *SQT = cast<SequentialType>(*GTI);
226 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
235 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
236 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
237 /// pointer to copy results into and BytesLeft is the number of bytes left in
238 /// the CurPtr buffer. TD is the target data.
239 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
240 unsigned char *CurPtr, unsigned BytesLeft,
241 const TargetData &TD) {
242 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
243 "Out of range access");
245 // If this element is zero or undefined, we can just return since *CurPtr is
247 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
250 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
251 if (CI->getBitWidth() > 64 ||
252 (CI->getBitWidth() & 7) != 0)
255 uint64_t Val = CI->getZExtValue();
256 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
258 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
259 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
265 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
266 if (CFP->getType()->isDoubleTy()) {
267 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
268 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
270 if (CFP->getType()->isFloatTy()){
271 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
272 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
277 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
278 const StructLayout *SL = TD.getStructLayout(CS->getType());
279 unsigned Index = SL->getElementContainingOffset(ByteOffset);
280 uint64_t CurEltOffset = SL->getElementOffset(Index);
281 ByteOffset -= CurEltOffset;
284 // If the element access is to the element itself and not to tail padding,
285 // read the bytes from the element.
286 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
288 if (ByteOffset < EltSize &&
289 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
295 // Check to see if we read from the last struct element, if so we're done.
296 if (Index == CS->getType()->getNumElements())
299 // If we read all of the bytes we needed from this element we're done.
300 uint64_t NextEltOffset = SL->getElementOffset(Index);
302 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
305 // Move to the next element of the struct.
306 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
307 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
309 CurEltOffset = NextEltOffset;
314 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
315 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
316 uint64_t Index = ByteOffset / EltSize;
317 uint64_t Offset = ByteOffset - Index * EltSize;
318 for (; Index != CA->getType()->getNumElements(); ++Index) {
319 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
322 if (EltSize >= BytesLeft)
326 BytesLeft -= EltSize;
332 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
333 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
334 uint64_t Index = ByteOffset / EltSize;
335 uint64_t Offset = ByteOffset - Index * EltSize;
336 for (; Index != CV->getType()->getNumElements(); ++Index) {
337 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
340 if (EltSize >= BytesLeft)
344 BytesLeft -= EltSize;
350 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(C)) {
351 uint64_t EltSize = CDS->getElementByteSize();
352 uint64_t Index = ByteOffset / EltSize;
353 uint64_t Offset = ByteOffset - Index * EltSize;
354 for (unsigned e = CDS->getNumElements(); Index != e; ++Index) {
355 if (!ReadDataFromGlobal(CDS->getElementAsConstant(Index), Offset, CurPtr,
358 if (EltSize >= BytesLeft)
362 BytesLeft -= EltSize;
368 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
369 if (CE->getOpcode() == Instruction::IntToPtr &&
370 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
371 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
375 // Otherwise, unknown initializer type.
379 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
380 const TargetData &TD) {
381 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
382 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
384 // If this isn't an integer load we can't fold it directly.
386 // If this is a float/double load, we can try folding it as an int32/64 load
387 // and then bitcast the result. This can be useful for union cases. Note
388 // that address spaces don't matter here since we're not going to result in
389 // an actual new load.
391 if (LoadTy->isFloatTy())
392 MapTy = Type::getInt32PtrTy(C->getContext());
393 else if (LoadTy->isDoubleTy())
394 MapTy = Type::getInt64PtrTy(C->getContext());
395 else if (LoadTy->isVectorTy()) {
396 MapTy = IntegerType::get(C->getContext(),
397 TD.getTypeAllocSizeInBits(LoadTy));
398 MapTy = PointerType::getUnqual(MapTy);
402 C = FoldBitCast(C, MapTy, TD);
403 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
404 return FoldBitCast(Res, LoadTy, TD);
408 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
409 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
413 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
416 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
417 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
418 !GV->getInitializer()->getType()->isSized())
421 // If we're loading off the beginning of the global, some bytes may be valid,
422 // but we don't try to handle this.
423 if (Offset < 0) return 0;
425 // If we're not accessing anything in this constant, the result is undefined.
426 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
427 return UndefValue::get(IntType);
429 unsigned char RawBytes[32] = {0};
430 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
434 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
435 for (unsigned i = 1; i != BytesLoaded; ++i) {
437 ResultVal |= RawBytes[BytesLoaded-1-i];
440 return ConstantInt::get(IntType->getContext(), ResultVal);
443 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
444 /// produce if it is constant and determinable. If this is not determinable,
446 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
447 const TargetData *TD) {
448 // First, try the easy cases:
449 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
450 if (GV->isConstant() && GV->hasDefinitiveInitializer())
451 return GV->getInitializer();
453 // If the loaded value isn't a constant expr, we can't handle it.
454 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
457 if (CE->getOpcode() == Instruction::GetElementPtr) {
458 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
459 if (GV->isConstant() && GV->hasDefinitiveInitializer())
461 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
465 // Instead of loading constant c string, use corresponding integer value
466 // directly if string length is small enough.
468 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
469 unsigned StrLen = Str.length();
470 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
471 unsigned NumBits = Ty->getPrimitiveSizeInBits();
472 // Replace load with immediate integer if the result is an integer or fp
474 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
475 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
476 APInt StrVal(NumBits, 0);
477 APInt SingleChar(NumBits, 0);
478 if (TD->isLittleEndian()) {
479 for (signed i = StrLen-1; i >= 0; i--) {
480 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
481 StrVal = (StrVal << 8) | SingleChar;
484 for (unsigned i = 0; i < StrLen; i++) {
485 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
486 StrVal = (StrVal << 8) | SingleChar;
488 // Append NULL at the end.
490 StrVal = (StrVal << 8) | SingleChar;
493 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
494 if (Ty->isFloatingPointTy())
495 Res = ConstantExpr::getBitCast(Res, Ty);
500 // If this load comes from anywhere in a constant global, and if the global
501 // is all undef or zero, we know what it loads.
502 if (GlobalVariable *GV =
503 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
504 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
505 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
506 if (GV->getInitializer()->isNullValue())
507 return Constant::getNullValue(ResTy);
508 if (isa<UndefValue>(GV->getInitializer()))
509 return UndefValue::get(ResTy);
513 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
514 // currently don't do any of this for big endian systems. It can be
515 // generalized in the future if someone is interested.
516 if (TD && TD->isLittleEndian())
517 return FoldReinterpretLoadFromConstPtr(CE, *TD);
521 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
522 if (LI->isVolatile()) return 0;
524 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
525 return ConstantFoldLoadFromConstPtr(C, TD);
530 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
531 /// Attempt to symbolically evaluate the result of a binary operator merging
532 /// these together. If target data info is available, it is provided as TD,
533 /// otherwise TD is null.
534 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
535 Constant *Op1, const TargetData *TD){
538 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
539 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
543 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
544 // constant. This happens frequently when iterating over a global array.
545 if (Opc == Instruction::Sub && TD) {
546 GlobalValue *GV1, *GV2;
547 int64_t Offs1, Offs2;
549 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
550 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
552 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
553 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
560 /// CastGEPIndices - If array indices are not pointer-sized integers,
561 /// explicitly cast them so that they aren't implicitly casted by the
563 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
564 Type *ResultTy, const TargetData *TD,
565 const TargetLibraryInfo *TLI) {
567 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
570 SmallVector<Constant*, 32> NewIdxs;
571 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
573 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
574 Ops.slice(1, i-1)))) &&
575 Ops[i]->getType() != IntPtrTy) {
577 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
583 NewIdxs.push_back(Ops[i]);
588 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
589 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
590 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
595 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
596 /// constant expression, do so.
597 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
598 Type *ResultTy, const TargetData *TD,
599 const TargetLibraryInfo *TLI) {
600 Constant *Ptr = Ops[0];
601 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
602 !Ptr->getType()->isPointerTy())
605 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
607 // If this is a constant expr gep that is effectively computing an
608 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
609 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
610 if (!isa<ConstantInt>(Ops[i])) {
612 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
613 // "inttoptr (sub (ptrtoint Ptr), V)"
614 if (Ops.size() == 2 &&
615 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
616 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
617 assert((CE == 0 || CE->getType() == IntPtrTy) &&
618 "CastGEPIndices didn't canonicalize index types!");
619 if (CE && CE->getOpcode() == Instruction::Sub &&
620 CE->getOperand(0)->isNullValue()) {
621 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
622 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
623 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
624 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
625 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
632 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
634 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
635 makeArrayRef((Value **)Ops.data() + 1,
637 Ptr = cast<Constant>(Ptr->stripPointerCasts());
639 // If this is a GEP of a GEP, fold it all into a single GEP.
640 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
641 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
643 // Do not try the incorporate the sub-GEP if some index is not a number.
644 bool AllConstantInt = true;
645 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
646 if (!isa<ConstantInt>(NestedOps[i])) {
647 AllConstantInt = false;
653 Ptr = cast<Constant>(GEP->getOperand(0));
654 Offset += APInt(BitWidth,
655 TD->getIndexedOffset(Ptr->getType(), NestedOps));
656 Ptr = cast<Constant>(Ptr->stripPointerCasts());
659 // If the base value for this address is a literal integer value, fold the
660 // getelementptr to the resulting integer value casted to the pointer type.
661 APInt BasePtr(BitWidth, 0);
662 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
663 if (CE->getOpcode() == Instruction::IntToPtr)
664 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
665 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
666 if (Ptr->isNullValue() || BasePtr != 0) {
667 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
668 return ConstantExpr::getIntToPtr(C, ResultTy);
671 // Otherwise form a regular getelementptr. Recompute the indices so that
672 // we eliminate over-indexing of the notional static type array bounds.
673 // This makes it easy to determine if the getelementptr is "inbounds".
674 // Also, this helps GlobalOpt do SROA on GlobalVariables.
675 Type *Ty = Ptr->getType();
676 SmallVector<Constant*, 32> NewIdxs;
678 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
679 if (ATy->isPointerTy()) {
680 // The only pointer indexing we'll do is on the first index of the GEP.
681 if (!NewIdxs.empty())
684 // Only handle pointers to sized types, not pointers to functions.
685 if (!ATy->getElementType()->isSized())
689 // Determine which element of the array the offset points into.
690 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
691 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
693 // The element size is 0. This may be [0 x Ty]*, so just use a zero
694 // index for this level and proceed to the next level to see if it can
695 // accommodate the offset.
696 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
698 // The element size is non-zero divide the offset by the element
699 // size (rounding down), to compute the index at this level.
700 APInt NewIdx = Offset.udiv(ElemSize);
701 Offset -= NewIdx * ElemSize;
702 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
704 Ty = ATy->getElementType();
705 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
706 // Determine which field of the struct the offset points into. The
707 // getZExtValue is at least as safe as the StructLayout API because we
708 // know the offset is within the struct at this point.
709 const StructLayout &SL = *TD->getStructLayout(STy);
710 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
711 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
713 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
714 Ty = STy->getTypeAtIndex(ElIdx);
716 // We've reached some non-indexable type.
719 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
721 // If we haven't used up the entire offset by descending the static
722 // type, then the offset is pointing into the middle of an indivisible
723 // member, so we can't simplify it.
729 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
730 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
731 "Computed GetElementPtr has unexpected type!");
733 // If we ended up indexing a member with a type that doesn't match
734 // the type of what the original indices indexed, add a cast.
735 if (Ty != cast<PointerType>(ResultTy)->getElementType())
736 C = FoldBitCast(C, ResultTy, *TD);
743 //===----------------------------------------------------------------------===//
744 // Constant Folding public APIs
745 //===----------------------------------------------------------------------===//
747 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
748 /// If successful, the constant result is returned, if not, null is returned.
749 /// Note that this fails if not all of the operands are constant. Otherwise,
750 /// this function can only fail when attempting to fold instructions like loads
751 /// and stores, which have no constant expression form.
752 Constant *llvm::ConstantFoldInstruction(Instruction *I,
753 const TargetData *TD,
754 const TargetLibraryInfo *TLI) {
755 // Handle PHI nodes quickly here...
756 if (PHINode *PN = dyn_cast<PHINode>(I)) {
757 Constant *CommonValue = 0;
759 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
760 Value *Incoming = PN->getIncomingValue(i);
761 // If the incoming value is undef then skip it. Note that while we could
762 // skip the value if it is equal to the phi node itself we choose not to
763 // because that would break the rule that constant folding only applies if
764 // all operands are constants.
765 if (isa<UndefValue>(Incoming))
767 // If the incoming value is not a constant, or is a different constant to
768 // the one we saw previously, then give up.
769 Constant *C = dyn_cast<Constant>(Incoming);
770 if (!C || (CommonValue && C != CommonValue))
775 // If we reach here, all incoming values are the same constant or undef.
776 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
779 // Scan the operand list, checking to see if they are all constants, if so,
780 // hand off to ConstantFoldInstOperands.
781 SmallVector<Constant*, 8> Ops;
782 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
783 if (Constant *Op = dyn_cast<Constant>(*i))
786 return 0; // All operands not constant!
788 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
789 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
792 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
793 return ConstantFoldLoadInst(LI, TD);
795 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
796 return ConstantExpr::getInsertValue(
797 cast<Constant>(IVI->getAggregateOperand()),
798 cast<Constant>(IVI->getInsertedValueOperand()),
801 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
802 return ConstantExpr::getExtractValue(
803 cast<Constant>(EVI->getAggregateOperand()),
806 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
809 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
810 /// using the specified TargetData. If successful, the constant result is
811 /// result is returned, if not, null is returned.
812 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
813 const TargetData *TD,
814 const TargetLibraryInfo *TLI) {
815 SmallVector<Constant*, 8> Ops;
816 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
818 Constant *NewC = cast<Constant>(*i);
819 // Recursively fold the ConstantExpr's operands.
820 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
821 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
826 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
828 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
831 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
832 /// specified opcode and operands. If successful, the constant result is
833 /// returned, if not, null is returned. Note that this function can fail when
834 /// attempting to fold instructions like loads and stores, which have no
835 /// constant expression form.
837 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
838 /// information, due to only being passed an opcode and operands. Constant
839 /// folding using this function strips this information.
841 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
842 ArrayRef<Constant *> Ops,
843 const TargetData *TD,
844 const TargetLibraryInfo *TLI) {
845 // Handle easy binops first.
846 if (Instruction::isBinaryOp(Opcode)) {
847 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
848 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
851 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
856 case Instruction::ICmp:
857 case Instruction::FCmp: assert(0 && "Invalid for compares");
858 case Instruction::Call:
859 if (Function *F = dyn_cast<Function>(Ops.back()))
860 if (canConstantFoldCallTo(F))
861 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
863 case Instruction::PtrToInt:
864 // If the input is a inttoptr, eliminate the pair. This requires knowing
865 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
866 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
867 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
868 Constant *Input = CE->getOperand(0);
869 unsigned InWidth = Input->getType()->getScalarSizeInBits();
870 if (TD->getPointerSizeInBits() < InWidth) {
872 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
873 TD->getPointerSizeInBits()));
874 Input = ConstantExpr::getAnd(Input, Mask);
876 // Do a zext or trunc to get to the dest size.
877 return ConstantExpr::getIntegerCast(Input, DestTy, false);
880 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
881 case Instruction::IntToPtr:
882 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
883 // the int size is >= the ptr size. This requires knowing the width of a
884 // pointer, so it can't be done in ConstantExpr::getCast.
885 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
887 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
888 CE->getOpcode() == Instruction::PtrToInt)
889 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
891 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
892 case Instruction::Trunc:
893 case Instruction::ZExt:
894 case Instruction::SExt:
895 case Instruction::FPTrunc:
896 case Instruction::FPExt:
897 case Instruction::UIToFP:
898 case Instruction::SIToFP:
899 case Instruction::FPToUI:
900 case Instruction::FPToSI:
901 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
902 case Instruction::BitCast:
904 return FoldBitCast(Ops[0], DestTy, *TD);
905 return ConstantExpr::getBitCast(Ops[0], DestTy);
906 case Instruction::Select:
907 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
908 case Instruction::ExtractElement:
909 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
910 case Instruction::InsertElement:
911 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
912 case Instruction::ShuffleVector:
913 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
914 case Instruction::GetElementPtr:
915 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
917 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
920 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
924 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
925 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
926 /// returns a constant expression of the specified operands.
928 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
929 Constant *Ops0, Constant *Ops1,
930 const TargetData *TD,
931 const TargetLibraryInfo *TLI) {
932 // fold: icmp (inttoptr x), null -> icmp x, 0
933 // fold: icmp (ptrtoint x), 0 -> icmp x, null
934 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
935 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
937 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
938 // around to know if bit truncation is happening.
939 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
940 if (TD && Ops1->isNullValue()) {
941 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
942 if (CE0->getOpcode() == Instruction::IntToPtr) {
943 // Convert the integer value to the right size to ensure we get the
944 // proper extension or truncation.
945 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
947 Constant *Null = Constant::getNullValue(C->getType());
948 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
951 // Only do this transformation if the int is intptrty in size, otherwise
952 // there is a truncation or extension that we aren't modeling.
953 if (CE0->getOpcode() == Instruction::PtrToInt &&
954 CE0->getType() == IntPtrTy) {
955 Constant *C = CE0->getOperand(0);
956 Constant *Null = Constant::getNullValue(C->getType());
957 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
961 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
962 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
963 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
965 if (CE0->getOpcode() == Instruction::IntToPtr) {
966 // Convert the integer value to the right size to ensure we get the
967 // proper extension or truncation.
968 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
970 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
972 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
975 // Only do this transformation if the int is intptrty in size, otherwise
976 // there is a truncation or extension that we aren't modeling.
977 if ((CE0->getOpcode() == Instruction::PtrToInt &&
978 CE0->getType() == IntPtrTy &&
979 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
980 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
981 CE1->getOperand(0), TD, TLI);
985 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
986 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
987 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
988 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
990 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
993 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
996 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
997 Constant *Ops[] = { LHS, RHS };
998 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1002 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1006 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1007 /// getelementptr constantexpr, return the constant value being addressed by the
1008 /// constant expression, or null if something is funny and we can't decide.
1009 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1011 if (!CE->getOperand(1)->isNullValue())
1012 return 0; // Do not allow stepping over the value!
1014 SmallVector<Constant*, 8> Indices(CE->getNumOperands()-2);
1015 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i)
1016 Indices[i-2] = CE->getOperand(i);
1017 return ConstantFoldLoadThroughGEPIndices(C, Indices);
1020 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1021 /// indices (with an *implied* zero pointer index that is not in the list),
1022 /// return the constant value being addressed by a virtual load, or null if
1023 /// something is funny and we can't decide.
1024 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1025 ArrayRef<Constant*> Indices) {
1026 // Loop over all of the operands, tracking down which value we are
1028 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1029 ConstantInt *Idx = dyn_cast<ConstantInt>(Indices[i]);
1030 if (Idx == 0) return 0;
1032 uint64_t IdxVal = Idx->getZExtValue();
1034 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1035 C = CS->getOperand(IdxVal);
1036 } else if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(C)){
1037 C = CAZ->getElementValue(Idx);
1038 } else if (UndefValue *UV = dyn_cast<UndefValue>(C)) {
1039 C = UV->getElementValue(Idx);
1040 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
1041 if (IdxVal >= CA->getType()->getNumElements())
1043 C = CA->getOperand(IdxVal);
1044 } else if (ConstantDataSequential *CDS=dyn_cast<ConstantDataSequential>(C)){
1045 if (IdxVal >= CDS->getNumElements())
1047 C = CDS->getElementAsConstant(IdxVal);
1048 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1049 if (IdxVal >= CV->getType()->getNumElements())
1051 C = CV->getOperand(IdxVal);
1060 //===----------------------------------------------------------------------===//
1061 // Constant Folding for Calls
1064 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1065 /// the specified function.
1067 llvm::canConstantFoldCallTo(const Function *F) {
1068 switch (F->getIntrinsicID()) {
1069 case Intrinsic::sqrt:
1070 case Intrinsic::pow:
1071 case Intrinsic::powi:
1072 case Intrinsic::bswap:
1073 case Intrinsic::ctpop:
1074 case Intrinsic::ctlz:
1075 case Intrinsic::cttz:
1076 case Intrinsic::sadd_with_overflow:
1077 case Intrinsic::uadd_with_overflow:
1078 case Intrinsic::ssub_with_overflow:
1079 case Intrinsic::usub_with_overflow:
1080 case Intrinsic::smul_with_overflow:
1081 case Intrinsic::umul_with_overflow:
1082 case Intrinsic::convert_from_fp16:
1083 case Intrinsic::convert_to_fp16:
1084 case Intrinsic::x86_sse_cvtss2si:
1085 case Intrinsic::x86_sse_cvtss2si64:
1086 case Intrinsic::x86_sse_cvttss2si:
1087 case Intrinsic::x86_sse_cvttss2si64:
1088 case Intrinsic::x86_sse2_cvtsd2si:
1089 case Intrinsic::x86_sse2_cvtsd2si64:
1090 case Intrinsic::x86_sse2_cvttsd2si:
1091 case Intrinsic::x86_sse2_cvttsd2si64:
1098 if (!F->hasName()) return false;
1099 StringRef Name = F->getName();
1101 // In these cases, the check of the length is required. We don't want to
1102 // return true for a name like "cos\0blah" which strcmp would return equal to
1103 // "cos", but has length 8.
1105 default: return false;
1107 return Name == "acos" || Name == "asin" ||
1108 Name == "atan" || Name == "atan2";
1110 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1112 return Name == "exp" || Name == "exp2";
1114 return Name == "fabs" || Name == "fmod" || Name == "floor";
1116 return Name == "log" || Name == "log10";
1118 return Name == "pow";
1120 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1121 Name == "sinf" || Name == "sqrtf";
1123 return Name == "tan" || Name == "tanh";
1127 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1129 sys::llvm_fenv_clearexcept();
1131 if (sys::llvm_fenv_testexcept()) {
1132 sys::llvm_fenv_clearexcept();
1136 if (Ty->isFloatTy())
1137 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1138 if (Ty->isDoubleTy())
1139 return ConstantFP::get(Ty->getContext(), APFloat(V));
1140 llvm_unreachable("Can only constant fold float/double");
1143 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1144 double V, double W, Type *Ty) {
1145 sys::llvm_fenv_clearexcept();
1147 if (sys::llvm_fenv_testexcept()) {
1148 sys::llvm_fenv_clearexcept();
1152 if (Ty->isFloatTy())
1153 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1154 if (Ty->isDoubleTy())
1155 return ConstantFP::get(Ty->getContext(), APFloat(V));
1156 llvm_unreachable("Can only constant fold float/double");
1159 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1160 /// conversion of a constant floating point. If roundTowardZero is false, the
1161 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1162 /// the behavior of the non-truncating SSE instructions in the default rounding
1163 /// mode. The desired integer type Ty is used to select how many bits are
1164 /// available for the result. Returns null if the conversion cannot be
1165 /// performed, otherwise returns the Constant value resulting from the
1167 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1169 assert(Op && "Called with NULL operand");
1170 APFloat Val(Op->getValueAPF());
1172 // All of these conversion intrinsics form an integer of at most 64bits.
1173 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1174 assert(ResultWidth <= 64 &&
1175 "Can only constant fold conversions to 64 and 32 bit ints");
1178 bool isExact = false;
1179 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1180 : APFloat::rmNearestTiesToEven;
1181 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1182 /*isSigned=*/true, mode,
1184 if (status != APFloat::opOK && status != APFloat::opInexact)
1186 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1189 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1190 /// with the specified arguments, returning null if unsuccessful.
1192 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1193 const TargetLibraryInfo *TLI) {
1194 if (!F->hasName()) return 0;
1195 StringRef Name = F->getName();
1197 Type *Ty = F->getReturnType();
1198 if (Operands.size() == 1) {
1199 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1200 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1201 APFloat Val(Op->getValueAPF());
1204 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1206 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1211 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1214 /// We only fold functions with finite arguments. Folding NaN and inf is
1215 /// likely to be aborted with an exception anyway, and some host libms
1216 /// have known errors raising exceptions.
1217 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1220 /// Currently APFloat versions of these functions do not exist, so we use
1221 /// the host native double versions. Float versions are not called
1222 /// directly but for all these it is true (float)(f((double)arg)) ==
1223 /// f(arg). Long double not supported yet.
1224 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1225 Op->getValueAPF().convertToDouble();
1228 if (Name == "acos" && TLI->has(LibFunc::acos))
1229 return ConstantFoldFP(acos, V, Ty);
1230 else if (Name == "asin" && TLI->has(LibFunc::asin))
1231 return ConstantFoldFP(asin, V, Ty);
1232 else if (Name == "atan" && TLI->has(LibFunc::atan))
1233 return ConstantFoldFP(atan, V, Ty);
1236 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1237 return ConstantFoldFP(ceil, V, Ty);
1238 else if (Name == "cos" && TLI->has(LibFunc::cos))
1239 return ConstantFoldFP(cos, V, Ty);
1240 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1241 return ConstantFoldFP(cosh, V, Ty);
1242 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1243 return ConstantFoldFP(cos, V, Ty);
1246 if (Name == "exp" && TLI->has(LibFunc::exp))
1247 return ConstantFoldFP(exp, V, Ty);
1249 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1250 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1252 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1256 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1257 return ConstantFoldFP(fabs, V, Ty);
1258 else if (Name == "floor" && TLI->has(LibFunc::floor))
1259 return ConstantFoldFP(floor, V, Ty);
1262 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1263 return ConstantFoldFP(log, V, Ty);
1264 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1265 return ConstantFoldFP(log10, V, Ty);
1266 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1267 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1269 return ConstantFoldFP(sqrt, V, Ty);
1271 return Constant::getNullValue(Ty);
1275 if (Name == "sin" && TLI->has(LibFunc::sin))
1276 return ConstantFoldFP(sin, V, Ty);
1277 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1278 return ConstantFoldFP(sinh, V, Ty);
1279 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1280 return ConstantFoldFP(sqrt, V, Ty);
1281 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1282 return ConstantFoldFP(sqrt, V, Ty);
1283 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1284 return ConstantFoldFP(sin, V, Ty);
1287 if (Name == "tan" && TLI->has(LibFunc::tan))
1288 return ConstantFoldFP(tan, V, Ty);
1289 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1290 return ConstantFoldFP(tanh, V, Ty);
1298 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1299 switch (F->getIntrinsicID()) {
1300 case Intrinsic::bswap:
1301 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1302 case Intrinsic::ctpop:
1303 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1304 case Intrinsic::convert_from_fp16: {
1305 APFloat Val(Op->getValue());
1308 APFloat::opStatus status =
1309 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1311 // Conversion is always precise.
1313 assert(status == APFloat::opOK && !lost &&
1314 "Precision lost during fp16 constfolding");
1316 return ConstantFP::get(F->getContext(), Val);
1323 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1324 switch (F->getIntrinsicID()) {
1326 case Intrinsic::x86_sse_cvtss2si:
1327 case Intrinsic::x86_sse_cvtss2si64:
1328 case Intrinsic::x86_sse2_cvtsd2si:
1329 case Intrinsic::x86_sse2_cvtsd2si64:
1330 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1331 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1332 case Intrinsic::x86_sse_cvttss2si:
1333 case Intrinsic::x86_sse_cvttss2si64:
1334 case Intrinsic::x86_sse2_cvttsd2si:
1335 case Intrinsic::x86_sse2_cvttsd2si64:
1336 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1337 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1341 if (isa<UndefValue>(Operands[0])) {
1342 if (F->getIntrinsicID() == Intrinsic::bswap)
1350 if (Operands.size() == 2) {
1351 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1352 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1354 double Op1V = Ty->isFloatTy() ?
1355 (double)Op1->getValueAPF().convertToFloat() :
1356 Op1->getValueAPF().convertToDouble();
1357 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1358 if (Op2->getType() != Op1->getType())
1361 double Op2V = Ty->isFloatTy() ?
1362 (double)Op2->getValueAPF().convertToFloat():
1363 Op2->getValueAPF().convertToDouble();
1365 if (F->getIntrinsicID() == Intrinsic::pow) {
1366 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1370 if (Name == "pow" && TLI->has(LibFunc::pow))
1371 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1372 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1373 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1374 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1375 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1376 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1377 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1378 return ConstantFP::get(F->getContext(),
1379 APFloat((float)std::pow((float)Op1V,
1380 (int)Op2C->getZExtValue())));
1381 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1382 return ConstantFP::get(F->getContext(),
1383 APFloat((double)std::pow((double)Op1V,
1384 (int)Op2C->getZExtValue())));
1389 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1390 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1391 switch (F->getIntrinsicID()) {
1393 case Intrinsic::sadd_with_overflow:
1394 case Intrinsic::uadd_with_overflow:
1395 case Intrinsic::ssub_with_overflow:
1396 case Intrinsic::usub_with_overflow:
1397 case Intrinsic::smul_with_overflow:
1398 case Intrinsic::umul_with_overflow: {
1401 switch (F->getIntrinsicID()) {
1402 default: assert(0 && "Invalid case");
1403 case Intrinsic::sadd_with_overflow:
1404 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1406 case Intrinsic::uadd_with_overflow:
1407 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1409 case Intrinsic::ssub_with_overflow:
1410 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1412 case Intrinsic::usub_with_overflow:
1413 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1415 case Intrinsic::smul_with_overflow:
1416 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1418 case Intrinsic::umul_with_overflow:
1419 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1423 ConstantInt::get(F->getContext(), Res),
1424 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1426 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1428 case Intrinsic::cttz:
1429 // FIXME: This should check for Op2 == 1, and become unreachable if
1431 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1432 case Intrinsic::ctlz:
1433 // FIXME: This should check for Op2 == 1, and become unreachable if
1435 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());