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 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
59 return ConstantExpr::getBitCast(C, DestTy);
61 unsigned NumSrcElts = CDV->getType()->getNumElements();
63 Type *SrcEltTy = CDV->getType()->getElementType();
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (SrcEltTy->isFloatingPointTy()) {
68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
71 // Ask VMCore to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
73 CDV = cast<ConstantDataVector>(C);
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
79 APInt Result(IT->getBitWidth(), 0);
80 for (unsigned i = 0; i != NumSrcElts; ++i) {
82 if (TD.isLittleEndian())
83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
85 Result |= CDV->getElementAsInteger(i);
88 return ConstantInt::get(IT, Result);
91 // The code below only handles casts to vectors currently.
92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
94 return ConstantExpr::getBitCast(C, DestTy);
96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
97 // vector so the code below can handle it uniformly.
98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
99 Constant *Ops = C; // don't take the address of C!
100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
103 // If this is a bitcast from constant vector -> vector, fold it.
104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
105 return ConstantExpr::getBitCast(C, DestTy);
107 // If the element types match, VMCore can fold it.
108 unsigned NumDstElt = DestVTy->getNumElements();
109 unsigned NumSrcElt = C->getType()->getVectorNumElements();
110 if (NumDstElt == NumSrcElt)
111 return ConstantExpr::getBitCast(C, DestTy);
113 Type *SrcEltTy = C->getType()->getVectorElementType();
114 Type *DstEltTy = DestVTy->getElementType();
116 // Otherwise, we're changing the number of elements in a vector, which
117 // requires endianness information to do the right thing. For example,
118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
119 // folds to (little endian):
120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
121 // and to (big endian):
122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
124 // First thing is first. We only want to think about integer here, so if
125 // we have something in FP form, recast it as integer.
126 if (DstEltTy->isFloatingPointTy()) {
127 // Fold to an vector of integers with same size as our FP type.
128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
131 // Recursively handle this integer conversion, if possible.
132 C = FoldBitCast(C, DestIVTy, TD);
134 // Finally, VMCore can handle this now that #elts line up.
135 return ConstantExpr::getBitCast(C, DestTy);
138 // Okay, we know the destination is integer, if the input is FP, convert
139 // it to integer first.
140 if (SrcEltTy->isFloatingPointTy()) {
141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
144 // Ask VMCore to do the conversion now that #elts line up.
145 C = ConstantExpr::getBitCast(C, SrcIVTy);
146 // If VMCore wasn't able to fold it, bail out.
147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
148 !isa<ConstantDataVector>(C))
152 // Now we know that the input and output vectors are both integer vectors
153 // of the same size, and that their #elements is not the same. Do the
154 // conversion here, which depends on whether the input or output has
156 bool isLittleEndian = TD.isLittleEndian();
158 SmallVector<Constant*, 32> Result;
159 if (NumDstElt < NumSrcElt) {
160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
161 Constant *Zero = Constant::getNullValue(DstEltTy);
162 unsigned Ratio = NumSrcElt/NumDstElt;
163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
165 for (unsigned i = 0; i != NumDstElt; ++i) {
166 // Build each element of the result.
167 Constant *Elt = Zero;
168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
169 for (unsigned j = 0; j != Ratio; ++j) {
170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
171 if (!Src) // Reject constantexpr elements.
172 return ConstantExpr::getBitCast(C, DestTy);
174 // Zero extend the element to the right size.
175 Src = ConstantExpr::getZExt(Src, Elt->getType());
177 // Shift it to the right place, depending on endianness.
178 Src = ConstantExpr::getShl(Src,
179 ConstantInt::get(Src->getType(), ShiftAmt));
180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
183 Elt = ConstantExpr::getOr(Elt, Src);
185 Result.push_back(Elt);
187 return ConstantVector::get(Result);
190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
191 unsigned Ratio = NumDstElt/NumSrcElt;
192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
194 // Loop over each source value, expanding into multiple results.
195 for (unsigned i = 0; i != NumSrcElt; ++i) {
196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
197 if (!Src) // Reject constantexpr elements.
198 return ConstantExpr::getBitCast(C, DestTy);
200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
201 for (unsigned j = 0; j != Ratio; ++j) {
202 // Shift the piece of the value into the right place, depending on
204 Constant *Elt = ConstantExpr::getLShr(Src,
205 ConstantInt::get(Src->getType(), ShiftAmt));
206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
208 // Truncate and remember this piece.
209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
213 return ConstantVector::get(Result);
217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
218 /// from a global, return the global and the constant. Because of
219 /// constantexprs, this function is recursive.
220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
221 int64_t &Offset, const TargetData &TD) {
222 // Trivial case, constant is the global.
223 if ((GV = dyn_cast<GlobalValue>(C))) {
228 // Otherwise, if this isn't a constant expr, bail out.
229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
230 if (!CE) return false;
232 // Look through ptr->int and ptr->ptr casts.
233 if (CE->getOpcode() == Instruction::PtrToInt ||
234 CE->getOpcode() == Instruction::BitCast)
235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
238 if (CE->getOpcode() == Instruction::GetElementPtr) {
239 // Cannot compute this if the element type of the pointer is missing size
241 if (!cast<PointerType>(CE->getOperand(0)->getType())
242 ->getElementType()->isSized())
245 // If the base isn't a global+constant, we aren't either.
246 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
249 // Otherwise, add any offset that our operands provide.
250 gep_type_iterator GTI = gep_type_begin(CE);
251 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
252 i != e; ++i, ++GTI) {
253 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
254 if (!CI) return false; // Index isn't a simple constant?
255 if (CI->isZero()) continue; // Not adding anything.
257 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
259 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
261 SequentialType *SQT = cast<SequentialType>(*GTI);
262 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
271 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
272 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
273 /// pointer to copy results into and BytesLeft is the number of bytes left in
274 /// the CurPtr buffer. TD is the target data.
275 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
276 unsigned char *CurPtr, unsigned BytesLeft,
277 const TargetData &TD) {
278 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
279 "Out of range access");
281 // If this element is zero or undefined, we can just return since *CurPtr is
283 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
286 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
287 if (CI->getBitWidth() > 64 ||
288 (CI->getBitWidth() & 7) != 0)
291 uint64_t Val = CI->getZExtValue();
292 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
294 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
295 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
301 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
302 if (CFP->getType()->isDoubleTy()) {
303 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
304 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
306 if (CFP->getType()->isFloatTy()){
307 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
308 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
313 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
314 const StructLayout *SL = TD.getStructLayout(CS->getType());
315 unsigned Index = SL->getElementContainingOffset(ByteOffset);
316 uint64_t CurEltOffset = SL->getElementOffset(Index);
317 ByteOffset -= CurEltOffset;
320 // If the element access is to the element itself and not to tail padding,
321 // read the bytes from the element.
322 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
324 if (ByteOffset < EltSize &&
325 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
331 // Check to see if we read from the last struct element, if so we're done.
332 if (Index == CS->getType()->getNumElements())
335 // If we read all of the bytes we needed from this element we're done.
336 uint64_t NextEltOffset = SL->getElementOffset(Index);
338 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
341 // Move to the next element of the struct.
342 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
343 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
345 CurEltOffset = NextEltOffset;
350 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
351 isa<ConstantDataSequential>(C)) {
352 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
353 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
354 uint64_t Index = ByteOffset / EltSize;
355 uint64_t Offset = ByteOffset - Index * EltSize;
357 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
358 NumElts = AT->getNumElements();
360 NumElts = cast<VectorType>(C->getType())->getNumElements();
362 for (; Index != NumElts; ++Index) {
363 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
367 uint64_t BytesWritten = EltSize - Offset;
368 assert(BytesWritten <= EltSize && "Not indexing into this element?");
369 if (BytesWritten >= BytesLeft)
373 BytesLeft -= BytesWritten;
374 CurPtr += BytesWritten;
379 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
380 if (CE->getOpcode() == Instruction::IntToPtr &&
381 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
382 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
386 // Otherwise, unknown initializer type.
390 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
391 const TargetData &TD) {
392 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
393 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
395 // If this isn't an integer load we can't fold it directly.
397 // If this is a float/double load, we can try folding it as an int32/64 load
398 // and then bitcast the result. This can be useful for union cases. Note
399 // that address spaces don't matter here since we're not going to result in
400 // an actual new load.
402 if (LoadTy->isFloatTy())
403 MapTy = Type::getInt32PtrTy(C->getContext());
404 else if (LoadTy->isDoubleTy())
405 MapTy = Type::getInt64PtrTy(C->getContext());
406 else if (LoadTy->isVectorTy()) {
407 MapTy = IntegerType::get(C->getContext(),
408 TD.getTypeAllocSizeInBits(LoadTy));
409 MapTy = PointerType::getUnqual(MapTy);
413 C = FoldBitCast(C, MapTy, TD);
414 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
415 return FoldBitCast(Res, LoadTy, TD);
419 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
420 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
424 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
427 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
428 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
429 !GV->getInitializer()->getType()->isSized())
432 // If we're loading off the beginning of the global, some bytes may be valid,
433 // but we don't try to handle this.
434 if (Offset < 0) return 0;
436 // If we're not accessing anything in this constant, the result is undefined.
437 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
438 return UndefValue::get(IntType);
440 unsigned char RawBytes[32] = {0};
441 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
445 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
446 for (unsigned i = 1; i != BytesLoaded; ++i) {
448 ResultVal |= RawBytes[BytesLoaded-1-i];
451 return ConstantInt::get(IntType->getContext(), ResultVal);
454 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
455 /// produce if it is constant and determinable. If this is not determinable,
457 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
458 const TargetData *TD) {
459 // First, try the easy cases:
460 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
461 if (GV->isConstant() && GV->hasDefinitiveInitializer())
462 return GV->getInitializer();
464 // If the loaded value isn't a constant expr, we can't handle it.
465 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
468 if (CE->getOpcode() == Instruction::GetElementPtr) {
469 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
470 if (GV->isConstant() && GV->hasDefinitiveInitializer())
472 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
476 // Instead of loading constant c string, use corresponding integer value
477 // directly if string length is small enough.
479 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
480 unsigned StrLen = Str.size();
481 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
482 unsigned NumBits = Ty->getPrimitiveSizeInBits();
483 // Replace load with immediate integer if the result is an integer or fp
485 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
486 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
487 APInt StrVal(NumBits, 0);
488 APInt SingleChar(NumBits, 0);
489 if (TD->isLittleEndian()) {
490 for (signed i = StrLen-1; i >= 0; i--) {
491 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
492 StrVal = (StrVal << 8) | SingleChar;
495 for (unsigned i = 0; i < StrLen; i++) {
496 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
497 StrVal = (StrVal << 8) | SingleChar;
499 // Append NULL at the end.
501 StrVal = (StrVal << 8) | SingleChar;
504 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
505 if (Ty->isFloatingPointTy())
506 Res = ConstantExpr::getBitCast(Res, Ty);
511 // If this load comes from anywhere in a constant global, and if the global
512 // is all undef or zero, we know what it loads.
513 if (GlobalVariable *GV =
514 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
515 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
516 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
517 if (GV->getInitializer()->isNullValue())
518 return Constant::getNullValue(ResTy);
519 if (isa<UndefValue>(GV->getInitializer()))
520 return UndefValue::get(ResTy);
524 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
525 // currently don't do any of this for big endian systems. It can be
526 // generalized in the future if someone is interested.
527 if (TD && TD->isLittleEndian())
528 return FoldReinterpretLoadFromConstPtr(CE, *TD);
532 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
533 if (LI->isVolatile()) return 0;
535 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
536 return ConstantFoldLoadFromConstPtr(C, TD);
541 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
542 /// Attempt to symbolically evaluate the result of a binary operator merging
543 /// these together. If target data info is available, it is provided as TD,
544 /// otherwise TD is null.
545 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
546 Constant *Op1, const TargetData *TD){
549 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
550 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
554 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
555 // constant. This happens frequently when iterating over a global array.
556 if (Opc == Instruction::Sub && TD) {
557 GlobalValue *GV1, *GV2;
558 int64_t Offs1, Offs2;
560 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
561 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
563 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
564 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
571 /// CastGEPIndices - If array indices are not pointer-sized integers,
572 /// explicitly cast them so that they aren't implicitly casted by the
574 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
575 Type *ResultTy, const TargetData *TD,
576 const TargetLibraryInfo *TLI) {
578 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
581 SmallVector<Constant*, 32> NewIdxs;
582 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
584 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
585 Ops.slice(1, i-1)))) &&
586 Ops[i]->getType() != IntPtrTy) {
588 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
594 NewIdxs.push_back(Ops[i]);
599 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
600 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
601 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
606 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
607 /// constant expression, do so.
608 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
609 Type *ResultTy, const TargetData *TD,
610 const TargetLibraryInfo *TLI) {
611 Constant *Ptr = Ops[0];
612 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
613 !Ptr->getType()->isPointerTy())
616 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
618 // If this is a constant expr gep that is effectively computing an
619 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
620 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
621 if (!isa<ConstantInt>(Ops[i])) {
623 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
624 // "inttoptr (sub (ptrtoint Ptr), V)"
625 if (Ops.size() == 2 &&
626 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
627 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
628 assert((CE == 0 || CE->getType() == IntPtrTy) &&
629 "CastGEPIndices didn't canonicalize index types!");
630 if (CE && CE->getOpcode() == Instruction::Sub &&
631 CE->getOperand(0)->isNullValue()) {
632 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
633 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
634 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
635 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
636 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
643 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
645 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
646 makeArrayRef((Value **)Ops.data() + 1,
648 Ptr = cast<Constant>(Ptr->stripPointerCasts());
650 // If this is a GEP of a GEP, fold it all into a single GEP.
651 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
652 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
654 // Do not try the incorporate the sub-GEP if some index is not a number.
655 bool AllConstantInt = true;
656 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
657 if (!isa<ConstantInt>(NestedOps[i])) {
658 AllConstantInt = false;
664 Ptr = cast<Constant>(GEP->getOperand(0));
665 Offset += APInt(BitWidth,
666 TD->getIndexedOffset(Ptr->getType(), NestedOps));
667 Ptr = cast<Constant>(Ptr->stripPointerCasts());
670 // If the base value for this address is a literal integer value, fold the
671 // getelementptr to the resulting integer value casted to the pointer type.
672 APInt BasePtr(BitWidth, 0);
673 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
674 if (CE->getOpcode() == Instruction::IntToPtr)
675 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
676 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
677 if (Ptr->isNullValue() || BasePtr != 0) {
678 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
679 return ConstantExpr::getIntToPtr(C, ResultTy);
682 // Otherwise form a regular getelementptr. Recompute the indices so that
683 // we eliminate over-indexing of the notional static type array bounds.
684 // This makes it easy to determine if the getelementptr is "inbounds".
685 // Also, this helps GlobalOpt do SROA on GlobalVariables.
686 Type *Ty = Ptr->getType();
687 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
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 // If we end up with an offset that isn't valid for this struct type, we
719 // can't re-form this GEP in a regular form, so bail out. The pointer
720 // operand likely went through casts that are necessary to make the GEP
722 const StructLayout &SL = *TD->getStructLayout(STy);
723 if (Offset.uge(SL.getSizeInBytes()))
726 // Determine which field of the struct the offset points into. The
727 // getZExtValue is fine as we've already ensured that the offset is
728 // within the range representable by the StructLayout API.
729 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
730 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
732 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
733 Ty = STy->getTypeAtIndex(ElIdx);
735 // We've reached some non-indexable type.
738 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
740 // If we haven't used up the entire offset by descending the static
741 // type, then the offset is pointing into the middle of an indivisible
742 // member, so we can't simplify it.
748 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
749 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
750 "Computed GetElementPtr has unexpected type!");
752 // If we ended up indexing a member with a type that doesn't match
753 // the type of what the original indices indexed, add a cast.
754 if (Ty != cast<PointerType>(ResultTy)->getElementType())
755 C = FoldBitCast(C, ResultTy, *TD);
762 //===----------------------------------------------------------------------===//
763 // Constant Folding public APIs
764 //===----------------------------------------------------------------------===//
766 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
767 /// If successful, the constant result is returned, if not, null is returned.
768 /// Note that this fails if not all of the operands are constant. Otherwise,
769 /// this function can only fail when attempting to fold instructions like loads
770 /// and stores, which have no constant expression form.
771 Constant *llvm::ConstantFoldInstruction(Instruction *I,
772 const TargetData *TD,
773 const TargetLibraryInfo *TLI) {
774 // Handle PHI nodes quickly here...
775 if (PHINode *PN = dyn_cast<PHINode>(I)) {
776 Constant *CommonValue = 0;
778 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
779 Value *Incoming = PN->getIncomingValue(i);
780 // If the incoming value is undef then skip it. Note that while we could
781 // skip the value if it is equal to the phi node itself we choose not to
782 // because that would break the rule that constant folding only applies if
783 // all operands are constants.
784 if (isa<UndefValue>(Incoming))
786 // If the incoming value is not a constant, then give up.
787 Constant *C = dyn_cast<Constant>(Incoming);
790 // Fold the PHI's operands.
791 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
792 C = ConstantFoldConstantExpression(NewC, TD, TLI);
793 // If the incoming value is a different constant to
794 // the one we saw previously, then give up.
795 if (CommonValue && C != CommonValue)
801 // If we reach here, all incoming values are the same constant or undef.
802 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
805 // Scan the operand list, checking to see if they are all constants, if so,
806 // hand off to ConstantFoldInstOperands.
807 SmallVector<Constant*, 8> Ops;
808 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
809 Constant *Op = dyn_cast<Constant>(*i);
811 return 0; // All operands not constant!
813 // Fold the Instruction's operands.
814 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
815 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
820 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
821 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
824 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
825 return ConstantFoldLoadInst(LI, TD);
827 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
828 return ConstantExpr::getInsertValue(
829 cast<Constant>(IVI->getAggregateOperand()),
830 cast<Constant>(IVI->getInsertedValueOperand()),
833 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
834 return ConstantExpr::getExtractValue(
835 cast<Constant>(EVI->getAggregateOperand()),
838 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
841 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
842 /// using the specified TargetData. If successful, the constant result is
843 /// result is returned, if not, null is returned.
844 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
845 const TargetData *TD,
846 const TargetLibraryInfo *TLI) {
847 SmallVector<Constant*, 8> Ops;
848 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
850 Constant *NewC = cast<Constant>(*i);
851 // Recursively fold the ConstantExpr's operands.
852 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
853 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
858 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
860 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
863 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
864 /// specified opcode and operands. If successful, the constant result is
865 /// returned, if not, null is returned. Note that this function can fail when
866 /// attempting to fold instructions like loads and stores, which have no
867 /// constant expression form.
869 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
870 /// information, due to only being passed an opcode and operands. Constant
871 /// folding using this function strips this information.
873 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
874 ArrayRef<Constant *> Ops,
875 const TargetData *TD,
876 const TargetLibraryInfo *TLI) {
877 // Handle easy binops first.
878 if (Instruction::isBinaryOp(Opcode)) {
879 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
880 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
883 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
888 case Instruction::ICmp:
889 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
890 case Instruction::Call:
891 if (Function *F = dyn_cast<Function>(Ops.back()))
892 if (canConstantFoldCallTo(F))
893 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
895 case Instruction::PtrToInt:
896 // If the input is a inttoptr, eliminate the pair. This requires knowing
897 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
898 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
899 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
900 Constant *Input = CE->getOperand(0);
901 unsigned InWidth = Input->getType()->getScalarSizeInBits();
902 if (TD->getPointerSizeInBits() < InWidth) {
904 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
905 TD->getPointerSizeInBits()));
906 Input = ConstantExpr::getAnd(Input, Mask);
908 // Do a zext or trunc to get to the dest size.
909 return ConstantExpr::getIntegerCast(Input, DestTy, false);
912 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
913 case Instruction::IntToPtr:
914 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
915 // the int size is >= the ptr size. This requires knowing the width of a
916 // pointer, so it can't be done in ConstantExpr::getCast.
917 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
919 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
920 CE->getOpcode() == Instruction::PtrToInt)
921 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
923 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
924 case Instruction::Trunc:
925 case Instruction::ZExt:
926 case Instruction::SExt:
927 case Instruction::FPTrunc:
928 case Instruction::FPExt:
929 case Instruction::UIToFP:
930 case Instruction::SIToFP:
931 case Instruction::FPToUI:
932 case Instruction::FPToSI:
933 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
934 case Instruction::BitCast:
936 return FoldBitCast(Ops[0], DestTy, *TD);
937 return ConstantExpr::getBitCast(Ops[0], DestTy);
938 case Instruction::Select:
939 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
940 case Instruction::ExtractElement:
941 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
942 case Instruction::InsertElement:
943 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
944 case Instruction::ShuffleVector:
945 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
946 case Instruction::GetElementPtr:
947 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
949 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
952 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
956 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
957 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
958 /// returns a constant expression of the specified operands.
960 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
961 Constant *Ops0, Constant *Ops1,
962 const TargetData *TD,
963 const TargetLibraryInfo *TLI) {
964 // fold: icmp (inttoptr x), null -> icmp x, 0
965 // fold: icmp (ptrtoint x), 0 -> icmp x, null
966 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
967 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
969 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
970 // around to know if bit truncation is happening.
971 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
972 if (TD && Ops1->isNullValue()) {
973 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
974 if (CE0->getOpcode() == Instruction::IntToPtr) {
975 // Convert the integer value to the right size to ensure we get the
976 // proper extension or truncation.
977 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
979 Constant *Null = Constant::getNullValue(C->getType());
980 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
983 // Only do this transformation if the int is intptrty in size, otherwise
984 // there is a truncation or extension that we aren't modeling.
985 if (CE0->getOpcode() == Instruction::PtrToInt &&
986 CE0->getType() == IntPtrTy) {
987 Constant *C = CE0->getOperand(0);
988 Constant *Null = Constant::getNullValue(C->getType());
989 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
993 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
994 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
995 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
997 if (CE0->getOpcode() == Instruction::IntToPtr) {
998 // Convert the integer value to the right size to ensure we get the
999 // proper extension or truncation.
1000 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1002 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1004 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1007 // Only do this transformation if the int is intptrty in size, otherwise
1008 // there is a truncation or extension that we aren't modeling.
1009 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1010 CE0->getType() == IntPtrTy &&
1011 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1012 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1013 CE1->getOperand(0), TD, TLI);
1017 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1018 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1019 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1020 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1022 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1025 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1028 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1029 Constant *Ops[] = { LHS, RHS };
1030 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1034 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1038 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1039 /// getelementptr constantexpr, return the constant value being addressed by the
1040 /// constant expression, or null if something is funny and we can't decide.
1041 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1043 if (!CE->getOperand(1)->isNullValue())
1044 return 0; // Do not allow stepping over the value!
1046 // Loop over all of the operands, tracking down which value we are
1048 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1049 C = C->getAggregateElement(CE->getOperand(i));
1050 if (C == 0) return 0;
1055 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1056 /// indices (with an *implied* zero pointer index that is not in the list),
1057 /// return the constant value being addressed by a virtual load, or null if
1058 /// something is funny and we can't decide.
1059 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1060 ArrayRef<Constant*> Indices) {
1061 // Loop over all of the operands, tracking down which value we are
1063 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1064 C = C->getAggregateElement(Indices[i]);
1065 if (C == 0) return 0;
1071 //===----------------------------------------------------------------------===//
1072 // Constant Folding for Calls
1075 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1076 /// the specified function.
1078 llvm::canConstantFoldCallTo(const Function *F) {
1079 switch (F->getIntrinsicID()) {
1080 case Intrinsic::sqrt:
1081 case Intrinsic::pow:
1082 case Intrinsic::powi:
1083 case Intrinsic::bswap:
1084 case Intrinsic::ctpop:
1085 case Intrinsic::ctlz:
1086 case Intrinsic::cttz:
1087 case Intrinsic::sadd_with_overflow:
1088 case Intrinsic::uadd_with_overflow:
1089 case Intrinsic::ssub_with_overflow:
1090 case Intrinsic::usub_with_overflow:
1091 case Intrinsic::smul_with_overflow:
1092 case Intrinsic::umul_with_overflow:
1093 case Intrinsic::convert_from_fp16:
1094 case Intrinsic::convert_to_fp16:
1095 case Intrinsic::x86_sse_cvtss2si:
1096 case Intrinsic::x86_sse_cvtss2si64:
1097 case Intrinsic::x86_sse_cvttss2si:
1098 case Intrinsic::x86_sse_cvttss2si64:
1099 case Intrinsic::x86_sse2_cvtsd2si:
1100 case Intrinsic::x86_sse2_cvtsd2si64:
1101 case Intrinsic::x86_sse2_cvttsd2si:
1102 case Intrinsic::x86_sse2_cvttsd2si64:
1109 if (!F->hasName()) return false;
1110 StringRef Name = F->getName();
1112 // In these cases, the check of the length is required. We don't want to
1113 // return true for a name like "cos\0blah" which strcmp would return equal to
1114 // "cos", but has length 8.
1116 default: return false;
1118 return Name == "acos" || Name == "asin" ||
1119 Name == "atan" || Name == "atan2";
1121 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1123 return Name == "exp" || Name == "exp2";
1125 return Name == "fabs" || Name == "fmod" || Name == "floor";
1127 return Name == "log" || Name == "log10";
1129 return Name == "pow";
1131 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1132 Name == "sinf" || Name == "sqrtf";
1134 return Name == "tan" || Name == "tanh";
1138 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1140 sys::llvm_fenv_clearexcept();
1142 if (sys::llvm_fenv_testexcept()) {
1143 sys::llvm_fenv_clearexcept();
1147 if (Ty->isFloatTy())
1148 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1149 if (Ty->isDoubleTy())
1150 return ConstantFP::get(Ty->getContext(), APFloat(V));
1151 llvm_unreachable("Can only constant fold float/double");
1154 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1155 double V, double W, Type *Ty) {
1156 sys::llvm_fenv_clearexcept();
1158 if (sys::llvm_fenv_testexcept()) {
1159 sys::llvm_fenv_clearexcept();
1163 if (Ty->isFloatTy())
1164 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1165 if (Ty->isDoubleTy())
1166 return ConstantFP::get(Ty->getContext(), APFloat(V));
1167 llvm_unreachable("Can only constant fold float/double");
1170 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1171 /// conversion of a constant floating point. If roundTowardZero is false, the
1172 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1173 /// the behavior of the non-truncating SSE instructions in the default rounding
1174 /// mode. The desired integer type Ty is used to select how many bits are
1175 /// available for the result. Returns null if the conversion cannot be
1176 /// performed, otherwise returns the Constant value resulting from the
1178 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1179 bool roundTowardZero, Type *Ty) {
1180 // All of these conversion intrinsics form an integer of at most 64bits.
1181 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1182 assert(ResultWidth <= 64 &&
1183 "Can only constant fold conversions to 64 and 32 bit ints");
1186 bool isExact = false;
1187 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1188 : APFloat::rmNearestTiesToEven;
1189 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1190 /*isSigned=*/true, mode,
1192 if (status != APFloat::opOK && status != APFloat::opInexact)
1194 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1197 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1198 /// with the specified arguments, returning null if unsuccessful.
1200 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1201 const TargetLibraryInfo *TLI) {
1202 if (!F->hasName()) return 0;
1203 StringRef Name = F->getName();
1205 Type *Ty = F->getReturnType();
1206 if (Operands.size() == 1) {
1207 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1208 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1209 APFloat Val(Op->getValueAPF());
1212 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1214 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1219 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1222 /// We only fold functions with finite arguments. Folding NaN and inf is
1223 /// likely to be aborted with an exception anyway, and some host libms
1224 /// have known errors raising exceptions.
1225 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1228 /// Currently APFloat versions of these functions do not exist, so we use
1229 /// the host native double versions. Float versions are not called
1230 /// directly but for all these it is true (float)(f((double)arg)) ==
1231 /// f(arg). Long double not supported yet.
1232 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1233 Op->getValueAPF().convertToDouble();
1236 if (Name == "acos" && TLI->has(LibFunc::acos))
1237 return ConstantFoldFP(acos, V, Ty);
1238 else if (Name == "asin" && TLI->has(LibFunc::asin))
1239 return ConstantFoldFP(asin, V, Ty);
1240 else if (Name == "atan" && TLI->has(LibFunc::atan))
1241 return ConstantFoldFP(atan, V, Ty);
1244 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1245 return ConstantFoldFP(ceil, V, Ty);
1246 else if (Name == "cos" && TLI->has(LibFunc::cos))
1247 return ConstantFoldFP(cos, V, Ty);
1248 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1249 return ConstantFoldFP(cosh, V, Ty);
1250 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1251 return ConstantFoldFP(cos, V, Ty);
1254 if (Name == "exp" && TLI->has(LibFunc::exp))
1255 return ConstantFoldFP(exp, V, Ty);
1257 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1258 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1260 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1264 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1265 return ConstantFoldFP(fabs, V, Ty);
1266 else if (Name == "floor" && TLI->has(LibFunc::floor))
1267 return ConstantFoldFP(floor, V, Ty);
1270 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1271 return ConstantFoldFP(log, V, Ty);
1272 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1273 return ConstantFoldFP(log10, V, Ty);
1274 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1275 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1277 return ConstantFoldFP(sqrt, V, Ty);
1279 return Constant::getNullValue(Ty);
1283 if (Name == "sin" && TLI->has(LibFunc::sin))
1284 return ConstantFoldFP(sin, V, Ty);
1285 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1286 return ConstantFoldFP(sinh, V, Ty);
1287 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1288 return ConstantFoldFP(sqrt, V, Ty);
1289 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1290 return ConstantFoldFP(sqrt, V, Ty);
1291 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1292 return ConstantFoldFP(sin, V, Ty);
1295 if (Name == "tan" && TLI->has(LibFunc::tan))
1296 return ConstantFoldFP(tan, V, Ty);
1297 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1298 return ConstantFoldFP(tanh, V, Ty);
1306 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1307 switch (F->getIntrinsicID()) {
1308 case Intrinsic::bswap:
1309 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1310 case Intrinsic::ctpop:
1311 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1312 case Intrinsic::convert_from_fp16: {
1313 APFloat Val(Op->getValue());
1316 APFloat::opStatus status =
1317 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1319 // Conversion is always precise.
1321 assert(status == APFloat::opOK && !lost &&
1322 "Precision lost during fp16 constfolding");
1324 return ConstantFP::get(F->getContext(), Val);
1331 // Support ConstantVector in case we have an Undef in the top.
1332 if (isa<ConstantVector>(Operands[0]) ||
1333 isa<ConstantDataVector>(Operands[0])) {
1334 Constant *Op = cast<Constant>(Operands[0]);
1335 switch (F->getIntrinsicID()) {
1337 case Intrinsic::x86_sse_cvtss2si:
1338 case Intrinsic::x86_sse_cvtss2si64:
1339 case Intrinsic::x86_sse2_cvtsd2si:
1340 case Intrinsic::x86_sse2_cvtsd2si64:
1341 if (ConstantFP *FPOp =
1342 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1343 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1344 /*roundTowardZero=*/false, Ty);
1345 case Intrinsic::x86_sse_cvttss2si:
1346 case Intrinsic::x86_sse_cvttss2si64:
1347 case Intrinsic::x86_sse2_cvttsd2si:
1348 case Intrinsic::x86_sse2_cvttsd2si64:
1349 if (ConstantFP *FPOp =
1350 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1351 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1352 /*roundTowardZero=*/true, Ty);
1356 if (isa<UndefValue>(Operands[0])) {
1357 if (F->getIntrinsicID() == Intrinsic::bswap)
1365 if (Operands.size() == 2) {
1366 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1367 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1369 double Op1V = Ty->isFloatTy() ?
1370 (double)Op1->getValueAPF().convertToFloat() :
1371 Op1->getValueAPF().convertToDouble();
1372 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1373 if (Op2->getType() != Op1->getType())
1376 double Op2V = Ty->isFloatTy() ?
1377 (double)Op2->getValueAPF().convertToFloat():
1378 Op2->getValueAPF().convertToDouble();
1380 if (F->getIntrinsicID() == Intrinsic::pow) {
1381 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1385 if (Name == "pow" && TLI->has(LibFunc::pow))
1386 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1387 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1388 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1389 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1390 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1391 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1392 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1393 return ConstantFP::get(F->getContext(),
1394 APFloat((float)std::pow((float)Op1V,
1395 (int)Op2C->getZExtValue())));
1396 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1397 return ConstantFP::get(F->getContext(),
1398 APFloat((double)std::pow((double)Op1V,
1399 (int)Op2C->getZExtValue())));
1404 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1405 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1406 switch (F->getIntrinsicID()) {
1408 case Intrinsic::sadd_with_overflow:
1409 case Intrinsic::uadd_with_overflow:
1410 case Intrinsic::ssub_with_overflow:
1411 case Intrinsic::usub_with_overflow:
1412 case Intrinsic::smul_with_overflow:
1413 case Intrinsic::umul_with_overflow: {
1416 switch (F->getIntrinsicID()) {
1417 default: llvm_unreachable("Invalid case");
1418 case Intrinsic::sadd_with_overflow:
1419 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1421 case Intrinsic::uadd_with_overflow:
1422 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1424 case Intrinsic::ssub_with_overflow:
1425 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1427 case Intrinsic::usub_with_overflow:
1428 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1430 case Intrinsic::smul_with_overflow:
1431 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1433 case Intrinsic::umul_with_overflow:
1434 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1438 ConstantInt::get(F->getContext(), Res),
1439 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1441 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1443 case Intrinsic::cttz:
1444 // FIXME: This should check for Op2 == 1, and become unreachable if
1446 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1447 case Intrinsic::ctlz:
1448 // FIXME: This should check for Op2 == 1, and become unreachable if
1450 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());