1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines routines for folding instructions into constants.
12 // Also, to supplement the basic IR ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // DataLayout information. These functions cannot go in IR due to library
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/StringMap.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/FEnv.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// DataLayout. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const DataLayout &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
55 // Handle a vector->integer cast.
56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
57 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
59 return ConstantExpr::getBitCast(C, DestTy);
61 unsigned NumSrcElts = CDV->getType()->getNumElements();
63 Type *SrcEltTy = CDV->getType()->getElementType();
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (SrcEltTy->isFloatingPointTy()) {
68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
71 // Ask IR to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
73 CDV = cast<ConstantDataVector>(C);
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
79 APInt Result(IT->getBitWidth(), 0);
80 for (unsigned i = 0; i != NumSrcElts; ++i) {
82 if (TD.isLittleEndian())
83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
85 Result |= CDV->getElementAsInteger(i);
88 return ConstantInt::get(IT, Result);
91 // The code below only handles casts to vectors currently.
92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
94 return ConstantExpr::getBitCast(C, DestTy);
96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
97 // vector so the code below can handle it uniformly.
98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
99 Constant *Ops = C; // don't take the address of C!
100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
103 // If this is a bitcast from constant vector -> vector, fold it.
104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
105 return ConstantExpr::getBitCast(C, DestTy);
107 // If the element types match, IR can fold it.
108 unsigned NumDstElt = DestVTy->getNumElements();
109 unsigned NumSrcElt = C->getType()->getVectorNumElements();
110 if (NumDstElt == NumSrcElt)
111 return ConstantExpr::getBitCast(C, DestTy);
113 Type *SrcEltTy = C->getType()->getVectorElementType();
114 Type *DstEltTy = DestVTy->getElementType();
116 // Otherwise, we're changing the number of elements in a vector, which
117 // requires endianness information to do the right thing. For example,
118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
119 // folds to (little endian):
120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
121 // and to (big endian):
122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
124 // First thing is first. We only want to think about integer here, so if
125 // we have something in FP form, recast it as integer.
126 if (DstEltTy->isFloatingPointTy()) {
127 // Fold to an vector of integers with same size as our FP type.
128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
131 // Recursively handle this integer conversion, if possible.
132 C = FoldBitCast(C, DestIVTy, TD);
134 // Finally, IR can handle this now that #elts line up.
135 return ConstantExpr::getBitCast(C, DestTy);
138 // Okay, we know the destination is integer, if the input is FP, convert
139 // it to integer first.
140 if (SrcEltTy->isFloatingPointTy()) {
141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
144 // Ask IR to do the conversion now that #elts line up.
145 C = ConstantExpr::getBitCast(C, SrcIVTy);
146 // If IR wasn't able to fold it, bail out.
147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
148 !isa<ConstantDataVector>(C))
152 // Now we know that the input and output vectors are both integer vectors
153 // of the same size, and that their #elements is not the same. Do the
154 // conversion here, which depends on whether the input or output has
156 bool isLittleEndian = TD.isLittleEndian();
158 SmallVector<Constant*, 32> Result;
159 if (NumDstElt < NumSrcElt) {
160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
161 Constant *Zero = Constant::getNullValue(DstEltTy);
162 unsigned Ratio = NumSrcElt/NumDstElt;
163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
165 for (unsigned i = 0; i != NumDstElt; ++i) {
166 // Build each element of the result.
167 Constant *Elt = Zero;
168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
169 for (unsigned j = 0; j != Ratio; ++j) {
170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
171 if (!Src) // Reject constantexpr elements.
172 return ConstantExpr::getBitCast(C, DestTy);
174 // Zero extend the element to the right size.
175 Src = ConstantExpr::getZExt(Src, Elt->getType());
177 // Shift it to the right place, depending on endianness.
178 Src = ConstantExpr::getShl(Src,
179 ConstantInt::get(Src->getType(), ShiftAmt));
180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
183 Elt = ConstantExpr::getOr(Elt, Src);
185 Result.push_back(Elt);
187 return ConstantVector::get(Result);
190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
191 unsigned Ratio = NumDstElt/NumSrcElt;
192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
194 // Loop over each source value, expanding into multiple results.
195 for (unsigned i = 0; i != NumSrcElt; ++i) {
196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
197 if (!Src) // Reject constantexpr elements.
198 return ConstantExpr::getBitCast(C, DestTy);
200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
201 for (unsigned j = 0; j != Ratio; ++j) {
202 // Shift the piece of the value into the right place, depending on
204 Constant *Elt = ConstantExpr::getLShr(Src,
205 ConstantInt::get(Src->getType(), ShiftAmt));
206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
208 // Truncate and remember this piece.
209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
213 return ConstantVector::get(Result);
217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
218 /// from a global, return the global and the constant. Because of
219 /// constantexprs, this function is recursive.
220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
221 APInt &Offset, const DataLayout &TD) {
222 // Trivial case, constant is the global.
223 if ((GV = dyn_cast<GlobalValue>(C))) {
224 Offset.clearAllBits();
228 // Otherwise, if this isn't a constant expr, bail out.
229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
230 if (!CE) return false;
232 // Look through ptr->int and ptr->ptr casts.
233 if (CE->getOpcode() == Instruction::PtrToInt ||
234 CE->getOpcode() == Instruction::BitCast)
235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
238 if (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)) {
260 APInt(Offset.getBitWidth(),
261 TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()));
263 SequentialType *SQT = cast<SequentialType>(*GTI);
264 Offset += APInt(Offset.getBitWidth(),
265 TD.getTypeAllocSize(SQT->getElementType()) *
275 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
276 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
277 /// pointer to copy results into and BytesLeft is the number of bytes left in
278 /// the CurPtr buffer. TD is the target data.
279 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
280 unsigned char *CurPtr, unsigned BytesLeft,
281 const DataLayout &TD) {
282 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
283 "Out of range access");
285 // If this element is zero or undefined, we can just return since *CurPtr is
287 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
290 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
291 if (CI->getBitWidth() > 64 ||
292 (CI->getBitWidth() & 7) != 0)
295 uint64_t Val = CI->getZExtValue();
296 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
298 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
300 if (!TD.isLittleEndian())
301 n = IntBytes - n - 1;
302 CurPtr[i] = (unsigned char)(Val >> (n * 8));
308 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
309 if (CFP->getType()->isDoubleTy()) {
310 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
311 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
313 if (CFP->getType()->isFloatTy()){
314 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
315 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
320 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
321 const StructLayout *SL = TD.getStructLayout(CS->getType());
322 unsigned Index = SL->getElementContainingOffset(ByteOffset);
323 uint64_t CurEltOffset = SL->getElementOffset(Index);
324 ByteOffset -= CurEltOffset;
327 // If the element access is to the element itself and not to tail padding,
328 // read the bytes from the element.
329 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
331 if (ByteOffset < EltSize &&
332 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
338 // Check to see if we read from the last struct element, if so we're done.
339 if (Index == CS->getType()->getNumElements())
342 // If we read all of the bytes we needed from this element we're done.
343 uint64_t NextEltOffset = SL->getElementOffset(Index);
345 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
348 // Move to the next element of the struct.
349 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
350 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
352 CurEltOffset = NextEltOffset;
357 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
358 isa<ConstantDataSequential>(C)) {
359 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
360 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
361 uint64_t Index = ByteOffset / EltSize;
362 uint64_t Offset = ByteOffset - Index * EltSize;
364 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
365 NumElts = AT->getNumElements();
367 NumElts = cast<VectorType>(C->getType())->getNumElements();
369 for (; Index != NumElts; ++Index) {
370 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
374 uint64_t BytesWritten = EltSize - Offset;
375 assert(BytesWritten <= EltSize && "Not indexing into this element?");
376 if (BytesWritten >= BytesLeft)
380 BytesLeft -= BytesWritten;
381 CurPtr += BytesWritten;
386 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
387 if (CE->getOpcode() == Instruction::IntToPtr &&
388 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
389 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
393 // Otherwise, unknown initializer type.
397 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
398 const DataLayout &TD) {
399 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
400 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
402 // If this isn't an integer load we can't fold it directly.
404 // If this is a float/double load, we can try folding it as an int32/64 load
405 // and then bitcast the result. This can be useful for union cases. Note
406 // that address spaces don't matter here since we're not going to result in
407 // an actual new load.
409 if (LoadTy->isFloatTy())
410 MapTy = Type::getInt32PtrTy(C->getContext());
411 else if (LoadTy->isDoubleTy())
412 MapTy = Type::getInt64PtrTy(C->getContext());
413 else if (LoadTy->isVectorTy()) {
414 MapTy = IntegerType::get(C->getContext(),
415 TD.getTypeAllocSizeInBits(LoadTy));
416 MapTy = PointerType::getUnqual(MapTy);
420 C = FoldBitCast(C, MapTy, TD);
421 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
422 return FoldBitCast(Res, LoadTy, TD);
426 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
427 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
430 APInt Offset(TD.getPointerSizeInBits(), 0);
431 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
434 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
435 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
436 !GV->getInitializer()->getType()->isSized())
439 // If we're loading off the beginning of the global, some bytes may be valid,
440 // but we don't try to handle this.
441 if (Offset.isNegative()) return 0;
443 // If we're not accessing anything in this constant, the result is undefined.
444 if (Offset.getZExtValue() >=
445 TD.getTypeAllocSize(GV->getInitializer()->getType()))
446 return UndefValue::get(IntType);
448 unsigned char RawBytes[32] = {0};
449 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
453 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
454 if (TD.isLittleEndian()) {
455 ResultVal = RawBytes[BytesLoaded - 1];
456 for (unsigned i = 1; i != BytesLoaded; ++i) {
458 ResultVal |= RawBytes[BytesLoaded-1-i];
461 ResultVal = RawBytes[0];
462 for (unsigned i = 1; i != BytesLoaded; ++i) {
464 ResultVal |= RawBytes[i];
468 return ConstantInt::get(IntType->getContext(), ResultVal);
471 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
472 /// produce if it is constant and determinable. If this is not determinable,
474 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
475 const DataLayout *TD) {
476 // First, try the easy cases:
477 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
478 if (GV->isConstant() && GV->hasDefinitiveInitializer())
479 return GV->getInitializer();
481 // If the loaded value isn't a constant expr, we can't handle it.
482 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
485 if (CE->getOpcode() == Instruction::GetElementPtr) {
486 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
487 if (GV->isConstant() && GV->hasDefinitiveInitializer())
489 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
493 // Instead of loading constant c string, use corresponding integer value
494 // directly if string length is small enough.
496 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
497 unsigned StrLen = Str.size();
498 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
499 unsigned NumBits = Ty->getPrimitiveSizeInBits();
500 // Replace load with immediate integer if the result is an integer or fp
502 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
503 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
504 APInt StrVal(NumBits, 0);
505 APInt SingleChar(NumBits, 0);
506 if (TD->isLittleEndian()) {
507 for (signed i = StrLen-1; i >= 0; i--) {
508 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
509 StrVal = (StrVal << 8) | SingleChar;
512 for (unsigned i = 0; i < StrLen; i++) {
513 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
514 StrVal = (StrVal << 8) | SingleChar;
516 // Append NULL at the end.
518 StrVal = (StrVal << 8) | SingleChar;
521 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
522 if (Ty->isFloatingPointTy())
523 Res = ConstantExpr::getBitCast(Res, Ty);
528 // If this load comes from anywhere in a constant global, and if the global
529 // is all undef or zero, we know what it loads.
530 if (GlobalVariable *GV =
531 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
532 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
533 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
534 if (GV->getInitializer()->isNullValue())
535 return Constant::getNullValue(ResTy);
536 if (isa<UndefValue>(GV->getInitializer()))
537 return UndefValue::get(ResTy);
541 // Try hard to fold loads from bitcasted strange and non-type-safe things.
543 return FoldReinterpretLoadFromConstPtr(CE, *TD);
547 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
548 if (LI->isVolatile()) return 0;
550 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
551 return ConstantFoldLoadFromConstPtr(C, TD);
556 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
557 /// Attempt to symbolically evaluate the result of a binary operator merging
558 /// these together. If target data info is available, it is provided as TD,
559 /// otherwise TD is null.
560 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
561 Constant *Op1, const DataLayout *TD){
564 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
565 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
569 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
570 // constant. This happens frequently when iterating over a global array.
571 if (Opc == Instruction::Sub && TD) {
572 GlobalValue *GV1, *GV2;
573 APInt Offs1(TD->getPointerSizeInBits(), 0),
574 Offs2(TD->getPointerSizeInBits(), 0);
576 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
577 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
579 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
580 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
587 /// CastGEPIndices - If array indices are not pointer-sized integers,
588 /// explicitly cast them so that they aren't implicitly casted by the
590 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
591 Type *ResultTy, const DataLayout *TD,
592 const TargetLibraryInfo *TLI) {
594 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
597 SmallVector<Constant*, 32> NewIdxs;
598 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
600 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
601 Ops.slice(1, i-1)))) &&
602 Ops[i]->getType() != IntPtrTy) {
604 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
610 NewIdxs.push_back(Ops[i]);
615 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
616 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
617 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
622 /// Strip the pointer casts, but preserve the address space information.
623 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
624 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
625 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
626 Ptr = cast<Constant>(Ptr->stripPointerCasts());
627 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
629 // Preserve the address space number of the pointer.
630 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
631 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
632 OldPtrTy->getAddressSpace());
633 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy);
638 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
639 /// constant expression, do so.
640 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
641 Type *ResultTy, const DataLayout *TD,
642 const TargetLibraryInfo *TLI) {
643 Constant *Ptr = Ops[0];
644 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
645 !Ptr->getType()->isPointerTy())
648 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
650 // If this is a constant expr gep that is effectively computing an
651 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
652 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
653 if (!isa<ConstantInt>(Ops[i])) {
655 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
656 // "inttoptr (sub (ptrtoint Ptr), V)"
657 if (Ops.size() == 2 &&
658 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
659 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
660 assert((CE == 0 || CE->getType() == IntPtrTy) &&
661 "CastGEPIndices didn't canonicalize index types!");
662 if (CE && CE->getOpcode() == Instruction::Sub &&
663 CE->getOperand(0)->isNullValue()) {
664 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
665 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
666 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
667 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
668 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
675 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
677 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
678 makeArrayRef((Value *const*)
681 Ptr = StripPtrCastKeepAS(Ptr);
683 // If this is a GEP of a GEP, fold it all into a single GEP.
684 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
685 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
687 // Do not try the incorporate the sub-GEP if some index is not a number.
688 bool AllConstantInt = true;
689 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
690 if (!isa<ConstantInt>(NestedOps[i])) {
691 AllConstantInt = false;
697 Ptr = cast<Constant>(GEP->getOperand(0));
698 Offset += APInt(BitWidth,
699 TD->getIndexedOffset(Ptr->getType(), NestedOps));
700 Ptr = StripPtrCastKeepAS(Ptr);
703 // If the base value for this address is a literal integer value, fold the
704 // getelementptr to the resulting integer value casted to the pointer type.
705 APInt BasePtr(BitWidth, 0);
706 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
707 if (CE->getOpcode() == Instruction::IntToPtr)
708 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
709 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
710 if (Ptr->isNullValue() || BasePtr != 0) {
711 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
712 return ConstantExpr::getIntToPtr(C, ResultTy);
715 // Otherwise form a regular getelementptr. Recompute the indices so that
716 // we eliminate over-indexing of the notional static type array bounds.
717 // This makes it easy to determine if the getelementptr is "inbounds".
718 // Also, this helps GlobalOpt do SROA on GlobalVariables.
719 Type *Ty = Ptr->getType();
720 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
721 SmallVector<Constant*, 32> NewIdxs;
723 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
724 if (ATy->isPointerTy()) {
725 // The only pointer indexing we'll do is on the first index of the GEP.
726 if (!NewIdxs.empty())
729 // Only handle pointers to sized types, not pointers to functions.
730 if (!ATy->getElementType()->isSized())
734 // Determine which element of the array the offset points into.
735 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
736 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
738 // The element size is 0. This may be [0 x Ty]*, so just use a zero
739 // index for this level and proceed to the next level to see if it can
740 // accommodate the offset.
741 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
743 // The element size is non-zero divide the offset by the element
744 // size (rounding down), to compute the index at this level.
745 APInt NewIdx = Offset.udiv(ElemSize);
746 Offset -= NewIdx * ElemSize;
747 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
749 Ty = ATy->getElementType();
750 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
751 // If we end up with an offset that isn't valid for this struct type, we
752 // can't re-form this GEP in a regular form, so bail out. The pointer
753 // operand likely went through casts that are necessary to make the GEP
755 const StructLayout &SL = *TD->getStructLayout(STy);
756 if (Offset.uge(SL.getSizeInBytes()))
759 // Determine which field of the struct the offset points into. The
760 // getZExtValue is fine as we've already ensured that the offset is
761 // within the range representable by the StructLayout API.
762 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
763 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
765 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
766 Ty = STy->getTypeAtIndex(ElIdx);
768 // We've reached some non-indexable type.
771 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
773 // If we haven't used up the entire offset by descending the static
774 // type, then the offset is pointing into the middle of an indivisible
775 // member, so we can't simplify it.
781 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
782 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
783 "Computed GetElementPtr has unexpected type!");
785 // If we ended up indexing a member with a type that doesn't match
786 // the type of what the original indices indexed, add a cast.
787 if (Ty != cast<PointerType>(ResultTy)->getElementType())
788 C = FoldBitCast(C, ResultTy, *TD);
795 //===----------------------------------------------------------------------===//
796 // Constant Folding public APIs
797 //===----------------------------------------------------------------------===//
799 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
800 /// If successful, the constant result is returned, if not, null is returned.
801 /// Note that this fails if not all of the operands are constant. Otherwise,
802 /// this function can only fail when attempting to fold instructions like loads
803 /// and stores, which have no constant expression form.
804 Constant *llvm::ConstantFoldInstruction(Instruction *I,
805 const DataLayout *TD,
806 const TargetLibraryInfo *TLI) {
807 // Handle PHI nodes quickly here...
808 if (PHINode *PN = dyn_cast<PHINode>(I)) {
809 Constant *CommonValue = 0;
811 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
812 Value *Incoming = PN->getIncomingValue(i);
813 // If the incoming value is undef then skip it. Note that while we could
814 // skip the value if it is equal to the phi node itself we choose not to
815 // because that would break the rule that constant folding only applies if
816 // all operands are constants.
817 if (isa<UndefValue>(Incoming))
819 // If the incoming value is not a constant, then give up.
820 Constant *C = dyn_cast<Constant>(Incoming);
823 // Fold the PHI's operands.
824 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
825 C = ConstantFoldConstantExpression(NewC, TD, TLI);
826 // If the incoming value is a different constant to
827 // the one we saw previously, then give up.
828 if (CommonValue && C != CommonValue)
834 // If we reach here, all incoming values are the same constant or undef.
835 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
838 // Scan the operand list, checking to see if they are all constants, if so,
839 // hand off to ConstantFoldInstOperands.
840 SmallVector<Constant*, 8> Ops;
841 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
842 Constant *Op = dyn_cast<Constant>(*i);
844 return 0; // All operands not constant!
846 // Fold the Instruction's operands.
847 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
848 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
853 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
854 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
857 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
858 return ConstantFoldLoadInst(LI, TD);
860 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
861 return ConstantExpr::getInsertValue(
862 cast<Constant>(IVI->getAggregateOperand()),
863 cast<Constant>(IVI->getInsertedValueOperand()),
866 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
867 return ConstantExpr::getExtractValue(
868 cast<Constant>(EVI->getAggregateOperand()),
871 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
874 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
875 /// using the specified DataLayout. If successful, the constant result is
876 /// result is returned, if not, null is returned.
877 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
878 const DataLayout *TD,
879 const TargetLibraryInfo *TLI) {
880 SmallVector<Constant*, 8> Ops;
881 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
883 Constant *NewC = cast<Constant>(*i);
884 // Recursively fold the ConstantExpr's operands.
885 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
886 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
891 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
893 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
896 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
897 /// specified opcode and operands. If successful, the constant result is
898 /// returned, if not, null is returned. Note that this function can fail when
899 /// attempting to fold instructions like loads and stores, which have no
900 /// constant expression form.
902 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
903 /// information, due to only being passed an opcode and operands. Constant
904 /// folding using this function strips this information.
906 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
907 ArrayRef<Constant *> Ops,
908 const DataLayout *TD,
909 const TargetLibraryInfo *TLI) {
910 // Handle easy binops first.
911 if (Instruction::isBinaryOp(Opcode)) {
912 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
913 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
916 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
921 case Instruction::ICmp:
922 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
923 case Instruction::Call:
924 if (Function *F = dyn_cast<Function>(Ops.back()))
925 if (canConstantFoldCallTo(F))
926 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
928 case Instruction::PtrToInt:
929 // If the input is a inttoptr, eliminate the pair. This requires knowing
930 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
931 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
932 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
933 Constant *Input = CE->getOperand(0);
934 unsigned InWidth = Input->getType()->getScalarSizeInBits();
935 if (TD->getPointerSizeInBits() < InWidth) {
937 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
938 TD->getPointerSizeInBits()));
939 Input = ConstantExpr::getAnd(Input, Mask);
941 // Do a zext or trunc to get to the dest size.
942 return ConstantExpr::getIntegerCast(Input, DestTy, false);
945 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
946 case Instruction::IntToPtr:
947 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
948 // the int size is >= the ptr size. This requires knowing the width of a
949 // pointer, so it can't be done in ConstantExpr::getCast.
950 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
952 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
953 CE->getOpcode() == Instruction::PtrToInt)
954 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
956 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
957 case Instruction::Trunc:
958 case Instruction::ZExt:
959 case Instruction::SExt:
960 case Instruction::FPTrunc:
961 case Instruction::FPExt:
962 case Instruction::UIToFP:
963 case Instruction::SIToFP:
964 case Instruction::FPToUI:
965 case Instruction::FPToSI:
966 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
967 case Instruction::BitCast:
969 return FoldBitCast(Ops[0], DestTy, *TD);
970 return ConstantExpr::getBitCast(Ops[0], DestTy);
971 case Instruction::Select:
972 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
973 case Instruction::ExtractElement:
974 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
975 case Instruction::InsertElement:
976 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
977 case Instruction::ShuffleVector:
978 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
979 case Instruction::GetElementPtr:
980 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
982 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
985 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
989 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
990 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
991 /// returns a constant expression of the specified operands.
993 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
994 Constant *Ops0, Constant *Ops1,
995 const DataLayout *TD,
996 const TargetLibraryInfo *TLI) {
997 // fold: icmp (inttoptr x), null -> icmp x, 0
998 // fold: icmp (ptrtoint x), 0 -> icmp x, null
999 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1000 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1002 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
1003 // around to know if bit truncation is happening.
1004 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1005 if (TD && Ops1->isNullValue()) {
1006 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1007 if (CE0->getOpcode() == Instruction::IntToPtr) {
1008 // Convert the integer value to the right size to ensure we get the
1009 // proper extension or truncation.
1010 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1012 Constant *Null = Constant::getNullValue(C->getType());
1013 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1016 // Only do this transformation if the int is intptrty in size, otherwise
1017 // there is a truncation or extension that we aren't modeling.
1018 if (CE0->getOpcode() == Instruction::PtrToInt &&
1019 CE0->getType() == IntPtrTy) {
1020 Constant *C = CE0->getOperand(0);
1021 Constant *Null = Constant::getNullValue(C->getType());
1022 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1026 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1027 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1028 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1030 if (CE0->getOpcode() == Instruction::IntToPtr) {
1031 // Convert the integer value to the right size to ensure we get the
1032 // proper extension or truncation.
1033 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1035 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1037 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1040 // Only do this transformation if the int is intptrty in size, otherwise
1041 // there is a truncation or extension that we aren't modeling.
1042 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1043 CE0->getType() == IntPtrTy &&
1044 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1045 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1046 CE1->getOperand(0), TD, TLI);
1050 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1051 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1052 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1053 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1055 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1058 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1061 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1062 Constant *Ops[] = { LHS, RHS };
1063 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1067 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1071 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1072 /// getelementptr constantexpr, return the constant value being addressed by the
1073 /// constant expression, or null if something is funny and we can't decide.
1074 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1076 if (!CE->getOperand(1)->isNullValue())
1077 return 0; // Do not allow stepping over the value!
1079 // Loop over all of the operands, tracking down which value we are
1081 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1082 C = C->getAggregateElement(CE->getOperand(i));
1083 if (C == 0) return 0;
1088 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1089 /// indices (with an *implied* zero pointer index that is not in the list),
1090 /// return the constant value being addressed by a virtual load, or null if
1091 /// something is funny and we can't decide.
1092 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1093 ArrayRef<Constant*> Indices) {
1094 // Loop over all of the operands, tracking down which value we are
1096 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1097 C = C->getAggregateElement(Indices[i]);
1098 if (C == 0) return 0;
1104 //===----------------------------------------------------------------------===//
1105 // Constant Folding for Calls
1108 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1109 /// the specified function.
1111 llvm::canConstantFoldCallTo(const Function *F) {
1112 switch (F->getIntrinsicID()) {
1113 case Intrinsic::sqrt:
1114 case Intrinsic::pow:
1115 case Intrinsic::powi:
1116 case Intrinsic::bswap:
1117 case Intrinsic::ctpop:
1118 case Intrinsic::ctlz:
1119 case Intrinsic::cttz:
1120 case Intrinsic::sadd_with_overflow:
1121 case Intrinsic::uadd_with_overflow:
1122 case Intrinsic::ssub_with_overflow:
1123 case Intrinsic::usub_with_overflow:
1124 case Intrinsic::smul_with_overflow:
1125 case Intrinsic::umul_with_overflow:
1126 case Intrinsic::convert_from_fp16:
1127 case Intrinsic::convert_to_fp16:
1128 case Intrinsic::x86_sse_cvtss2si:
1129 case Intrinsic::x86_sse_cvtss2si64:
1130 case Intrinsic::x86_sse_cvttss2si:
1131 case Intrinsic::x86_sse_cvttss2si64:
1132 case Intrinsic::x86_sse2_cvtsd2si:
1133 case Intrinsic::x86_sse2_cvtsd2si64:
1134 case Intrinsic::x86_sse2_cvttsd2si:
1135 case Intrinsic::x86_sse2_cvttsd2si64:
1142 if (!F->hasName()) return false;
1143 StringRef Name = F->getName();
1145 // In these cases, the check of the length is required. We don't want to
1146 // return true for a name like "cos\0blah" which strcmp would return equal to
1147 // "cos", but has length 8.
1149 default: return false;
1151 return Name == "acos" || Name == "asin" ||
1152 Name == "atan" || Name == "atan2";
1154 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1156 return Name == "exp" || Name == "exp2";
1158 return Name == "fabs" || Name == "fmod" || Name == "floor";
1160 return Name == "log" || Name == "log10";
1162 return Name == "pow";
1164 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1165 Name == "sinf" || Name == "sqrtf";
1167 return Name == "tan" || Name == "tanh";
1171 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1173 sys::llvm_fenv_clearexcept();
1175 if (sys::llvm_fenv_testexcept()) {
1176 sys::llvm_fenv_clearexcept();
1180 if (Ty->isFloatTy())
1181 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1182 if (Ty->isDoubleTy())
1183 return ConstantFP::get(Ty->getContext(), APFloat(V));
1184 llvm_unreachable("Can only constant fold float/double");
1187 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1188 double V, double W, Type *Ty) {
1189 sys::llvm_fenv_clearexcept();
1191 if (sys::llvm_fenv_testexcept()) {
1192 sys::llvm_fenv_clearexcept();
1196 if (Ty->isFloatTy())
1197 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1198 if (Ty->isDoubleTy())
1199 return ConstantFP::get(Ty->getContext(), APFloat(V));
1200 llvm_unreachable("Can only constant fold float/double");
1203 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1204 /// conversion of a constant floating point. If roundTowardZero is false, the
1205 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1206 /// the behavior of the non-truncating SSE instructions in the default rounding
1207 /// mode. The desired integer type Ty is used to select how many bits are
1208 /// available for the result. Returns null if the conversion cannot be
1209 /// performed, otherwise returns the Constant value resulting from the
1211 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1212 bool roundTowardZero, Type *Ty) {
1213 // All of these conversion intrinsics form an integer of at most 64bits.
1214 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1215 assert(ResultWidth <= 64 &&
1216 "Can only constant fold conversions to 64 and 32 bit ints");
1219 bool isExact = false;
1220 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1221 : APFloat::rmNearestTiesToEven;
1222 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1223 /*isSigned=*/true, mode,
1225 if (status != APFloat::opOK && status != APFloat::opInexact)
1227 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1230 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1231 /// with the specified arguments, returning null if unsuccessful.
1233 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1234 const TargetLibraryInfo *TLI) {
1235 if (!F->hasName()) return 0;
1236 StringRef Name = F->getName();
1238 Type *Ty = F->getReturnType();
1239 if (Operands.size() == 1) {
1240 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1241 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1242 APFloat Val(Op->getValueAPF());
1245 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1247 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1252 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1255 /// We only fold functions with finite arguments. Folding NaN and inf is
1256 /// likely to be aborted with an exception anyway, and some host libms
1257 /// have known errors raising exceptions.
1258 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1261 /// Currently APFloat versions of these functions do not exist, so we use
1262 /// the host native double versions. Float versions are not called
1263 /// directly but for all these it is true (float)(f((double)arg)) ==
1264 /// f(arg). Long double not supported yet.
1265 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1266 Op->getValueAPF().convertToDouble();
1269 if (Name == "acos" && TLI->has(LibFunc::acos))
1270 return ConstantFoldFP(acos, V, Ty);
1271 else if (Name == "asin" && TLI->has(LibFunc::asin))
1272 return ConstantFoldFP(asin, V, Ty);
1273 else if (Name == "atan" && TLI->has(LibFunc::atan))
1274 return ConstantFoldFP(atan, V, Ty);
1277 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1278 return ConstantFoldFP(ceil, V, Ty);
1279 else if (Name == "cos" && TLI->has(LibFunc::cos))
1280 return ConstantFoldFP(cos, V, Ty);
1281 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1282 return ConstantFoldFP(cosh, V, Ty);
1283 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1284 return ConstantFoldFP(cos, V, Ty);
1287 if (Name == "exp" && TLI->has(LibFunc::exp))
1288 return ConstantFoldFP(exp, V, Ty);
1290 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1291 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1293 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1297 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1298 return ConstantFoldFP(fabs, V, Ty);
1299 else if (Name == "floor" && TLI->has(LibFunc::floor))
1300 return ConstantFoldFP(floor, V, Ty);
1303 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1304 return ConstantFoldFP(log, V, Ty);
1305 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1306 return ConstantFoldFP(log10, V, Ty);
1307 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1308 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1310 return ConstantFoldFP(sqrt, V, Ty);
1312 return Constant::getNullValue(Ty);
1316 if (Name == "sin" && TLI->has(LibFunc::sin))
1317 return ConstantFoldFP(sin, V, Ty);
1318 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1319 return ConstantFoldFP(sinh, V, Ty);
1320 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1321 return ConstantFoldFP(sqrt, V, Ty);
1322 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1323 return ConstantFoldFP(sqrt, V, Ty);
1324 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1325 return ConstantFoldFP(sin, V, Ty);
1328 if (Name == "tan" && TLI->has(LibFunc::tan))
1329 return ConstantFoldFP(tan, V, Ty);
1330 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1331 return ConstantFoldFP(tanh, V, Ty);
1339 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1340 switch (F->getIntrinsicID()) {
1341 case Intrinsic::bswap:
1342 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1343 case Intrinsic::ctpop:
1344 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1345 case Intrinsic::convert_from_fp16: {
1346 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1349 APFloat::opStatus status =
1350 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1352 // Conversion is always precise.
1354 assert(status == APFloat::opOK && !lost &&
1355 "Precision lost during fp16 constfolding");
1357 return ConstantFP::get(F->getContext(), Val);
1364 // Support ConstantVector in case we have an Undef in the top.
1365 if (isa<ConstantVector>(Operands[0]) ||
1366 isa<ConstantDataVector>(Operands[0])) {
1367 Constant *Op = cast<Constant>(Operands[0]);
1368 switch (F->getIntrinsicID()) {
1370 case Intrinsic::x86_sse_cvtss2si:
1371 case Intrinsic::x86_sse_cvtss2si64:
1372 case Intrinsic::x86_sse2_cvtsd2si:
1373 case Intrinsic::x86_sse2_cvtsd2si64:
1374 if (ConstantFP *FPOp =
1375 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1376 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1377 /*roundTowardZero=*/false, Ty);
1378 case Intrinsic::x86_sse_cvttss2si:
1379 case Intrinsic::x86_sse_cvttss2si64:
1380 case Intrinsic::x86_sse2_cvttsd2si:
1381 case Intrinsic::x86_sse2_cvttsd2si64:
1382 if (ConstantFP *FPOp =
1383 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1384 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1385 /*roundTowardZero=*/true, Ty);
1389 if (isa<UndefValue>(Operands[0])) {
1390 if (F->getIntrinsicID() == Intrinsic::bswap)
1398 if (Operands.size() == 2) {
1399 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1400 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1402 double Op1V = Ty->isFloatTy() ?
1403 (double)Op1->getValueAPF().convertToFloat() :
1404 Op1->getValueAPF().convertToDouble();
1405 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1406 if (Op2->getType() != Op1->getType())
1409 double Op2V = Ty->isFloatTy() ?
1410 (double)Op2->getValueAPF().convertToFloat():
1411 Op2->getValueAPF().convertToDouble();
1413 if (F->getIntrinsicID() == Intrinsic::pow) {
1414 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1418 if (Name == "pow" && TLI->has(LibFunc::pow))
1419 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1420 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1421 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1422 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1423 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1424 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1425 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1426 return ConstantFP::get(F->getContext(),
1427 APFloat((float)std::pow((float)Op1V,
1428 (int)Op2C->getZExtValue())));
1429 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1430 return ConstantFP::get(F->getContext(),
1431 APFloat((double)std::pow((double)Op1V,
1432 (int)Op2C->getZExtValue())));
1437 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1438 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1439 switch (F->getIntrinsicID()) {
1441 case Intrinsic::sadd_with_overflow:
1442 case Intrinsic::uadd_with_overflow:
1443 case Intrinsic::ssub_with_overflow:
1444 case Intrinsic::usub_with_overflow:
1445 case Intrinsic::smul_with_overflow:
1446 case Intrinsic::umul_with_overflow: {
1449 switch (F->getIntrinsicID()) {
1450 default: llvm_unreachable("Invalid case");
1451 case Intrinsic::sadd_with_overflow:
1452 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1454 case Intrinsic::uadd_with_overflow:
1455 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1457 case Intrinsic::ssub_with_overflow:
1458 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1460 case Intrinsic::usub_with_overflow:
1461 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1463 case Intrinsic::smul_with_overflow:
1464 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1466 case Intrinsic::umul_with_overflow:
1467 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1471 ConstantInt::get(F->getContext(), Res),
1472 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1474 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1476 case Intrinsic::cttz:
1477 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1478 return UndefValue::get(Ty);
1479 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1480 case Intrinsic::ctlz:
1481 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1482 return UndefValue::get(Ty);
1483 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());