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/ADT/SmallVector.h"
30 #include "llvm/ADT/StringMap.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/FEnv.h"
39 //===----------------------------------------------------------------------===//
40 // Constant Folding internal helper functions
41 //===----------------------------------------------------------------------===//
43 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
44 /// TargetData. This always returns a non-null constant, but it may be a
45 /// ConstantExpr if unfoldable.
46 static Constant *FoldBitCast(Constant *C, Type *DestTy,
47 const TargetData &TD) {
48 // Catch the obvious splat cases.
49 if (C->isNullValue()) return Constant::getNullValue(DestTy);
50 if (C->isAllOnesValue()) return Constant::getAllOnesValue(DestTy);
52 // The code below only handles casts to vectors currently.
53 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
55 return ConstantExpr::getBitCast(C, DestTy);
57 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
58 // vector so the code below can handle it uniformly.
59 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
60 Constant *Ops = C; // don't take the address of C!
61 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
64 // If this is a bitcast from constant vector -> vector, fold it.
65 ConstantVector *CV = dyn_cast<ConstantVector>(C);
67 return ConstantExpr::getBitCast(C, DestTy);
69 // If the element types match, VMCore can fold it.
70 unsigned NumDstElt = DestVTy->getNumElements();
71 unsigned NumSrcElt = CV->getNumOperands();
72 if (NumDstElt == NumSrcElt)
73 return ConstantExpr::getBitCast(C, DestTy);
75 Type *SrcEltTy = CV->getType()->getElementType();
76 Type *DstEltTy = DestVTy->getElementType();
78 // Otherwise, we're changing the number of elements in a vector, which
79 // requires endianness information to do the right thing. For example,
80 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
81 // folds to (little endian):
82 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
83 // and to (big endian):
84 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
86 // First thing is first. We only want to think about integer here, so if
87 // we have something in FP form, recast it as integer.
88 if (DstEltTy->isFloatingPointTy()) {
89 // Fold to an vector of integers with same size as our FP type.
90 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
92 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
93 // Recursively handle this integer conversion, if possible.
94 C = FoldBitCast(C, DestIVTy, TD);
95 if (!C) return ConstantExpr::getBitCast(C, DestTy);
97 // Finally, VMCore can handle this now that #elts line up.
98 return ConstantExpr::getBitCast(C, DestTy);
101 // Okay, we know the destination is integer, if the input is FP, convert
102 // it to integer first.
103 if (SrcEltTy->isFloatingPointTy()) {
104 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
106 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
107 // Ask VMCore to do the conversion now that #elts line up.
108 C = ConstantExpr::getBitCast(C, SrcIVTy);
109 CV = dyn_cast<ConstantVector>(C);
110 if (!CV) // If VMCore wasn't able to fold it, bail out.
114 // Now we know that the input and output vectors are both integer vectors
115 // of the same size, and that their #elements is not the same. Do the
116 // conversion here, which depends on whether the input or output has
118 bool isLittleEndian = TD.isLittleEndian();
120 SmallVector<Constant*, 32> Result;
121 if (NumDstElt < NumSrcElt) {
122 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
123 Constant *Zero = Constant::getNullValue(DstEltTy);
124 unsigned Ratio = NumSrcElt/NumDstElt;
125 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
127 for (unsigned i = 0; i != NumDstElt; ++i) {
128 // Build each element of the result.
129 Constant *Elt = Zero;
130 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
131 for (unsigned j = 0; j != Ratio; ++j) {
132 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
133 if (!Src) // Reject constantexpr elements.
134 return ConstantExpr::getBitCast(C, DestTy);
136 // Zero extend the element to the right size.
137 Src = ConstantExpr::getZExt(Src, Elt->getType());
139 // Shift it to the right place, depending on endianness.
140 Src = ConstantExpr::getShl(Src,
141 ConstantInt::get(Src->getType(), ShiftAmt));
142 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
145 Elt = ConstantExpr::getOr(Elt, Src);
147 Result.push_back(Elt);
150 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
151 unsigned Ratio = NumDstElt/NumSrcElt;
152 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
154 // Loop over each source value, expanding into multiple results.
155 for (unsigned i = 0; i != NumSrcElt; ++i) {
156 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
157 if (!Src) // Reject constantexpr elements.
158 return ConstantExpr::getBitCast(C, DestTy);
160 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
161 for (unsigned j = 0; j != Ratio; ++j) {
162 // Shift the piece of the value into the right place, depending on
164 Constant *Elt = ConstantExpr::getLShr(Src,
165 ConstantInt::get(Src->getType(), ShiftAmt));
166 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
168 // Truncate and remember this piece.
169 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
174 return ConstantVector::get(Result);
178 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
179 /// from a global, return the global and the constant. Because of
180 /// constantexprs, this function is recursive.
181 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
182 int64_t &Offset, const TargetData &TD) {
183 // Trivial case, constant is the global.
184 if ((GV = dyn_cast<GlobalValue>(C))) {
189 // Otherwise, if this isn't a constant expr, bail out.
190 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
191 if (!CE) return false;
193 // Look through ptr->int and ptr->ptr casts.
194 if (CE->getOpcode() == Instruction::PtrToInt ||
195 CE->getOpcode() == Instruction::BitCast)
196 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
198 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
199 if (CE->getOpcode() == Instruction::GetElementPtr) {
200 // Cannot compute this if the element type of the pointer is missing size
202 if (!cast<PointerType>(CE->getOperand(0)->getType())
203 ->getElementType()->isSized())
206 // If the base isn't a global+constant, we aren't either.
207 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
210 // Otherwise, add any offset that our operands provide.
211 gep_type_iterator GTI = gep_type_begin(CE);
212 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
213 i != e; ++i, ++GTI) {
214 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
215 if (!CI) return false; // Index isn't a simple constant?
216 if (CI->isZero()) continue; // Not adding anything.
218 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
220 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
222 SequentialType *SQT = cast<SequentialType>(*GTI);
223 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
232 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
233 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
234 /// pointer to copy results into and BytesLeft is the number of bytes left in
235 /// the CurPtr buffer. TD is the target data.
236 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
237 unsigned char *CurPtr, unsigned BytesLeft,
238 const TargetData &TD) {
239 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
240 "Out of range access");
242 // If this element is zero or undefined, we can just return since *CurPtr is
244 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
247 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
248 if (CI->getBitWidth() > 64 ||
249 (CI->getBitWidth() & 7) != 0)
252 uint64_t Val = CI->getZExtValue();
253 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
255 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
256 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
262 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
263 if (CFP->getType()->isDoubleTy()) {
264 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
265 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
267 if (CFP->getType()->isFloatTy()){
268 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
269 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
274 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
275 const StructLayout *SL = TD.getStructLayout(CS->getType());
276 unsigned Index = SL->getElementContainingOffset(ByteOffset);
277 uint64_t CurEltOffset = SL->getElementOffset(Index);
278 ByteOffset -= CurEltOffset;
281 // If the element access is to the element itself and not to tail padding,
282 // read the bytes from the element.
283 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
285 if (ByteOffset < EltSize &&
286 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
292 // Check to see if we read from the last struct element, if so we're done.
293 if (Index == CS->getType()->getNumElements())
296 // If we read all of the bytes we needed from this element we're done.
297 uint64_t NextEltOffset = SL->getElementOffset(Index);
299 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
302 // Move to the next element of the struct.
303 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
304 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
306 CurEltOffset = NextEltOffset;
311 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
312 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
313 uint64_t Index = ByteOffset / EltSize;
314 uint64_t Offset = ByteOffset - Index * EltSize;
315 for (; Index != CA->getType()->getNumElements(); ++Index) {
316 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
319 if (EltSize >= BytesLeft)
323 BytesLeft -= EltSize;
329 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
330 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
331 uint64_t Index = ByteOffset / EltSize;
332 uint64_t Offset = ByteOffset - Index * EltSize;
333 for (; Index != CV->getType()->getNumElements(); ++Index) {
334 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
337 if (EltSize >= BytesLeft)
341 BytesLeft -= EltSize;
347 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
348 if (CE->getOpcode() == Instruction::IntToPtr &&
349 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
350 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
354 // Otherwise, unknown initializer type.
358 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
359 const TargetData &TD) {
360 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
361 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
363 // If this isn't an integer load we can't fold it directly.
365 // If this is a float/double load, we can try folding it as an int32/64 load
366 // and then bitcast the result. This can be useful for union cases. Note
367 // that address spaces don't matter here since we're not going to result in
368 // an actual new load.
370 if (LoadTy->isFloatTy())
371 MapTy = Type::getInt32PtrTy(C->getContext());
372 else if (LoadTy->isDoubleTy())
373 MapTy = Type::getInt64PtrTy(C->getContext());
374 else if (LoadTy->isVectorTy()) {
375 MapTy = IntegerType::get(C->getContext(),
376 TD.getTypeAllocSizeInBits(LoadTy));
377 MapTy = PointerType::getUnqual(MapTy);
381 C = FoldBitCast(C, MapTy, TD);
382 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
383 return FoldBitCast(Res, LoadTy, TD);
387 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
388 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
392 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
395 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
396 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
397 !GV->getInitializer()->getType()->isSized())
400 // If we're loading off the beginning of the global, some bytes may be valid,
401 // but we don't try to handle this.
402 if (Offset < 0) return 0;
404 // If we're not accessing anything in this constant, the result is undefined.
405 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
406 return UndefValue::get(IntType);
408 unsigned char RawBytes[32] = {0};
409 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
413 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
414 for (unsigned i = 1; i != BytesLoaded; ++i) {
416 ResultVal |= RawBytes[BytesLoaded-1-i];
419 return ConstantInt::get(IntType->getContext(), ResultVal);
422 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
423 /// produce if it is constant and determinable. If this is not determinable,
425 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
426 const TargetData *TD) {
427 // First, try the easy cases:
428 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
429 if (GV->isConstant() && GV->hasDefinitiveInitializer())
430 return GV->getInitializer();
432 // If the loaded value isn't a constant expr, we can't handle it.
433 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
436 if (CE->getOpcode() == Instruction::GetElementPtr) {
437 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
438 if (GV->isConstant() && GV->hasDefinitiveInitializer())
440 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
444 // Instead of loading constant c string, use corresponding integer value
445 // directly if string length is small enough.
447 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
448 unsigned StrLen = Str.length();
449 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
450 unsigned NumBits = Ty->getPrimitiveSizeInBits();
451 // Replace load with immediate integer if the result is an integer or fp
453 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
454 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
455 APInt StrVal(NumBits, 0);
456 APInt SingleChar(NumBits, 0);
457 if (TD->isLittleEndian()) {
458 for (signed i = StrLen-1; i >= 0; i--) {
459 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
460 StrVal = (StrVal << 8) | SingleChar;
463 for (unsigned i = 0; i < StrLen; i++) {
464 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
465 StrVal = (StrVal << 8) | SingleChar;
467 // Append NULL at the end.
469 StrVal = (StrVal << 8) | SingleChar;
472 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
473 if (Ty->isFloatingPointTy())
474 Res = ConstantExpr::getBitCast(Res, Ty);
479 // If this load comes from anywhere in a constant global, and if the global
480 // is all undef or zero, we know what it loads.
481 if (GlobalVariable *GV =
482 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
483 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
484 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
485 if (GV->getInitializer()->isNullValue())
486 return Constant::getNullValue(ResTy);
487 if (isa<UndefValue>(GV->getInitializer()))
488 return UndefValue::get(ResTy);
492 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
493 // currently don't do any of this for big endian systems. It can be
494 // generalized in the future if someone is interested.
495 if (TD && TD->isLittleEndian())
496 return FoldReinterpretLoadFromConstPtr(CE, *TD);
500 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
501 if (LI->isVolatile()) return 0;
503 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
504 return ConstantFoldLoadFromConstPtr(C, TD);
509 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
510 /// Attempt to symbolically evaluate the result of a binary operator merging
511 /// these together. If target data info is available, it is provided as TD,
512 /// otherwise TD is null.
513 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
514 Constant *Op1, const TargetData *TD){
517 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
518 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
522 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
523 // constant. This happens frequently when iterating over a global array.
524 if (Opc == Instruction::Sub && TD) {
525 GlobalValue *GV1, *GV2;
526 int64_t Offs1, Offs2;
528 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
529 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
531 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
532 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
539 /// CastGEPIndices - If array indices are not pointer-sized integers,
540 /// explicitly cast them so that they aren't implicitly casted by the
542 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
544 const TargetData *TD) {
546 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
549 SmallVector<Constant*, 32> NewIdxs;
550 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
552 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
553 Ops.slice(1, i-1)))) &&
554 Ops[i]->getType() != IntPtrTy) {
556 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
562 NewIdxs.push_back(Ops[i]);
567 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
568 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
569 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
574 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
575 /// constant expression, do so.
576 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
578 const TargetData *TD) {
579 Constant *Ptr = Ops[0];
580 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
583 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
585 // If this is a constant expr gep that is effectively computing an
586 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
587 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
588 if (!isa<ConstantInt>(Ops[i])) {
590 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
591 // "inttoptr (sub (ptrtoint Ptr), V)"
592 if (Ops.size() == 2 &&
593 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
594 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
595 assert((CE == 0 || CE->getType() == IntPtrTy) &&
596 "CastGEPIndices didn't canonicalize index types!");
597 if (CE && CE->getOpcode() == Instruction::Sub &&
598 CE->getOperand(0)->isNullValue()) {
599 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
600 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
601 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
602 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
603 Res = ConstantFoldConstantExpression(ResCE, TD);
610 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
612 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
613 makeArrayRef((Value **)Ops.data() + 1,
615 Ptr = cast<Constant>(Ptr->stripPointerCasts());
617 // If this is a GEP of a GEP, fold it all into a single GEP.
618 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
619 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
621 // Do not try the incorporate the sub-GEP if some index is not a number.
622 bool AllConstantInt = true;
623 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
624 if (!isa<ConstantInt>(NestedOps[i])) {
625 AllConstantInt = false;
631 Ptr = cast<Constant>(GEP->getOperand(0));
632 Offset += APInt(BitWidth,
633 TD->getIndexedOffset(Ptr->getType(), NestedOps));
634 Ptr = cast<Constant>(Ptr->stripPointerCasts());
637 // If the base value for this address is a literal integer value, fold the
638 // getelementptr to the resulting integer value casted to the pointer type.
639 APInt BasePtr(BitWidth, 0);
640 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
641 if (CE->getOpcode() == Instruction::IntToPtr)
642 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
643 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
644 if (Ptr->isNullValue() || BasePtr != 0) {
645 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
646 return ConstantExpr::getIntToPtr(C, ResultTy);
649 // Otherwise form a regular getelementptr. Recompute the indices so that
650 // we eliminate over-indexing of the notional static type array bounds.
651 // This makes it easy to determine if the getelementptr is "inbounds".
652 // Also, this helps GlobalOpt do SROA on GlobalVariables.
653 Type *Ty = Ptr->getType();
654 SmallVector<Constant*, 32> NewIdxs;
656 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
657 if (ATy->isPointerTy()) {
658 // The only pointer indexing we'll do is on the first index of the GEP.
659 if (!NewIdxs.empty())
662 // Only handle pointers to sized types, not pointers to functions.
663 if (!ATy->getElementType()->isSized())
667 // Determine which element of the array the offset points into.
668 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
669 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
671 // The element size is 0. This may be [0 x Ty]*, so just use a zero
672 // index for this level and proceed to the next level to see if it can
673 // accommodate the offset.
674 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
676 // The element size is non-zero divide the offset by the element
677 // size (rounding down), to compute the index at this level.
678 APInt NewIdx = Offset.udiv(ElemSize);
679 Offset -= NewIdx * ElemSize;
680 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
682 Ty = ATy->getElementType();
683 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
684 // Determine which field of the struct the offset points into. The
685 // getZExtValue is at least as safe as the StructLayout API because we
686 // know the offset is within the struct at this point.
687 const StructLayout &SL = *TD->getStructLayout(STy);
688 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
689 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
691 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
692 Ty = STy->getTypeAtIndex(ElIdx);
694 // We've reached some non-indexable type.
697 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
699 // If we haven't used up the entire offset by descending the static
700 // type, then the offset is pointing into the middle of an indivisible
701 // member, so we can't simplify it.
707 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
708 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
709 "Computed GetElementPtr has unexpected type!");
711 // If we ended up indexing a member with a type that doesn't match
712 // the type of what the original indices indexed, add a cast.
713 if (Ty != cast<PointerType>(ResultTy)->getElementType())
714 C = FoldBitCast(C, ResultTy, *TD);
721 //===----------------------------------------------------------------------===//
722 // Constant Folding public APIs
723 //===----------------------------------------------------------------------===//
725 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
726 /// If successful, the constant result is returned, if not, null is returned.
727 /// Note that this fails if not all of the operands are constant. Otherwise,
728 /// this function can only fail when attempting to fold instructions like loads
729 /// and stores, which have no constant expression form.
730 Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
731 // Handle PHI nodes quickly here...
732 if (PHINode *PN = dyn_cast<PHINode>(I)) {
733 Constant *CommonValue = 0;
735 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
736 Value *Incoming = PN->getIncomingValue(i);
737 // If the incoming value is undef then skip it. Note that while we could
738 // skip the value if it is equal to the phi node itself we choose not to
739 // because that would break the rule that constant folding only applies if
740 // all operands are constants.
741 if (isa<UndefValue>(Incoming))
743 // If the incoming value is not a constant, or is a different constant to
744 // the one we saw previously, then give up.
745 Constant *C = dyn_cast<Constant>(Incoming);
746 if (!C || (CommonValue && C != CommonValue))
751 // If we reach here, all incoming values are the same constant or undef.
752 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
755 // Scan the operand list, checking to see if they are all constants, if so,
756 // hand off to ConstantFoldInstOperands.
757 SmallVector<Constant*, 8> Ops;
758 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
759 if (Constant *Op = dyn_cast<Constant>(*i))
762 return 0; // All operands not constant!
764 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
765 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
768 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
769 return ConstantFoldLoadInst(LI, TD);
771 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
772 return ConstantExpr::getInsertValue(
773 cast<Constant>(IVI->getAggregateOperand()),
774 cast<Constant>(IVI->getInsertedValueOperand()),
777 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
778 return ConstantExpr::getExtractValue(
779 cast<Constant>(EVI->getAggregateOperand()),
782 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD);
785 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
786 /// using the specified TargetData. If successful, the constant result is
787 /// result is returned, if not, null is returned.
788 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
789 const TargetData *TD) {
790 SmallVector<Constant*, 8> Ops;
791 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
793 Constant *NewC = cast<Constant>(*i);
794 // Recursively fold the ConstantExpr's operands.
795 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
796 NewC = ConstantFoldConstantExpression(NewCE, TD);
801 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
803 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD);
806 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
807 /// specified opcode and operands. If successful, the constant result is
808 /// returned, if not, null is returned. Note that this function can fail when
809 /// attempting to fold instructions like loads and stores, which have no
810 /// constant expression form.
812 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
813 /// information, due to only being passed an opcode and operands. Constant
814 /// folding using this function strips this information.
816 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
817 ArrayRef<Constant *> Ops,
818 const TargetData *TD) {
819 // Handle easy binops first.
820 if (Instruction::isBinaryOp(Opcode)) {
821 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
822 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
825 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
830 case Instruction::ICmp:
831 case Instruction::FCmp: assert(0 && "Invalid for compares");
832 case Instruction::Call:
833 if (Function *F = dyn_cast<Function>(Ops.back()))
834 if (canConstantFoldCallTo(F))
835 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1));
837 case Instruction::PtrToInt:
838 // If the input is a inttoptr, eliminate the pair. This requires knowing
839 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
840 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
841 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
842 Constant *Input = CE->getOperand(0);
843 unsigned InWidth = Input->getType()->getScalarSizeInBits();
844 if (TD->getPointerSizeInBits() < InWidth) {
846 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
847 TD->getPointerSizeInBits()));
848 Input = ConstantExpr::getAnd(Input, Mask);
850 // Do a zext or trunc to get to the dest size.
851 return ConstantExpr::getIntegerCast(Input, DestTy, false);
854 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
855 case Instruction::IntToPtr:
856 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
857 // the int size is >= the ptr size. This requires knowing the width of a
858 // pointer, so it can't be done in ConstantExpr::getCast.
859 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
861 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
862 CE->getOpcode() == Instruction::PtrToInt)
863 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
865 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
866 case Instruction::Trunc:
867 case Instruction::ZExt:
868 case Instruction::SExt:
869 case Instruction::FPTrunc:
870 case Instruction::FPExt:
871 case Instruction::UIToFP:
872 case Instruction::SIToFP:
873 case Instruction::FPToUI:
874 case Instruction::FPToSI:
875 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
876 case Instruction::BitCast:
878 return FoldBitCast(Ops[0], DestTy, *TD);
879 return ConstantExpr::getBitCast(Ops[0], DestTy);
880 case Instruction::Select:
881 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
882 case Instruction::ExtractElement:
883 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
884 case Instruction::InsertElement:
885 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
886 case Instruction::ShuffleVector:
887 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
888 case Instruction::GetElementPtr:
889 if (Constant *C = CastGEPIndices(Ops, DestTy, TD))
891 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD))
894 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
898 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
899 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
900 /// returns a constant expression of the specified operands.
902 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
903 Constant *Ops0, Constant *Ops1,
904 const TargetData *TD) {
905 // fold: icmp (inttoptr x), null -> icmp x, 0
906 // fold: icmp (ptrtoint x), 0 -> icmp x, null
907 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
908 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
910 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
911 // around to know if bit truncation is happening.
912 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
913 if (TD && Ops1->isNullValue()) {
914 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
915 if (CE0->getOpcode() == Instruction::IntToPtr) {
916 // Convert the integer value to the right size to ensure we get the
917 // proper extension or truncation.
918 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
920 Constant *Null = Constant::getNullValue(C->getType());
921 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
924 // Only do this transformation if the int is intptrty in size, otherwise
925 // there is a truncation or extension that we aren't modeling.
926 if (CE0->getOpcode() == Instruction::PtrToInt &&
927 CE0->getType() == IntPtrTy) {
928 Constant *C = CE0->getOperand(0);
929 Constant *Null = Constant::getNullValue(C->getType());
930 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
934 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
935 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
936 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
938 if (CE0->getOpcode() == Instruction::IntToPtr) {
939 // Convert the integer value to the right size to ensure we get the
940 // proper extension or truncation.
941 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
943 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
945 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
948 // Only do this transformation if the int is intptrty in size, otherwise
949 // there is a truncation or extension that we aren't modeling.
950 if ((CE0->getOpcode() == Instruction::PtrToInt &&
951 CE0->getType() == IntPtrTy &&
952 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
953 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
954 CE1->getOperand(0), TD);
958 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
959 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
960 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
961 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
963 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
965 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
967 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
968 Constant *Ops[] = { LHS, RHS };
969 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD);
973 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
977 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
978 /// getelementptr constantexpr, return the constant value being addressed by the
979 /// constant expression, or null if something is funny and we can't decide.
980 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
982 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
983 return 0; // Do not allow stepping over the value!
985 // Loop over all of the operands, tracking down which value we are
987 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
988 for (++I; I != E; ++I)
989 if (StructType *STy = dyn_cast<StructType>(*I)) {
990 ConstantInt *CU = cast<ConstantInt>(I.getOperand());
991 assert(CU->getZExtValue() < STy->getNumElements() &&
992 "Struct index out of range!");
993 unsigned El = (unsigned)CU->getZExtValue();
994 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
995 C = CS->getOperand(El);
996 } else if (isa<ConstantAggregateZero>(C)) {
997 C = Constant::getNullValue(STy->getElementType(El));
998 } else if (isa<UndefValue>(C)) {
999 C = UndefValue::get(STy->getElementType(El));
1003 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
1004 if (ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
1005 if (CI->getZExtValue() >= ATy->getNumElements())
1007 if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
1008 C = CA->getOperand(CI->getZExtValue());
1009 else if (isa<ConstantAggregateZero>(C))
1010 C = Constant::getNullValue(ATy->getElementType());
1011 else if (isa<UndefValue>(C))
1012 C = UndefValue::get(ATy->getElementType());
1015 } else if (VectorType *VTy = dyn_cast<VectorType>(*I)) {
1016 if (CI->getZExtValue() >= VTy->getNumElements())
1018 if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
1019 C = CP->getOperand(CI->getZExtValue());
1020 else if (isa<ConstantAggregateZero>(C))
1021 C = Constant::getNullValue(VTy->getElementType());
1022 else if (isa<UndefValue>(C))
1023 C = UndefValue::get(VTy->getElementType());
1036 //===----------------------------------------------------------------------===//
1037 // Constant Folding for Calls
1040 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1041 /// the specified function.
1043 llvm::canConstantFoldCallTo(const Function *F) {
1044 switch (F->getIntrinsicID()) {
1045 case Intrinsic::sqrt:
1046 case Intrinsic::powi:
1047 case Intrinsic::bswap:
1048 case Intrinsic::ctpop:
1049 case Intrinsic::ctlz:
1050 case Intrinsic::cttz:
1051 case Intrinsic::sadd_with_overflow:
1052 case Intrinsic::uadd_with_overflow:
1053 case Intrinsic::ssub_with_overflow:
1054 case Intrinsic::usub_with_overflow:
1055 case Intrinsic::smul_with_overflow:
1056 case Intrinsic::umul_with_overflow:
1057 case Intrinsic::convert_from_fp16:
1058 case Intrinsic::convert_to_fp16:
1059 case Intrinsic::x86_sse_cvtss2si:
1060 case Intrinsic::x86_sse_cvtss2si64:
1061 case Intrinsic::x86_sse_cvttss2si:
1062 case Intrinsic::x86_sse_cvttss2si64:
1063 case Intrinsic::x86_sse2_cvtsd2si:
1064 case Intrinsic::x86_sse2_cvtsd2si64:
1065 case Intrinsic::x86_sse2_cvttsd2si:
1066 case Intrinsic::x86_sse2_cvttsd2si64:
1073 if (!F->hasName()) return false;
1074 StringRef Name = F->getName();
1076 // In these cases, the check of the length is required. We don't want to
1077 // return true for a name like "cos\0blah" which strcmp would return equal to
1078 // "cos", but has length 8.
1080 default: return false;
1082 return Name == "acos" || Name == "asin" ||
1083 Name == "atan" || Name == "atan2";
1085 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1087 return Name == "exp" || Name == "exp2";
1089 return Name == "fabs" || Name == "fmod" || Name == "floor";
1091 return Name == "log" || Name == "log10";
1093 return Name == "pow";
1095 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1096 Name == "sinf" || Name == "sqrtf";
1098 return Name == "tan" || Name == "tanh";
1102 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1104 sys::llvm_fenv_clearexcept();
1106 if (sys::llvm_fenv_testexcept()) {
1107 sys::llvm_fenv_clearexcept();
1111 if (Ty->isFloatTy())
1112 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1113 if (Ty->isDoubleTy())
1114 return ConstantFP::get(Ty->getContext(), APFloat(V));
1115 llvm_unreachable("Can only constant fold float/double");
1116 return 0; // dummy return to suppress warning
1119 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1120 double V, double W, Type *Ty) {
1121 sys::llvm_fenv_clearexcept();
1123 if (sys::llvm_fenv_testexcept()) {
1124 sys::llvm_fenv_clearexcept();
1128 if (Ty->isFloatTy())
1129 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1130 if (Ty->isDoubleTy())
1131 return ConstantFP::get(Ty->getContext(), APFloat(V));
1132 llvm_unreachable("Can only constant fold float/double");
1133 return 0; // dummy return to suppress warning
1136 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1137 /// conversion of a constant floating point. If roundTowardZero is false, the
1138 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1139 /// the behavior of the non-truncating SSE instructions in the default rounding
1140 /// mode. The desired integer type Ty is used to select how many bits are
1141 /// available for the result. Returns null if the conversion cannot be
1142 /// performed, otherwise returns the Constant value resulting from the
1144 static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero,
1146 assert(Op && "Called with NULL operand");
1147 APFloat Val(Op->getValueAPF());
1149 // All of these conversion intrinsics form an integer of at most 64bits.
1150 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1151 assert(ResultWidth <= 64 &&
1152 "Can only constant fold conversions to 64 and 32 bit ints");
1155 bool isExact = false;
1156 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1157 : APFloat::rmNearestTiesToEven;
1158 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1159 /*isSigned=*/true, mode,
1161 if (status != APFloat::opOK && status != APFloat::opInexact)
1163 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1166 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1167 /// with the specified arguments, returning null if unsuccessful.
1169 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands) {
1170 if (!F->hasName()) return 0;
1171 StringRef Name = F->getName();
1173 Type *Ty = F->getReturnType();
1174 if (Operands.size() == 1) {
1175 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1176 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1177 APFloat Val(Op->getValueAPF());
1180 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1182 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1185 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1188 /// We only fold functions with finite arguments. Folding NaN and inf is
1189 /// likely to be aborted with an exception anyway, and some host libms
1190 /// have known errors raising exceptions.
1191 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1194 /// Currently APFloat versions of these functions do not exist, so we use
1195 /// the host native double versions. Float versions are not called
1196 /// directly but for all these it is true (float)(f((double)arg)) ==
1197 /// f(arg). Long double not supported yet.
1198 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1199 Op->getValueAPF().convertToDouble();
1203 return ConstantFoldFP(acos, V, Ty);
1204 else if (Name == "asin")
1205 return ConstantFoldFP(asin, V, Ty);
1206 else if (Name == "atan")
1207 return ConstantFoldFP(atan, V, Ty);
1211 return ConstantFoldFP(ceil, V, Ty);
1212 else if (Name == "cos")
1213 return ConstantFoldFP(cos, V, Ty);
1214 else if (Name == "cosh")
1215 return ConstantFoldFP(cosh, V, Ty);
1216 else if (Name == "cosf")
1217 return ConstantFoldFP(cos, V, Ty);
1221 return ConstantFoldFP(exp, V, Ty);
1223 if (Name == "exp2") {
1224 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1226 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1231 return ConstantFoldFP(fabs, V, Ty);
1232 else if (Name == "floor")
1233 return ConstantFoldFP(floor, V, Ty);
1236 if (Name == "log" && V > 0)
1237 return ConstantFoldFP(log, V, Ty);
1238 else if (Name == "log10" && V > 0)
1239 return ConstantFoldFP(log10, V, Ty);
1240 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1241 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1243 return ConstantFoldFP(sqrt, V, Ty);
1245 return Constant::getNullValue(Ty);
1250 return ConstantFoldFP(sin, V, Ty);
1251 else if (Name == "sinh")
1252 return ConstantFoldFP(sinh, V, Ty);
1253 else if (Name == "sqrt" && V >= 0)
1254 return ConstantFoldFP(sqrt, V, Ty);
1255 else if (Name == "sqrtf" && V >= 0)
1256 return ConstantFoldFP(sqrt, V, Ty);
1257 else if (Name == "sinf")
1258 return ConstantFoldFP(sin, V, Ty);
1262 return ConstantFoldFP(tan, V, Ty);
1263 else if (Name == "tanh")
1264 return ConstantFoldFP(tanh, V, Ty);
1272 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1273 switch (F->getIntrinsicID()) {
1274 case Intrinsic::bswap:
1275 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1276 case Intrinsic::ctpop:
1277 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1278 case Intrinsic::cttz:
1279 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
1280 case Intrinsic::ctlz:
1281 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
1282 case Intrinsic::convert_from_fp16: {
1283 APFloat Val(Op->getValue());
1286 APFloat::opStatus status =
1287 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1289 // Conversion is always precise.
1291 assert(status == APFloat::opOK && !lost &&
1292 "Precision lost during fp16 constfolding");
1294 return ConstantFP::get(F->getContext(), Val);
1301 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) {
1302 switch (F->getIntrinsicID()) {
1304 case Intrinsic::x86_sse_cvtss2si:
1305 case Intrinsic::x86_sse_cvtss2si64:
1306 case Intrinsic::x86_sse2_cvtsd2si:
1307 case Intrinsic::x86_sse2_cvtsd2si64:
1308 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1309 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty);
1310 case Intrinsic::x86_sse_cvttss2si:
1311 case Intrinsic::x86_sse_cvttss2si64:
1312 case Intrinsic::x86_sse2_cvttsd2si:
1313 case Intrinsic::x86_sse2_cvttsd2si64:
1314 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0)))
1315 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty);
1319 if (isa<UndefValue>(Operands[0])) {
1320 if (F->getIntrinsicID() == Intrinsic::bswap)
1328 if (Operands.size() == 2) {
1329 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1330 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1332 double Op1V = Ty->isFloatTy() ?
1333 (double)Op1->getValueAPF().convertToFloat() :
1334 Op1->getValueAPF().convertToDouble();
1335 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1336 if (Op2->getType() != Op1->getType())
1339 double Op2V = Ty->isFloatTy() ?
1340 (double)Op2->getValueAPF().convertToFloat():
1341 Op2->getValueAPF().convertToDouble();
1344 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1346 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1347 if (Name == "atan2")
1348 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1349 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1350 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1351 return ConstantFP::get(F->getContext(),
1352 APFloat((float)std::pow((float)Op1V,
1353 (int)Op2C->getZExtValue())));
1354 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1355 return ConstantFP::get(F->getContext(),
1356 APFloat((double)std::pow((double)Op1V,
1357 (int)Op2C->getZExtValue())));
1363 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1364 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1365 switch (F->getIntrinsicID()) {
1367 case Intrinsic::sadd_with_overflow:
1368 case Intrinsic::uadd_with_overflow:
1369 case Intrinsic::ssub_with_overflow:
1370 case Intrinsic::usub_with_overflow:
1371 case Intrinsic::smul_with_overflow:
1372 case Intrinsic::umul_with_overflow: {
1375 switch (F->getIntrinsicID()) {
1376 default: assert(0 && "Invalid case");
1377 case Intrinsic::sadd_with_overflow:
1378 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1380 case Intrinsic::uadd_with_overflow:
1381 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1383 case Intrinsic::ssub_with_overflow:
1384 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1386 case Intrinsic::usub_with_overflow:
1387 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1389 case Intrinsic::smul_with_overflow:
1390 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1392 case Intrinsic::umul_with_overflow:
1393 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1397 ConstantInt::get(F->getContext(), Res),
1398 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1400 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);