1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
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 implements folding of constants for LLVM. This implements the
11 // (internal) ConstantFold.h interface, which is used by the
12 // ConstantExpr::get* methods to automatically fold constants when possible.
14 // The current constant folding implementation is implemented in two pieces: the
15 // pieces that don't need TargetData, and the pieces that do. This is to avoid
16 // a dependence in VMCore on Target.
18 //===----------------------------------------------------------------------===//
20 #include "ConstantFold.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Support/MathExtras.h"
36 //===----------------------------------------------------------------------===//
37 // ConstantFold*Instruction Implementations
38 //===----------------------------------------------------------------------===//
40 /// BitCastConstantVector - Convert the specified ConstantVector node to the
41 /// specified vector type. At this point, we know that the elements of the
42 /// input vector constant are all simple integer or FP values.
43 static Constant *BitCastConstantVector(ConstantVector *CV,
44 const VectorType *DstTy) {
45 // If this cast changes element count then we can't handle it here:
46 // doing so requires endianness information. This should be handled by
47 // Analysis/ConstantFolding.cpp
48 unsigned NumElts = DstTy->getNumElements();
49 if (NumElts != CV->getNumOperands())
52 // Check to verify that all elements of the input are simple.
53 for (unsigned i = 0; i != NumElts; ++i) {
54 if (!isa<ConstantInt>(CV->getOperand(i)) &&
55 !isa<ConstantFP>(CV->getOperand(i)))
59 // Bitcast each element now.
60 std::vector<Constant*> Result;
61 const Type *DstEltTy = DstTy->getElementType();
62 for (unsigned i = 0; i != NumElts; ++i)
63 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i),
65 return ConstantVector::get(Result);
68 /// This function determines which opcode to use to fold two constant cast
69 /// expressions together. It uses CastInst::isEliminableCastPair to determine
70 /// the opcode. Consequently its just a wrapper around that function.
71 /// @brief Determine if it is valid to fold a cast of a cast
74 unsigned opc, ///< opcode of the second cast constant expression
75 ConstantExpr *Op, ///< the first cast constant expression
76 const Type *DstTy ///< desintation type of the first cast
78 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
79 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
80 assert(CastInst::isCast(opc) && "Invalid cast opcode");
82 // The the types and opcodes for the two Cast constant expressions
83 const Type *SrcTy = Op->getOperand(0)->getType();
84 const Type *MidTy = Op->getType();
85 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
86 Instruction::CastOps secondOp = Instruction::CastOps(opc);
88 // Let CastInst::isEliminableCastPair do the heavy lifting.
89 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
90 Type::getInt64Ty(DstTy->getContext()));
93 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
94 const Type *SrcTy = V->getType();
96 return V; // no-op cast
98 // Check to see if we are casting a pointer to an aggregate to a pointer to
99 // the first element. If so, return the appropriate GEP instruction.
100 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
101 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
102 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
103 SmallVector<Value*, 8> IdxList;
105 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
106 IdxList.push_back(Zero);
107 const Type *ElTy = PTy->getElementType();
108 while (ElTy != DPTy->getElementType()) {
109 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
110 if (STy->getNumElements() == 0) break;
111 ElTy = STy->getElementType(0);
112 IdxList.push_back(Zero);
113 } else if (const SequentialType *STy =
114 dyn_cast<SequentialType>(ElTy)) {
115 if (ElTy->isPointerTy()) break; // Can't index into pointers!
116 ElTy = STy->getElementType();
117 IdxList.push_back(Zero);
123 if (ElTy == DPTy->getElementType())
124 // This GEP is inbounds because all indices are zero.
125 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0],
129 // Handle casts from one vector constant to another. We know that the src
130 // and dest type have the same size (otherwise its an illegal cast).
131 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
132 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
133 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
134 "Not cast between same sized vectors!");
136 // First, check for null. Undef is already handled.
137 if (isa<ConstantAggregateZero>(V))
138 return Constant::getNullValue(DestTy);
140 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
141 return BitCastConstantVector(CV, DestPTy);
144 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
145 // This allows for other simplifications (although some of them
146 // can only be handled by Analysis/ConstantFolding.cpp).
147 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
148 return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy);
151 // Finally, implement bitcast folding now. The code below doesn't handle
153 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
154 return ConstantPointerNull::get(cast<PointerType>(DestTy));
156 // Handle integral constant input.
157 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
158 if (DestTy->isIntegerTy())
159 // Integral -> Integral. This is a no-op because the bit widths must
160 // be the same. Consequently, we just fold to V.
163 if (DestTy->isFloatingPointTy())
164 return ConstantFP::get(DestTy->getContext(),
165 APFloat(CI->getValue(),
166 !DestTy->isPPC_FP128Ty()));
168 // Otherwise, can't fold this (vector?)
172 // Handle ConstantFP input: FP -> Integral.
173 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
174 return ConstantInt::get(FP->getContext(),
175 FP->getValueAPF().bitcastToAPInt());
181 /// ExtractConstantBytes - V is an integer constant which only has a subset of
182 /// its bytes used. The bytes used are indicated by ByteStart (which is the
183 /// first byte used, counting from the least significant byte) and ByteSize,
184 /// which is the number of bytes used.
186 /// This function analyzes the specified constant to see if the specified byte
187 /// range can be returned as a simplified constant. If so, the constant is
188 /// returned, otherwise null is returned.
190 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
192 assert(C->getType()->isIntegerTy() &&
193 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
194 "Non-byte sized integer input");
195 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
196 assert(ByteSize && "Must be accessing some piece");
197 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
198 assert(ByteSize != CSize && "Should not extract everything");
200 // Constant Integers are simple.
201 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
202 APInt V = CI->getValue();
204 V = V.lshr(ByteStart*8);
205 V = V.trunc(ByteSize*8);
206 return ConstantInt::get(CI->getContext(), V);
209 // In the input is a constant expr, we might be able to recursively simplify.
210 // If not, we definitely can't do anything.
211 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
212 if (CE == 0) return 0;
214 switch (CE->getOpcode()) {
216 case Instruction::Or: {
217 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
223 if (RHSC->isAllOnesValue())
226 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
229 return ConstantExpr::getOr(LHS, RHS);
231 case Instruction::And: {
232 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
237 if (RHS->isNullValue())
240 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
243 return ConstantExpr::getAnd(LHS, RHS);
245 case Instruction::LShr: {
246 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
249 unsigned ShAmt = Amt->getZExtValue();
250 // Cannot analyze non-byte shifts.
251 if ((ShAmt & 7) != 0)
255 // If the extract is known to be all zeros, return zero.
256 if (ByteStart >= CSize-ShAmt)
257 return Constant::getNullValue(IntegerType::get(CE->getContext(),
259 // If the extract is known to be fully in the input, extract it.
260 if (ByteStart+ByteSize+ShAmt <= CSize)
261 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
263 // TODO: Handle the 'partially zero' case.
267 case Instruction::Shl: {
268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
271 unsigned ShAmt = Amt->getZExtValue();
272 // Cannot analyze non-byte shifts.
273 if ((ShAmt & 7) != 0)
277 // If the extract is known to be all zeros, return zero.
278 if (ByteStart+ByteSize <= ShAmt)
279 return Constant::getNullValue(IntegerType::get(CE->getContext(),
281 // If the extract is known to be fully in the input, extract it.
282 if (ByteStart >= ShAmt)
283 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
285 // TODO: Handle the 'partially zero' case.
289 case Instruction::ZExt: {
290 unsigned SrcBitSize =
291 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
293 // If extracting something that is completely zero, return 0.
294 if (ByteStart*8 >= SrcBitSize)
295 return Constant::getNullValue(IntegerType::get(CE->getContext(),
298 // If exactly extracting the input, return it.
299 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
300 return CE->getOperand(0);
302 // If extracting something completely in the input, if if the input is a
303 // multiple of 8 bits, recurse.
304 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
305 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
307 // Otherwise, if extracting a subset of the input, which is not multiple of
308 // 8 bits, do a shift and trunc to get the bits.
309 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
310 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
311 Constant *Res = CE->getOperand(0);
313 Res = ConstantExpr::getLShr(Res,
314 ConstantInt::get(Res->getType(), ByteStart*8));
315 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
319 // TODO: Handle the 'partially zero' case.
325 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
326 /// on Ty, with any known factors factored out. If Folded is false,
327 /// return null if no factoring was possible, to avoid endlessly
328 /// bouncing an unfoldable expression back into the top-level folder.
330 static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy,
332 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
333 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
334 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
335 return ConstantExpr::getNUWMul(E, N);
338 if (const StructType *STy = dyn_cast<StructType>(Ty))
339 if (!STy->isPacked()) {
340 unsigned NumElems = STy->getNumElements();
341 // An empty struct has size zero.
343 return ConstantExpr::getNullValue(DestTy);
344 // Check for a struct with all members having the same size.
345 Constant *MemberSize =
346 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
348 for (unsigned i = 1; i != NumElems; ++i)
350 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
355 Constant *N = ConstantInt::get(DestTy, NumElems);
356 return ConstantExpr::getNUWMul(MemberSize, N);
360 // Pointer size doesn't depend on the pointee type, so canonicalize them
361 // to an arbitrary pointee.
362 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
363 if (!PTy->getElementType()->isIntegerTy(1))
365 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
366 PTy->getAddressSpace()),
369 // If there's no interesting folding happening, bail so that we don't create
370 // a constant that looks like it needs folding but really doesn't.
374 // Base case: Get a regular sizeof expression.
375 Constant *C = ConstantExpr::getSizeOf(Ty);
376 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
382 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
383 /// on Ty, with any known factors factored out. If Folded is false,
384 /// return null if no factoring was possible, to avoid endlessly
385 /// bouncing an unfoldable expression back into the top-level folder.
387 static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy,
389 // The alignment of an array is equal to the alignment of the
390 // array element. Note that this is not always true for vectors.
391 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
392 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
393 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
400 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
401 // Packed structs always have an alignment of 1.
403 return ConstantInt::get(DestTy, 1);
405 // Otherwise, struct alignment is the maximum alignment of any member.
406 // Without target data, we can't compare much, but we can check to see
407 // if all the members have the same alignment.
408 unsigned NumElems = STy->getNumElements();
409 // An empty struct has minimal alignment.
411 return ConstantInt::get(DestTy, 1);
412 // Check for a struct with all members having the same alignment.
413 Constant *MemberAlign =
414 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
416 for (unsigned i = 1; i != NumElems; ++i)
417 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
425 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
426 // to an arbitrary pointee.
427 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
428 if (!PTy->getElementType()->isIntegerTy(1))
430 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
432 PTy->getAddressSpace()),
435 // If there's no interesting folding happening, bail so that we don't create
436 // a constant that looks like it needs folding but really doesn't.
440 // Base case: Get a regular alignof expression.
441 Constant *C = ConstantExpr::getAlignOf(Ty);
442 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
448 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
449 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
450 /// return null if no factoring was possible, to avoid endlessly
451 /// bouncing an unfoldable expression back into the top-level folder.
453 static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo,
456 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
457 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
460 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
461 return ConstantExpr::getNUWMul(E, N);
464 if (const StructType *STy = dyn_cast<StructType>(Ty))
465 if (!STy->isPacked()) {
466 unsigned NumElems = STy->getNumElements();
467 // An empty struct has no members.
470 // Check for a struct with all members having the same size.
471 Constant *MemberSize =
472 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
474 for (unsigned i = 1; i != NumElems; ++i)
476 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
481 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
486 return ConstantExpr::getNUWMul(MemberSize, N);
490 // If there's no interesting folding happening, bail so that we don't create
491 // a constant that looks like it needs folding but really doesn't.
495 // Base case: Get a regular offsetof expression.
496 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
497 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
503 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
504 const Type *DestTy) {
505 if (isa<UndefValue>(V)) {
506 // zext(undef) = 0, because the top bits will be zero.
507 // sext(undef) = 0, because the top bits will all be the same.
508 // [us]itofp(undef) = 0, because the result value is bounded.
509 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
510 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
511 return Constant::getNullValue(DestTy);
512 return UndefValue::get(DestTy);
514 // No compile-time operations on this type yet.
515 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
518 // If the cast operand is a constant expression, there's a few things we can
519 // do to try to simplify it.
520 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
522 // Try hard to fold cast of cast because they are often eliminable.
523 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
524 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
525 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
526 // If all of the indexes in the GEP are null values, there is no pointer
527 // adjustment going on. We might as well cast the source pointer.
528 bool isAllNull = true;
529 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
530 if (!CE->getOperand(i)->isNullValue()) {
535 // This is casting one pointer type to another, always BitCast
536 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
540 // If the cast operand is a constant vector, perform the cast by
541 // operating on each element. In the cast of bitcasts, the element
542 // count may be mismatched; don't attempt to handle that here.
543 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
544 if (DestTy->isVectorTy() &&
545 cast<VectorType>(DestTy)->getNumElements() ==
546 CV->getType()->getNumElements()) {
547 std::vector<Constant*> res;
548 const VectorType *DestVecTy = cast<VectorType>(DestTy);
549 const Type *DstEltTy = DestVecTy->getElementType();
550 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
551 res.push_back(ConstantExpr::getCast(opc,
552 CV->getOperand(i), DstEltTy));
553 return ConstantVector::get(DestVecTy, res);
556 // We actually have to do a cast now. Perform the cast according to the
560 llvm_unreachable("Failed to cast constant expression");
561 case Instruction::FPTrunc:
562 case Instruction::FPExt:
563 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
565 APFloat Val = FPC->getValueAPF();
566 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
567 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
568 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
569 DestTy->isFP128Ty() ? APFloat::IEEEquad :
571 APFloat::rmNearestTiesToEven, &ignored);
572 return ConstantFP::get(V->getContext(), Val);
574 return 0; // Can't fold.
575 case Instruction::FPToUI:
576 case Instruction::FPToSI:
577 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
578 const APFloat &V = FPC->getValueAPF();
581 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
582 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
583 APFloat::rmTowardZero, &ignored);
584 APInt Val(DestBitWidth, 2, x);
585 return ConstantInt::get(FPC->getContext(), Val);
587 return 0; // Can't fold.
588 case Instruction::IntToPtr: //always treated as unsigned
589 if (V->isNullValue()) // Is it an integral null value?
590 return ConstantPointerNull::get(cast<PointerType>(DestTy));
591 return 0; // Other pointer types cannot be casted
592 case Instruction::PtrToInt: // always treated as unsigned
593 // Is it a null pointer value?
594 if (V->isNullValue())
595 return ConstantInt::get(DestTy, 0);
596 // If this is a sizeof-like expression, pull out multiplications by
597 // known factors to expose them to subsequent folding. If it's an
598 // alignof-like expression, factor out known factors.
599 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
600 if (CE->getOpcode() == Instruction::GetElementPtr &&
601 CE->getOperand(0)->isNullValue()) {
603 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
604 if (CE->getNumOperands() == 2) {
605 // Handle a sizeof-like expression.
606 Constant *Idx = CE->getOperand(1);
607 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
608 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
609 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
612 return ConstantExpr::getMul(C, Idx);
614 } else if (CE->getNumOperands() == 3 &&
615 CE->getOperand(1)->isNullValue()) {
616 // Handle an alignof-like expression.
617 if (const StructType *STy = dyn_cast<StructType>(Ty))
618 if (!STy->isPacked()) {
619 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
621 STy->getNumElements() == 2 &&
622 STy->getElementType(0)->isIntegerTy(1)) {
623 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
626 // Handle an offsetof-like expression.
627 if (Ty->isStructTy() || Ty->isArrayTy()) {
628 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
634 // Other pointer types cannot be casted
636 case Instruction::UIToFP:
637 case Instruction::SIToFP:
638 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
639 APInt api = CI->getValue();
640 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true);
641 (void)apf.convertFromAPInt(api,
642 opc==Instruction::SIToFP,
643 APFloat::rmNearestTiesToEven);
644 return ConstantFP::get(V->getContext(), apf);
647 case Instruction::ZExt:
648 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
649 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
650 return ConstantInt::get(V->getContext(),
651 CI->getValue().zext(BitWidth));
654 case Instruction::SExt:
655 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
656 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
657 return ConstantInt::get(V->getContext(),
658 CI->getValue().sext(BitWidth));
661 case Instruction::Trunc: {
662 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
663 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
664 return ConstantInt::get(V->getContext(),
665 CI->getValue().trunc(DestBitWidth));
668 // The input must be a constantexpr. See if we can simplify this based on
669 // the bytes we are demanding. Only do this if the source and dest are an
670 // even multiple of a byte.
671 if ((DestBitWidth & 7) == 0 &&
672 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
673 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
678 case Instruction::BitCast:
679 return FoldBitCast(V, DestTy);
683 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
684 Constant *V1, Constant *V2) {
685 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
686 return CB->getZExtValue() ? V1 : V2;
688 if (isa<UndefValue>(V1)) return V2;
689 if (isa<UndefValue>(V2)) return V1;
690 if (isa<UndefValue>(Cond)) return V1;
691 if (V1 == V2) return V1;
695 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
697 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
698 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
699 if (Val->isNullValue()) // ee(zero, x) -> zero
700 return Constant::getNullValue(
701 cast<VectorType>(Val->getType())->getElementType());
703 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
704 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
705 return CVal->getOperand(CIdx->getZExtValue());
706 } else if (isa<UndefValue>(Idx)) {
707 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
708 return CVal->getOperand(0);
714 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
717 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
719 APInt idxVal = CIdx->getValue();
720 if (isa<UndefValue>(Val)) {
721 // Insertion of scalar constant into vector undef
722 // Optimize away insertion of undef
723 if (isa<UndefValue>(Elt))
725 // Otherwise break the aggregate undef into multiple undefs and do
728 cast<VectorType>(Val->getType())->getNumElements();
729 std::vector<Constant*> Ops;
731 for (unsigned i = 0; i < numOps; ++i) {
733 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
736 return ConstantVector::get(Ops);
738 if (isa<ConstantAggregateZero>(Val)) {
739 // Insertion of scalar constant into vector aggregate zero
740 // Optimize away insertion of zero
741 if (Elt->isNullValue())
743 // Otherwise break the aggregate zero into multiple zeros and do
746 cast<VectorType>(Val->getType())->getNumElements();
747 std::vector<Constant*> Ops;
749 for (unsigned i = 0; i < numOps; ++i) {
751 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
754 return ConstantVector::get(Ops);
756 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
757 // Insertion of scalar constant into vector constant
758 std::vector<Constant*> Ops;
759 Ops.reserve(CVal->getNumOperands());
760 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
762 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
765 return ConstantVector::get(Ops);
771 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
772 /// return the specified element value. Otherwise return null.
773 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
774 if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
775 return CV->getOperand(EltNo);
777 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
778 if (isa<ConstantAggregateZero>(C))
779 return Constant::getNullValue(EltTy);
780 if (isa<UndefValue>(C))
781 return UndefValue::get(EltTy);
785 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
788 // Undefined shuffle mask -> undefined value.
789 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
791 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
792 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
793 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
795 // Loop over the shuffle mask, evaluating each element.
796 SmallVector<Constant*, 32> Result;
797 for (unsigned i = 0; i != MaskNumElts; ++i) {
798 Constant *InElt = GetVectorElement(Mask, i);
799 if (InElt == 0) return 0;
801 if (isa<UndefValue>(InElt))
802 InElt = UndefValue::get(EltTy);
803 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
804 unsigned Elt = CI->getZExtValue();
805 if (Elt >= SrcNumElts*2)
806 InElt = UndefValue::get(EltTy);
807 else if (Elt >= SrcNumElts)
808 InElt = GetVectorElement(V2, Elt - SrcNumElts);
810 InElt = GetVectorElement(V1, Elt);
811 if (InElt == 0) return 0;
816 Result.push_back(InElt);
819 return ConstantVector::get(&Result[0], Result.size());
822 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
823 const unsigned *Idxs,
825 // Base case: no indices, so return the entire value.
829 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
830 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
834 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
836 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
840 // Otherwise recurse.
841 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
842 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs),
845 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
846 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs),
848 ConstantVector *CV = cast<ConstantVector>(Agg);
849 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs),
853 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
855 const unsigned *Idxs,
857 // Base case: no indices, so replace the entire value.
861 if (isa<UndefValue>(Agg)) {
862 // Insertion of constant into aggregate undef
863 // Optimize away insertion of undef.
864 if (isa<UndefValue>(Val))
867 // Otherwise break the aggregate undef into multiple undefs and do
869 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
871 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
872 numOps = AR->getNumElements();
874 numOps = cast<StructType>(AggTy)->getNumElements();
876 std::vector<Constant*> Ops(numOps);
877 for (unsigned i = 0; i < numOps; ++i) {
878 const Type *MemberTy = AggTy->getTypeAtIndex(i);
881 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
882 Val, Idxs+1, NumIdx-1) :
883 UndefValue::get(MemberTy);
887 if (const StructType* ST = dyn_cast<StructType>(AggTy))
888 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
889 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
892 if (isa<ConstantAggregateZero>(Agg)) {
893 // Insertion of constant into aggregate zero
894 // Optimize away insertion of zero.
895 if (Val->isNullValue())
898 // Otherwise break the aggregate zero into multiple zeros and do
900 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
902 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
903 numOps = AR->getNumElements();
905 numOps = cast<StructType>(AggTy)->getNumElements();
907 std::vector<Constant*> Ops(numOps);
908 for (unsigned i = 0; i < numOps; ++i) {
909 const Type *MemberTy = AggTy->getTypeAtIndex(i);
912 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
913 Val, Idxs+1, NumIdx-1) :
914 Constant::getNullValue(MemberTy);
918 if (const StructType *ST = dyn_cast<StructType>(AggTy))
919 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
920 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
923 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
924 // Insertion of constant into aggregate constant.
925 std::vector<Constant*> Ops(Agg->getNumOperands());
926 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
927 Constant *Op = cast<Constant>(Agg->getOperand(i));
929 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1);
933 if (const StructType* ST = dyn_cast<StructType>(Agg->getType()))
934 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
935 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
942 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
943 Constant *C1, Constant *C2) {
944 // No compile-time operations on this type yet.
945 if (C1->getType()->isPPC_FP128Ty())
948 // Handle UndefValue up front.
949 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
951 case Instruction::Xor:
952 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
953 // Handle undef ^ undef -> 0 special case. This is a common
955 return Constant::getNullValue(C1->getType());
957 case Instruction::Add:
958 case Instruction::Sub:
959 return UndefValue::get(C1->getType());
960 case Instruction::Mul:
961 case Instruction::And:
962 return Constant::getNullValue(C1->getType());
963 case Instruction::UDiv:
964 case Instruction::SDiv:
965 case Instruction::URem:
966 case Instruction::SRem:
967 if (!isa<UndefValue>(C2)) // undef / X -> 0
968 return Constant::getNullValue(C1->getType());
969 return C2; // X / undef -> undef
970 case Instruction::Or: // X | undef -> -1
971 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
972 return Constant::getAllOnesValue(PTy);
973 return Constant::getAllOnesValue(C1->getType());
974 case Instruction::LShr:
975 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
976 return C1; // undef lshr undef -> undef
977 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
979 case Instruction::AShr:
980 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
981 return Constant::getAllOnesValue(C1->getType());
982 else if (isa<UndefValue>(C1))
983 return C1; // undef ashr undef -> undef
985 return C1; // X ashr undef --> X
986 case Instruction::Shl:
987 // undef << X -> 0 or X << undef -> 0
988 return Constant::getNullValue(C1->getType());
992 // Handle simplifications when the RHS is a constant int.
993 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
995 case Instruction::Add:
996 if (CI2->equalsInt(0)) return C1; // X + 0 == X
998 case Instruction::Sub:
999 if (CI2->equalsInt(0)) return C1; // X - 0 == X
1001 case Instruction::Mul:
1002 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
1003 if (CI2->equalsInt(1))
1004 return C1; // X * 1 == X
1006 case Instruction::UDiv:
1007 case Instruction::SDiv:
1008 if (CI2->equalsInt(1))
1009 return C1; // X / 1 == X
1010 if (CI2->equalsInt(0))
1011 return UndefValue::get(CI2->getType()); // X / 0 == undef
1013 case Instruction::URem:
1014 case Instruction::SRem:
1015 if (CI2->equalsInt(1))
1016 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1017 if (CI2->equalsInt(0))
1018 return UndefValue::get(CI2->getType()); // X % 0 == undef
1020 case Instruction::And:
1021 if (CI2->isZero()) return C2; // X & 0 == 0
1022 if (CI2->isAllOnesValue())
1023 return C1; // X & -1 == X
1025 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1026 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1027 if (CE1->getOpcode() == Instruction::ZExt) {
1028 unsigned DstWidth = CI2->getType()->getBitWidth();
1030 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1031 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1032 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1036 // If and'ing the address of a global with a constant, fold it.
1037 if (CE1->getOpcode() == Instruction::PtrToInt &&
1038 isa<GlobalValue>(CE1->getOperand(0))) {
1039 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1041 // Functions are at least 4-byte aligned.
1042 unsigned GVAlign = GV->getAlignment();
1043 if (isa<Function>(GV))
1044 GVAlign = std::max(GVAlign, 4U);
1047 unsigned DstWidth = CI2->getType()->getBitWidth();
1048 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1049 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1051 // If checking bits we know are clear, return zero.
1052 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1053 return Constant::getNullValue(CI2->getType());
1058 case Instruction::Or:
1059 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1060 if (CI2->isAllOnesValue())
1061 return C2; // X | -1 == -1
1063 case Instruction::Xor:
1064 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1066 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1067 switch (CE1->getOpcode()) {
1069 case Instruction::ICmp:
1070 case Instruction::FCmp:
1071 // cmp pred ^ true -> cmp !pred
1072 assert(CI2->equalsInt(1));
1073 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1074 pred = CmpInst::getInversePredicate(pred);
1075 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1076 CE1->getOperand(1));
1080 case Instruction::AShr:
1081 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1082 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1083 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1084 return ConstantExpr::getLShr(C1, C2);
1087 } else if (isa<ConstantInt>(C1)) {
1088 // If C1 is a ConstantInt and C2 is not, swap the operands.
1089 if (Instruction::isCommutative(Opcode))
1090 return ConstantExpr::get(Opcode, C2, C1);
1093 // At this point we know neither constant is an UndefValue.
1094 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1095 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1096 using namespace APIntOps;
1097 const APInt &C1V = CI1->getValue();
1098 const APInt &C2V = CI2->getValue();
1102 case Instruction::Add:
1103 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1104 case Instruction::Sub:
1105 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1106 case Instruction::Mul:
1107 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1108 case Instruction::UDiv:
1109 assert(!CI2->isNullValue() && "Div by zero handled above");
1110 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1111 case Instruction::SDiv:
1112 assert(!CI2->isNullValue() && "Div by zero handled above");
1113 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1114 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1115 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1116 case Instruction::URem:
1117 assert(!CI2->isNullValue() && "Div by zero handled above");
1118 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1119 case Instruction::SRem:
1120 assert(!CI2->isNullValue() && "Div by zero handled above");
1121 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1122 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1123 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1124 case Instruction::And:
1125 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1126 case Instruction::Or:
1127 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1128 case Instruction::Xor:
1129 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1130 case Instruction::Shl: {
1131 uint32_t shiftAmt = C2V.getZExtValue();
1132 if (shiftAmt < C1V.getBitWidth())
1133 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1135 return UndefValue::get(C1->getType()); // too big shift is undef
1137 case Instruction::LShr: {
1138 uint32_t shiftAmt = C2V.getZExtValue();
1139 if (shiftAmt < C1V.getBitWidth())
1140 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1142 return UndefValue::get(C1->getType()); // too big shift is undef
1144 case Instruction::AShr: {
1145 uint32_t shiftAmt = C2V.getZExtValue();
1146 if (shiftAmt < C1V.getBitWidth())
1147 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1149 return UndefValue::get(C1->getType()); // too big shift is undef
1155 case Instruction::SDiv:
1156 case Instruction::UDiv:
1157 case Instruction::URem:
1158 case Instruction::SRem:
1159 case Instruction::LShr:
1160 case Instruction::AShr:
1161 case Instruction::Shl:
1162 if (CI1->equalsInt(0)) return C1;
1167 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1168 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1169 APFloat C1V = CFP1->getValueAPF();
1170 APFloat C2V = CFP2->getValueAPF();
1171 APFloat C3V = C1V; // copy for modification
1175 case Instruction::FAdd:
1176 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1177 return ConstantFP::get(C1->getContext(), C3V);
1178 case Instruction::FSub:
1179 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1180 return ConstantFP::get(C1->getContext(), C3V);
1181 case Instruction::FMul:
1182 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1183 return ConstantFP::get(C1->getContext(), C3V);
1184 case Instruction::FDiv:
1185 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1186 return ConstantFP::get(C1->getContext(), C3V);
1187 case Instruction::FRem:
1188 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1189 return ConstantFP::get(C1->getContext(), C3V);
1192 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1193 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1194 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1195 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1196 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1197 std::vector<Constant*> Res;
1198 const Type* EltTy = VTy->getElementType();
1204 case Instruction::Add:
1205 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1206 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1207 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1208 Res.push_back(ConstantExpr::getAdd(C1, C2));
1210 return ConstantVector::get(Res);
1211 case Instruction::FAdd:
1212 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1213 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1214 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1215 Res.push_back(ConstantExpr::getFAdd(C1, C2));
1217 return ConstantVector::get(Res);
1218 case Instruction::Sub:
1219 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1220 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1221 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1222 Res.push_back(ConstantExpr::getSub(C1, C2));
1224 return ConstantVector::get(Res);
1225 case Instruction::FSub:
1226 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1227 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1228 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1229 Res.push_back(ConstantExpr::getFSub(C1, C2));
1231 return ConstantVector::get(Res);
1232 case Instruction::Mul:
1233 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1234 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1235 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1236 Res.push_back(ConstantExpr::getMul(C1, C2));
1238 return ConstantVector::get(Res);
1239 case Instruction::FMul:
1240 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1241 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1242 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1243 Res.push_back(ConstantExpr::getFMul(C1, C2));
1245 return ConstantVector::get(Res);
1246 case Instruction::UDiv:
1247 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1248 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1249 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1250 Res.push_back(ConstantExpr::getUDiv(C1, C2));
1252 return ConstantVector::get(Res);
1253 case Instruction::SDiv:
1254 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1255 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1256 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1257 Res.push_back(ConstantExpr::getSDiv(C1, C2));
1259 return ConstantVector::get(Res);
1260 case Instruction::FDiv:
1261 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1262 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1263 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1264 Res.push_back(ConstantExpr::getFDiv(C1, C2));
1266 return ConstantVector::get(Res);
1267 case Instruction::URem:
1268 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1269 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1270 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1271 Res.push_back(ConstantExpr::getURem(C1, C2));
1273 return ConstantVector::get(Res);
1274 case Instruction::SRem:
1275 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1276 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1277 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1278 Res.push_back(ConstantExpr::getSRem(C1, C2));
1280 return ConstantVector::get(Res);
1281 case Instruction::FRem:
1282 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1283 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1284 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1285 Res.push_back(ConstantExpr::getFRem(C1, C2));
1287 return ConstantVector::get(Res);
1288 case Instruction::And:
1289 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1290 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1291 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1292 Res.push_back(ConstantExpr::getAnd(C1, C2));
1294 return ConstantVector::get(Res);
1295 case Instruction::Or:
1296 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1297 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1298 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1299 Res.push_back(ConstantExpr::getOr(C1, C2));
1301 return ConstantVector::get(Res);
1302 case Instruction::Xor:
1303 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1304 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1305 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1306 Res.push_back(ConstantExpr::getXor(C1, C2));
1308 return ConstantVector::get(Res);
1309 case Instruction::LShr:
1310 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1311 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1312 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1313 Res.push_back(ConstantExpr::getLShr(C1, C2));
1315 return ConstantVector::get(Res);
1316 case Instruction::AShr:
1317 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1318 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1319 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1320 Res.push_back(ConstantExpr::getAShr(C1, C2));
1322 return ConstantVector::get(Res);
1323 case Instruction::Shl:
1324 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1325 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1326 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1327 Res.push_back(ConstantExpr::getShl(C1, C2));
1329 return ConstantVector::get(Res);
1334 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1335 // There are many possible foldings we could do here. We should probably
1336 // at least fold add of a pointer with an integer into the appropriate
1337 // getelementptr. This will improve alias analysis a bit.
1339 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1341 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1342 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1343 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1344 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1346 } else if (isa<ConstantExpr>(C2)) {
1347 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1348 // other way if possible.
1349 if (Instruction::isCommutative(Opcode))
1350 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1353 // i1 can be simplified in many cases.
1354 if (C1->getType()->isIntegerTy(1)) {
1356 case Instruction::Add:
1357 case Instruction::Sub:
1358 return ConstantExpr::getXor(C1, C2);
1359 case Instruction::Mul:
1360 return ConstantExpr::getAnd(C1, C2);
1361 case Instruction::Shl:
1362 case Instruction::LShr:
1363 case Instruction::AShr:
1364 // We can assume that C2 == 0. If it were one the result would be
1365 // undefined because the shift value is as large as the bitwidth.
1367 case Instruction::SDiv:
1368 case Instruction::UDiv:
1369 // We can assume that C2 == 1. If it were zero the result would be
1370 // undefined through division by zero.
1372 case Instruction::URem:
1373 case Instruction::SRem:
1374 // We can assume that C2 == 1. If it were zero the result would be
1375 // undefined through division by zero.
1376 return ConstantInt::getFalse(C1->getContext());
1382 // We don't know how to fold this.
1386 /// isZeroSizedType - This type is zero sized if its an array or structure of
1387 /// zero sized types. The only leaf zero sized type is an empty structure.
1388 static bool isMaybeZeroSizedType(const Type *Ty) {
1389 if (Ty->isOpaqueTy()) return true; // Can't say.
1390 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1392 // If all of elements have zero size, this does too.
1393 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1394 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1397 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1398 return isMaybeZeroSizedType(ATy->getElementType());
1403 /// IdxCompare - Compare the two constants as though they were getelementptr
1404 /// indices. This allows coersion of the types to be the same thing.
1406 /// If the two constants are the "same" (after coersion), return 0. If the
1407 /// first is less than the second, return -1, if the second is less than the
1408 /// first, return 1. If the constants are not integral, return -2.
1410 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
1411 if (C1 == C2) return 0;
1413 // Ok, we found a different index. If they are not ConstantInt, we can't do
1414 // anything with them.
1415 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1416 return -2; // don't know!
1418 // Ok, we have two differing integer indices. Sign extend them to be the same
1419 // type. Long is always big enough, so we use it.
1420 if (!C1->getType()->isIntegerTy(64))
1421 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1423 if (!C2->getType()->isIntegerTy(64))
1424 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1426 if (C1 == C2) return 0; // They are equal
1428 // If the type being indexed over is really just a zero sized type, there is
1429 // no pointer difference being made here.
1430 if (isMaybeZeroSizedType(ElTy))
1431 return -2; // dunno.
1433 // If they are really different, now that they are the same type, then we
1434 // found a difference!
1435 if (cast<ConstantInt>(C1)->getSExtValue() <
1436 cast<ConstantInt>(C2)->getSExtValue())
1442 /// evaluateFCmpRelation - This function determines if there is anything we can
1443 /// decide about the two constants provided. This doesn't need to handle simple
1444 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1445 /// If we can determine that the two constants have a particular relation to
1446 /// each other, we should return the corresponding FCmpInst predicate,
1447 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1448 /// ConstantFoldCompareInstruction.
1450 /// To simplify this code we canonicalize the relation so that the first
1451 /// operand is always the most "complex" of the two. We consider ConstantFP
1452 /// to be the simplest, and ConstantExprs to be the most complex.
1453 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1454 assert(V1->getType() == V2->getType() &&
1455 "Cannot compare values of different types!");
1457 // No compile-time operations on this type yet.
1458 if (V1->getType()->isPPC_FP128Ty())
1459 return FCmpInst::BAD_FCMP_PREDICATE;
1461 // Handle degenerate case quickly
1462 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1464 if (!isa<ConstantExpr>(V1)) {
1465 if (!isa<ConstantExpr>(V2)) {
1466 // We distilled thisUse the standard constant folder for a few cases
1468 R = dyn_cast<ConstantInt>(
1469 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1470 if (R && !R->isZero())
1471 return FCmpInst::FCMP_OEQ;
1472 R = dyn_cast<ConstantInt>(
1473 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1474 if (R && !R->isZero())
1475 return FCmpInst::FCMP_OLT;
1476 R = dyn_cast<ConstantInt>(
1477 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1478 if (R && !R->isZero())
1479 return FCmpInst::FCMP_OGT;
1481 // Nothing more we can do
1482 return FCmpInst::BAD_FCMP_PREDICATE;
1485 // If the first operand is simple and second is ConstantExpr, swap operands.
1486 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1487 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1488 return FCmpInst::getSwappedPredicate(SwappedRelation);
1490 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1491 // constantexpr or a simple constant.
1492 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1493 switch (CE1->getOpcode()) {
1494 case Instruction::FPTrunc:
1495 case Instruction::FPExt:
1496 case Instruction::UIToFP:
1497 case Instruction::SIToFP:
1498 // We might be able to do something with these but we don't right now.
1504 // There are MANY other foldings that we could perform here. They will
1505 // probably be added on demand, as they seem needed.
1506 return FCmpInst::BAD_FCMP_PREDICATE;
1509 /// evaluateICmpRelation - This function determines if there is anything we can
1510 /// decide about the two constants provided. This doesn't need to handle simple
1511 /// things like integer comparisons, but should instead handle ConstantExprs
1512 /// and GlobalValues. If we can determine that the two constants have a
1513 /// particular relation to each other, we should return the corresponding ICmp
1514 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1516 /// To simplify this code we canonicalize the relation so that the first
1517 /// operand is always the most "complex" of the two. We consider simple
1518 /// constants (like ConstantInt) to be the simplest, followed by
1519 /// GlobalValues, followed by ConstantExpr's (the most complex).
1521 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1523 assert(V1->getType() == V2->getType() &&
1524 "Cannot compare different types of values!");
1525 if (V1 == V2) return ICmpInst::ICMP_EQ;
1527 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1528 !isa<BlockAddress>(V1)) {
1529 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1530 !isa<BlockAddress>(V2)) {
1531 // We distilled this down to a simple case, use the standard constant
1534 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1535 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1536 if (R && !R->isZero())
1538 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1539 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1540 if (R && !R->isZero())
1542 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1543 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1544 if (R && !R->isZero())
1547 // If we couldn't figure it out, bail.
1548 return ICmpInst::BAD_ICMP_PREDICATE;
1551 // If the first operand is simple, swap operands.
1552 ICmpInst::Predicate SwappedRelation =
1553 evaluateICmpRelation(V2, V1, isSigned);
1554 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1555 return ICmpInst::getSwappedPredicate(SwappedRelation);
1557 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1558 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1559 ICmpInst::Predicate SwappedRelation =
1560 evaluateICmpRelation(V2, V1, isSigned);
1561 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1562 return ICmpInst::getSwappedPredicate(SwappedRelation);
1563 return ICmpInst::BAD_ICMP_PREDICATE;
1566 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1567 // constant (which, since the types must match, means that it's a
1568 // ConstantPointerNull).
1569 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1570 // Don't try to decide equality of aliases.
1571 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1572 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1573 return ICmpInst::ICMP_NE;
1574 } else if (isa<BlockAddress>(V2)) {
1575 return ICmpInst::ICMP_NE; // Globals never equal labels.
1577 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1578 // GlobalVals can never be null unless they have external weak linkage.
1579 // We don't try to evaluate aliases here.
1580 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1581 return ICmpInst::ICMP_NE;
1583 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1584 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1585 ICmpInst::Predicate SwappedRelation =
1586 evaluateICmpRelation(V2, V1, isSigned);
1587 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1588 return ICmpInst::getSwappedPredicate(SwappedRelation);
1589 return ICmpInst::BAD_ICMP_PREDICATE;
1592 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1593 // constant (which, since the types must match, means that it is a
1594 // ConstantPointerNull).
1595 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1596 // Block address in another function can't equal this one, but block
1597 // addresses in the current function might be the same if blocks are
1599 if (BA2->getFunction() != BA->getFunction())
1600 return ICmpInst::ICMP_NE;
1602 // Block addresses aren't null, don't equal the address of globals.
1603 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1604 "Canonicalization guarantee!");
1605 return ICmpInst::ICMP_NE;
1608 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1609 // constantexpr, a global, block address, or a simple constant.
1610 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1611 Constant *CE1Op0 = CE1->getOperand(0);
1613 switch (CE1->getOpcode()) {
1614 case Instruction::Trunc:
1615 case Instruction::FPTrunc:
1616 case Instruction::FPExt:
1617 case Instruction::FPToUI:
1618 case Instruction::FPToSI:
1619 break; // We can't evaluate floating point casts or truncations.
1621 case Instruction::UIToFP:
1622 case Instruction::SIToFP:
1623 case Instruction::BitCast:
1624 case Instruction::ZExt:
1625 case Instruction::SExt:
1626 // If the cast is not actually changing bits, and the second operand is a
1627 // null pointer, do the comparison with the pre-casted value.
1628 if (V2->isNullValue() &&
1629 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1630 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1631 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1632 return evaluateICmpRelation(CE1Op0,
1633 Constant::getNullValue(CE1Op0->getType()),
1638 case Instruction::GetElementPtr:
1639 // Ok, since this is a getelementptr, we know that the constant has a
1640 // pointer type. Check the various cases.
1641 if (isa<ConstantPointerNull>(V2)) {
1642 // If we are comparing a GEP to a null pointer, check to see if the base
1643 // of the GEP equals the null pointer.
1644 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1645 if (GV->hasExternalWeakLinkage())
1646 // Weak linkage GVals could be zero or not. We're comparing that
1647 // to null pointer so its greater-or-equal
1648 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1650 // If its not weak linkage, the GVal must have a non-zero address
1651 // so the result is greater-than
1652 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1653 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1654 // If we are indexing from a null pointer, check to see if we have any
1655 // non-zero indices.
1656 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1657 if (!CE1->getOperand(i)->isNullValue())
1658 // Offsetting from null, must not be equal.
1659 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1660 // Only zero indexes from null, must still be zero.
1661 return ICmpInst::ICMP_EQ;
1663 // Otherwise, we can't really say if the first operand is null or not.
1664 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1665 if (isa<ConstantPointerNull>(CE1Op0)) {
1666 if (GV2->hasExternalWeakLinkage())
1667 // Weak linkage GVals could be zero or not. We're comparing it to
1668 // a null pointer, so its less-or-equal
1669 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1671 // If its not weak linkage, the GVal must have a non-zero address
1672 // so the result is less-than
1673 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1674 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1676 // If this is a getelementptr of the same global, then it must be
1677 // different. Because the types must match, the getelementptr could
1678 // only have at most one index, and because we fold getelementptr's
1679 // with a single zero index, it must be nonzero.
1680 assert(CE1->getNumOperands() == 2 &&
1681 !CE1->getOperand(1)->isNullValue() &&
1682 "Suprising getelementptr!");
1683 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1685 // If they are different globals, we don't know what the value is,
1686 // but they can't be equal.
1687 return ICmpInst::ICMP_NE;
1691 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1692 Constant *CE2Op0 = CE2->getOperand(0);
1694 // There are MANY other foldings that we could perform here. They will
1695 // probably be added on demand, as they seem needed.
1696 switch (CE2->getOpcode()) {
1698 case Instruction::GetElementPtr:
1699 // By far the most common case to handle is when the base pointers are
1700 // obviously to the same or different globals.
1701 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1702 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1703 return ICmpInst::ICMP_NE;
1704 // Ok, we know that both getelementptr instructions are based on the
1705 // same global. From this, we can precisely determine the relative
1706 // ordering of the resultant pointers.
1709 // The logic below assumes that the result of the comparison
1710 // can be determined by finding the first index that differs.
1711 // This doesn't work if there is over-indexing in any
1712 // subsequent indices, so check for that case first.
1713 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1714 !CE2->isGEPWithNoNotionalOverIndexing())
1715 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1717 // Compare all of the operands the GEP's have in common.
1718 gep_type_iterator GTI = gep_type_begin(CE1);
1719 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1721 switch (IdxCompare(CE1->getOperand(i),
1722 CE2->getOperand(i), GTI.getIndexedType())) {
1723 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1724 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1725 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1728 // Ok, we ran out of things they have in common. If any leftovers
1729 // are non-zero then we have a difference, otherwise we are equal.
1730 for (; i < CE1->getNumOperands(); ++i)
1731 if (!CE1->getOperand(i)->isNullValue()) {
1732 if (isa<ConstantInt>(CE1->getOperand(i)))
1733 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1735 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1738 for (; i < CE2->getNumOperands(); ++i)
1739 if (!CE2->getOperand(i)->isNullValue()) {
1740 if (isa<ConstantInt>(CE2->getOperand(i)))
1741 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1743 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1745 return ICmpInst::ICMP_EQ;
1754 return ICmpInst::BAD_ICMP_PREDICATE;
1757 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1758 Constant *C1, Constant *C2) {
1759 const Type *ResultTy;
1760 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1761 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1762 VT->getNumElements());
1764 ResultTy = Type::getInt1Ty(C1->getContext());
1766 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1767 if (pred == FCmpInst::FCMP_FALSE)
1768 return Constant::getNullValue(ResultTy);
1770 if (pred == FCmpInst::FCMP_TRUE)
1771 return Constant::getAllOnesValue(ResultTy);
1773 // Handle some degenerate cases first
1774 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1775 // For EQ and NE, we can always pick a value for the undef to make the
1776 // predicate pass or fail, so we can return undef.
1777 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)))
1778 return UndefValue::get(ResultTy);
1779 // Otherwise, pick the same value as the non-undef operand, and fold
1780 // it to true or false.
1781 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1784 // No compile-time operations on this type yet.
1785 if (C1->getType()->isPPC_FP128Ty())
1788 // icmp eq/ne(null,GV) -> false/true
1789 if (C1->isNullValue()) {
1790 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1791 // Don't try to evaluate aliases. External weak GV can be null.
1792 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1793 if (pred == ICmpInst::ICMP_EQ)
1794 return ConstantInt::getFalse(C1->getContext());
1795 else if (pred == ICmpInst::ICMP_NE)
1796 return ConstantInt::getTrue(C1->getContext());
1798 // icmp eq/ne(GV,null) -> false/true
1799 } else if (C2->isNullValue()) {
1800 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1801 // Don't try to evaluate aliases. External weak GV can be null.
1802 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1803 if (pred == ICmpInst::ICMP_EQ)
1804 return ConstantInt::getFalse(C1->getContext());
1805 else if (pred == ICmpInst::ICMP_NE)
1806 return ConstantInt::getTrue(C1->getContext());
1810 // If the comparison is a comparison between two i1's, simplify it.
1811 if (C1->getType()->isIntegerTy(1)) {
1813 case ICmpInst::ICMP_EQ:
1814 if (isa<ConstantInt>(C2))
1815 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1816 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1817 case ICmpInst::ICMP_NE:
1818 return ConstantExpr::getXor(C1, C2);
1824 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1825 APInt V1 = cast<ConstantInt>(C1)->getValue();
1826 APInt V2 = cast<ConstantInt>(C2)->getValue();
1828 default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1829 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1830 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1831 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1832 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1833 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1834 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1835 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1836 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1837 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1838 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1840 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1841 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1842 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1843 APFloat::cmpResult R = C1V.compare(C2V);
1845 default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1846 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1847 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1848 case FCmpInst::FCMP_UNO:
1849 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1850 case FCmpInst::FCMP_ORD:
1851 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1852 case FCmpInst::FCMP_UEQ:
1853 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1854 R==APFloat::cmpEqual);
1855 case FCmpInst::FCMP_OEQ:
1856 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1857 case FCmpInst::FCMP_UNE:
1858 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1859 case FCmpInst::FCMP_ONE:
1860 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1861 R==APFloat::cmpGreaterThan);
1862 case FCmpInst::FCMP_ULT:
1863 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1864 R==APFloat::cmpLessThan);
1865 case FCmpInst::FCMP_OLT:
1866 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1867 case FCmpInst::FCMP_UGT:
1868 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1869 R==APFloat::cmpGreaterThan);
1870 case FCmpInst::FCMP_OGT:
1871 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1872 case FCmpInst::FCMP_ULE:
1873 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1874 case FCmpInst::FCMP_OLE:
1875 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1876 R==APFloat::cmpEqual);
1877 case FCmpInst::FCMP_UGE:
1878 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1879 case FCmpInst::FCMP_OGE:
1880 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1881 R==APFloat::cmpEqual);
1883 } else if (C1->getType()->isVectorTy()) {
1884 SmallVector<Constant*, 16> C1Elts, C2Elts;
1885 C1->getVectorElements(C1Elts);
1886 C2->getVectorElements(C2Elts);
1887 if (C1Elts.empty() || C2Elts.empty())
1890 // If we can constant fold the comparison of each element, constant fold
1891 // the whole vector comparison.
1892 SmallVector<Constant*, 4> ResElts;
1893 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1894 // Compare the elements, producing an i1 result or constant expr.
1895 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1897 return ConstantVector::get(&ResElts[0], ResElts.size());
1900 if (C1->getType()->isFloatingPointTy()) {
1901 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1902 switch (evaluateFCmpRelation(C1, C2)) {
1903 default: llvm_unreachable("Unknown relation!");
1904 case FCmpInst::FCMP_UNO:
1905 case FCmpInst::FCMP_ORD:
1906 case FCmpInst::FCMP_UEQ:
1907 case FCmpInst::FCMP_UNE:
1908 case FCmpInst::FCMP_ULT:
1909 case FCmpInst::FCMP_UGT:
1910 case FCmpInst::FCMP_ULE:
1911 case FCmpInst::FCMP_UGE:
1912 case FCmpInst::FCMP_TRUE:
1913 case FCmpInst::FCMP_FALSE:
1914 case FCmpInst::BAD_FCMP_PREDICATE:
1915 break; // Couldn't determine anything about these constants.
1916 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1917 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1918 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1919 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1921 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1922 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1923 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1924 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1926 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1927 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1928 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1929 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1931 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1932 // We can only partially decide this relation.
1933 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1935 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1938 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1939 // We can only partially decide this relation.
1940 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1942 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1945 case ICmpInst::ICMP_NE: // We know that C1 != C2
1946 // We can only partially decide this relation.
1947 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1949 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1954 // If we evaluated the result, return it now.
1956 return ConstantInt::get(ResultTy, Result);
1959 // Evaluate the relation between the two constants, per the predicate.
1960 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1961 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1962 default: llvm_unreachable("Unknown relational!");
1963 case ICmpInst::BAD_ICMP_PREDICATE:
1964 break; // Couldn't determine anything about these constants.
1965 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1966 // If we know the constants are equal, we can decide the result of this
1967 // computation precisely.
1968 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1970 case ICmpInst::ICMP_ULT:
1972 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1974 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1978 case ICmpInst::ICMP_SLT:
1980 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1982 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1986 case ICmpInst::ICMP_UGT:
1988 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1990 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1994 case ICmpInst::ICMP_SGT:
1996 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1998 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2002 case ICmpInst::ICMP_ULE:
2003 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2004 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2006 case ICmpInst::ICMP_SLE:
2007 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2008 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2010 case ICmpInst::ICMP_UGE:
2011 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2012 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2014 case ICmpInst::ICMP_SGE:
2015 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2016 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2018 case ICmpInst::ICMP_NE:
2019 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2020 if (pred == ICmpInst::ICMP_NE) Result = 1;
2024 // If we evaluated the result, return it now.
2026 return ConstantInt::get(ResultTy, Result);
2028 // If the right hand side is a bitcast, try using its inverse to simplify
2029 // it by moving it to the left hand side. We can't do this if it would turn
2030 // a vector compare into a scalar compare or visa versa.
2031 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2032 Constant *CE2Op0 = CE2->getOperand(0);
2033 if (CE2->getOpcode() == Instruction::BitCast &&
2034 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
2035 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2036 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2040 // If the left hand side is an extension, try eliminating it.
2041 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2042 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
2043 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
2044 Constant *CE1Op0 = CE1->getOperand(0);
2045 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2046 if (CE1Inverse == CE1Op0) {
2047 // Check whether we can safely truncate the right hand side.
2048 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2049 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2050 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2056 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2057 (C1->isNullValue() && !C2->isNullValue())) {
2058 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2059 // other way if possible.
2060 // Also, if C1 is null and C2 isn't, flip them around.
2061 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2062 return ConstantExpr::getICmp(pred, C2, C1);
2068 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2070 template<typename IndexTy>
2071 static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) {
2072 // No indices means nothing that could be out of bounds.
2073 if (NumIdx == 0) return true;
2075 // If the first index is zero, it's in bounds.
2076 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2078 // If the first index is one and all the rest are zero, it's in bounds,
2079 // by the one-past-the-end rule.
2080 if (!cast<ConstantInt>(Idxs[0])->isOne())
2082 for (unsigned i = 1, e = NumIdx; i != e; ++i)
2083 if (!cast<Constant>(Idxs[i])->isNullValue())
2088 template<typename IndexTy>
2089 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
2091 IndexTy const *Idxs,
2093 Constant *Idx0 = cast<Constant>(Idxs[0]);
2095 (NumIdx == 1 && Idx0->isNullValue()))
2098 if (isa<UndefValue>(C)) {
2099 const PointerType *Ptr = cast<PointerType>(C->getType());
2100 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, Idxs+NumIdx);
2101 assert(Ty != 0 && "Invalid indices for GEP!");
2102 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2105 if (C->isNullValue()) {
2107 for (unsigned i = 0, e = NumIdx; i != e; ++i)
2108 if (!cast<Constant>(Idxs[i])->isNullValue()) {
2113 const PointerType *Ptr = cast<PointerType>(C->getType());
2114 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs,
2116 assert(Ty != 0 && "Invalid indices for GEP!");
2117 return ConstantPointerNull::get(
2118 PointerType::get(Ty,Ptr->getAddressSpace()));
2122 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2123 // Combine Indices - If the source pointer to this getelementptr instruction
2124 // is a getelementptr instruction, combine the indices of the two
2125 // getelementptr instructions into a single instruction.
2127 if (CE->getOpcode() == Instruction::GetElementPtr) {
2128 const Type *LastTy = 0;
2129 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2133 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) {
2134 SmallVector<Value*, 16> NewIndices;
2135 NewIndices.reserve(NumIdx + CE->getNumOperands());
2136 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2137 NewIndices.push_back(CE->getOperand(i));
2139 // Add the last index of the source with the first index of the new GEP.
2140 // Make sure to handle the case when they are actually different types.
2141 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2142 // Otherwise it must be an array.
2143 if (!Idx0->isNullValue()) {
2144 const Type *IdxTy = Combined->getType();
2145 if (IdxTy != Idx0->getType()) {
2146 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2147 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2148 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2149 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2152 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2156 NewIndices.push_back(Combined);
2157 NewIndices.append(Idxs+1, Idxs+NumIdx);
2158 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ?
2159 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0),
2161 NewIndices.size()) :
2162 ConstantExpr::getGetElementPtr(CE->getOperand(0),
2168 // Implement folding of:
2169 // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*),
2171 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
2173 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
2174 if (const PointerType *SPT =
2175 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2176 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2177 if (const ArrayType *CAT =
2178 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2179 if (CAT->getElementType() == SAT->getElementType())
2181 ConstantExpr::getInBoundsGetElementPtr(
2182 (Constant*)CE->getOperand(0), Idxs, NumIdx) :
2183 ConstantExpr::getGetElementPtr(
2184 (Constant*)CE->getOperand(0), Idxs, NumIdx);
2188 // Check to see if any array indices are not within the corresponding
2189 // notional array bounds. If so, try to determine if they can be factored
2190 // out into preceding dimensions.
2191 bool Unknown = false;
2192 SmallVector<Constant *, 8> NewIdxs;
2193 const Type *Ty = C->getType();
2194 const Type *Prev = 0;
2195 for (unsigned i = 0; i != NumIdx;
2196 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2197 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2198 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2199 if (ATy->getNumElements() <= INT64_MAX &&
2200 ATy->getNumElements() != 0 &&
2201 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2202 if (isa<SequentialType>(Prev)) {
2203 // It's out of range, but we can factor it into the prior
2205 NewIdxs.resize(NumIdx);
2206 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2207 ATy->getNumElements());
2208 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2210 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2211 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2213 // Before adding, extend both operands to i64 to avoid
2214 // overflow trouble.
2215 if (!PrevIdx->getType()->isIntegerTy(64))
2216 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2217 Type::getInt64Ty(Div->getContext()));
2218 if (!Div->getType()->isIntegerTy(64))
2219 Div = ConstantExpr::getSExt(Div,
2220 Type::getInt64Ty(Div->getContext()));
2222 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2224 // It's out of range, but the prior dimension is a struct
2225 // so we can't do anything about it.
2230 // We don't know if it's in range or not.
2235 // If we did any factoring, start over with the adjusted indices.
2236 if (!NewIdxs.empty()) {
2237 for (unsigned i = 0; i != NumIdx; ++i)
2238 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2240 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(),
2242 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size());
2245 // If all indices are known integers and normalized, we can do a simple
2246 // check for the "inbounds" property.
2247 if (!Unknown && !inBounds &&
2248 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx))
2249 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx);
2254 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2256 Constant* const *Idxs,
2258 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);
2261 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2265 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);