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 // template-based folder for simple primitive constants like ConstantInt, and
16 // the special case hackery that we use to symbolically evaluate expressions
17 // that use ConstantExprs.
19 //===----------------------------------------------------------------------===//
21 #include "ConstantFold.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Function.h"
26 #include "llvm/GlobalAlias.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/ManagedStatic.h"
31 #include "llvm/Support/MathExtras.h"
35 //===----------------------------------------------------------------------===//
36 // ConstantFold*Instruction Implementations
37 //===----------------------------------------------------------------------===//
39 /// BitCastConstantVector - Convert the specified ConstantVector node to the
40 /// specified vector type. At this point, we know that the elements of the
41 /// input vector constant are all simple integer or FP values.
42 static Constant *BitCastConstantVector(ConstantVector *CV,
43 const VectorType *DstTy) {
44 // If this cast changes element count then we can't handle it here:
45 // doing so requires endianness information. This should be handled by
46 // Analysis/ConstantFolding.cpp
47 unsigned NumElts = DstTy->getNumElements();
48 if (NumElts != CV->getNumOperands())
51 // Check to verify that all elements of the input are simple.
52 for (unsigned i = 0; i != NumElts; ++i) {
53 if (!isa<ConstantInt>(CV->getOperand(i)) &&
54 !isa<ConstantFP>(CV->getOperand(i)))
58 // Bitcast each element now.
59 std::vector<Constant*> Result;
60 const Type *DstEltTy = DstTy->getElementType();
61 for (unsigned i = 0; i != NumElts; ++i)
62 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), DstEltTy));
63 return ConstantVector::get(Result);
66 /// This function determines which opcode to use to fold two constant cast
67 /// expressions together. It uses CastInst::isEliminableCastPair to determine
68 /// the opcode. Consequently its just a wrapper around that function.
69 /// @brief Determine if it is valid to fold a cast of a cast
72 unsigned opc, ///< opcode of the second cast constant expression
73 const ConstantExpr*Op, ///< the first cast constant expression
74 const Type *DstTy ///< desintation type of the first cast
76 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
77 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
78 assert(CastInst::isCast(opc) && "Invalid cast opcode");
80 // The the types and opcodes for the two Cast constant expressions
81 const Type *SrcTy = Op->getOperand(0)->getType();
82 const Type *MidTy = Op->getType();
83 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
84 Instruction::CastOps secondOp = Instruction::CastOps(opc);
86 // Let CastInst::isEliminableCastPair do the heavy lifting.
87 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
91 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
92 const Type *SrcTy = V->getType();
94 return V; // no-op cast
96 // Check to see if we are casting a pointer to an aggregate to a pointer to
97 // the first element. If so, return the appropriate GEP instruction.
98 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
99 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
100 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
101 SmallVector<Value*, 8> IdxList;
102 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
103 const Type *ElTy = PTy->getElementType();
104 while (ElTy != DPTy->getElementType()) {
105 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
106 if (STy->getNumElements() == 0) break;
107 ElTy = STy->getElementType(0);
108 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
109 } else if (const SequentialType *STy =
110 dyn_cast<SequentialType>(ElTy)) {
111 if (isa<PointerType>(ElTy)) break; // Can't index into pointers!
112 ElTy = STy->getElementType();
113 IdxList.push_back(IdxList[0]);
119 if (ElTy == DPTy->getElementType())
120 return ConstantExpr::getGetElementPtr(V, &IdxList[0], IdxList.size());
123 // Handle casts from one vector constant to another. We know that the src
124 // and dest type have the same size (otherwise its an illegal cast).
125 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
126 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
127 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
128 "Not cast between same sized vectors!");
129 // First, check for null. Undef is already handled.
130 if (isa<ConstantAggregateZero>(V))
131 return Constant::getNullValue(DestTy);
133 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
134 return BitCastConstantVector(CV, DestPTy);
138 // Finally, implement bitcast folding now. The code below doesn't handle
140 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
141 return ConstantPointerNull::get(cast<PointerType>(DestTy));
143 // Handle integral constant input.
144 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
145 if (DestTy->isInteger())
146 // Integral -> Integral. This is a no-op because the bit widths must
147 // be the same. Consequently, we just fold to V.
150 if (DestTy->isFloatingPoint()) {
151 assert((DestTy == Type::DoubleTy || DestTy == Type::FloatTy) &&
153 return ConstantFP::get(APFloat(CI->getValue()));
155 // Otherwise, can't fold this (vector?)
159 // Handle ConstantFP input.
160 if (const ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
162 if (DestTy == Type::Int32Ty) {
163 return ConstantInt::get(FP->getValueAPF().bitcastToAPInt());
165 assert(DestTy == Type::Int64Ty && "only support f32/f64 for now!");
166 return ConstantInt::get(FP->getValueAPF().bitcastToAPInt());
173 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V,
174 const Type *DestTy) {
175 if (isa<UndefValue>(V)) {
176 // zext(undef) = 0, because the top bits will be zero.
177 // sext(undef) = 0, because the top bits will all be the same.
178 // [us]itofp(undef) = 0, because the result value is bounded.
179 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
180 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
181 return Constant::getNullValue(DestTy);
182 return UndefValue::get(DestTy);
184 // No compile-time operations on this type yet.
185 if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty)
188 // If the cast operand is a constant expression, there's a few things we can
189 // do to try to simplify it.
190 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
192 // Try hard to fold cast of cast because they are often eliminable.
193 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
194 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
195 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
196 // If all of the indexes in the GEP are null values, there is no pointer
197 // adjustment going on. We might as well cast the source pointer.
198 bool isAllNull = true;
199 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
200 if (!CE->getOperand(i)->isNullValue()) {
205 // This is casting one pointer type to another, always BitCast
206 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
210 // We actually have to do a cast now. Perform the cast according to the
213 case Instruction::FPTrunc:
214 case Instruction::FPExt:
215 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
216 APFloat Val = FPC->getValueAPF();
217 Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle :
218 DestTy == Type::DoubleTy ? APFloat::IEEEdouble :
219 DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended :
220 DestTy == Type::FP128Ty ? APFloat::IEEEquad :
222 APFloat::rmNearestTiesToEven);
223 return ConstantFP::get(Val);
225 return 0; // Can't fold.
226 case Instruction::FPToUI:
227 case Instruction::FPToSI:
228 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
229 const APFloat &V = FPC->getValueAPF();
231 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
232 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
233 APFloat::rmTowardZero);
234 APInt Val(DestBitWidth, 2, x);
235 return ConstantInt::get(Val);
237 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
238 std::vector<Constant*> res;
239 const VectorType *DestVecTy = cast<VectorType>(DestTy);
240 const Type *DstEltTy = DestVecTy->getElementType();
241 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
242 res.push_back(ConstantExpr::getCast(opc, CV->getOperand(i), DstEltTy));
243 return ConstantVector::get(DestVecTy, res);
245 return 0; // Can't fold.
246 case Instruction::IntToPtr: //always treated as unsigned
247 if (V->isNullValue()) // Is it an integral null value?
248 return ConstantPointerNull::get(cast<PointerType>(DestTy));
249 return 0; // Other pointer types cannot be casted
250 case Instruction::PtrToInt: // always treated as unsigned
251 if (V->isNullValue()) // is it a null pointer value?
252 return ConstantInt::get(DestTy, 0);
253 return 0; // Other pointer types cannot be casted
254 case Instruction::UIToFP:
255 case Instruction::SIToFP:
256 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
257 APInt api = CI->getValue();
258 const uint64_t zero[] = {0, 0};
259 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(),
261 (void)apf.convertFromAPInt(api,
262 opc==Instruction::SIToFP,
263 APFloat::rmNearestTiesToEven);
264 return ConstantFP::get(apf);
266 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
267 std::vector<Constant*> res;
268 const VectorType *DestVecTy = cast<VectorType>(DestTy);
269 const Type *DstEltTy = DestVecTy->getElementType();
270 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
271 res.push_back(ConstantExpr::getCast(opc, CV->getOperand(i), DstEltTy));
272 return ConstantVector::get(DestVecTy, res);
275 case Instruction::ZExt:
276 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
277 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
278 APInt Result(CI->getValue());
279 Result.zext(BitWidth);
280 return ConstantInt::get(Result);
283 case Instruction::SExt:
284 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
285 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
286 APInt Result(CI->getValue());
287 Result.sext(BitWidth);
288 return ConstantInt::get(Result);
291 case Instruction::Trunc:
292 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
293 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
294 APInt Result(CI->getValue());
295 Result.trunc(BitWidth);
296 return ConstantInt::get(Result);
299 case Instruction::BitCast:
300 return FoldBitCast(const_cast<Constant*>(V), DestTy);
302 assert(!"Invalid CE CastInst opcode");
306 assert(0 && "Failed to cast constant expression");
310 Constant *llvm::ConstantFoldSelectInstruction(const Constant *Cond,
312 const Constant *V2) {
313 if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
314 return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2);
316 if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2);
317 if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1);
318 if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1);
319 if (V1 == V2) return const_cast<Constant*>(V1);
323 Constant *llvm::ConstantFoldExtractElementInstruction(const Constant *Val,
324 const Constant *Idx) {
325 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
326 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
327 if (Val->isNullValue()) // ee(zero, x) -> zero
328 return Constant::getNullValue(
329 cast<VectorType>(Val->getType())->getElementType());
331 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
332 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
333 return CVal->getOperand(CIdx->getZExtValue());
334 } else if (isa<UndefValue>(Idx)) {
335 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
336 return CVal->getOperand(0);
342 Constant *llvm::ConstantFoldInsertElementInstruction(const Constant *Val,
344 const Constant *Idx) {
345 const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
347 APInt idxVal = CIdx->getValue();
348 if (isa<UndefValue>(Val)) {
349 // Insertion of scalar constant into vector undef
350 // Optimize away insertion of undef
351 if (isa<UndefValue>(Elt))
352 return const_cast<Constant*>(Val);
353 // Otherwise break the aggregate undef into multiple undefs and do
356 cast<VectorType>(Val->getType())->getNumElements();
357 std::vector<Constant*> Ops;
359 for (unsigned i = 0; i < numOps; ++i) {
361 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
362 Ops.push_back(const_cast<Constant*>(Op));
364 return ConstantVector::get(Ops);
366 if (isa<ConstantAggregateZero>(Val)) {
367 // Insertion of scalar constant into vector aggregate zero
368 // Optimize away insertion of zero
369 if (Elt->isNullValue())
370 return const_cast<Constant*>(Val);
371 // Otherwise break the aggregate zero into multiple zeros and do
374 cast<VectorType>(Val->getType())->getNumElements();
375 std::vector<Constant*> Ops;
377 for (unsigned i = 0; i < numOps; ++i) {
379 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
380 Ops.push_back(const_cast<Constant*>(Op));
382 return ConstantVector::get(Ops);
384 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
385 // Insertion of scalar constant into vector constant
386 std::vector<Constant*> Ops;
387 Ops.reserve(CVal->getNumOperands());
388 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
390 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
391 Ops.push_back(const_cast<Constant*>(Op));
393 return ConstantVector::get(Ops);
399 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
400 /// return the specified element value. Otherwise return null.
401 static Constant *GetVectorElement(const Constant *C, unsigned EltNo) {
402 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C))
403 return CV->getOperand(EltNo);
405 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
406 if (isa<ConstantAggregateZero>(C))
407 return Constant::getNullValue(EltTy);
408 if (isa<UndefValue>(C))
409 return UndefValue::get(EltTy);
413 Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1,
415 const Constant *Mask) {
416 // Undefined shuffle mask -> undefined value.
417 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
419 unsigned NumElts = cast<VectorType>(V1->getType())->getNumElements();
420 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
422 // Loop over the shuffle mask, evaluating each element.
423 SmallVector<Constant*, 32> Result;
424 for (unsigned i = 0; i != NumElts; ++i) {
425 Constant *InElt = GetVectorElement(Mask, i);
426 if (InElt == 0) return 0;
428 if (isa<UndefValue>(InElt))
429 InElt = UndefValue::get(EltTy);
430 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
431 unsigned Elt = CI->getZExtValue();
432 if (Elt >= NumElts*2)
433 InElt = UndefValue::get(EltTy);
434 else if (Elt >= NumElts)
435 InElt = GetVectorElement(V2, Elt-NumElts);
437 InElt = GetVectorElement(V1, Elt);
438 if (InElt == 0) return 0;
443 Result.push_back(InElt);
446 return ConstantVector::get(&Result[0], Result.size());
449 Constant *llvm::ConstantFoldExtractValueInstruction(const Constant *Agg,
450 const unsigned *Idxs,
452 // Base case: no indices, so return the entire value.
454 return const_cast<Constant *>(Agg);
456 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
457 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
461 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
463 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
467 // Otherwise recurse.
468 return ConstantFoldExtractValueInstruction(Agg->getOperand(*Idxs),
472 Constant *llvm::ConstantFoldInsertValueInstruction(const Constant *Agg,
474 const unsigned *Idxs,
476 // Base case: no indices, so replace the entire value.
478 return const_cast<Constant *>(Val);
480 if (isa<UndefValue>(Agg)) {
481 // Insertion of constant into aggregate undef
482 // Optimize away insertion of undef
483 if (isa<UndefValue>(Val))
484 return const_cast<Constant*>(Agg);
485 // Otherwise break the aggregate undef into multiple undefs and do
487 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
489 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
490 numOps = AR->getNumElements();
492 numOps = cast<StructType>(AggTy)->getNumElements();
493 std::vector<Constant*> Ops(numOps);
494 for (unsigned i = 0; i < numOps; ++i) {
495 const Type *MemberTy = AggTy->getTypeAtIndex(i);
498 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
499 Val, Idxs+1, NumIdx-1) :
500 UndefValue::get(MemberTy);
501 Ops[i] = const_cast<Constant*>(Op);
503 if (isa<StructType>(AggTy))
504 return ConstantStruct::get(Ops);
506 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
508 if (isa<ConstantAggregateZero>(Agg)) {
509 // Insertion of constant into aggregate zero
510 // Optimize away insertion of zero
511 if (Val->isNullValue())
512 return const_cast<Constant*>(Agg);
513 // Otherwise break the aggregate zero into multiple zeros and do
515 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
517 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
518 numOps = AR->getNumElements();
520 numOps = cast<StructType>(AggTy)->getNumElements();
521 std::vector<Constant*> Ops(numOps);
522 for (unsigned i = 0; i < numOps; ++i) {
523 const Type *MemberTy = AggTy->getTypeAtIndex(i);
526 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
527 Val, Idxs+1, NumIdx-1) :
528 Constant::getNullValue(MemberTy);
529 Ops[i] = const_cast<Constant*>(Op);
531 if (isa<StructType>(AggTy))
532 return ConstantStruct::get(Ops);
534 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
536 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
537 // Insertion of constant into aggregate constant
538 std::vector<Constant*> Ops(Agg->getNumOperands());
539 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
542 ConstantFoldInsertValueInstruction(Agg->getOperand(i),
543 Val, Idxs+1, NumIdx-1) :
545 Ops[i] = const_cast<Constant*>(Op);
548 if (isa<StructType>(Agg->getType()))
549 C = ConstantStruct::get(Ops);
551 C = ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
558 /// EvalVectorOp - Given two vector constants and a function pointer, apply the
559 /// function pointer to each element pair, producing a new ConstantVector
560 /// constant. Either or both of V1 and V2 may be NULL, meaning a
561 /// ConstantAggregateZero operand.
562 static Constant *EvalVectorOp(const ConstantVector *V1,
563 const ConstantVector *V2,
564 const VectorType *VTy,
565 Constant *(*FP)(Constant*, Constant*)) {
566 std::vector<Constant*> Res;
567 const Type *EltTy = VTy->getElementType();
568 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
569 const Constant *C1 = V1 ? V1->getOperand(i) : Constant::getNullValue(EltTy);
570 const Constant *C2 = V2 ? V2->getOperand(i) : Constant::getNullValue(EltTy);
571 Res.push_back(FP(const_cast<Constant*>(C1),
572 const_cast<Constant*>(C2)));
574 return ConstantVector::get(Res);
577 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
579 const Constant *C2) {
580 // No compile-time operations on this type yet.
581 if (C1->getType() == Type::PPC_FP128Ty)
584 // Handle UndefValue up front
585 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
587 case Instruction::Xor:
588 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
589 // Handle undef ^ undef -> 0 special case. This is a common
591 return Constant::getNullValue(C1->getType());
593 case Instruction::Add:
594 case Instruction::Sub:
595 return UndefValue::get(C1->getType());
596 case Instruction::Mul:
597 case Instruction::And:
598 return Constant::getNullValue(C1->getType());
599 case Instruction::UDiv:
600 case Instruction::SDiv:
601 case Instruction::FDiv:
602 case Instruction::URem:
603 case Instruction::SRem:
604 case Instruction::FRem:
605 if (!isa<UndefValue>(C2)) // undef / X -> 0
606 return Constant::getNullValue(C1->getType());
607 return const_cast<Constant*>(C2); // X / undef -> undef
608 case Instruction::Or: // X | undef -> -1
609 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
610 return ConstantVector::getAllOnesValue(PTy);
611 return ConstantInt::getAllOnesValue(C1->getType());
612 case Instruction::LShr:
613 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
614 return const_cast<Constant*>(C1); // undef lshr undef -> undef
615 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
617 case Instruction::AShr:
618 if (!isa<UndefValue>(C2))
619 return const_cast<Constant*>(C1); // undef ashr X --> undef
620 else if (isa<UndefValue>(C1))
621 return const_cast<Constant*>(C1); // undef ashr undef -> undef
623 return const_cast<Constant*>(C1); // X ashr undef --> X
624 case Instruction::Shl:
625 // undef << X -> 0 or X << undef -> 0
626 return Constant::getNullValue(C1->getType());
630 // Handle simplifications of the RHS when a constant int.
631 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
633 case Instruction::Add:
634 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X + 0 == X
636 case Instruction::Sub:
637 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X - 0 == X
639 case Instruction::Mul:
640 if (CI2->equalsInt(0)) return const_cast<Constant*>(C2); // X * 0 == 0
641 if (CI2->equalsInt(1))
642 return const_cast<Constant*>(C1); // X * 1 == X
644 case Instruction::UDiv:
645 case Instruction::SDiv:
646 if (CI2->equalsInt(1))
647 return const_cast<Constant*>(C1); // X / 1 == X
649 case Instruction::URem:
650 case Instruction::SRem:
651 if (CI2->equalsInt(1))
652 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
654 case Instruction::And:
655 if (CI2->isZero()) return const_cast<Constant*>(C2); // X & 0 == 0
656 if (CI2->isAllOnesValue())
657 return const_cast<Constant*>(C1); // X & -1 == X
659 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
660 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
661 if (CE1->getOpcode() == Instruction::ZExt) {
662 unsigned DstWidth = CI2->getType()->getBitWidth();
664 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
665 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
666 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
667 return const_cast<Constant*>(C1);
670 // If and'ing the address of a global with a constant, fold it.
671 if (CE1->getOpcode() == Instruction::PtrToInt &&
672 isa<GlobalValue>(CE1->getOperand(0))) {
673 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
675 // Functions are at least 4-byte aligned.
676 unsigned GVAlign = GV->getAlignment();
677 if (isa<Function>(GV))
678 GVAlign = std::max(GVAlign, 4U);
681 unsigned DstWidth = CI2->getType()->getBitWidth();
682 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
683 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
685 // If checking bits we know are clear, return zero.
686 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
687 return Constant::getNullValue(CI2->getType());
692 case Instruction::Or:
693 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X | 0 == X
694 if (CI2->isAllOnesValue())
695 return const_cast<Constant*>(C2); // X | -1 == -1
697 case Instruction::Xor:
698 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X ^ 0 == X
700 case Instruction::AShr:
701 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
702 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
703 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
704 return ConstantExpr::getLShr(const_cast<Constant*>(C1),
705 const_cast<Constant*>(C2));
710 // At this point we know neither constant is an UndefValue.
711 if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
712 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
713 using namespace APIntOps;
714 const APInt &C1V = CI1->getValue();
715 const APInt &C2V = CI2->getValue();
719 case Instruction::Add:
720 return ConstantInt::get(C1V + C2V);
721 case Instruction::Sub:
722 return ConstantInt::get(C1V - C2V);
723 case Instruction::Mul:
724 return ConstantInt::get(C1V * C2V);
725 case Instruction::UDiv:
726 if (CI2->isNullValue())
727 return 0; // X / 0 -> can't fold
728 return ConstantInt::get(C1V.udiv(C2V));
729 case Instruction::SDiv:
730 if (CI2->isNullValue())
731 return 0; // X / 0 -> can't fold
732 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
733 return 0; // MIN_INT / -1 -> overflow
734 return ConstantInt::get(C1V.sdiv(C2V));
735 case Instruction::URem:
736 if (C2->isNullValue())
737 return 0; // X / 0 -> can't fold
738 return ConstantInt::get(C1V.urem(C2V));
739 case Instruction::SRem:
740 if (CI2->isNullValue())
741 return 0; // X % 0 -> can't fold
742 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
743 return 0; // MIN_INT % -1 -> overflow
744 return ConstantInt::get(C1V.srem(C2V));
745 case Instruction::And:
746 return ConstantInt::get(C1V & C2V);
747 case Instruction::Or:
748 return ConstantInt::get(C1V | C2V);
749 case Instruction::Xor:
750 return ConstantInt::get(C1V ^ C2V);
751 case Instruction::Shl: {
752 uint32_t shiftAmt = C2V.getZExtValue();
753 if (shiftAmt < C1V.getBitWidth())
754 return ConstantInt::get(C1V.shl(shiftAmt));
756 return UndefValue::get(C1->getType()); // too big shift is undef
758 case Instruction::LShr: {
759 uint32_t shiftAmt = C2V.getZExtValue();
760 if (shiftAmt < C1V.getBitWidth())
761 return ConstantInt::get(C1V.lshr(shiftAmt));
763 return UndefValue::get(C1->getType()); // too big shift is undef
765 case Instruction::AShr: {
766 uint32_t shiftAmt = C2V.getZExtValue();
767 if (shiftAmt < C1V.getBitWidth())
768 return ConstantInt::get(C1V.ashr(shiftAmt));
770 return UndefValue::get(C1->getType()); // too big shift is undef
774 } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
775 if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
776 APFloat C1V = CFP1->getValueAPF();
777 APFloat C2V = CFP2->getValueAPF();
778 APFloat C3V = C1V; // copy for modification
782 case Instruction::Add:
783 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
784 return ConstantFP::get(C3V);
785 case Instruction::Sub:
786 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
787 return ConstantFP::get(C3V);
788 case Instruction::Mul:
789 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
790 return ConstantFP::get(C3V);
791 case Instruction::FDiv:
792 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
793 return ConstantFP::get(C3V);
794 case Instruction::FRem:
796 // IEEE 754, Section 7.1, #5
797 if (CFP1->getType() == Type::DoubleTy)
798 return ConstantFP::get(APFloat(std::numeric_limits<double>::
800 if (CFP1->getType() == Type::FloatTy)
801 return ConstantFP::get(APFloat(std::numeric_limits<float>::
805 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
806 return ConstantFP::get(C3V);
809 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
810 const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
811 const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
812 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
813 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
817 case Instruction::Add:
818 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAdd);
819 case Instruction::Sub:
820 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSub);
821 case Instruction::Mul:
822 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getMul);
823 case Instruction::UDiv:
824 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getUDiv);
825 case Instruction::SDiv:
826 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSDiv);
827 case Instruction::FDiv:
828 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFDiv);
829 case Instruction::URem:
830 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getURem);
831 case Instruction::SRem:
832 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSRem);
833 case Instruction::FRem:
834 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFRem);
835 case Instruction::And:
836 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAnd);
837 case Instruction::Or:
838 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getOr);
839 case Instruction::Xor:
840 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getXor);
845 if (isa<ConstantExpr>(C1)) {
846 // There are many possible foldings we could do here. We should probably
847 // at least fold add of a pointer with an integer into the appropriate
848 // getelementptr. This will improve alias analysis a bit.
849 } else if (isa<ConstantExpr>(C2)) {
850 // If C2 is a constant expr and C1 isn't, flop them around and fold the
851 // other way if possible.
853 case Instruction::Add:
854 case Instruction::Mul:
855 case Instruction::And:
856 case Instruction::Or:
857 case Instruction::Xor:
858 // No change of opcode required.
859 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
861 case Instruction::Shl:
862 case Instruction::LShr:
863 case Instruction::AShr:
864 case Instruction::Sub:
865 case Instruction::SDiv:
866 case Instruction::UDiv:
867 case Instruction::FDiv:
868 case Instruction::URem:
869 case Instruction::SRem:
870 case Instruction::FRem:
871 default: // These instructions cannot be flopped around.
876 // We don't know how to fold this.
880 /// isZeroSizedType - This type is zero sized if its an array or structure of
881 /// zero sized types. The only leaf zero sized type is an empty structure.
882 static bool isMaybeZeroSizedType(const Type *Ty) {
883 if (isa<OpaqueType>(Ty)) return true; // Can't say.
884 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
886 // If all of elements have zero size, this does too.
887 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
888 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
891 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
892 return isMaybeZeroSizedType(ATy->getElementType());
897 /// IdxCompare - Compare the two constants as though they were getelementptr
898 /// indices. This allows coersion of the types to be the same thing.
900 /// If the two constants are the "same" (after coersion), return 0. If the
901 /// first is less than the second, return -1, if the second is less than the
902 /// first, return 1. If the constants are not integral, return -2.
904 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
905 if (C1 == C2) return 0;
907 // Ok, we found a different index. If they are not ConstantInt, we can't do
908 // anything with them.
909 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
910 return -2; // don't know!
912 // Ok, we have two differing integer indices. Sign extend them to be the same
913 // type. Long is always big enough, so we use it.
914 if (C1->getType() != Type::Int64Ty)
915 C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
917 if (C2->getType() != Type::Int64Ty)
918 C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
920 if (C1 == C2) return 0; // They are equal
922 // If the type being indexed over is really just a zero sized type, there is
923 // no pointer difference being made here.
924 if (isMaybeZeroSizedType(ElTy))
927 // If they are really different, now that they are the same type, then we
928 // found a difference!
929 if (cast<ConstantInt>(C1)->getSExtValue() <
930 cast<ConstantInt>(C2)->getSExtValue())
936 /// evaluateFCmpRelation - This function determines if there is anything we can
937 /// decide about the two constants provided. This doesn't need to handle simple
938 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
939 /// If we can determine that the two constants have a particular relation to
940 /// each other, we should return the corresponding FCmpInst predicate,
941 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
942 /// ConstantFoldCompareInstruction.
944 /// To simplify this code we canonicalize the relation so that the first
945 /// operand is always the most "complex" of the two. We consider ConstantFP
946 /// to be the simplest, and ConstantExprs to be the most complex.
947 static FCmpInst::Predicate evaluateFCmpRelation(const Constant *V1,
948 const Constant *V2) {
949 assert(V1->getType() == V2->getType() &&
950 "Cannot compare values of different types!");
952 // No compile-time operations on this type yet.
953 if (V1->getType() == Type::PPC_FP128Ty)
954 return FCmpInst::BAD_FCMP_PREDICATE;
956 // Handle degenerate case quickly
957 if (V1 == V2) return FCmpInst::FCMP_OEQ;
959 if (!isa<ConstantExpr>(V1)) {
960 if (!isa<ConstantExpr>(V2)) {
961 // We distilled thisUse the standard constant folder for a few cases
963 Constant *C1 = const_cast<Constant*>(V1);
964 Constant *C2 = const_cast<Constant*>(V2);
965 R = dyn_cast<ConstantInt>(
966 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2));
967 if (R && !R->isZero())
968 return FCmpInst::FCMP_OEQ;
969 R = dyn_cast<ConstantInt>(
970 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2));
971 if (R && !R->isZero())
972 return FCmpInst::FCMP_OLT;
973 R = dyn_cast<ConstantInt>(
974 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2));
975 if (R && !R->isZero())
976 return FCmpInst::FCMP_OGT;
978 // Nothing more we can do
979 return FCmpInst::BAD_FCMP_PREDICATE;
982 // If the first operand is simple and second is ConstantExpr, swap operands.
983 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
984 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
985 return FCmpInst::getSwappedPredicate(SwappedRelation);
987 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
988 // constantexpr or a simple constant.
989 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
990 switch (CE1->getOpcode()) {
991 case Instruction::FPTrunc:
992 case Instruction::FPExt:
993 case Instruction::UIToFP:
994 case Instruction::SIToFP:
995 // We might be able to do something with these but we don't right now.
1001 // There are MANY other foldings that we could perform here. They will
1002 // probably be added on demand, as they seem needed.
1003 return FCmpInst::BAD_FCMP_PREDICATE;
1006 /// evaluateICmpRelation - This function determines if there is anything we can
1007 /// decide about the two constants provided. This doesn't need to handle simple
1008 /// things like integer comparisons, but should instead handle ConstantExprs
1009 /// and GlobalValues. If we can determine that the two constants have a
1010 /// particular relation to each other, we should return the corresponding ICmp
1011 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1013 /// To simplify this code we canonicalize the relation so that the first
1014 /// operand is always the most "complex" of the two. We consider simple
1015 /// constants (like ConstantInt) to be the simplest, followed by
1016 /// GlobalValues, followed by ConstantExpr's (the most complex).
1018 static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1,
1021 assert(V1->getType() == V2->getType() &&
1022 "Cannot compare different types of values!");
1023 if (V1 == V2) return ICmpInst::ICMP_EQ;
1025 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) {
1026 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) {
1027 // We distilled this down to a simple case, use the standard constant
1030 Constant *C1 = const_cast<Constant*>(V1);
1031 Constant *C2 = const_cast<Constant*>(V2);
1032 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1033 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1034 if (R && !R->isZero())
1036 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1037 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1038 if (R && !R->isZero())
1040 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1041 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1042 if (R && !R->isZero())
1045 // If we couldn't figure it out, bail.
1046 return ICmpInst::BAD_ICMP_PREDICATE;
1049 // If the first operand is simple, swap operands.
1050 ICmpInst::Predicate SwappedRelation =
1051 evaluateICmpRelation(V2, V1, isSigned);
1052 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1053 return ICmpInst::getSwappedPredicate(SwappedRelation);
1055 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) {
1056 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1057 ICmpInst::Predicate SwappedRelation =
1058 evaluateICmpRelation(V2, V1, isSigned);
1059 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1060 return ICmpInst::getSwappedPredicate(SwappedRelation);
1062 return ICmpInst::BAD_ICMP_PREDICATE;
1065 // Now we know that the RHS is a GlobalValue or simple constant,
1066 // which (since the types must match) means that it's a ConstantPointerNull.
1067 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
1068 // Don't try to decide equality of aliases.
1069 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2))
1070 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage())
1071 return ICmpInst::ICMP_NE;
1073 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1074 // GlobalVals can never be null. Don't try to evaluate aliases.
1075 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1))
1076 return ICmpInst::ICMP_NE;
1079 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1080 // constantexpr, a CPR, or a simple constant.
1081 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1082 const Constant *CE1Op0 = CE1->getOperand(0);
1084 switch (CE1->getOpcode()) {
1085 case Instruction::Trunc:
1086 case Instruction::FPTrunc:
1087 case Instruction::FPExt:
1088 case Instruction::FPToUI:
1089 case Instruction::FPToSI:
1090 break; // We can't evaluate floating point casts or truncations.
1092 case Instruction::UIToFP:
1093 case Instruction::SIToFP:
1094 case Instruction::BitCast:
1095 case Instruction::ZExt:
1096 case Instruction::SExt:
1097 // If the cast is not actually changing bits, and the second operand is a
1098 // null pointer, do the comparison with the pre-casted value.
1099 if (V2->isNullValue() &&
1100 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
1101 bool sgnd = isSigned;
1102 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1103 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1104 return evaluateICmpRelation(CE1Op0,
1105 Constant::getNullValue(CE1Op0->getType()),
1109 // If the dest type is a pointer type, and the RHS is a constantexpr cast
1110 // from the same type as the src of the LHS, evaluate the inputs. This is
1111 // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)",
1112 // which happens a lot in compilers with tagged integers.
1113 if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2))
1114 if (CE2->isCast() && isa<PointerType>(CE1->getType()) &&
1115 CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() &&
1116 CE1->getOperand(0)->getType()->isInteger()) {
1117 bool sgnd = isSigned;
1118 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1119 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1120 return evaluateICmpRelation(CE1->getOperand(0), CE2->getOperand(0),
1125 case Instruction::GetElementPtr:
1126 // Ok, since this is a getelementptr, we know that the constant has a
1127 // pointer type. Check the various cases.
1128 if (isa<ConstantPointerNull>(V2)) {
1129 // If we are comparing a GEP to a null pointer, check to see if the base
1130 // of the GEP equals the null pointer.
1131 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1132 if (GV->hasExternalWeakLinkage())
1133 // Weak linkage GVals could be zero or not. We're comparing that
1134 // to null pointer so its greater-or-equal
1135 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1137 // If its not weak linkage, the GVal must have a non-zero address
1138 // so the result is greater-than
1139 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1140 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1141 // If we are indexing from a null pointer, check to see if we have any
1142 // non-zero indices.
1143 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1144 if (!CE1->getOperand(i)->isNullValue())
1145 // Offsetting from null, must not be equal.
1146 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1147 // Only zero indexes from null, must still be zero.
1148 return ICmpInst::ICMP_EQ;
1150 // Otherwise, we can't really say if the first operand is null or not.
1151 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
1152 if (isa<ConstantPointerNull>(CE1Op0)) {
1153 if (CPR2->hasExternalWeakLinkage())
1154 // Weak linkage GVals could be zero or not. We're comparing it to
1155 // a null pointer, so its less-or-equal
1156 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1158 // If its not weak linkage, the GVal must have a non-zero address
1159 // so the result is less-than
1160 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1161 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) {
1163 // If this is a getelementptr of the same global, then it must be
1164 // different. Because the types must match, the getelementptr could
1165 // only have at most one index, and because we fold getelementptr's
1166 // with a single zero index, it must be nonzero.
1167 assert(CE1->getNumOperands() == 2 &&
1168 !CE1->getOperand(1)->isNullValue() &&
1169 "Suprising getelementptr!");
1170 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1172 // If they are different globals, we don't know what the value is,
1173 // but they can't be equal.
1174 return ICmpInst::ICMP_NE;
1178 const ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1179 const Constant *CE2Op0 = CE2->getOperand(0);
1181 // There are MANY other foldings that we could perform here. They will
1182 // probably be added on demand, as they seem needed.
1183 switch (CE2->getOpcode()) {
1185 case Instruction::GetElementPtr:
1186 // By far the most common case to handle is when the base pointers are
1187 // obviously to the same or different globals.
1188 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1189 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1190 return ICmpInst::ICMP_NE;
1191 // Ok, we know that both getelementptr instructions are based on the
1192 // same global. From this, we can precisely determine the relative
1193 // ordering of the resultant pointers.
1196 // Compare all of the operands the GEP's have in common.
1197 gep_type_iterator GTI = gep_type_begin(CE1);
1198 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1200 switch (IdxCompare(CE1->getOperand(i), CE2->getOperand(i),
1201 GTI.getIndexedType())) {
1202 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1203 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1204 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1207 // Ok, we ran out of things they have in common. If any leftovers
1208 // are non-zero then we have a difference, otherwise we are equal.
1209 for (; i < CE1->getNumOperands(); ++i)
1210 if (!CE1->getOperand(i)->isNullValue()) {
1211 if (isa<ConstantInt>(CE1->getOperand(i)))
1212 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1214 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1217 for (; i < CE2->getNumOperands(); ++i)
1218 if (!CE2->getOperand(i)->isNullValue()) {
1219 if (isa<ConstantInt>(CE2->getOperand(i)))
1220 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1222 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1224 return ICmpInst::ICMP_EQ;
1233 return ICmpInst::BAD_ICMP_PREDICATE;
1236 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1238 const Constant *C2) {
1239 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1240 if (pred == FCmpInst::FCMP_FALSE) {
1241 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1242 return Constant::getNullValue(VectorType::getInteger(VT));
1244 return ConstantInt::getFalse();
1247 if (pred == FCmpInst::FCMP_TRUE) {
1248 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1249 return Constant::getAllOnesValue(VectorType::getInteger(VT));
1251 return ConstantInt::getTrue();
1254 // Handle some degenerate cases first
1255 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1256 // vicmp/vfcmp -> [vector] undef
1257 if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType()))
1258 return UndefValue::get(VectorType::getInteger(VTy));
1260 // icmp/fcmp -> i1 undef
1261 return UndefValue::get(Type::Int1Ty);
1264 // No compile-time operations on this type yet.
1265 if (C1->getType() == Type::PPC_FP128Ty)
1268 // icmp eq/ne(null,GV) -> false/true
1269 if (C1->isNullValue()) {
1270 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1271 // Don't try to evaluate aliases. External weak GV can be null.
1272 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1273 if (pred == ICmpInst::ICMP_EQ)
1274 return ConstantInt::getFalse();
1275 else if (pred == ICmpInst::ICMP_NE)
1276 return ConstantInt::getTrue();
1278 // icmp eq/ne(GV,null) -> false/true
1279 } else if (C2->isNullValue()) {
1280 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1281 // Don't try to evaluate aliases. External weak GV can be null.
1282 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1283 if (pred == ICmpInst::ICMP_EQ)
1284 return ConstantInt::getFalse();
1285 else if (pred == ICmpInst::ICMP_NE)
1286 return ConstantInt::getTrue();
1290 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1291 APInt V1 = cast<ConstantInt>(C1)->getValue();
1292 APInt V2 = cast<ConstantInt>(C2)->getValue();
1294 default: assert(0 && "Invalid ICmp Predicate"); return 0;
1295 case ICmpInst::ICMP_EQ: return ConstantInt::get(Type::Int1Ty, V1 == V2);
1296 case ICmpInst::ICMP_NE: return ConstantInt::get(Type::Int1Ty, V1 != V2);
1297 case ICmpInst::ICMP_SLT:return ConstantInt::get(Type::Int1Ty, V1.slt(V2));
1298 case ICmpInst::ICMP_SGT:return ConstantInt::get(Type::Int1Ty, V1.sgt(V2));
1299 case ICmpInst::ICMP_SLE:return ConstantInt::get(Type::Int1Ty, V1.sle(V2));
1300 case ICmpInst::ICMP_SGE:return ConstantInt::get(Type::Int1Ty, V1.sge(V2));
1301 case ICmpInst::ICMP_ULT:return ConstantInt::get(Type::Int1Ty, V1.ult(V2));
1302 case ICmpInst::ICMP_UGT:return ConstantInt::get(Type::Int1Ty, V1.ugt(V2));
1303 case ICmpInst::ICMP_ULE:return ConstantInt::get(Type::Int1Ty, V1.ule(V2));
1304 case ICmpInst::ICMP_UGE:return ConstantInt::get(Type::Int1Ty, V1.uge(V2));
1306 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1307 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1308 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1309 APFloat::cmpResult R = C1V.compare(C2V);
1311 default: assert(0 && "Invalid FCmp Predicate"); return 0;
1312 case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse();
1313 case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue();
1314 case FCmpInst::FCMP_UNO:
1315 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered);
1316 case FCmpInst::FCMP_ORD:
1317 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpUnordered);
1318 case FCmpInst::FCMP_UEQ:
1319 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1320 R==APFloat::cmpEqual);
1321 case FCmpInst::FCMP_OEQ:
1322 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpEqual);
1323 case FCmpInst::FCMP_UNE:
1324 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpEqual);
1325 case FCmpInst::FCMP_ONE:
1326 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1327 R==APFloat::cmpGreaterThan);
1328 case FCmpInst::FCMP_ULT:
1329 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1330 R==APFloat::cmpLessThan);
1331 case FCmpInst::FCMP_OLT:
1332 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan);
1333 case FCmpInst::FCMP_UGT:
1334 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1335 R==APFloat::cmpGreaterThan);
1336 case FCmpInst::FCMP_OGT:
1337 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan);
1338 case FCmpInst::FCMP_ULE:
1339 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpGreaterThan);
1340 case FCmpInst::FCMP_OLE:
1341 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1342 R==APFloat::cmpEqual);
1343 case FCmpInst::FCMP_UGE:
1344 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpLessThan);
1345 case FCmpInst::FCMP_OGE:
1346 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan ||
1347 R==APFloat::cmpEqual);
1349 } else if (isa<VectorType>(C1->getType())) {
1350 SmallVector<Constant*, 16> C1Elts, C2Elts;
1351 C1->getVectorElements(C1Elts);
1352 C2->getVectorElements(C2Elts);
1354 // If we can constant fold the comparison of each element, constant fold
1355 // the whole vector comparison.
1356 SmallVector<Constant*, 4> ResElts;
1357 const Type *InEltTy = C1Elts[0]->getType();
1358 bool isFP = InEltTy->isFloatingPoint();
1359 const Type *ResEltTy = InEltTy;
1361 ResEltTy = IntegerType::get(InEltTy->getPrimitiveSizeInBits());
1363 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1364 // Compare the elements, producing an i1 result or constant expr.
1367 C = ConstantExpr::getFCmp(pred, C1Elts[i], C2Elts[i]);
1369 C = ConstantExpr::getICmp(pred, C1Elts[i], C2Elts[i]);
1371 // If it is a bool or undef result, convert to the dest type.
1372 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
1374 ResElts.push_back(Constant::getNullValue(ResEltTy));
1376 ResElts.push_back(Constant::getAllOnesValue(ResEltTy));
1377 } else if (isa<UndefValue>(C)) {
1378 ResElts.push_back(UndefValue::get(ResEltTy));
1384 if (ResElts.size() == C1Elts.size())
1385 return ConstantVector::get(&ResElts[0], ResElts.size());
1388 if (C1->getType()->isFloatingPoint()) {
1389 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1390 switch (evaluateFCmpRelation(C1, C2)) {
1391 default: assert(0 && "Unknown relation!");
1392 case FCmpInst::FCMP_UNO:
1393 case FCmpInst::FCMP_ORD:
1394 case FCmpInst::FCMP_UEQ:
1395 case FCmpInst::FCMP_UNE:
1396 case FCmpInst::FCMP_ULT:
1397 case FCmpInst::FCMP_UGT:
1398 case FCmpInst::FCMP_ULE:
1399 case FCmpInst::FCMP_UGE:
1400 case FCmpInst::FCMP_TRUE:
1401 case FCmpInst::FCMP_FALSE:
1402 case FCmpInst::BAD_FCMP_PREDICATE:
1403 break; // Couldn't determine anything about these constants.
1404 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1405 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1406 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1407 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1409 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1410 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1411 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1412 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1414 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1415 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1416 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1417 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1419 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1420 // We can only partially decide this relation.
1421 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1423 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1426 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1427 // We can only partially decide this relation.
1428 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1430 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1433 case ICmpInst::ICMP_NE: // We know that C1 != C2
1434 // We can only partially decide this relation.
1435 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1437 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1442 // If we evaluated the result, return it now.
1444 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) {
1446 return Constant::getNullValue(VectorType::getInteger(VT));
1448 return Constant::getAllOnesValue(VectorType::getInteger(VT));
1450 return ConstantInt::get(Type::Int1Ty, Result);
1454 // Evaluate the relation between the two constants, per the predicate.
1455 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1456 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1457 default: assert(0 && "Unknown relational!");
1458 case ICmpInst::BAD_ICMP_PREDICATE:
1459 break; // Couldn't determine anything about these constants.
1460 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1461 // If we know the constants are equal, we can decide the result of this
1462 // computation precisely.
1463 Result = (pred == ICmpInst::ICMP_EQ ||
1464 pred == ICmpInst::ICMP_ULE ||
1465 pred == ICmpInst::ICMP_SLE ||
1466 pred == ICmpInst::ICMP_UGE ||
1467 pred == ICmpInst::ICMP_SGE);
1469 case ICmpInst::ICMP_ULT:
1470 // If we know that C1 < C2, we can decide the result of this computation
1472 Result = (pred == ICmpInst::ICMP_ULT ||
1473 pred == ICmpInst::ICMP_NE ||
1474 pred == ICmpInst::ICMP_ULE);
1476 case ICmpInst::ICMP_SLT:
1477 // If we know that C1 < C2, we can decide the result of this computation
1479 Result = (pred == ICmpInst::ICMP_SLT ||
1480 pred == ICmpInst::ICMP_NE ||
1481 pred == ICmpInst::ICMP_SLE);
1483 case ICmpInst::ICMP_UGT:
1484 // If we know that C1 > C2, we can decide the result of this computation
1486 Result = (pred == ICmpInst::ICMP_UGT ||
1487 pred == ICmpInst::ICMP_NE ||
1488 pred == ICmpInst::ICMP_UGE);
1490 case ICmpInst::ICMP_SGT:
1491 // If we know that C1 > C2, we can decide the result of this computation
1493 Result = (pred == ICmpInst::ICMP_SGT ||
1494 pred == ICmpInst::ICMP_NE ||
1495 pred == ICmpInst::ICMP_SGE);
1497 case ICmpInst::ICMP_ULE:
1498 // If we know that C1 <= C2, we can only partially decide this relation.
1499 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1500 if (pred == ICmpInst::ICMP_ULT) Result = 1;
1502 case ICmpInst::ICMP_SLE:
1503 // If we know that C1 <= C2, we can only partially decide this relation.
1504 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1505 if (pred == ICmpInst::ICMP_SLT) Result = 1;
1508 case ICmpInst::ICMP_UGE:
1509 // If we know that C1 >= C2, we can only partially decide this relation.
1510 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1511 if (pred == ICmpInst::ICMP_UGT) Result = 1;
1513 case ICmpInst::ICMP_SGE:
1514 // If we know that C1 >= C2, we can only partially decide this relation.
1515 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1516 if (pred == ICmpInst::ICMP_SGT) Result = 1;
1519 case ICmpInst::ICMP_NE:
1520 // If we know that C1 != C2, we can only partially decide this relation.
1521 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1522 if (pred == ICmpInst::ICMP_NE) Result = 1;
1526 // If we evaluated the result, return it now.
1528 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) {
1530 return Constant::getNullValue(VT);
1532 return Constant::getAllOnesValue(VT);
1534 return ConstantInt::get(Type::Int1Ty, Result);
1537 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) {
1538 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1539 // other way if possible.
1541 case ICmpInst::ICMP_EQ:
1542 case ICmpInst::ICMP_NE:
1543 // No change of predicate required.
1544 return ConstantFoldCompareInstruction(pred, C2, C1);
1546 case ICmpInst::ICMP_ULT:
1547 case ICmpInst::ICMP_SLT:
1548 case ICmpInst::ICMP_UGT:
1549 case ICmpInst::ICMP_SGT:
1550 case ICmpInst::ICMP_ULE:
1551 case ICmpInst::ICMP_SLE:
1552 case ICmpInst::ICMP_UGE:
1553 case ICmpInst::ICMP_SGE:
1554 // Change the predicate as necessary to swap the operands.
1555 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1556 return ConstantFoldCompareInstruction(pred, C2, C1);
1558 default: // These predicates cannot be flopped around.
1566 Constant *llvm::ConstantFoldGetElementPtr(const Constant *C,
1567 Constant* const *Idxs,
1570 (NumIdx == 1 && Idxs[0]->isNullValue()))
1571 return const_cast<Constant*>(C);
1573 if (isa<UndefValue>(C)) {
1574 const PointerType *Ptr = cast<PointerType>(C->getType());
1575 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
1577 (Value **)Idxs+NumIdx);
1578 assert(Ty != 0 && "Invalid indices for GEP!");
1579 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1582 Constant *Idx0 = Idxs[0];
1583 if (C->isNullValue()) {
1585 for (unsigned i = 0, e = NumIdx; i != e; ++i)
1586 if (!Idxs[i]->isNullValue()) {
1591 const PointerType *Ptr = cast<PointerType>(C->getType());
1592 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
1594 (Value**)Idxs+NumIdx);
1595 assert(Ty != 0 && "Invalid indices for GEP!");
1597 ConstantPointerNull::get(PointerType::get(Ty,Ptr->getAddressSpace()));
1601 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) {
1602 // Combine Indices - If the source pointer to this getelementptr instruction
1603 // is a getelementptr instruction, combine the indices of the two
1604 // getelementptr instructions into a single instruction.
1606 if (CE->getOpcode() == Instruction::GetElementPtr) {
1607 const Type *LastTy = 0;
1608 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1612 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
1613 SmallVector<Value*, 16> NewIndices;
1614 NewIndices.reserve(NumIdx + CE->getNumOperands());
1615 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1616 NewIndices.push_back(CE->getOperand(i));
1618 // Add the last index of the source with the first index of the new GEP.
1619 // Make sure to handle the case when they are actually different types.
1620 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1621 // Otherwise it must be an array.
1622 if (!Idx0->isNullValue()) {
1623 const Type *IdxTy = Combined->getType();
1624 if (IdxTy != Idx0->getType()) {
1625 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty);
1626 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined,
1628 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1631 ConstantExpr::get(Instruction::Add, Idx0, Combined);
1635 NewIndices.push_back(Combined);
1636 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
1637 return ConstantExpr::getGetElementPtr(CE->getOperand(0), &NewIndices[0],
1642 // Implement folding of:
1643 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*),
1645 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
1647 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
1648 if (const PointerType *SPT =
1649 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
1650 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
1651 if (const ArrayType *CAT =
1652 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
1653 if (CAT->getElementType() == SAT->getElementType())
1654 return ConstantExpr::getGetElementPtr(
1655 (Constant*)CE->getOperand(0), Idxs, NumIdx);
1658 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1)
1659 // Into: inttoptr (i64 0 to i8*)
1660 // This happens with pointers to member functions in C++.
1661 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 &&
1662 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) &&
1663 cast<PointerType>(CE->getType())->getElementType() == Type::Int8Ty) {
1664 Constant *Base = CE->getOperand(0);
1665 Constant *Offset = Idxs[0];
1667 // Convert the smaller integer to the larger type.
1668 if (Offset->getType()->getPrimitiveSizeInBits() <
1669 Base->getType()->getPrimitiveSizeInBits())
1670 Offset = ConstantExpr::getSExt(Offset, Base->getType());
1671 else if (Base->getType()->getPrimitiveSizeInBits() <
1672 Offset->getType()->getPrimitiveSizeInBits())
1673 Base = ConstantExpr::getZExt(Base, Base->getType());
1675 Base = ConstantExpr::getAdd(Base, Offset);
1676 return ConstantExpr::getIntToPtr(Base, CE->getType());