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/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), DstEltTy));
64 return ConstantVector::get(Result);
67 /// This function determines which opcode to use to fold two constant cast
68 /// expressions together. It uses CastInst::isEliminableCastPair to determine
69 /// the opcode. Consequently its just a wrapper around that function.
70 /// @brief Determine if it is valid to fold a cast of a cast
73 unsigned opc, ///< opcode of the second cast constant expression
74 const ConstantExpr*Op, ///< the first cast constant expression
75 const Type *DstTy ///< desintation type of the first cast
77 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
78 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
79 assert(CastInst::isCast(opc) && "Invalid cast opcode");
81 // The the types and opcodes for the two Cast constant expressions
82 const Type *SrcTy = Op->getOperand(0)->getType();
83 const Type *MidTy = Op->getType();
84 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
85 Instruction::CastOps secondOp = Instruction::CastOps(opc);
87 // Let CastInst::isEliminableCastPair do the heavy lifting.
88 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
92 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
93 const Type *SrcTy = V->getType();
95 return V; // no-op cast
97 // Check to see if we are casting a pointer to an aggregate to a pointer to
98 // the first element. If so, return the appropriate GEP instruction.
99 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
100 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
101 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
102 SmallVector<Value*, 8> IdxList;
103 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
104 const Type *ElTy = PTy->getElementType();
105 while (ElTy != DPTy->getElementType()) {
106 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
107 if (STy->getNumElements() == 0) break;
108 ElTy = STy->getElementType(0);
109 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
110 } else if (const SequentialType *STy =
111 dyn_cast<SequentialType>(ElTy)) {
112 if (isa<PointerType>(ElTy)) break; // Can't index into pointers!
113 ElTy = STy->getElementType();
114 IdxList.push_back(IdxList[0]);
120 if (ElTy == DPTy->getElementType())
121 return ConstantExpr::getGetElementPtr(V, &IdxList[0], IdxList.size());
124 // Handle casts from one vector constant to another. We know that the src
125 // and dest type have the same size (otherwise its an illegal cast).
126 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
127 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
128 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
129 "Not cast between same sized vectors!");
131 // First, check for null. Undef is already handled.
132 if (isa<ConstantAggregateZero>(V))
133 return Constant::getNullValue(DestTy);
135 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
136 return BitCastConstantVector(CV, DestPTy);
139 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
140 // This allows for other simplifications (although some of them
141 // can only be handled by Analysis/ConstantFolding.cpp).
142 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
143 return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy);
146 // Finally, implement bitcast folding now. The code below doesn't handle
148 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
149 return ConstantPointerNull::get(cast<PointerType>(DestTy));
151 // Handle integral constant input.
152 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
153 if (DestTy->isInteger())
154 // Integral -> Integral. This is a no-op because the bit widths must
155 // be the same. Consequently, we just fold to V.
158 if (DestTy->isFloatingPoint())
159 return ConstantFP::get(APFloat(CI->getValue(),
160 DestTy != Type::PPC_FP128Ty));
162 // Otherwise, can't fold this (vector?)
166 // Handle ConstantFP input.
167 if (const ConstantFP *FP = dyn_cast<ConstantFP>(V))
169 return ConstantInt::get(FP->getValueAPF().bitcastToAPInt());
175 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V,
176 const Type *DestTy) {
177 if (isa<UndefValue>(V)) {
178 // zext(undef) = 0, because the top bits will be zero.
179 // sext(undef) = 0, because the top bits will all be the same.
180 // [us]itofp(undef) = 0, because the result value is bounded.
181 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
182 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
183 return Constant::getNullValue(DestTy);
184 return UndefValue::get(DestTy);
186 // No compile-time operations on this type yet.
187 if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty)
190 // If the cast operand is a constant expression, there's a few things we can
191 // do to try to simplify it.
192 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
194 // Try hard to fold cast of cast because they are often eliminable.
195 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
196 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
197 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
198 // If all of the indexes in the GEP are null values, there is no pointer
199 // adjustment going on. We might as well cast the source pointer.
200 bool isAllNull = true;
201 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
202 if (!CE->getOperand(i)->isNullValue()) {
207 // This is casting one pointer type to another, always BitCast
208 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
212 // If the cast operand is a constant vector, perform the cast by
213 // operating on each element. In the cast of bitcasts, the element
214 // count may be mismatched; don't attempt to handle that here.
215 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V))
216 if (isa<VectorType>(DestTy) &&
217 cast<VectorType>(DestTy)->getNumElements() ==
218 CV->getType()->getNumElements()) {
219 std::vector<Constant*> res;
220 const VectorType *DestVecTy = cast<VectorType>(DestTy);
221 const Type *DstEltTy = DestVecTy->getElementType();
222 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
223 res.push_back(ConstantExpr::getCast(opc,
224 CV->getOperand(i), DstEltTy));
225 return ConstantVector::get(DestVecTy, res);
228 // We actually have to do a cast now. Perform the cast according to the
231 case Instruction::FPTrunc:
232 case Instruction::FPExt:
233 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
235 APFloat Val = FPC->getValueAPF();
236 Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle :
237 DestTy == Type::DoubleTy ? APFloat::IEEEdouble :
238 DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended :
239 DestTy == Type::FP128Ty ? APFloat::IEEEquad :
241 APFloat::rmNearestTiesToEven, &ignored);
242 return ConstantFP::get(Val);
244 return 0; // Can't fold.
245 case Instruction::FPToUI:
246 case Instruction::FPToSI:
247 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
248 const APFloat &V = FPC->getValueAPF();
251 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
252 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
253 APFloat::rmTowardZero, &ignored);
254 APInt Val(DestBitWidth, 2, x);
255 return ConstantInt::get(Val);
257 return 0; // Can't fold.
258 case Instruction::IntToPtr: //always treated as unsigned
259 if (V->isNullValue()) // Is it an integral null value?
260 return ConstantPointerNull::get(cast<PointerType>(DestTy));
261 return 0; // Other pointer types cannot be casted
262 case Instruction::PtrToInt: // always treated as unsigned
263 if (V->isNullValue()) // is it a null pointer value?
264 return ConstantInt::get(DestTy, 0);
265 return 0; // Other pointer types cannot be casted
266 case Instruction::UIToFP:
267 case Instruction::SIToFP:
268 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
269 APInt api = CI->getValue();
270 const uint64_t zero[] = {0, 0};
271 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(),
273 (void)apf.convertFromAPInt(api,
274 opc==Instruction::SIToFP,
275 APFloat::rmNearestTiesToEven);
276 return ConstantFP::get(apf);
279 case Instruction::ZExt:
280 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
281 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
282 APInt Result(CI->getValue());
283 Result.zext(BitWidth);
284 return ConstantInt::get(Result);
287 case Instruction::SExt:
288 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
289 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
290 APInt Result(CI->getValue());
291 Result.sext(BitWidth);
292 return ConstantInt::get(Result);
295 case Instruction::Trunc:
296 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
297 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
298 APInt Result(CI->getValue());
299 Result.trunc(BitWidth);
300 return ConstantInt::get(Result);
303 case Instruction::BitCast:
304 return FoldBitCast(const_cast<Constant*>(V), DestTy);
306 assert(!"Invalid CE CastInst opcode");
310 LLVM_UNREACHABLE("Failed to cast constant expression");
314 Constant *llvm::ConstantFoldSelectInstruction(const Constant *Cond,
316 const Constant *V2) {
317 if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
318 return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2);
320 if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2);
321 if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1);
322 if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1);
323 if (V1 == V2) return const_cast<Constant*>(V1);
327 Constant *llvm::ConstantFoldExtractElementInstruction(const Constant *Val,
328 const Constant *Idx) {
329 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
330 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
331 if (Val->isNullValue()) // ee(zero, x) -> zero
332 return Constant::getNullValue(
333 cast<VectorType>(Val->getType())->getElementType());
335 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
336 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
337 return CVal->getOperand(CIdx->getZExtValue());
338 } else if (isa<UndefValue>(Idx)) {
339 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
340 return CVal->getOperand(0);
346 Constant *llvm::ConstantFoldInsertElementInstruction(const Constant *Val,
348 const Constant *Idx) {
349 const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
351 APInt idxVal = CIdx->getValue();
352 if (isa<UndefValue>(Val)) {
353 // Insertion of scalar constant into vector undef
354 // Optimize away insertion of undef
355 if (isa<UndefValue>(Elt))
356 return const_cast<Constant*>(Val);
357 // Otherwise break the aggregate undef into multiple undefs and do
360 cast<VectorType>(Val->getType())->getNumElements();
361 std::vector<Constant*> Ops;
363 for (unsigned i = 0; i < numOps; ++i) {
365 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
366 Ops.push_back(const_cast<Constant*>(Op));
368 return ConstantVector::get(Ops);
370 if (isa<ConstantAggregateZero>(Val)) {
371 // Insertion of scalar constant into vector aggregate zero
372 // Optimize away insertion of zero
373 if (Elt->isNullValue())
374 return const_cast<Constant*>(Val);
375 // Otherwise break the aggregate zero into multiple zeros and do
378 cast<VectorType>(Val->getType())->getNumElements();
379 std::vector<Constant*> Ops;
381 for (unsigned i = 0; i < numOps; ++i) {
383 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
384 Ops.push_back(const_cast<Constant*>(Op));
386 return ConstantVector::get(Ops);
388 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
389 // Insertion of scalar constant into vector constant
390 std::vector<Constant*> Ops;
391 Ops.reserve(CVal->getNumOperands());
392 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
394 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
395 Ops.push_back(const_cast<Constant*>(Op));
397 return ConstantVector::get(Ops);
403 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
404 /// return the specified element value. Otherwise return null.
405 static Constant *GetVectorElement(const Constant *C, unsigned EltNo) {
406 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C))
407 return CV->getOperand(EltNo);
409 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
410 if (isa<ConstantAggregateZero>(C))
411 return Constant::getNullValue(EltTy);
412 if (isa<UndefValue>(C))
413 return UndefValue::get(EltTy);
417 Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1,
419 const Constant *Mask) {
420 // Undefined shuffle mask -> undefined value.
421 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
423 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
424 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
425 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
427 // Loop over the shuffle mask, evaluating each element.
428 SmallVector<Constant*, 32> Result;
429 for (unsigned i = 0; i != MaskNumElts; ++i) {
430 Constant *InElt = GetVectorElement(Mask, i);
431 if (InElt == 0) return 0;
433 if (isa<UndefValue>(InElt))
434 InElt = UndefValue::get(EltTy);
435 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
436 unsigned Elt = CI->getZExtValue();
437 if (Elt >= SrcNumElts*2)
438 InElt = UndefValue::get(EltTy);
439 else if (Elt >= SrcNumElts)
440 InElt = GetVectorElement(V2, Elt - SrcNumElts);
442 InElt = GetVectorElement(V1, Elt);
443 if (InElt == 0) return 0;
448 Result.push_back(InElt);
451 return ConstantVector::get(&Result[0], Result.size());
454 Constant *llvm::ConstantFoldExtractValueInstruction(const Constant *Agg,
455 const unsigned *Idxs,
457 // Base case: no indices, so return the entire value.
459 return const_cast<Constant *>(Agg);
461 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
462 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
466 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
468 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
472 // Otherwise recurse.
473 return ConstantFoldExtractValueInstruction(Agg->getOperand(*Idxs),
477 Constant *llvm::ConstantFoldInsertValueInstruction(const Constant *Agg,
479 const unsigned *Idxs,
481 // Base case: no indices, so replace the entire value.
483 return const_cast<Constant *>(Val);
485 if (isa<UndefValue>(Agg)) {
486 // Insertion of constant into aggregate undef
487 // Optimize away insertion of undef
488 if (isa<UndefValue>(Val))
489 return const_cast<Constant*>(Agg);
490 // Otherwise break the aggregate undef into multiple undefs and do
492 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
494 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
495 numOps = AR->getNumElements();
497 numOps = cast<StructType>(AggTy)->getNumElements();
498 std::vector<Constant*> Ops(numOps);
499 for (unsigned i = 0; i < numOps; ++i) {
500 const Type *MemberTy = AggTy->getTypeAtIndex(i);
503 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
504 Val, Idxs+1, NumIdx-1) :
505 UndefValue::get(MemberTy);
506 Ops[i] = const_cast<Constant*>(Op);
508 if (isa<StructType>(AggTy))
509 return ConstantStruct::get(Ops);
511 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
513 if (isa<ConstantAggregateZero>(Agg)) {
514 // Insertion of constant into aggregate zero
515 // Optimize away insertion of zero
516 if (Val->isNullValue())
517 return const_cast<Constant*>(Agg);
518 // Otherwise break the aggregate zero into multiple zeros and do
520 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
522 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
523 numOps = AR->getNumElements();
525 numOps = cast<StructType>(AggTy)->getNumElements();
526 std::vector<Constant*> Ops(numOps);
527 for (unsigned i = 0; i < numOps; ++i) {
528 const Type *MemberTy = AggTy->getTypeAtIndex(i);
531 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
532 Val, Idxs+1, NumIdx-1) :
533 Constant::getNullValue(MemberTy);
534 Ops[i] = const_cast<Constant*>(Op);
536 if (isa<StructType>(AggTy))
537 return ConstantStruct::get(Ops);
539 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
541 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
542 // Insertion of constant into aggregate constant
543 std::vector<Constant*> Ops(Agg->getNumOperands());
544 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
547 ConstantFoldInsertValueInstruction(Agg->getOperand(i),
548 Val, Idxs+1, NumIdx-1) :
550 Ops[i] = const_cast<Constant*>(Op);
553 if (isa<StructType>(Agg->getType()))
554 C = ConstantStruct::get(Ops);
556 C = ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
563 /// EvalVectorOp - Given two vector constants and a function pointer, apply the
564 /// function pointer to each element pair, producing a new ConstantVector
565 /// constant. Either or both of V1 and V2 may be NULL, meaning a
566 /// ConstantAggregateZero operand.
567 static Constant *EvalVectorOp(const ConstantVector *V1,
568 const ConstantVector *V2,
569 const VectorType *VTy,
570 Constant *(*FP)(Constant*, Constant*)) {
571 std::vector<Constant*> Res;
572 const Type *EltTy = VTy->getElementType();
573 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
574 const Constant *C1 = V1 ? V1->getOperand(i) : Constant::getNullValue(EltTy);
575 const Constant *C2 = V2 ? V2->getOperand(i) : Constant::getNullValue(EltTy);
576 Res.push_back(FP(const_cast<Constant*>(C1),
577 const_cast<Constant*>(C2)));
579 return ConstantVector::get(Res);
582 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
584 const Constant *C2) {
585 // No compile-time operations on this type yet.
586 if (C1->getType() == Type::PPC_FP128Ty)
589 // Handle UndefValue up front
590 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
592 case Instruction::Xor:
593 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
594 // Handle undef ^ undef -> 0 special case. This is a common
596 return Constant::getNullValue(C1->getType());
598 case Instruction::Add:
599 case Instruction::Sub:
600 return UndefValue::get(C1->getType());
601 case Instruction::Mul:
602 case Instruction::And:
603 return Constant::getNullValue(C1->getType());
604 case Instruction::UDiv:
605 case Instruction::SDiv:
606 case Instruction::URem:
607 case Instruction::SRem:
608 if (!isa<UndefValue>(C2)) // undef / X -> 0
609 return Constant::getNullValue(C1->getType());
610 return const_cast<Constant*>(C2); // X / undef -> undef
611 case Instruction::Or: // X | undef -> -1
612 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
613 return ConstantVector::getAllOnesValue(PTy);
614 return ConstantInt::getAllOnesValue(C1->getType());
615 case Instruction::LShr:
616 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
617 return const_cast<Constant*>(C1); // undef lshr undef -> undef
618 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
620 case Instruction::AShr:
621 if (!isa<UndefValue>(C2))
622 return const_cast<Constant*>(C1); // undef ashr X --> undef
623 else if (isa<UndefValue>(C1))
624 return const_cast<Constant*>(C1); // undef ashr undef -> undef
626 return const_cast<Constant*>(C1); // X ashr undef --> X
627 case Instruction::Shl:
628 // undef << X -> 0 or X << undef -> 0
629 return Constant::getNullValue(C1->getType());
633 // Handle simplifications when the RHS is a constant int.
634 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
636 case Instruction::Add:
637 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X + 0 == X
639 case Instruction::Sub:
640 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X - 0 == X
642 case Instruction::Mul:
643 if (CI2->equalsInt(0)) return const_cast<Constant*>(C2); // X * 0 == 0
644 if (CI2->equalsInt(1))
645 return const_cast<Constant*>(C1); // X * 1 == X
647 case Instruction::UDiv:
648 case Instruction::SDiv:
649 if (CI2->equalsInt(1))
650 return const_cast<Constant*>(C1); // X / 1 == X
651 if (CI2->equalsInt(0))
652 return UndefValue::get(CI2->getType()); // X / 0 == undef
654 case Instruction::URem:
655 case Instruction::SRem:
656 if (CI2->equalsInt(1))
657 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
658 if (CI2->equalsInt(0))
659 return UndefValue::get(CI2->getType()); // X % 0 == undef
661 case Instruction::And:
662 if (CI2->isZero()) return const_cast<Constant*>(C2); // X & 0 == 0
663 if (CI2->isAllOnesValue())
664 return const_cast<Constant*>(C1); // X & -1 == X
666 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
667 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
668 if (CE1->getOpcode() == Instruction::ZExt) {
669 unsigned DstWidth = CI2->getType()->getBitWidth();
671 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
672 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
673 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
674 return const_cast<Constant*>(C1);
677 // If and'ing the address of a global with a constant, fold it.
678 if (CE1->getOpcode() == Instruction::PtrToInt &&
679 isa<GlobalValue>(CE1->getOperand(0))) {
680 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
682 // Functions are at least 4-byte aligned.
683 unsigned GVAlign = GV->getAlignment();
684 if (isa<Function>(GV))
685 GVAlign = std::max(GVAlign, 4U);
688 unsigned DstWidth = CI2->getType()->getBitWidth();
689 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
690 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
692 // If checking bits we know are clear, return zero.
693 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
694 return Constant::getNullValue(CI2->getType());
699 case Instruction::Or:
700 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X | 0 == X
701 if (CI2->isAllOnesValue())
702 return const_cast<Constant*>(C2); // X | -1 == -1
704 case Instruction::Xor:
705 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X ^ 0 == X
707 case Instruction::AShr:
708 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
709 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
710 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
711 return ConstantExpr::getLShr(const_cast<Constant*>(C1),
712 const_cast<Constant*>(C2));
717 // At this point we know neither constant is an UndefValue.
718 if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
719 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
720 using namespace APIntOps;
721 const APInt &C1V = CI1->getValue();
722 const APInt &C2V = CI2->getValue();
726 case Instruction::Add:
727 return ConstantInt::get(C1V + C2V);
728 case Instruction::Sub:
729 return ConstantInt::get(C1V - C2V);
730 case Instruction::Mul:
731 return ConstantInt::get(C1V * C2V);
732 case Instruction::UDiv:
733 assert(!CI2->isNullValue() && "Div by zero handled above");
734 return ConstantInt::get(C1V.udiv(C2V));
735 case Instruction::SDiv:
736 assert(!CI2->isNullValue() && "Div by zero handled above");
737 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
738 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
739 return ConstantInt::get(C1V.sdiv(C2V));
740 case Instruction::URem:
741 assert(!CI2->isNullValue() && "Div by zero handled above");
742 return ConstantInt::get(C1V.urem(C2V));
743 case Instruction::SRem:
744 assert(!CI2->isNullValue() && "Div by zero handled above");
745 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
746 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
747 return ConstantInt::get(C1V.srem(C2V));
748 case Instruction::And:
749 return ConstantInt::get(C1V & C2V);
750 case Instruction::Or:
751 return ConstantInt::get(C1V | C2V);
752 case Instruction::Xor:
753 return ConstantInt::get(C1V ^ C2V);
754 case Instruction::Shl: {
755 uint32_t shiftAmt = C2V.getZExtValue();
756 if (shiftAmt < C1V.getBitWidth())
757 return ConstantInt::get(C1V.shl(shiftAmt));
759 return UndefValue::get(C1->getType()); // too big shift is undef
761 case Instruction::LShr: {
762 uint32_t shiftAmt = C2V.getZExtValue();
763 if (shiftAmt < C1V.getBitWidth())
764 return ConstantInt::get(C1V.lshr(shiftAmt));
766 return UndefValue::get(C1->getType()); // too big shift is undef
768 case Instruction::AShr: {
769 uint32_t shiftAmt = C2V.getZExtValue();
770 if (shiftAmt < C1V.getBitWidth())
771 return ConstantInt::get(C1V.ashr(shiftAmt));
773 return UndefValue::get(C1->getType()); // too big shift is undef
779 case Instruction::SDiv:
780 case Instruction::UDiv:
781 case Instruction::URem:
782 case Instruction::SRem:
783 case Instruction::LShr:
784 case Instruction::AShr:
785 case Instruction::Shl:
786 if (CI1->equalsInt(0)) return const_cast<Constant*>(C1);
791 } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
792 if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
793 APFloat C1V = CFP1->getValueAPF();
794 APFloat C2V = CFP2->getValueAPF();
795 APFloat C3V = C1V; // copy for modification
799 case Instruction::FAdd:
800 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
801 return ConstantFP::get(C3V);
802 case Instruction::FSub:
803 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
804 return ConstantFP::get(C3V);
805 case Instruction::FMul:
806 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
807 return ConstantFP::get(C3V);
808 case Instruction::FDiv:
809 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
810 return ConstantFP::get(C3V);
811 case Instruction::FRem:
812 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
813 return ConstantFP::get(C3V);
816 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
817 const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
818 const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
819 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
820 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
824 case Instruction::Add:
825 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAdd);
826 case Instruction::FAdd:
827 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFAdd);
828 case Instruction::Sub:
829 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSub);
830 case Instruction::FSub:
831 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFSub);
832 case Instruction::Mul:
833 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getMul);
834 case Instruction::FMul:
835 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFMul);
836 case Instruction::UDiv:
837 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getUDiv);
838 case Instruction::SDiv:
839 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSDiv);
840 case Instruction::FDiv:
841 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFDiv);
842 case Instruction::URem:
843 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getURem);
844 case Instruction::SRem:
845 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSRem);
846 case Instruction::FRem:
847 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFRem);
848 case Instruction::And:
849 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAnd);
850 case Instruction::Or:
851 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getOr);
852 case Instruction::Xor:
853 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getXor);
854 case Instruction::LShr:
855 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getLShr);
856 case Instruction::AShr:
857 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAShr);
858 case Instruction::Shl:
859 return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getShl);
864 if (isa<ConstantExpr>(C1)) {
865 // There are many possible foldings we could do here. We should probably
866 // at least fold add of a pointer with an integer into the appropriate
867 // getelementptr. This will improve alias analysis a bit.
868 } else if (isa<ConstantExpr>(C2)) {
869 // If C2 is a constant expr and C1 isn't, flop them around and fold the
870 // other way if possible.
872 case Instruction::Add:
873 case Instruction::FAdd:
874 case Instruction::Mul:
875 case Instruction::FMul:
876 case Instruction::And:
877 case Instruction::Or:
878 case Instruction::Xor:
879 // No change of opcode required.
880 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
882 case Instruction::Shl:
883 case Instruction::LShr:
884 case Instruction::AShr:
885 case Instruction::Sub:
886 case Instruction::FSub:
887 case Instruction::SDiv:
888 case Instruction::UDiv:
889 case Instruction::FDiv:
890 case Instruction::URem:
891 case Instruction::SRem:
892 case Instruction::FRem:
893 default: // These instructions cannot be flopped around.
898 // We don't know how to fold this.
902 /// isZeroSizedType - This type is zero sized if its an array or structure of
903 /// zero sized types. The only leaf zero sized type is an empty structure.
904 static bool isMaybeZeroSizedType(const Type *Ty) {
905 if (isa<OpaqueType>(Ty)) return true; // Can't say.
906 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
908 // If all of elements have zero size, this does too.
909 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
910 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
913 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
914 return isMaybeZeroSizedType(ATy->getElementType());
919 /// IdxCompare - Compare the two constants as though they were getelementptr
920 /// indices. This allows coersion of the types to be the same thing.
922 /// If the two constants are the "same" (after coersion), return 0. If the
923 /// first is less than the second, return -1, if the second is less than the
924 /// first, return 1. If the constants are not integral, return -2.
926 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
927 if (C1 == C2) return 0;
929 // Ok, we found a different index. If they are not ConstantInt, we can't do
930 // anything with them.
931 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
932 return -2; // don't know!
934 // Ok, we have two differing integer indices. Sign extend them to be the same
935 // type. Long is always big enough, so we use it.
936 if (C1->getType() != Type::Int64Ty)
937 C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
939 if (C2->getType() != Type::Int64Ty)
940 C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
942 if (C1 == C2) return 0; // They are equal
944 // If the type being indexed over is really just a zero sized type, there is
945 // no pointer difference being made here.
946 if (isMaybeZeroSizedType(ElTy))
949 // If they are really different, now that they are the same type, then we
950 // found a difference!
951 if (cast<ConstantInt>(C1)->getSExtValue() <
952 cast<ConstantInt>(C2)->getSExtValue())
958 /// evaluateFCmpRelation - This function determines if there is anything we can
959 /// decide about the two constants provided. This doesn't need to handle simple
960 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
961 /// If we can determine that the two constants have a particular relation to
962 /// each other, we should return the corresponding FCmpInst predicate,
963 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
964 /// ConstantFoldCompareInstruction.
966 /// To simplify this code we canonicalize the relation so that the first
967 /// operand is always the most "complex" of the two. We consider ConstantFP
968 /// to be the simplest, and ConstantExprs to be the most complex.
969 static FCmpInst::Predicate evaluateFCmpRelation(const Constant *V1,
970 const Constant *V2) {
971 assert(V1->getType() == V2->getType() &&
972 "Cannot compare values of different types!");
974 // No compile-time operations on this type yet.
975 if (V1->getType() == Type::PPC_FP128Ty)
976 return FCmpInst::BAD_FCMP_PREDICATE;
978 // Handle degenerate case quickly
979 if (V1 == V2) return FCmpInst::FCMP_OEQ;
981 if (!isa<ConstantExpr>(V1)) {
982 if (!isa<ConstantExpr>(V2)) {
983 // We distilled thisUse the standard constant folder for a few cases
985 Constant *C1 = const_cast<Constant*>(V1);
986 Constant *C2 = const_cast<Constant*>(V2);
987 R = dyn_cast<ConstantInt>(
988 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2));
989 if (R && !R->isZero())
990 return FCmpInst::FCMP_OEQ;
991 R = dyn_cast<ConstantInt>(
992 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2));
993 if (R && !R->isZero())
994 return FCmpInst::FCMP_OLT;
995 R = dyn_cast<ConstantInt>(
996 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2));
997 if (R && !R->isZero())
998 return FCmpInst::FCMP_OGT;
1000 // Nothing more we can do
1001 return FCmpInst::BAD_FCMP_PREDICATE;
1004 // If the first operand is simple and second is ConstantExpr, swap operands.
1005 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1006 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1007 return FCmpInst::getSwappedPredicate(SwappedRelation);
1009 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1010 // constantexpr or a simple constant.
1011 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1012 switch (CE1->getOpcode()) {
1013 case Instruction::FPTrunc:
1014 case Instruction::FPExt:
1015 case Instruction::UIToFP:
1016 case Instruction::SIToFP:
1017 // We might be able to do something with these but we don't right now.
1023 // There are MANY other foldings that we could perform here. They will
1024 // probably be added on demand, as they seem needed.
1025 return FCmpInst::BAD_FCMP_PREDICATE;
1028 /// evaluateICmpRelation - This function determines if there is anything we can
1029 /// decide about the two constants provided. This doesn't need to handle simple
1030 /// things like integer comparisons, but should instead handle ConstantExprs
1031 /// and GlobalValues. If we can determine that the two constants have a
1032 /// particular relation to each other, we should return the corresponding ICmp
1033 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1035 /// To simplify this code we canonicalize the relation so that the first
1036 /// operand is always the most "complex" of the two. We consider simple
1037 /// constants (like ConstantInt) to be the simplest, followed by
1038 /// GlobalValues, followed by ConstantExpr's (the most complex).
1040 static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1,
1043 assert(V1->getType() == V2->getType() &&
1044 "Cannot compare different types of values!");
1045 if (V1 == V2) return ICmpInst::ICMP_EQ;
1047 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) {
1048 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) {
1049 // We distilled this down to a simple case, use the standard constant
1052 Constant *C1 = const_cast<Constant*>(V1);
1053 Constant *C2 = const_cast<Constant*>(V2);
1054 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1055 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1056 if (R && !R->isZero())
1058 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1059 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1060 if (R && !R->isZero())
1062 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1063 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
1064 if (R && !R->isZero())
1067 // If we couldn't figure it out, bail.
1068 return ICmpInst::BAD_ICMP_PREDICATE;
1071 // If the first operand is simple, swap operands.
1072 ICmpInst::Predicate SwappedRelation =
1073 evaluateICmpRelation(V2, V1, isSigned);
1074 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1075 return ICmpInst::getSwappedPredicate(SwappedRelation);
1077 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) {
1078 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1079 ICmpInst::Predicate SwappedRelation =
1080 evaluateICmpRelation(V2, V1, isSigned);
1081 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1082 return ICmpInst::getSwappedPredicate(SwappedRelation);
1084 return ICmpInst::BAD_ICMP_PREDICATE;
1087 // Now we know that the RHS is a GlobalValue or simple constant,
1088 // which (since the types must match) means that it's a ConstantPointerNull.
1089 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
1090 // Don't try to decide equality of aliases.
1091 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2))
1092 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage())
1093 return ICmpInst::ICMP_NE;
1095 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1096 // GlobalVals can never be null. Don't try to evaluate aliases.
1097 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1))
1098 return ICmpInst::ICMP_NE;
1101 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1102 // constantexpr, a CPR, or a simple constant.
1103 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1104 const Constant *CE1Op0 = CE1->getOperand(0);
1106 switch (CE1->getOpcode()) {
1107 case Instruction::Trunc:
1108 case Instruction::FPTrunc:
1109 case Instruction::FPExt:
1110 case Instruction::FPToUI:
1111 case Instruction::FPToSI:
1112 break; // We can't evaluate floating point casts or truncations.
1114 case Instruction::UIToFP:
1115 case Instruction::SIToFP:
1116 case Instruction::BitCast:
1117 case Instruction::ZExt:
1118 case Instruction::SExt:
1119 // If the cast is not actually changing bits, and the second operand is a
1120 // null pointer, do the comparison with the pre-casted value.
1121 if (V2->isNullValue() &&
1122 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
1123 bool sgnd = isSigned;
1124 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1125 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1126 return evaluateICmpRelation(CE1Op0,
1127 Constant::getNullValue(CE1Op0->getType()),
1131 // If the dest type is a pointer type, and the RHS is a constantexpr cast
1132 // from the same type as the src of the LHS, evaluate the inputs. This is
1133 // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)",
1134 // which happens a lot in compilers with tagged integers.
1135 if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2))
1136 if (CE2->isCast() && isa<PointerType>(CE1->getType()) &&
1137 CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() &&
1138 CE1->getOperand(0)->getType()->isInteger()) {
1139 bool sgnd = isSigned;
1140 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1141 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1142 return evaluateICmpRelation(CE1->getOperand(0), CE2->getOperand(0),
1147 case Instruction::GetElementPtr:
1148 // Ok, since this is a getelementptr, we know that the constant has a
1149 // pointer type. Check the various cases.
1150 if (isa<ConstantPointerNull>(V2)) {
1151 // If we are comparing a GEP to a null pointer, check to see if the base
1152 // of the GEP equals the null pointer.
1153 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1154 if (GV->hasExternalWeakLinkage())
1155 // Weak linkage GVals could be zero or not. We're comparing that
1156 // to null pointer so its greater-or-equal
1157 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1159 // If its not weak linkage, the GVal must have a non-zero address
1160 // so the result is greater-than
1161 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1162 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1163 // If we are indexing from a null pointer, check to see if we have any
1164 // non-zero indices.
1165 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1166 if (!CE1->getOperand(i)->isNullValue())
1167 // Offsetting from null, must not be equal.
1168 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1169 // Only zero indexes from null, must still be zero.
1170 return ICmpInst::ICMP_EQ;
1172 // Otherwise, we can't really say if the first operand is null or not.
1173 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
1174 if (isa<ConstantPointerNull>(CE1Op0)) {
1175 if (CPR2->hasExternalWeakLinkage())
1176 // Weak linkage GVals could be zero or not. We're comparing it to
1177 // a null pointer, so its less-or-equal
1178 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1180 // If its not weak linkage, the GVal must have a non-zero address
1181 // so the result is less-than
1182 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1183 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) {
1185 // If this is a getelementptr of the same global, then it must be
1186 // different. Because the types must match, the getelementptr could
1187 // only have at most one index, and because we fold getelementptr's
1188 // with a single zero index, it must be nonzero.
1189 assert(CE1->getNumOperands() == 2 &&
1190 !CE1->getOperand(1)->isNullValue() &&
1191 "Suprising getelementptr!");
1192 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1194 // If they are different globals, we don't know what the value is,
1195 // but they can't be equal.
1196 return ICmpInst::ICMP_NE;
1200 const ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1201 const Constant *CE2Op0 = CE2->getOperand(0);
1203 // There are MANY other foldings that we could perform here. They will
1204 // probably be added on demand, as they seem needed.
1205 switch (CE2->getOpcode()) {
1207 case Instruction::GetElementPtr:
1208 // By far the most common case to handle is when the base pointers are
1209 // obviously to the same or different globals.
1210 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1211 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1212 return ICmpInst::ICMP_NE;
1213 // Ok, we know that both getelementptr instructions are based on the
1214 // same global. From this, we can precisely determine the relative
1215 // ordering of the resultant pointers.
1218 // Compare all of the operands the GEP's have in common.
1219 gep_type_iterator GTI = gep_type_begin(CE1);
1220 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1222 switch (IdxCompare(CE1->getOperand(i), CE2->getOperand(i),
1223 GTI.getIndexedType())) {
1224 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1225 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1226 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1229 // Ok, we ran out of things they have in common. If any leftovers
1230 // are non-zero then we have a difference, otherwise we are equal.
1231 for (; i < CE1->getNumOperands(); ++i)
1232 if (!CE1->getOperand(i)->isNullValue()) {
1233 if (isa<ConstantInt>(CE1->getOperand(i)))
1234 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1236 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1239 for (; i < CE2->getNumOperands(); ++i)
1240 if (!CE2->getOperand(i)->isNullValue()) {
1241 if (isa<ConstantInt>(CE2->getOperand(i)))
1242 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1244 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1246 return ICmpInst::ICMP_EQ;
1255 return ICmpInst::BAD_ICMP_PREDICATE;
1258 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1260 const Constant *C2) {
1261 const Type *ResultTy;
1262 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1263 ResultTy = VectorType::get(Type::Int1Ty, VT->getNumElements());
1265 ResultTy = Type::Int1Ty;
1267 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1268 if (pred == FCmpInst::FCMP_FALSE)
1269 return Constant::getNullValue(ResultTy);
1271 if (pred == FCmpInst::FCMP_TRUE)
1272 return Constant::getAllOnesValue(ResultTy);
1274 // Handle some degenerate cases first
1275 if (isa<UndefValue>(C1) || isa<UndefValue>(C2))
1276 return UndefValue::get(ResultTy);
1278 // No compile-time operations on this type yet.
1279 if (C1->getType() == Type::PPC_FP128Ty)
1282 // icmp eq/ne(null,GV) -> false/true
1283 if (C1->isNullValue()) {
1284 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1285 // Don't try to evaluate aliases. External weak GV can be null.
1286 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1287 if (pred == ICmpInst::ICMP_EQ)
1288 return ConstantInt::getFalse();
1289 else if (pred == ICmpInst::ICMP_NE)
1290 return ConstantInt::getTrue();
1292 // icmp eq/ne(GV,null) -> false/true
1293 } else if (C2->isNullValue()) {
1294 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1295 // Don't try to evaluate aliases. External weak GV can be null.
1296 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1297 if (pred == ICmpInst::ICMP_EQ)
1298 return ConstantInt::getFalse();
1299 else if (pred == ICmpInst::ICMP_NE)
1300 return ConstantInt::getTrue();
1304 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1305 APInt V1 = cast<ConstantInt>(C1)->getValue();
1306 APInt V2 = cast<ConstantInt>(C2)->getValue();
1308 default: LLVM_UNREACHABLE("Invalid ICmp Predicate"); return 0;
1309 case ICmpInst::ICMP_EQ: return ConstantInt::get(Type::Int1Ty, V1 == V2);
1310 case ICmpInst::ICMP_NE: return ConstantInt::get(Type::Int1Ty, V1 != V2);
1311 case ICmpInst::ICMP_SLT:return ConstantInt::get(Type::Int1Ty, V1.slt(V2));
1312 case ICmpInst::ICMP_SGT:return ConstantInt::get(Type::Int1Ty, V1.sgt(V2));
1313 case ICmpInst::ICMP_SLE:return ConstantInt::get(Type::Int1Ty, V1.sle(V2));
1314 case ICmpInst::ICMP_SGE:return ConstantInt::get(Type::Int1Ty, V1.sge(V2));
1315 case ICmpInst::ICMP_ULT:return ConstantInt::get(Type::Int1Ty, V1.ult(V2));
1316 case ICmpInst::ICMP_UGT:return ConstantInt::get(Type::Int1Ty, V1.ugt(V2));
1317 case ICmpInst::ICMP_ULE:return ConstantInt::get(Type::Int1Ty, V1.ule(V2));
1318 case ICmpInst::ICMP_UGE:return ConstantInt::get(Type::Int1Ty, V1.uge(V2));
1320 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1321 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1322 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1323 APFloat::cmpResult R = C1V.compare(C2V);
1325 default: LLVM_UNREACHABLE("Invalid FCmp Predicate"); return 0;
1326 case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse();
1327 case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue();
1328 case FCmpInst::FCMP_UNO:
1329 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered);
1330 case FCmpInst::FCMP_ORD:
1331 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpUnordered);
1332 case FCmpInst::FCMP_UEQ:
1333 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1334 R==APFloat::cmpEqual);
1335 case FCmpInst::FCMP_OEQ:
1336 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpEqual);
1337 case FCmpInst::FCMP_UNE:
1338 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpEqual);
1339 case FCmpInst::FCMP_ONE:
1340 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1341 R==APFloat::cmpGreaterThan);
1342 case FCmpInst::FCMP_ULT:
1343 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1344 R==APFloat::cmpLessThan);
1345 case FCmpInst::FCMP_OLT:
1346 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan);
1347 case FCmpInst::FCMP_UGT:
1348 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1349 R==APFloat::cmpGreaterThan);
1350 case FCmpInst::FCMP_OGT:
1351 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan);
1352 case FCmpInst::FCMP_ULE:
1353 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpGreaterThan);
1354 case FCmpInst::FCMP_OLE:
1355 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1356 R==APFloat::cmpEqual);
1357 case FCmpInst::FCMP_UGE:
1358 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpLessThan);
1359 case FCmpInst::FCMP_OGE:
1360 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan ||
1361 R==APFloat::cmpEqual);
1363 } else if (isa<VectorType>(C1->getType())) {
1364 SmallVector<Constant*, 16> C1Elts, C2Elts;
1365 C1->getVectorElements(C1Elts);
1366 C2->getVectorElements(C2Elts);
1368 // If we can constant fold the comparison of each element, constant fold
1369 // the whole vector comparison.
1370 SmallVector<Constant*, 4> ResElts;
1371 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1372 // Compare the elements, producing an i1 result or constant expr.
1373 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1375 return ConstantVector::get(&ResElts[0], ResElts.size());
1378 if (C1->getType()->isFloatingPoint()) {
1379 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1380 switch (evaluateFCmpRelation(C1, C2)) {
1381 default: LLVM_UNREACHABLE("Unknown relation!");
1382 case FCmpInst::FCMP_UNO:
1383 case FCmpInst::FCMP_ORD:
1384 case FCmpInst::FCMP_UEQ:
1385 case FCmpInst::FCMP_UNE:
1386 case FCmpInst::FCMP_ULT:
1387 case FCmpInst::FCMP_UGT:
1388 case FCmpInst::FCMP_ULE:
1389 case FCmpInst::FCMP_UGE:
1390 case FCmpInst::FCMP_TRUE:
1391 case FCmpInst::FCMP_FALSE:
1392 case FCmpInst::BAD_FCMP_PREDICATE:
1393 break; // Couldn't determine anything about these constants.
1394 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1395 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1396 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1397 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1399 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1400 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1401 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1402 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1404 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1405 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1406 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1407 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1409 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1410 // We can only partially decide this relation.
1411 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1413 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1416 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1417 // We can only partially decide this relation.
1418 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1420 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1423 case ICmpInst::ICMP_NE: // We know that C1 != C2
1424 // We can only partially decide this relation.
1425 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1427 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1432 // If we evaluated the result, return it now.
1434 return ConstantInt::get(Type::Int1Ty, Result);
1437 // Evaluate the relation between the two constants, per the predicate.
1438 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1439 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1440 default: LLVM_UNREACHABLE("Unknown relational!");
1441 case ICmpInst::BAD_ICMP_PREDICATE:
1442 break; // Couldn't determine anything about these constants.
1443 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1444 // If we know the constants are equal, we can decide the result of this
1445 // computation precisely.
1446 Result = (pred == ICmpInst::ICMP_EQ ||
1447 pred == ICmpInst::ICMP_ULE ||
1448 pred == ICmpInst::ICMP_SLE ||
1449 pred == ICmpInst::ICMP_UGE ||
1450 pred == ICmpInst::ICMP_SGE);
1452 case ICmpInst::ICMP_ULT:
1453 // If we know that C1 < C2, we can decide the result of this computation
1455 Result = (pred == ICmpInst::ICMP_ULT ||
1456 pred == ICmpInst::ICMP_NE ||
1457 pred == ICmpInst::ICMP_ULE);
1459 case ICmpInst::ICMP_SLT:
1460 // If we know that C1 < C2, we can decide the result of this computation
1462 Result = (pred == ICmpInst::ICMP_SLT ||
1463 pred == ICmpInst::ICMP_NE ||
1464 pred == ICmpInst::ICMP_SLE);
1466 case ICmpInst::ICMP_UGT:
1467 // If we know that C1 > C2, we can decide the result of this computation
1469 Result = (pred == ICmpInst::ICMP_UGT ||
1470 pred == ICmpInst::ICMP_NE ||
1471 pred == ICmpInst::ICMP_UGE);
1473 case ICmpInst::ICMP_SGT:
1474 // If we know that C1 > C2, we can decide the result of this computation
1476 Result = (pred == ICmpInst::ICMP_SGT ||
1477 pred == ICmpInst::ICMP_NE ||
1478 pred == ICmpInst::ICMP_SGE);
1480 case ICmpInst::ICMP_ULE:
1481 // If we know that C1 <= C2, we can only partially decide this relation.
1482 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1483 if (pred == ICmpInst::ICMP_ULT) Result = 1;
1485 case ICmpInst::ICMP_SLE:
1486 // If we know that C1 <= C2, we can only partially decide this relation.
1487 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1488 if (pred == ICmpInst::ICMP_SLT) Result = 1;
1491 case ICmpInst::ICMP_UGE:
1492 // If we know that C1 >= C2, we can only partially decide this relation.
1493 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1494 if (pred == ICmpInst::ICMP_UGT) Result = 1;
1496 case ICmpInst::ICMP_SGE:
1497 // If we know that C1 >= C2, we can only partially decide this relation.
1498 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1499 if (pred == ICmpInst::ICMP_SGT) Result = 1;
1502 case ICmpInst::ICMP_NE:
1503 // If we know that C1 != C2, we can only partially decide this relation.
1504 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1505 if (pred == ICmpInst::ICMP_NE) Result = 1;
1509 // If we evaluated the result, return it now.
1511 return ConstantInt::get(Type::Int1Ty, Result);
1513 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) {
1514 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1515 // other way if possible.
1517 case ICmpInst::ICMP_EQ:
1518 case ICmpInst::ICMP_NE:
1519 // No change of predicate required.
1520 return ConstantFoldCompareInstruction(pred, C2, C1);
1522 case ICmpInst::ICMP_ULT:
1523 case ICmpInst::ICMP_SLT:
1524 case ICmpInst::ICMP_UGT:
1525 case ICmpInst::ICMP_SGT:
1526 case ICmpInst::ICMP_ULE:
1527 case ICmpInst::ICMP_SLE:
1528 case ICmpInst::ICMP_UGE:
1529 case ICmpInst::ICMP_SGE:
1530 // Change the predicate as necessary to swap the operands.
1531 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1532 return ConstantFoldCompareInstruction(pred, C2, C1);
1534 default: // These predicates cannot be flopped around.
1542 Constant *llvm::ConstantFoldGetElementPtr(const Constant *C,
1543 Constant* const *Idxs,
1546 (NumIdx == 1 && Idxs[0]->isNullValue()))
1547 return const_cast<Constant*>(C);
1549 if (isa<UndefValue>(C)) {
1550 const PointerType *Ptr = cast<PointerType>(C->getType());
1551 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
1553 (Value **)Idxs+NumIdx);
1554 assert(Ty != 0 && "Invalid indices for GEP!");
1555 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1558 Constant *Idx0 = Idxs[0];
1559 if (C->isNullValue()) {
1561 for (unsigned i = 0, e = NumIdx; i != e; ++i)
1562 if (!Idxs[i]->isNullValue()) {
1567 const PointerType *Ptr = cast<PointerType>(C->getType());
1568 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
1570 (Value**)Idxs+NumIdx);
1571 assert(Ty != 0 && "Invalid indices for GEP!");
1573 ConstantPointerNull::get(PointerType::get(Ty,Ptr->getAddressSpace()));
1577 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) {
1578 // Combine Indices - If the source pointer to this getelementptr instruction
1579 // is a getelementptr instruction, combine the indices of the two
1580 // getelementptr instructions into a single instruction.
1582 if (CE->getOpcode() == Instruction::GetElementPtr) {
1583 const Type *LastTy = 0;
1584 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1588 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
1589 SmallVector<Value*, 16> NewIndices;
1590 NewIndices.reserve(NumIdx + CE->getNumOperands());
1591 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1592 NewIndices.push_back(CE->getOperand(i));
1594 // Add the last index of the source with the first index of the new GEP.
1595 // Make sure to handle the case when they are actually different types.
1596 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1597 // Otherwise it must be an array.
1598 if (!Idx0->isNullValue()) {
1599 const Type *IdxTy = Combined->getType();
1600 if (IdxTy != Idx0->getType()) {
1601 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty);
1602 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined,
1604 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1607 ConstantExpr::get(Instruction::Add, Idx0, Combined);
1611 NewIndices.push_back(Combined);
1612 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
1613 return ConstantExpr::getGetElementPtr(CE->getOperand(0), &NewIndices[0],
1618 // Implement folding of:
1619 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*),
1621 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
1623 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
1624 if (const PointerType *SPT =
1625 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
1626 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
1627 if (const ArrayType *CAT =
1628 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
1629 if (CAT->getElementType() == SAT->getElementType())
1630 return ConstantExpr::getGetElementPtr(
1631 (Constant*)CE->getOperand(0), Idxs, NumIdx);
1634 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1)
1635 // Into: inttoptr (i64 0 to i8*)
1636 // This happens with pointers to member functions in C++.
1637 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 &&
1638 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) &&
1639 cast<PointerType>(CE->getType())->getElementType() == Type::Int8Ty) {
1640 Constant *Base = CE->getOperand(0);
1641 Constant *Offset = Idxs[0];
1643 // Convert the smaller integer to the larger type.
1644 if (Offset->getType()->getPrimitiveSizeInBits() <
1645 Base->getType()->getPrimitiveSizeInBits())
1646 Offset = ConstantExpr::getSExt(Offset, Base->getType());
1647 else if (Base->getType()->getPrimitiveSizeInBits() <
1648 Offset->getType()->getPrimitiveSizeInBits())
1649 Base = ConstantExpr::getZExt(Base, Offset->getType());
1651 Base = ConstantExpr::getAdd(Base, Offset);
1652 return ConstantExpr::getIntToPtr(Base, CE->getType());