1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements folding of constants for LLVM. This implements the
11 // (internal) ConstantFold.h interface, which is used by the
12 // ConstantExpr::get* methods to automatically fold constants when possible.
14 // The current constant folding implementation is implemented in two pieces: the
15 // pieces that don't need TargetData, and the pieces that do. This is to avoid
16 // a dependence in VMCore on Target.
18 //===----------------------------------------------------------------------===//
20 #include "ConstantFold.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Support/MathExtras.h"
36 //===----------------------------------------------------------------------===//
37 // ConstantFold*Instruction Implementations
38 //===----------------------------------------------------------------------===//
40 /// BitCastConstantVector - Convert the specified ConstantVector node to the
41 /// specified vector type. At this point, we know that the elements of the
42 /// input vector constant are all simple integer or FP values.
43 static Constant *BitCastConstantVector(ConstantVector *CV,
44 const VectorType *DstTy) {
45 // If this cast changes element count then we can't handle it here:
46 // doing so requires endianness information. This should be handled by
47 // Analysis/ConstantFolding.cpp
48 unsigned NumElts = DstTy->getNumElements();
49 if (NumElts != CV->getNumOperands())
52 // Check to verify that all elements of the input are simple.
53 for (unsigned i = 0; i != NumElts; ++i) {
54 if (!isa<ConstantInt>(CV->getOperand(i)) &&
55 !isa<ConstantFP>(CV->getOperand(i)))
59 // Bitcast each element now.
60 std::vector<Constant*> Result;
61 const Type *DstEltTy = DstTy->getElementType();
62 for (unsigned i = 0; i != NumElts; ++i)
63 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i),
65 return ConstantVector::get(Result);
68 /// This function determines which opcode to use to fold two constant cast
69 /// expressions together. It uses CastInst::isEliminableCastPair to determine
70 /// the opcode. Consequently its just a wrapper around that function.
71 /// @brief Determine if it is valid to fold a cast of a cast
74 unsigned opc, ///< opcode of the second cast constant expression
75 ConstantExpr *Op, ///< the first cast constant expression
76 const Type *DstTy ///< desintation type of the first cast
78 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
79 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
80 assert(CastInst::isCast(opc) && "Invalid cast opcode");
82 // The the types and opcodes for the two Cast constant expressions
83 const Type *SrcTy = Op->getOperand(0)->getType();
84 const Type *MidTy = Op->getType();
85 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
86 Instruction::CastOps secondOp = Instruction::CastOps(opc);
88 // Let CastInst::isEliminableCastPair do the heavy lifting.
89 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
90 Type::getInt64Ty(DstTy->getContext()));
93 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
94 const Type *SrcTy = V->getType();
96 return V; // no-op cast
98 // Check to see if we are casting a pointer to an aggregate to a pointer to
99 // the first element. If so, return the appropriate GEP instruction.
100 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
101 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
102 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
103 SmallVector<Value*, 8> IdxList;
105 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
106 IdxList.push_back(Zero);
107 const Type *ElTy = PTy->getElementType();
108 while (ElTy != DPTy->getElementType()) {
109 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
110 if (STy->getNumElements() == 0) break;
111 ElTy = STy->getElementType(0);
112 IdxList.push_back(Zero);
113 } else if (const SequentialType *STy =
114 dyn_cast<SequentialType>(ElTy)) {
115 if (isa<PointerType>(ElTy)) break; // Can't index into pointers!
116 ElTy = STy->getElementType();
117 IdxList.push_back(Zero);
123 if (ElTy == DPTy->getElementType())
124 // This GEP is inbounds because all indices are zero.
125 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0],
129 // Handle casts from one vector constant to another. We know that the src
130 // and dest type have the same size (otherwise its an illegal cast).
131 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
132 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
133 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
134 "Not cast between same sized vectors!");
136 // First, check for null. Undef is already handled.
137 if (isa<ConstantAggregateZero>(V))
138 return Constant::getNullValue(DestTy);
140 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
141 return BitCastConstantVector(CV, DestPTy);
144 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
145 // This allows for other simplifications (although some of them
146 // can only be handled by Analysis/ConstantFolding.cpp).
147 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
148 return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy);
151 // Finally, implement bitcast folding now. The code below doesn't handle
153 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
154 return ConstantPointerNull::get(cast<PointerType>(DestTy));
156 // Handle integral constant input.
157 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
158 if (DestTy->isInteger())
159 // Integral -> Integral. This is a no-op because the bit widths must
160 // be the same. Consequently, we just fold to V.
163 if (DestTy->isFloatingPoint())
164 return ConstantFP::get(DestTy->getContext(),
165 APFloat(CI->getValue(),
166 !DestTy->isPPC_FP128Ty()));
168 // Otherwise, can't fold this (vector?)
172 // Handle ConstantFP input: FP -> Integral.
173 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
174 return ConstantInt::get(FP->getContext(),
175 FP->getValueAPF().bitcastToAPInt());
181 /// ExtractConstantBytes - V is an integer constant which only has a subset of
182 /// its bytes used. The bytes used are indicated by ByteStart (which is the
183 /// first byte used, counting from the least significant byte) and ByteSize,
184 /// which is the number of bytes used.
186 /// This function analyzes the specified constant to see if the specified byte
187 /// range can be returned as a simplified constant. If so, the constant is
188 /// returned, otherwise null is returned.
190 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
192 assert(isa<IntegerType>(C->getType()) &&
193 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
194 "Non-byte sized integer input");
195 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
196 assert(ByteSize && "Must be accessing some piece");
197 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
198 assert(ByteSize != CSize && "Should not extract everything");
200 // Constant Integers are simple.
201 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
202 APInt V = CI->getValue();
204 V = V.lshr(ByteStart*8);
206 return ConstantInt::get(CI->getContext(), V);
209 // In the input is a constant expr, we might be able to recursively simplify.
210 // If not, we definitely can't do anything.
211 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
212 if (CE == 0) return 0;
214 switch (CE->getOpcode()) {
216 case Instruction::Or: {
217 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
223 if (RHSC->isAllOnesValue())
226 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
229 return ConstantExpr::getOr(LHS, RHS);
231 case Instruction::And: {
232 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
237 if (RHS->isNullValue())
240 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
243 return ConstantExpr::getAnd(LHS, RHS);
245 case Instruction::LShr: {
246 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
249 unsigned ShAmt = Amt->getZExtValue();
250 // Cannot analyze non-byte shifts.
251 if ((ShAmt & 7) != 0)
255 // If the extract is known to be all zeros, return zero.
256 if (ByteStart >= CSize-ShAmt)
257 return Constant::getNullValue(IntegerType::get(CE->getContext(),
259 // If the extract is known to be fully in the input, extract it.
260 if (ByteStart+ByteSize+ShAmt <= CSize)
261 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
263 // TODO: Handle the 'partially zero' case.
267 case Instruction::Shl: {
268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
271 unsigned ShAmt = Amt->getZExtValue();
272 // Cannot analyze non-byte shifts.
273 if ((ShAmt & 7) != 0)
277 // If the extract is known to be all zeros, return zero.
278 if (ByteStart+ByteSize <= ShAmt)
279 return Constant::getNullValue(IntegerType::get(CE->getContext(),
281 // If the extract is known to be fully in the input, extract it.
282 if (ByteStart >= ShAmt)
283 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
285 // TODO: Handle the 'partially zero' case.
289 case Instruction::ZExt: {
290 unsigned SrcBitSize =
291 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
293 // If extracting something that is completely zero, return 0.
294 if (ByteStart*8 >= SrcBitSize)
295 return Constant::getNullValue(IntegerType::get(CE->getContext(),
298 // If exactly extracting the input, return it.
299 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
300 return CE->getOperand(0);
302 // If extracting something completely in the input, if if the input is a
303 // multiple of 8 bits, recurse.
304 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
305 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
307 // Otherwise, if extracting a subset of the input, which is not multiple of
308 // 8 bits, do a shift and trunc to get the bits.
309 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
310 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
311 Constant *Res = CE->getOperand(0);
313 Res = ConstantExpr::getLShr(Res,
314 ConstantInt::get(Res->getType(), ByteStart*8));
315 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
319 // TODO: Handle the 'partially zero' case.
325 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
326 /// on Ty, with any known factors factored out. If Folded is false,
327 /// return null if no factoring was possible, to avoid endlessly
328 /// bouncing an unfoldable expression back into the top-level folder.
330 static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy,
332 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
333 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
334 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
335 return ConstantExpr::getNUWMul(E, N);
337 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
338 Constant *N = ConstantInt::get(DestTy, VTy->getNumElements());
339 Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true);
340 return ConstantExpr::getNUWMul(E, N);
342 if (const StructType *STy = dyn_cast<StructType>(Ty))
343 if (!STy->isPacked()) {
344 unsigned NumElems = STy->getNumElements();
345 // An empty struct has size zero.
347 return ConstantExpr::getNullValue(DestTy);
348 // Check for a struct with all members having the same type.
349 const Type *MemberTy = STy->getElementType(0);
351 for (unsigned i = 1; i != NumElems; ++i)
352 if (MemberTy != STy->getElementType(i)) {
357 Constant *N = ConstantInt::get(DestTy, NumElems);
358 Constant *E = getFoldedSizeOf(MemberTy, DestTy, true);
359 return ConstantExpr::getNUWMul(E, N);
363 // If there's no interesting folding happening, bail so that we don't create
364 // a constant that looks like it needs folding but really doesn't.
368 // Base case: Get a regular sizeof expression.
369 Constant *C = ConstantExpr::getSizeOf(Ty);
370 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
376 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
377 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
378 /// return null if no factoring was possible, to avoid endlessly
379 /// bouncing an unfoldable expression back into the top-level folder.
381 static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo,
384 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
385 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
388 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
389 return ConstantExpr::getNUWMul(E, N);
391 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
392 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
395 Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true);
396 return ConstantExpr::getNUWMul(E, N);
398 if (const StructType *STy = dyn_cast<StructType>(Ty))
399 if (!STy->isPacked()) {
400 unsigned NumElems = STy->getNumElements();
401 // An empty struct has no members.
404 // Check for a struct with all members having the same type.
405 const Type *MemberTy = STy->getElementType(0);
407 for (unsigned i = 1; i != NumElems; ++i)
408 if (MemberTy != STy->getElementType(i)) {
413 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
418 Constant *E = getFoldedSizeOf(MemberTy, DestTy, true);
419 return ConstantExpr::getNUWMul(E, N);
423 // If there's no interesting folding happening, bail so that we don't create
424 // a constant that looks like it needs folding but really doesn't.
428 // Base case: Get a regular offsetof expression.
429 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
430 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
436 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
437 const Type *DestTy) {
438 if (isa<UndefValue>(V)) {
439 // zext(undef) = 0, because the top bits will be zero.
440 // sext(undef) = 0, because the top bits will all be the same.
441 // [us]itofp(undef) = 0, because the result value is bounded.
442 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
443 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
444 return Constant::getNullValue(DestTy);
445 return UndefValue::get(DestTy);
447 // No compile-time operations on this type yet.
448 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
451 // If the cast operand is a constant expression, there's a few things we can
452 // do to try to simplify it.
453 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
455 // Try hard to fold cast of cast because they are often eliminable.
456 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
457 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
458 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
459 // If all of the indexes in the GEP are null values, there is no pointer
460 // adjustment going on. We might as well cast the source pointer.
461 bool isAllNull = true;
462 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
463 if (!CE->getOperand(i)->isNullValue()) {
468 // This is casting one pointer type to another, always BitCast
469 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
473 // If the cast operand is a constant vector, perform the cast by
474 // operating on each element. In the cast of bitcasts, the element
475 // count may be mismatched; don't attempt to handle that here.
476 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
477 if (isa<VectorType>(DestTy) &&
478 cast<VectorType>(DestTy)->getNumElements() ==
479 CV->getType()->getNumElements()) {
480 std::vector<Constant*> res;
481 const VectorType *DestVecTy = cast<VectorType>(DestTy);
482 const Type *DstEltTy = DestVecTy->getElementType();
483 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
484 res.push_back(ConstantExpr::getCast(opc,
485 CV->getOperand(i), DstEltTy));
486 return ConstantVector::get(DestVecTy, res);
489 // We actually have to do a cast now. Perform the cast according to the
493 llvm_unreachable("Failed to cast constant expression");
494 case Instruction::FPTrunc:
495 case Instruction::FPExt:
496 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
498 APFloat Val = FPC->getValueAPF();
499 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
500 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
501 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
502 DestTy->isFP128Ty() ? APFloat::IEEEquad :
504 APFloat::rmNearestTiesToEven, &ignored);
505 return ConstantFP::get(V->getContext(), Val);
507 return 0; // Can't fold.
508 case Instruction::FPToUI:
509 case Instruction::FPToSI:
510 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
511 const APFloat &V = FPC->getValueAPF();
514 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
515 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
516 APFloat::rmTowardZero, &ignored);
517 APInt Val(DestBitWidth, 2, x);
518 return ConstantInt::get(FPC->getContext(), Val);
520 return 0; // Can't fold.
521 case Instruction::IntToPtr: //always treated as unsigned
522 if (V->isNullValue()) // Is it an integral null value?
523 return ConstantPointerNull::get(cast<PointerType>(DestTy));
524 return 0; // Other pointer types cannot be casted
525 case Instruction::PtrToInt: // always treated as unsigned
526 // Is it a null pointer value?
527 if (V->isNullValue())
528 return ConstantInt::get(DestTy, 0);
529 // If this is a sizeof-like expression, pull out multiplications by
530 // known factors to expose them to subsequent folding. If it's an
531 // alignof-like expression, factor out known factors.
532 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
533 if (CE->getOpcode() == Instruction::GetElementPtr &&
534 CE->getOperand(0)->isNullValue()) {
536 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
537 if (CE->getNumOperands() == 2) {
538 // Handle a sizeof-like expression.
539 Constant *Idx = CE->getOperand(1);
540 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
541 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
542 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
545 return ConstantExpr::getMul(C, Idx);
547 } else if (CE->getNumOperands() == 3 &&
548 CE->getOperand(1)->isNullValue()) {
549 // Handle an alignof-like expression.
550 if (const StructType *STy = dyn_cast<StructType>(Ty))
551 if (!STy->isPacked()) {
552 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
554 STy->getNumElements() == 2 &&
555 STy->getElementType(0)->isInteger(1)) {
556 // The alignment of an array is equal to the alignment of the
557 // array element. Note that this is not always true for vectors.
558 if (const ArrayType *ATy =
559 dyn_cast<ArrayType>(STy->getElementType(1))) {
560 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
561 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
567 // Packed structs always have an alignment of 1.
568 if (const StructType *InnerSTy =
569 dyn_cast<StructType>(STy->getElementType(1)))
570 if (InnerSTy->isPacked())
571 return ConstantInt::get(DestTy, 1);
574 // Handle an offsetof-like expression.
575 if (isa<StructType>(Ty) || isa<ArrayType>(Ty) || isa<VectorType>(Ty)){
576 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
582 // Other pointer types cannot be casted
584 case Instruction::UIToFP:
585 case Instruction::SIToFP:
586 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
587 APInt api = CI->getValue();
588 const uint64_t zero[] = {0, 0};
589 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(),
591 (void)apf.convertFromAPInt(api,
592 opc==Instruction::SIToFP,
593 APFloat::rmNearestTiesToEven);
594 return ConstantFP::get(V->getContext(), apf);
597 case Instruction::ZExt:
598 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
599 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
600 APInt Result(CI->getValue());
601 Result.zext(BitWidth);
602 return ConstantInt::get(V->getContext(), Result);
605 case Instruction::SExt:
606 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
607 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
608 APInt Result(CI->getValue());
609 Result.sext(BitWidth);
610 return ConstantInt::get(V->getContext(), Result);
613 case Instruction::Trunc: {
614 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
615 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
616 APInt Result(CI->getValue());
617 Result.trunc(DestBitWidth);
618 return ConstantInt::get(V->getContext(), Result);
621 // The input must be a constantexpr. See if we can simplify this based on
622 // the bytes we are demanding. Only do this if the source and dest are an
623 // even multiple of a byte.
624 if ((DestBitWidth & 7) == 0 &&
625 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
626 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
631 case Instruction::BitCast:
632 return FoldBitCast(V, DestTy);
636 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
637 Constant *V1, Constant *V2) {
638 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
639 return CB->getZExtValue() ? V1 : V2;
641 if (isa<UndefValue>(V1)) return V2;
642 if (isa<UndefValue>(V2)) return V1;
643 if (isa<UndefValue>(Cond)) return V1;
644 if (V1 == V2) return V1;
648 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
650 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
651 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
652 if (Val->isNullValue()) // ee(zero, x) -> zero
653 return Constant::getNullValue(
654 cast<VectorType>(Val->getType())->getElementType());
656 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
657 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
658 return CVal->getOperand(CIdx->getZExtValue());
659 } else if (isa<UndefValue>(Idx)) {
660 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
661 return CVal->getOperand(0);
667 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
670 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
672 APInt idxVal = CIdx->getValue();
673 if (isa<UndefValue>(Val)) {
674 // Insertion of scalar constant into vector undef
675 // Optimize away insertion of undef
676 if (isa<UndefValue>(Elt))
678 // Otherwise break the aggregate undef into multiple undefs and do
681 cast<VectorType>(Val->getType())->getNumElements();
682 std::vector<Constant*> Ops;
684 for (unsigned i = 0; i < numOps; ++i) {
686 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
689 return ConstantVector::get(Ops);
691 if (isa<ConstantAggregateZero>(Val)) {
692 // Insertion of scalar constant into vector aggregate zero
693 // Optimize away insertion of zero
694 if (Elt->isNullValue())
696 // Otherwise break the aggregate zero into multiple zeros and do
699 cast<VectorType>(Val->getType())->getNumElements();
700 std::vector<Constant*> Ops;
702 for (unsigned i = 0; i < numOps; ++i) {
704 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
707 return ConstantVector::get(Ops);
709 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
710 // Insertion of scalar constant into vector constant
711 std::vector<Constant*> Ops;
712 Ops.reserve(CVal->getNumOperands());
713 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
715 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
718 return ConstantVector::get(Ops);
724 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
725 /// return the specified element value. Otherwise return null.
726 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
727 if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
728 return CV->getOperand(EltNo);
730 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
731 if (isa<ConstantAggregateZero>(C))
732 return Constant::getNullValue(EltTy);
733 if (isa<UndefValue>(C))
734 return UndefValue::get(EltTy);
738 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
741 // Undefined shuffle mask -> undefined value.
742 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
744 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
745 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
746 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
748 // Loop over the shuffle mask, evaluating each element.
749 SmallVector<Constant*, 32> Result;
750 for (unsigned i = 0; i != MaskNumElts; ++i) {
751 Constant *InElt = GetVectorElement(Mask, i);
752 if (InElt == 0) return 0;
754 if (isa<UndefValue>(InElt))
755 InElt = UndefValue::get(EltTy);
756 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
757 unsigned Elt = CI->getZExtValue();
758 if (Elt >= SrcNumElts*2)
759 InElt = UndefValue::get(EltTy);
760 else if (Elt >= SrcNumElts)
761 InElt = GetVectorElement(V2, Elt - SrcNumElts);
763 InElt = GetVectorElement(V1, Elt);
764 if (InElt == 0) return 0;
769 Result.push_back(InElt);
772 return ConstantVector::get(&Result[0], Result.size());
775 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
776 const unsigned *Idxs,
778 // Base case: no indices, so return the entire value.
782 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
783 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
787 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
789 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
793 // Otherwise recurse.
794 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
795 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs),
798 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
799 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs),
801 ConstantVector *CV = cast<ConstantVector>(Agg);
802 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs),
806 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
808 const unsigned *Idxs,
810 // Base case: no indices, so replace the entire value.
814 if (isa<UndefValue>(Agg)) {
815 // Insertion of constant into aggregate undef
816 // Optimize away insertion of undef.
817 if (isa<UndefValue>(Val))
820 // Otherwise break the aggregate undef into multiple undefs and do
822 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
824 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
825 numOps = AR->getNumElements();
827 numOps = cast<StructType>(AggTy)->getNumElements();
829 std::vector<Constant*> Ops(numOps);
830 for (unsigned i = 0; i < numOps; ++i) {
831 const Type *MemberTy = AggTy->getTypeAtIndex(i);
834 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
835 Val, Idxs+1, NumIdx-1) :
836 UndefValue::get(MemberTy);
840 if (const StructType* ST = dyn_cast<StructType>(AggTy))
841 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
842 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
845 if (isa<ConstantAggregateZero>(Agg)) {
846 // Insertion of constant into aggregate zero
847 // Optimize away insertion of zero.
848 if (Val->isNullValue())
851 // Otherwise break the aggregate zero into multiple zeros and do
853 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
855 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
856 numOps = AR->getNumElements();
858 numOps = cast<StructType>(AggTy)->getNumElements();
860 std::vector<Constant*> Ops(numOps);
861 for (unsigned i = 0; i < numOps; ++i) {
862 const Type *MemberTy = AggTy->getTypeAtIndex(i);
865 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
866 Val, Idxs+1, NumIdx-1) :
867 Constant::getNullValue(MemberTy);
871 if (const StructType *ST = dyn_cast<StructType>(AggTy))
872 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
873 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
876 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
877 // Insertion of constant into aggregate constant.
878 std::vector<Constant*> Ops(Agg->getNumOperands());
879 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
880 Constant *Op = cast<Constant>(Agg->getOperand(i));
882 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1);
886 if (const StructType* ST = dyn_cast<StructType>(Agg->getType()))
887 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
888 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
895 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
896 Constant *C1, Constant *C2) {
897 // No compile-time operations on this type yet.
898 if (C1->getType()->isPPC_FP128Ty())
901 // Handle UndefValue up front.
902 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
904 case Instruction::Xor:
905 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
906 // Handle undef ^ undef -> 0 special case. This is a common
908 return Constant::getNullValue(C1->getType());
910 case Instruction::Add:
911 case Instruction::Sub:
912 return UndefValue::get(C1->getType());
913 case Instruction::Mul:
914 case Instruction::And:
915 return Constant::getNullValue(C1->getType());
916 case Instruction::UDiv:
917 case Instruction::SDiv:
918 case Instruction::URem:
919 case Instruction::SRem:
920 if (!isa<UndefValue>(C2)) // undef / X -> 0
921 return Constant::getNullValue(C1->getType());
922 return C2; // X / undef -> undef
923 case Instruction::Or: // X | undef -> -1
924 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
925 return Constant::getAllOnesValue(PTy);
926 return Constant::getAllOnesValue(C1->getType());
927 case Instruction::LShr:
928 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
929 return C1; // undef lshr undef -> undef
930 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
932 case Instruction::AShr:
933 if (!isa<UndefValue>(C2))
934 return C1; // undef ashr X --> undef
935 else if (isa<UndefValue>(C1))
936 return C1; // undef ashr undef -> undef
938 return C1; // X ashr undef --> X
939 case Instruction::Shl:
940 // undef << X -> 0 or X << undef -> 0
941 return Constant::getNullValue(C1->getType());
945 // Handle simplifications when the RHS is a constant int.
946 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
948 case Instruction::Add:
949 if (CI2->equalsInt(0)) return C1; // X + 0 == X
951 case Instruction::Sub:
952 if (CI2->equalsInt(0)) return C1; // X - 0 == X
954 case Instruction::Mul:
955 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
956 if (CI2->equalsInt(1))
957 return C1; // X * 1 == X
959 case Instruction::UDiv:
960 case Instruction::SDiv:
961 if (CI2->equalsInt(1))
962 return C1; // X / 1 == X
963 if (CI2->equalsInt(0))
964 return UndefValue::get(CI2->getType()); // X / 0 == undef
966 case Instruction::URem:
967 case Instruction::SRem:
968 if (CI2->equalsInt(1))
969 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
970 if (CI2->equalsInt(0))
971 return UndefValue::get(CI2->getType()); // X % 0 == undef
973 case Instruction::And:
974 if (CI2->isZero()) return C2; // X & 0 == 0
975 if (CI2->isAllOnesValue())
976 return C1; // X & -1 == X
978 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
979 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
980 if (CE1->getOpcode() == Instruction::ZExt) {
981 unsigned DstWidth = CI2->getType()->getBitWidth();
983 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
984 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
985 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
989 // If and'ing the address of a global with a constant, fold it.
990 if (CE1->getOpcode() == Instruction::PtrToInt &&
991 isa<GlobalValue>(CE1->getOperand(0))) {
992 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
994 // Functions are at least 4-byte aligned.
995 unsigned GVAlign = GV->getAlignment();
996 if (isa<Function>(GV))
997 GVAlign = std::max(GVAlign, 4U);
1000 unsigned DstWidth = CI2->getType()->getBitWidth();
1001 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1002 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1004 // If checking bits we know are clear, return zero.
1005 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1006 return Constant::getNullValue(CI2->getType());
1011 case Instruction::Or:
1012 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1013 if (CI2->isAllOnesValue())
1014 return C2; // X | -1 == -1
1016 case Instruction::Xor:
1017 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1019 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1020 switch (CE1->getOpcode()) {
1022 case Instruction::ICmp:
1023 case Instruction::FCmp:
1024 // cmp pred ^ true -> cmp !pred
1025 assert(CI2->equalsInt(1));
1026 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1027 pred = CmpInst::getInversePredicate(pred);
1028 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1029 CE1->getOperand(1));
1033 case Instruction::AShr:
1034 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1035 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1036 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1037 return ConstantExpr::getLShr(C1, C2);
1042 // At this point we know neither constant is an UndefValue.
1043 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1044 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1045 using namespace APIntOps;
1046 const APInt &C1V = CI1->getValue();
1047 const APInt &C2V = CI2->getValue();
1051 case Instruction::Add:
1052 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1053 case Instruction::Sub:
1054 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1055 case Instruction::Mul:
1056 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1057 case Instruction::UDiv:
1058 assert(!CI2->isNullValue() && "Div by zero handled above");
1059 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1060 case Instruction::SDiv:
1061 assert(!CI2->isNullValue() && "Div by zero handled above");
1062 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1063 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1064 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1065 case Instruction::URem:
1066 assert(!CI2->isNullValue() && "Div by zero handled above");
1067 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1068 case Instruction::SRem:
1069 assert(!CI2->isNullValue() && "Div by zero handled above");
1070 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1071 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1072 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1073 case Instruction::And:
1074 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1075 case Instruction::Or:
1076 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1077 case Instruction::Xor:
1078 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1079 case Instruction::Shl: {
1080 uint32_t shiftAmt = C2V.getZExtValue();
1081 if (shiftAmt < C1V.getBitWidth())
1082 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1084 return UndefValue::get(C1->getType()); // too big shift is undef
1086 case Instruction::LShr: {
1087 uint32_t shiftAmt = C2V.getZExtValue();
1088 if (shiftAmt < C1V.getBitWidth())
1089 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1091 return UndefValue::get(C1->getType()); // too big shift is undef
1093 case Instruction::AShr: {
1094 uint32_t shiftAmt = C2V.getZExtValue();
1095 if (shiftAmt < C1V.getBitWidth())
1096 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1098 return UndefValue::get(C1->getType()); // too big shift is undef
1104 case Instruction::SDiv:
1105 case Instruction::UDiv:
1106 case Instruction::URem:
1107 case Instruction::SRem:
1108 case Instruction::LShr:
1109 case Instruction::AShr:
1110 case Instruction::Shl:
1111 if (CI1->equalsInt(0)) return C1;
1116 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1117 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1118 APFloat C1V = CFP1->getValueAPF();
1119 APFloat C2V = CFP2->getValueAPF();
1120 APFloat C3V = C1V; // copy for modification
1124 case Instruction::FAdd:
1125 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1126 return ConstantFP::get(C1->getContext(), C3V);
1127 case Instruction::FSub:
1128 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1129 return ConstantFP::get(C1->getContext(), C3V);
1130 case Instruction::FMul:
1131 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1132 return ConstantFP::get(C1->getContext(), C3V);
1133 case Instruction::FDiv:
1134 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1135 return ConstantFP::get(C1->getContext(), C3V);
1136 case Instruction::FRem:
1137 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1138 return ConstantFP::get(C1->getContext(), C3V);
1141 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1142 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1143 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1144 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1145 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1146 std::vector<Constant*> Res;
1147 const Type* EltTy = VTy->getElementType();
1153 case Instruction::Add:
1154 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1155 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1156 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1157 Res.push_back(ConstantExpr::getAdd(C1, C2));
1159 return ConstantVector::get(Res);
1160 case Instruction::FAdd:
1161 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1162 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1163 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1164 Res.push_back(ConstantExpr::getFAdd(C1, C2));
1166 return ConstantVector::get(Res);
1167 case Instruction::Sub:
1168 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1169 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1170 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1171 Res.push_back(ConstantExpr::getSub(C1, C2));
1173 return ConstantVector::get(Res);
1174 case Instruction::FSub:
1175 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1176 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1177 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1178 Res.push_back(ConstantExpr::getFSub(C1, C2));
1180 return ConstantVector::get(Res);
1181 case Instruction::Mul:
1182 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1183 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1184 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1185 Res.push_back(ConstantExpr::getMul(C1, C2));
1187 return ConstantVector::get(Res);
1188 case Instruction::FMul:
1189 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1190 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1191 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1192 Res.push_back(ConstantExpr::getFMul(C1, C2));
1194 return ConstantVector::get(Res);
1195 case Instruction::UDiv:
1196 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1197 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1198 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1199 Res.push_back(ConstantExpr::getUDiv(C1, C2));
1201 return ConstantVector::get(Res);
1202 case Instruction::SDiv:
1203 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1204 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1205 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1206 Res.push_back(ConstantExpr::getSDiv(C1, C2));
1208 return ConstantVector::get(Res);
1209 case Instruction::FDiv:
1210 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1211 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1212 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1213 Res.push_back(ConstantExpr::getFDiv(C1, C2));
1215 return ConstantVector::get(Res);
1216 case Instruction::URem:
1217 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1218 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1219 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1220 Res.push_back(ConstantExpr::getURem(C1, C2));
1222 return ConstantVector::get(Res);
1223 case Instruction::SRem:
1224 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1225 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1226 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1227 Res.push_back(ConstantExpr::getSRem(C1, C2));
1229 return ConstantVector::get(Res);
1230 case Instruction::FRem:
1231 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1232 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1233 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1234 Res.push_back(ConstantExpr::getFRem(C1, C2));
1236 return ConstantVector::get(Res);
1237 case Instruction::And:
1238 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1239 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1240 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1241 Res.push_back(ConstantExpr::getAnd(C1, C2));
1243 return ConstantVector::get(Res);
1244 case Instruction::Or:
1245 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1246 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1247 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1248 Res.push_back(ConstantExpr::getOr(C1, C2));
1250 return ConstantVector::get(Res);
1251 case Instruction::Xor:
1252 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1253 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1254 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1255 Res.push_back(ConstantExpr::getXor(C1, C2));
1257 return ConstantVector::get(Res);
1258 case Instruction::LShr:
1259 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1260 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1261 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1262 Res.push_back(ConstantExpr::getLShr(C1, C2));
1264 return ConstantVector::get(Res);
1265 case Instruction::AShr:
1266 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1267 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1268 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1269 Res.push_back(ConstantExpr::getAShr(C1, C2));
1271 return ConstantVector::get(Res);
1272 case Instruction::Shl:
1273 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1274 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1275 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1276 Res.push_back(ConstantExpr::getShl(C1, C2));
1278 return ConstantVector::get(Res);
1283 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1284 // There are many possible foldings we could do here. We should probably
1285 // at least fold add of a pointer with an integer into the appropriate
1286 // getelementptr. This will improve alias analysis a bit.
1288 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1290 if (Instruction::isAssociative(Opcode, C1->getType()) &&
1291 CE1->getOpcode() == Opcode) {
1292 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1293 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1294 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1296 } else if (isa<ConstantExpr>(C2)) {
1297 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1298 // other way if possible.
1300 case Instruction::Add:
1301 case Instruction::FAdd:
1302 case Instruction::Mul:
1303 case Instruction::FMul:
1304 case Instruction::And:
1305 case Instruction::Or:
1306 case Instruction::Xor:
1307 // No change of opcode required.
1308 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1310 case Instruction::Shl:
1311 case Instruction::LShr:
1312 case Instruction::AShr:
1313 case Instruction::Sub:
1314 case Instruction::FSub:
1315 case Instruction::SDiv:
1316 case Instruction::UDiv:
1317 case Instruction::FDiv:
1318 case Instruction::URem:
1319 case Instruction::SRem:
1320 case Instruction::FRem:
1321 default: // These instructions cannot be flopped around.
1326 // i1 can be simplified in many cases.
1327 if (C1->getType()->isInteger(1)) {
1329 case Instruction::Add:
1330 case Instruction::Sub:
1331 return ConstantExpr::getXor(C1, C2);
1332 case Instruction::Mul:
1333 return ConstantExpr::getAnd(C1, C2);
1334 case Instruction::Shl:
1335 case Instruction::LShr:
1336 case Instruction::AShr:
1337 // We can assume that C2 == 0. If it were one the result would be
1338 // undefined because the shift value is as large as the bitwidth.
1340 case Instruction::SDiv:
1341 case Instruction::UDiv:
1342 // We can assume that C2 == 1. If it were zero the result would be
1343 // undefined through division by zero.
1345 case Instruction::URem:
1346 case Instruction::SRem:
1347 // We can assume that C2 == 1. If it were zero the result would be
1348 // undefined through division by zero.
1349 return ConstantInt::getFalse(C1->getContext());
1355 // We don't know how to fold this.
1359 /// isZeroSizedType - This type is zero sized if its an array or structure of
1360 /// zero sized types. The only leaf zero sized type is an empty structure.
1361 static bool isMaybeZeroSizedType(const Type *Ty) {
1362 if (isa<OpaqueType>(Ty)) return true; // Can't say.
1363 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1365 // If all of elements have zero size, this does too.
1366 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1367 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1370 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1371 return isMaybeZeroSizedType(ATy->getElementType());
1376 /// IdxCompare - Compare the two constants as though they were getelementptr
1377 /// indices. This allows coersion of the types to be the same thing.
1379 /// If the two constants are the "same" (after coersion), return 0. If the
1380 /// first is less than the second, return -1, if the second is less than the
1381 /// first, return 1. If the constants are not integral, return -2.
1383 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
1384 if (C1 == C2) return 0;
1386 // Ok, we found a different index. If they are not ConstantInt, we can't do
1387 // anything with them.
1388 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1389 return -2; // don't know!
1391 // Ok, we have two differing integer indices. Sign extend them to be the same
1392 // type. Long is always big enough, so we use it.
1393 if (!C1->getType()->isInteger(64))
1394 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1396 if (!C2->getType()->isInteger(64))
1397 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1399 if (C1 == C2) return 0; // They are equal
1401 // If the type being indexed over is really just a zero sized type, there is
1402 // no pointer difference being made here.
1403 if (isMaybeZeroSizedType(ElTy))
1404 return -2; // dunno.
1406 // If they are really different, now that they are the same type, then we
1407 // found a difference!
1408 if (cast<ConstantInt>(C1)->getSExtValue() <
1409 cast<ConstantInt>(C2)->getSExtValue())
1415 /// evaluateFCmpRelation - This function determines if there is anything we can
1416 /// decide about the two constants provided. This doesn't need to handle simple
1417 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1418 /// If we can determine that the two constants have a particular relation to
1419 /// each other, we should return the corresponding FCmpInst predicate,
1420 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1421 /// ConstantFoldCompareInstruction.
1423 /// To simplify this code we canonicalize the relation so that the first
1424 /// operand is always the most "complex" of the two. We consider ConstantFP
1425 /// to be the simplest, and ConstantExprs to be the most complex.
1426 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1427 assert(V1->getType() == V2->getType() &&
1428 "Cannot compare values of different types!");
1430 // No compile-time operations on this type yet.
1431 if (V1->getType()->isPPC_FP128Ty())
1432 return FCmpInst::BAD_FCMP_PREDICATE;
1434 // Handle degenerate case quickly
1435 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1437 if (!isa<ConstantExpr>(V1)) {
1438 if (!isa<ConstantExpr>(V2)) {
1439 // We distilled thisUse the standard constant folder for a few cases
1441 R = dyn_cast<ConstantInt>(
1442 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1443 if (R && !R->isZero())
1444 return FCmpInst::FCMP_OEQ;
1445 R = dyn_cast<ConstantInt>(
1446 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1447 if (R && !R->isZero())
1448 return FCmpInst::FCMP_OLT;
1449 R = dyn_cast<ConstantInt>(
1450 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1451 if (R && !R->isZero())
1452 return FCmpInst::FCMP_OGT;
1454 // Nothing more we can do
1455 return FCmpInst::BAD_FCMP_PREDICATE;
1458 // If the first operand is simple and second is ConstantExpr, swap operands.
1459 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1460 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1461 return FCmpInst::getSwappedPredicate(SwappedRelation);
1463 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1464 // constantexpr or a simple constant.
1465 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1466 switch (CE1->getOpcode()) {
1467 case Instruction::FPTrunc:
1468 case Instruction::FPExt:
1469 case Instruction::UIToFP:
1470 case Instruction::SIToFP:
1471 // We might be able to do something with these but we don't right now.
1477 // There are MANY other foldings that we could perform here. They will
1478 // probably be added on demand, as they seem needed.
1479 return FCmpInst::BAD_FCMP_PREDICATE;
1482 /// evaluateICmpRelation - This function determines if there is anything we can
1483 /// decide about the two constants provided. This doesn't need to handle simple
1484 /// things like integer comparisons, but should instead handle ConstantExprs
1485 /// and GlobalValues. If we can determine that the two constants have a
1486 /// particular relation to each other, we should return the corresponding ICmp
1487 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1489 /// To simplify this code we canonicalize the relation so that the first
1490 /// operand is always the most "complex" of the two. We consider simple
1491 /// constants (like ConstantInt) to be the simplest, followed by
1492 /// GlobalValues, followed by ConstantExpr's (the most complex).
1494 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1496 assert(V1->getType() == V2->getType() &&
1497 "Cannot compare different types of values!");
1498 if (V1 == V2) return ICmpInst::ICMP_EQ;
1500 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1501 !isa<BlockAddress>(V1)) {
1502 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1503 !isa<BlockAddress>(V2)) {
1504 // We distilled this down to a simple case, use the standard constant
1507 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1508 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1509 if (R && !R->isZero())
1511 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1512 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1513 if (R && !R->isZero())
1515 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1516 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1517 if (R && !R->isZero())
1520 // If we couldn't figure it out, bail.
1521 return ICmpInst::BAD_ICMP_PREDICATE;
1524 // If the first operand is simple, swap operands.
1525 ICmpInst::Predicate SwappedRelation =
1526 evaluateICmpRelation(V2, V1, isSigned);
1527 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1528 return ICmpInst::getSwappedPredicate(SwappedRelation);
1530 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1531 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1532 ICmpInst::Predicate SwappedRelation =
1533 evaluateICmpRelation(V2, V1, isSigned);
1534 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1535 return ICmpInst::getSwappedPredicate(SwappedRelation);
1536 return ICmpInst::BAD_ICMP_PREDICATE;
1539 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1540 // constant (which, since the types must match, means that it's a
1541 // ConstantPointerNull).
1542 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1543 // Don't try to decide equality of aliases.
1544 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1545 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1546 return ICmpInst::ICMP_NE;
1547 } else if (isa<BlockAddress>(V2)) {
1548 return ICmpInst::ICMP_NE; // Globals never equal labels.
1550 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1551 // GlobalVals can never be null unless they have external weak linkage.
1552 // We don't try to evaluate aliases here.
1553 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1554 return ICmpInst::ICMP_NE;
1556 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1557 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1558 ICmpInst::Predicate SwappedRelation =
1559 evaluateICmpRelation(V2, V1, isSigned);
1560 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1561 return ICmpInst::getSwappedPredicate(SwappedRelation);
1562 return ICmpInst::BAD_ICMP_PREDICATE;
1565 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1566 // constant (which, since the types must match, means that it is a
1567 // ConstantPointerNull).
1568 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1569 // Block address in another function can't equal this one, but block
1570 // addresses in the current function might be the same if blocks are
1572 if (BA2->getFunction() != BA->getFunction())
1573 return ICmpInst::ICMP_NE;
1575 // Block addresses aren't null, don't equal the address of globals.
1576 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1577 "Canonicalization guarantee!");
1578 return ICmpInst::ICMP_NE;
1581 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1582 // constantexpr, a global, block address, or a simple constant.
1583 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1584 Constant *CE1Op0 = CE1->getOperand(0);
1586 switch (CE1->getOpcode()) {
1587 case Instruction::Trunc:
1588 case Instruction::FPTrunc:
1589 case Instruction::FPExt:
1590 case Instruction::FPToUI:
1591 case Instruction::FPToSI:
1592 break; // We can't evaluate floating point casts or truncations.
1594 case Instruction::UIToFP:
1595 case Instruction::SIToFP:
1596 case Instruction::BitCast:
1597 case Instruction::ZExt:
1598 case Instruction::SExt:
1599 // If the cast is not actually changing bits, and the second operand is a
1600 // null pointer, do the comparison with the pre-casted value.
1601 if (V2->isNullValue() &&
1602 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
1603 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1604 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1605 return evaluateICmpRelation(CE1Op0,
1606 Constant::getNullValue(CE1Op0->getType()),
1611 case Instruction::GetElementPtr:
1612 // Ok, since this is a getelementptr, we know that the constant has a
1613 // pointer type. Check the various cases.
1614 if (isa<ConstantPointerNull>(V2)) {
1615 // If we are comparing a GEP to a null pointer, check to see if the base
1616 // of the GEP equals the null pointer.
1617 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1618 if (GV->hasExternalWeakLinkage())
1619 // Weak linkage GVals could be zero or not. We're comparing that
1620 // to null pointer so its greater-or-equal
1621 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1623 // If its not weak linkage, the GVal must have a non-zero address
1624 // so the result is greater-than
1625 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1626 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1627 // If we are indexing from a null pointer, check to see if we have any
1628 // non-zero indices.
1629 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1630 if (!CE1->getOperand(i)->isNullValue())
1631 // Offsetting from null, must not be equal.
1632 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1633 // Only zero indexes from null, must still be zero.
1634 return ICmpInst::ICMP_EQ;
1636 // Otherwise, we can't really say if the first operand is null or not.
1637 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1638 if (isa<ConstantPointerNull>(CE1Op0)) {
1639 if (GV2->hasExternalWeakLinkage())
1640 // Weak linkage GVals could be zero or not. We're comparing it to
1641 // a null pointer, so its less-or-equal
1642 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1644 // If its not weak linkage, the GVal must have a non-zero address
1645 // so the result is less-than
1646 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1647 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1649 // If this is a getelementptr of the same global, then it must be
1650 // different. Because the types must match, the getelementptr could
1651 // only have at most one index, and because we fold getelementptr's
1652 // with a single zero index, it must be nonzero.
1653 assert(CE1->getNumOperands() == 2 &&
1654 !CE1->getOperand(1)->isNullValue() &&
1655 "Suprising getelementptr!");
1656 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1658 // If they are different globals, we don't know what the value is,
1659 // but they can't be equal.
1660 return ICmpInst::ICMP_NE;
1664 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1665 Constant *CE2Op0 = CE2->getOperand(0);
1667 // There are MANY other foldings that we could perform here. They will
1668 // probably be added on demand, as they seem needed.
1669 switch (CE2->getOpcode()) {
1671 case Instruction::GetElementPtr:
1672 // By far the most common case to handle is when the base pointers are
1673 // obviously to the same or different globals.
1674 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1675 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1676 return ICmpInst::ICMP_NE;
1677 // Ok, we know that both getelementptr instructions are based on the
1678 // same global. From this, we can precisely determine the relative
1679 // ordering of the resultant pointers.
1682 // The logic below assumes that the result of the comparison
1683 // can be determined by finding the first index that differs.
1684 // This doesn't work if there is over-indexing in any
1685 // subsequent indices, so check for that case first.
1686 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1687 !CE2->isGEPWithNoNotionalOverIndexing())
1688 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1690 // Compare all of the operands the GEP's have in common.
1691 gep_type_iterator GTI = gep_type_begin(CE1);
1692 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1694 switch (IdxCompare(CE1->getOperand(i),
1695 CE2->getOperand(i), GTI.getIndexedType())) {
1696 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1697 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1698 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1701 // Ok, we ran out of things they have in common. If any leftovers
1702 // are non-zero then we have a difference, otherwise we are equal.
1703 for (; i < CE1->getNumOperands(); ++i)
1704 if (!CE1->getOperand(i)->isNullValue()) {
1705 if (isa<ConstantInt>(CE1->getOperand(i)))
1706 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1708 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1711 for (; i < CE2->getNumOperands(); ++i)
1712 if (!CE2->getOperand(i)->isNullValue()) {
1713 if (isa<ConstantInt>(CE2->getOperand(i)))
1714 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1716 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1718 return ICmpInst::ICMP_EQ;
1727 return ICmpInst::BAD_ICMP_PREDICATE;
1730 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1731 Constant *C1, Constant *C2) {
1732 const Type *ResultTy;
1733 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1734 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1735 VT->getNumElements());
1737 ResultTy = Type::getInt1Ty(C1->getContext());
1739 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1740 if (pred == FCmpInst::FCMP_FALSE)
1741 return Constant::getNullValue(ResultTy);
1743 if (pred == FCmpInst::FCMP_TRUE)
1744 return Constant::getAllOnesValue(ResultTy);
1746 // Handle some degenerate cases first
1747 if (isa<UndefValue>(C1) || isa<UndefValue>(C2))
1748 return UndefValue::get(ResultTy);
1750 // No compile-time operations on this type yet.
1751 if (C1->getType()->isPPC_FP128Ty())
1754 // icmp eq/ne(null,GV) -> false/true
1755 if (C1->isNullValue()) {
1756 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1757 // Don't try to evaluate aliases. External weak GV can be null.
1758 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1759 if (pred == ICmpInst::ICMP_EQ)
1760 return ConstantInt::getFalse(C1->getContext());
1761 else if (pred == ICmpInst::ICMP_NE)
1762 return ConstantInt::getTrue(C1->getContext());
1764 // icmp eq/ne(GV,null) -> false/true
1765 } else if (C2->isNullValue()) {
1766 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1767 // Don't try to evaluate aliases. External weak GV can be null.
1768 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1769 if (pred == ICmpInst::ICMP_EQ)
1770 return ConstantInt::getFalse(C1->getContext());
1771 else if (pred == ICmpInst::ICMP_NE)
1772 return ConstantInt::getTrue(C1->getContext());
1776 // If the comparison is a comparison between two i1's, simplify it.
1777 if (C1->getType()->isInteger(1)) {
1779 case ICmpInst::ICMP_EQ:
1780 if (isa<ConstantInt>(C2))
1781 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1782 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1783 case ICmpInst::ICMP_NE:
1784 return ConstantExpr::getXor(C1, C2);
1790 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1791 APInt V1 = cast<ConstantInt>(C1)->getValue();
1792 APInt V2 = cast<ConstantInt>(C2)->getValue();
1794 default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1795 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1796 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1797 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1798 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1799 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1800 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1801 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1802 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1803 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1804 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1806 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1807 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1808 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1809 APFloat::cmpResult R = C1V.compare(C2V);
1811 default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1812 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1813 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1814 case FCmpInst::FCMP_UNO:
1815 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1816 case FCmpInst::FCMP_ORD:
1817 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1818 case FCmpInst::FCMP_UEQ:
1819 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1820 R==APFloat::cmpEqual);
1821 case FCmpInst::FCMP_OEQ:
1822 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1823 case FCmpInst::FCMP_UNE:
1824 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1825 case FCmpInst::FCMP_ONE:
1826 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1827 R==APFloat::cmpGreaterThan);
1828 case FCmpInst::FCMP_ULT:
1829 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1830 R==APFloat::cmpLessThan);
1831 case FCmpInst::FCMP_OLT:
1832 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1833 case FCmpInst::FCMP_UGT:
1834 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1835 R==APFloat::cmpGreaterThan);
1836 case FCmpInst::FCMP_OGT:
1837 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1838 case FCmpInst::FCMP_ULE:
1839 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1840 case FCmpInst::FCMP_OLE:
1841 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1842 R==APFloat::cmpEqual);
1843 case FCmpInst::FCMP_UGE:
1844 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1845 case FCmpInst::FCMP_OGE:
1846 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1847 R==APFloat::cmpEqual);
1849 } else if (isa<VectorType>(C1->getType())) {
1850 SmallVector<Constant*, 16> C1Elts, C2Elts;
1851 C1->getVectorElements(C1Elts);
1852 C2->getVectorElements(C2Elts);
1853 if (C1Elts.empty() || C2Elts.empty())
1856 // If we can constant fold the comparison of each element, constant fold
1857 // the whole vector comparison.
1858 SmallVector<Constant*, 4> ResElts;
1859 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1860 // Compare the elements, producing an i1 result or constant expr.
1861 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1863 return ConstantVector::get(&ResElts[0], ResElts.size());
1866 if (C1->getType()->isFloatingPoint()) {
1867 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1868 switch (evaluateFCmpRelation(C1, C2)) {
1869 default: llvm_unreachable("Unknown relation!");
1870 case FCmpInst::FCMP_UNO:
1871 case FCmpInst::FCMP_ORD:
1872 case FCmpInst::FCMP_UEQ:
1873 case FCmpInst::FCMP_UNE:
1874 case FCmpInst::FCMP_ULT:
1875 case FCmpInst::FCMP_UGT:
1876 case FCmpInst::FCMP_ULE:
1877 case FCmpInst::FCMP_UGE:
1878 case FCmpInst::FCMP_TRUE:
1879 case FCmpInst::FCMP_FALSE:
1880 case FCmpInst::BAD_FCMP_PREDICATE:
1881 break; // Couldn't determine anything about these constants.
1882 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1883 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1884 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1885 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1887 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1888 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1889 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1890 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1892 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1893 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1894 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1895 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1897 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1898 // We can only partially decide this relation.
1899 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1901 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1904 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1905 // We can only partially decide this relation.
1906 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1908 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1911 case ICmpInst::ICMP_NE: // We know that C1 != C2
1912 // We can only partially decide this relation.
1913 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1915 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1920 // If we evaluated the result, return it now.
1922 return ConstantInt::get(ResultTy, Result);
1925 // Evaluate the relation between the two constants, per the predicate.
1926 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1927 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1928 default: llvm_unreachable("Unknown relational!");
1929 case ICmpInst::BAD_ICMP_PREDICATE:
1930 break; // Couldn't determine anything about these constants.
1931 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1932 // If we know the constants are equal, we can decide the result of this
1933 // computation precisely.
1934 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1936 case ICmpInst::ICMP_ULT:
1938 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1940 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1944 case ICmpInst::ICMP_SLT:
1946 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1948 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1952 case ICmpInst::ICMP_UGT:
1954 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1956 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1960 case ICmpInst::ICMP_SGT:
1962 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1964 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1968 case ICmpInst::ICMP_ULE:
1969 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1970 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1972 case ICmpInst::ICMP_SLE:
1973 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1974 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1976 case ICmpInst::ICMP_UGE:
1977 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1978 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1980 case ICmpInst::ICMP_SGE:
1981 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1982 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1984 case ICmpInst::ICMP_NE:
1985 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1986 if (pred == ICmpInst::ICMP_NE) Result = 1;
1990 // If we evaluated the result, return it now.
1992 return ConstantInt::get(ResultTy, Result);
1994 // If the right hand side is a bitcast, try using its inverse to simplify
1995 // it by moving it to the left hand side. We can't do this if it would turn
1996 // a vector compare into a scalar compare or visa versa.
1997 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1998 Constant *CE2Op0 = CE2->getOperand(0);
1999 if (CE2->getOpcode() == Instruction::BitCast &&
2000 isa<VectorType>(CE2->getType())==isa<VectorType>(CE2Op0->getType())) {
2001 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2002 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2006 // If the left hand side is an extension, try eliminating it.
2007 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2008 if (CE1->getOpcode() == Instruction::SExt ||
2009 CE1->getOpcode() == Instruction::ZExt) {
2010 Constant *CE1Op0 = CE1->getOperand(0);
2011 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2012 if (CE1Inverse == CE1Op0) {
2013 // Check whether we can safely truncate the right hand side.
2014 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2015 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2016 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2022 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2023 (C1->isNullValue() && !C2->isNullValue())) {
2024 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2025 // other way if possible.
2026 // Also, if C1 is null and C2 isn't, flip them around.
2028 case ICmpInst::ICMP_EQ:
2029 case ICmpInst::ICMP_NE:
2030 // No change of predicate required.
2031 return ConstantExpr::getICmp(pred, C2, C1);
2033 case ICmpInst::ICMP_ULT:
2034 case ICmpInst::ICMP_SLT:
2035 case ICmpInst::ICMP_UGT:
2036 case ICmpInst::ICMP_SGT:
2037 case ICmpInst::ICMP_ULE:
2038 case ICmpInst::ICMP_SLE:
2039 case ICmpInst::ICMP_UGE:
2040 case ICmpInst::ICMP_SGE:
2041 // Change the predicate as necessary to swap the operands.
2042 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2043 return ConstantExpr::getICmp(pred, C2, C1);
2045 default: // These predicates cannot be flopped around.
2053 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2055 static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) {
2056 // No indices means nothing that could be out of bounds.
2057 if (NumIdx == 0) return true;
2059 // If the first index is zero, it's in bounds.
2060 if (Idxs[0]->isNullValue()) return true;
2062 // If the first index is one and all the rest are zero, it's in bounds,
2063 // by the one-past-the-end rule.
2064 if (!cast<ConstantInt>(Idxs[0])->isOne())
2066 for (unsigned i = 1, e = NumIdx; i != e; ++i)
2067 if (!Idxs[i]->isNullValue())
2072 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2074 Constant* const *Idxs,
2077 (NumIdx == 1 && Idxs[0]->isNullValue()))
2080 if (isa<UndefValue>(C)) {
2081 const PointerType *Ptr = cast<PointerType>(C->getType());
2082 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
2084 (Value **)Idxs+NumIdx);
2085 assert(Ty != 0 && "Invalid indices for GEP!");
2086 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2089 Constant *Idx0 = Idxs[0];
2090 if (C->isNullValue()) {
2092 for (unsigned i = 0, e = NumIdx; i != e; ++i)
2093 if (!Idxs[i]->isNullValue()) {
2098 const PointerType *Ptr = cast<PointerType>(C->getType());
2099 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
2101 (Value**)Idxs+NumIdx);
2102 assert(Ty != 0 && "Invalid indices for GEP!");
2103 return ConstantPointerNull::get(
2104 PointerType::get(Ty,Ptr->getAddressSpace()));
2108 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2109 // Combine Indices - If the source pointer to this getelementptr instruction
2110 // is a getelementptr instruction, combine the indices of the two
2111 // getelementptr instructions into a single instruction.
2113 if (CE->getOpcode() == Instruction::GetElementPtr) {
2114 const Type *LastTy = 0;
2115 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2119 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
2120 SmallVector<Value*, 16> NewIndices;
2121 NewIndices.reserve(NumIdx + CE->getNumOperands());
2122 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2123 NewIndices.push_back(CE->getOperand(i));
2125 // Add the last index of the source with the first index of the new GEP.
2126 // Make sure to handle the case when they are actually different types.
2127 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2128 // Otherwise it must be an array.
2129 if (!Idx0->isNullValue()) {
2130 const Type *IdxTy = Combined->getType();
2131 if (IdxTy != Idx0->getType()) {
2132 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2133 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2134 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2135 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2138 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2142 NewIndices.push_back(Combined);
2143 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
2144 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ?
2145 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0),
2147 NewIndices.size()) :
2148 ConstantExpr::getGetElementPtr(CE->getOperand(0),
2154 // Implement folding of:
2155 // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*),
2157 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
2159 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
2160 if (const PointerType *SPT =
2161 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2162 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2163 if (const ArrayType *CAT =
2164 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2165 if (CAT->getElementType() == SAT->getElementType())
2167 ConstantExpr::getInBoundsGetElementPtr(
2168 (Constant*)CE->getOperand(0), Idxs, NumIdx) :
2169 ConstantExpr::getGetElementPtr(
2170 (Constant*)CE->getOperand(0), Idxs, NumIdx);
2174 // Check to see if any array indices are not within the corresponding
2175 // notional array bounds. If so, try to determine if they can be factored
2176 // out into preceding dimensions.
2177 bool Unknown = false;
2178 SmallVector<Constant *, 8> NewIdxs;
2179 const Type *Ty = C->getType();
2180 const Type *Prev = 0;
2181 for (unsigned i = 0; i != NumIdx;
2182 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2183 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2184 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2185 if (ATy->getNumElements() <= INT64_MAX &&
2186 ATy->getNumElements() != 0 &&
2187 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2188 if (isa<SequentialType>(Prev)) {
2189 // It's out of range, but we can factor it into the prior
2191 NewIdxs.resize(NumIdx);
2192 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2193 ATy->getNumElements());
2194 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2196 Constant *PrevIdx = Idxs[i-1];
2197 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2199 // Before adding, extend both operands to i64 to avoid
2200 // overflow trouble.
2201 if (!PrevIdx->getType()->isInteger(64))
2202 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2203 Type::getInt64Ty(Div->getContext()));
2204 if (!Div->getType()->isInteger(64))
2205 Div = ConstantExpr::getSExt(Div,
2206 Type::getInt64Ty(Div->getContext()));
2208 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2210 // It's out of range, but the prior dimension is a struct
2211 // so we can't do anything about it.
2216 // We don't know if it's in range or not.
2221 // If we did any factoring, start over with the adjusted indices.
2222 if (!NewIdxs.empty()) {
2223 for (unsigned i = 0; i != NumIdx; ++i)
2224 if (!NewIdxs[i]) NewIdxs[i] = Idxs[i];
2226 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(),
2228 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size());
2231 // If all indices are known integers and normalized, we can do a simple
2232 // check for the "inbounds" property.
2233 if (!Unknown && !inBounds &&
2234 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx))
2235 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx);