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 size.
349 Constant *MemberSize =
350 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
352 for (unsigned i = 1; i != NumElems; ++i)
354 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
359 Constant *N = ConstantInt::get(DestTy, NumElems);
360 return ConstantExpr::getNUWMul(MemberSize, N);
364 // If there's no interesting folding happening, bail so that we don't create
365 // a constant that looks like it needs folding but really doesn't.
369 // Base case: Get a regular sizeof expression.
370 Constant *C = ConstantExpr::getSizeOf(Ty);
371 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
377 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
378 /// on Ty, with any known factors factored out. If Folded is false,
379 /// return null if no factoring was possible, to avoid endlessly
380 /// bouncing an unfoldable expression back into the top-level folder.
382 static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy,
384 // The alignment of an array is equal to the alignment of the
385 // array element. Note that this is not always true for vectors.
386 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
387 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
388 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
395 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
396 // Packed structs always have an alignment of 1.
398 return ConstantInt::get(DestTy, 1);
400 // Otherwise, struct alignment is the maximum alignment of any member.
401 // Without target data, we can't compare much, but we can check to see
402 // if all the members have the same alignment.
403 unsigned NumElems = STy->getNumElements();
404 // An empty struct has minimal alignment.
406 return ConstantInt::get(DestTy, 1);
407 // Check for a struct with all members having the same alignment.
408 Constant *MemberAlign =
409 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
411 for (unsigned i = 1; i != NumElems; ++i)
412 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
420 // If there's no interesting folding happening, bail so that we don't create
421 // a constant that looks like it needs folding but really doesn't.
425 // Base case: Get a regular alignof expression.
426 Constant *C = ConstantExpr::getAlignOf(Ty);
427 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
433 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
434 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
435 /// return null if no factoring was possible, to avoid endlessly
436 /// bouncing an unfoldable expression back into the top-level folder.
438 static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo,
441 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
442 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
445 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
446 return ConstantExpr::getNUWMul(E, N);
448 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
449 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
452 Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true);
453 return ConstantExpr::getNUWMul(E, N);
455 if (const StructType *STy = dyn_cast<StructType>(Ty))
456 if (!STy->isPacked()) {
457 unsigned NumElems = STy->getNumElements();
458 // An empty struct has no members.
461 // Check for a struct with all members having the same size.
462 Constant *MemberSize =
463 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
465 for (unsigned i = 1; i != NumElems; ++i)
467 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
472 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
477 return ConstantExpr::getNUWMul(MemberSize, N);
481 // If there's no interesting folding happening, bail so that we don't create
482 // a constant that looks like it needs folding but really doesn't.
486 // Base case: Get a regular offsetof expression.
487 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
488 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
494 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
495 const Type *DestTy) {
496 if (isa<UndefValue>(V)) {
497 // zext(undef) = 0, because the top bits will be zero.
498 // sext(undef) = 0, because the top bits will all be the same.
499 // [us]itofp(undef) = 0, because the result value is bounded.
500 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
501 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
502 return Constant::getNullValue(DestTy);
503 return UndefValue::get(DestTy);
505 // No compile-time operations on this type yet.
506 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
509 // If the cast operand is a constant expression, there's a few things we can
510 // do to try to simplify it.
511 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
513 // Try hard to fold cast of cast because they are often eliminable.
514 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
515 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
516 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
517 // If all of the indexes in the GEP are null values, there is no pointer
518 // adjustment going on. We might as well cast the source pointer.
519 bool isAllNull = true;
520 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
521 if (!CE->getOperand(i)->isNullValue()) {
526 // This is casting one pointer type to another, always BitCast
527 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
531 // If the cast operand is a constant vector, perform the cast by
532 // operating on each element. In the cast of bitcasts, the element
533 // count may be mismatched; don't attempt to handle that here.
534 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
535 if (isa<VectorType>(DestTy) &&
536 cast<VectorType>(DestTy)->getNumElements() ==
537 CV->getType()->getNumElements()) {
538 std::vector<Constant*> res;
539 const VectorType *DestVecTy = cast<VectorType>(DestTy);
540 const Type *DstEltTy = DestVecTy->getElementType();
541 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
542 res.push_back(ConstantExpr::getCast(opc,
543 CV->getOperand(i), DstEltTy));
544 return ConstantVector::get(DestVecTy, res);
547 // We actually have to do a cast now. Perform the cast according to the
551 llvm_unreachable("Failed to cast constant expression");
552 case Instruction::FPTrunc:
553 case Instruction::FPExt:
554 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
556 APFloat Val = FPC->getValueAPF();
557 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
558 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
559 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
560 DestTy->isFP128Ty() ? APFloat::IEEEquad :
562 APFloat::rmNearestTiesToEven, &ignored);
563 return ConstantFP::get(V->getContext(), Val);
565 return 0; // Can't fold.
566 case Instruction::FPToUI:
567 case Instruction::FPToSI:
568 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
569 const APFloat &V = FPC->getValueAPF();
572 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
573 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
574 APFloat::rmTowardZero, &ignored);
575 APInt Val(DestBitWidth, 2, x);
576 return ConstantInt::get(FPC->getContext(), Val);
578 return 0; // Can't fold.
579 case Instruction::IntToPtr: //always treated as unsigned
580 if (V->isNullValue()) // Is it an integral null value?
581 return ConstantPointerNull::get(cast<PointerType>(DestTy));
582 return 0; // Other pointer types cannot be casted
583 case Instruction::PtrToInt: // always treated as unsigned
584 // Is it a null pointer value?
585 if (V->isNullValue())
586 return ConstantInt::get(DestTy, 0);
587 // If this is a sizeof-like expression, pull out multiplications by
588 // known factors to expose them to subsequent folding. If it's an
589 // alignof-like expression, factor out known factors.
590 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
591 if (CE->getOpcode() == Instruction::GetElementPtr &&
592 CE->getOperand(0)->isNullValue()) {
594 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
595 if (CE->getNumOperands() == 2) {
596 // Handle a sizeof-like expression.
597 Constant *Idx = CE->getOperand(1);
598 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
599 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
600 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
603 return ConstantExpr::getMul(C, Idx);
605 } else if (CE->getNumOperands() == 3 &&
606 CE->getOperand(1)->isNullValue()) {
607 // Handle an alignof-like expression.
608 if (const StructType *STy = dyn_cast<StructType>(Ty))
609 if (!STy->isPacked()) {
610 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
612 STy->getNumElements() == 2 &&
613 STy->getElementType(0)->isInteger(1)) {
614 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
617 // Handle an offsetof-like expression.
618 if (isa<StructType>(Ty) || isa<ArrayType>(Ty) || isa<VectorType>(Ty)){
619 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
625 // Other pointer types cannot be casted
627 case Instruction::UIToFP:
628 case Instruction::SIToFP:
629 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
630 APInt api = CI->getValue();
631 const uint64_t zero[] = {0, 0};
632 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(),
634 (void)apf.convertFromAPInt(api,
635 opc==Instruction::SIToFP,
636 APFloat::rmNearestTiesToEven);
637 return ConstantFP::get(V->getContext(), apf);
640 case Instruction::ZExt:
641 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
642 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
643 APInt Result(CI->getValue());
644 Result.zext(BitWidth);
645 return ConstantInt::get(V->getContext(), Result);
648 case Instruction::SExt:
649 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
650 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
651 APInt Result(CI->getValue());
652 Result.sext(BitWidth);
653 return ConstantInt::get(V->getContext(), Result);
656 case Instruction::Trunc: {
657 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
658 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
659 APInt Result(CI->getValue());
660 Result.trunc(DestBitWidth);
661 return ConstantInt::get(V->getContext(), Result);
664 // The input must be a constantexpr. See if we can simplify this based on
665 // the bytes we are demanding. Only do this if the source and dest are an
666 // even multiple of a byte.
667 if ((DestBitWidth & 7) == 0 &&
668 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
669 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
674 case Instruction::BitCast:
675 return FoldBitCast(V, DestTy);
679 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
680 Constant *V1, Constant *V2) {
681 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
682 return CB->getZExtValue() ? V1 : V2;
684 if (isa<UndefValue>(V1)) return V2;
685 if (isa<UndefValue>(V2)) return V1;
686 if (isa<UndefValue>(Cond)) return V1;
687 if (V1 == V2) return V1;
691 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
693 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
694 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
695 if (Val->isNullValue()) // ee(zero, x) -> zero
696 return Constant::getNullValue(
697 cast<VectorType>(Val->getType())->getElementType());
699 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
700 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
701 return CVal->getOperand(CIdx->getZExtValue());
702 } else if (isa<UndefValue>(Idx)) {
703 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
704 return CVal->getOperand(0);
710 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
713 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
715 APInt idxVal = CIdx->getValue();
716 if (isa<UndefValue>(Val)) {
717 // Insertion of scalar constant into vector undef
718 // Optimize away insertion of undef
719 if (isa<UndefValue>(Elt))
721 // Otherwise break the aggregate undef into multiple undefs and do
724 cast<VectorType>(Val->getType())->getNumElements();
725 std::vector<Constant*> Ops;
727 for (unsigned i = 0; i < numOps; ++i) {
729 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
732 return ConstantVector::get(Ops);
734 if (isa<ConstantAggregateZero>(Val)) {
735 // Insertion of scalar constant into vector aggregate zero
736 // Optimize away insertion of zero
737 if (Elt->isNullValue())
739 // Otherwise break the aggregate zero into multiple zeros and do
742 cast<VectorType>(Val->getType())->getNumElements();
743 std::vector<Constant*> Ops;
745 for (unsigned i = 0; i < numOps; ++i) {
747 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
750 return ConstantVector::get(Ops);
752 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
753 // Insertion of scalar constant into vector constant
754 std::vector<Constant*> Ops;
755 Ops.reserve(CVal->getNumOperands());
756 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
758 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
761 return ConstantVector::get(Ops);
767 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
768 /// return the specified element value. Otherwise return null.
769 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
770 if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
771 return CV->getOperand(EltNo);
773 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
774 if (isa<ConstantAggregateZero>(C))
775 return Constant::getNullValue(EltTy);
776 if (isa<UndefValue>(C))
777 return UndefValue::get(EltTy);
781 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
784 // Undefined shuffle mask -> undefined value.
785 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
787 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
788 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
789 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
791 // Loop over the shuffle mask, evaluating each element.
792 SmallVector<Constant*, 32> Result;
793 for (unsigned i = 0; i != MaskNumElts; ++i) {
794 Constant *InElt = GetVectorElement(Mask, i);
795 if (InElt == 0) return 0;
797 if (isa<UndefValue>(InElt))
798 InElt = UndefValue::get(EltTy);
799 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
800 unsigned Elt = CI->getZExtValue();
801 if (Elt >= SrcNumElts*2)
802 InElt = UndefValue::get(EltTy);
803 else if (Elt >= SrcNumElts)
804 InElt = GetVectorElement(V2, Elt - SrcNumElts);
806 InElt = GetVectorElement(V1, Elt);
807 if (InElt == 0) return 0;
812 Result.push_back(InElt);
815 return ConstantVector::get(&Result[0], Result.size());
818 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
819 const unsigned *Idxs,
821 // Base case: no indices, so return the entire value.
825 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
826 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
830 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
832 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
836 // Otherwise recurse.
837 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
838 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs),
841 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
842 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs),
844 ConstantVector *CV = cast<ConstantVector>(Agg);
845 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs),
849 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
851 const unsigned *Idxs,
853 // Base case: no indices, so replace the entire value.
857 if (isa<UndefValue>(Agg)) {
858 // Insertion of constant into aggregate undef
859 // Optimize away insertion of undef.
860 if (isa<UndefValue>(Val))
863 // Otherwise break the aggregate undef into multiple undefs and do
865 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
867 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
868 numOps = AR->getNumElements();
870 numOps = cast<StructType>(AggTy)->getNumElements();
872 std::vector<Constant*> Ops(numOps);
873 for (unsigned i = 0; i < numOps; ++i) {
874 const Type *MemberTy = AggTy->getTypeAtIndex(i);
877 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
878 Val, Idxs+1, NumIdx-1) :
879 UndefValue::get(MemberTy);
883 if (const StructType* ST = dyn_cast<StructType>(AggTy))
884 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
885 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
888 if (isa<ConstantAggregateZero>(Agg)) {
889 // Insertion of constant into aggregate zero
890 // Optimize away insertion of zero.
891 if (Val->isNullValue())
894 // Otherwise break the aggregate zero into multiple zeros and do
896 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
898 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
899 numOps = AR->getNumElements();
901 numOps = cast<StructType>(AggTy)->getNumElements();
903 std::vector<Constant*> Ops(numOps);
904 for (unsigned i = 0; i < numOps; ++i) {
905 const Type *MemberTy = AggTy->getTypeAtIndex(i);
908 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
909 Val, Idxs+1, NumIdx-1) :
910 Constant::getNullValue(MemberTy);
914 if (const StructType *ST = dyn_cast<StructType>(AggTy))
915 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
916 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
919 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
920 // Insertion of constant into aggregate constant.
921 std::vector<Constant*> Ops(Agg->getNumOperands());
922 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
923 Constant *Op = cast<Constant>(Agg->getOperand(i));
925 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1);
929 if (const StructType* ST = dyn_cast<StructType>(Agg->getType()))
930 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
931 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
938 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
939 Constant *C1, Constant *C2) {
940 // No compile-time operations on this type yet.
941 if (C1->getType()->isPPC_FP128Ty())
944 // Handle UndefValue up front.
945 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
947 case Instruction::Xor:
948 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
949 // Handle undef ^ undef -> 0 special case. This is a common
951 return Constant::getNullValue(C1->getType());
953 case Instruction::Add:
954 case Instruction::Sub:
955 return UndefValue::get(C1->getType());
956 case Instruction::Mul:
957 case Instruction::And:
958 return Constant::getNullValue(C1->getType());
959 case Instruction::UDiv:
960 case Instruction::SDiv:
961 case Instruction::URem:
962 case Instruction::SRem:
963 if (!isa<UndefValue>(C2)) // undef / X -> 0
964 return Constant::getNullValue(C1->getType());
965 return C2; // X / undef -> undef
966 case Instruction::Or: // X | undef -> -1
967 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
968 return Constant::getAllOnesValue(PTy);
969 return Constant::getAllOnesValue(C1->getType());
970 case Instruction::LShr:
971 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
972 return C1; // undef lshr undef -> undef
973 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
975 case Instruction::AShr:
976 if (!isa<UndefValue>(C2))
977 return C1; // undef ashr X --> undef
978 else if (isa<UndefValue>(C1))
979 return C1; // undef ashr undef -> undef
981 return C1; // X ashr undef --> X
982 case Instruction::Shl:
983 // undef << X -> 0 or X << undef -> 0
984 return Constant::getNullValue(C1->getType());
988 // Handle simplifications when the RHS is a constant int.
989 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
991 case Instruction::Add:
992 if (CI2->equalsInt(0)) return C1; // X + 0 == X
994 case Instruction::Sub:
995 if (CI2->equalsInt(0)) return C1; // X - 0 == X
997 case Instruction::Mul:
998 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
999 if (CI2->equalsInt(1))
1000 return C1; // X * 1 == X
1002 case Instruction::UDiv:
1003 case Instruction::SDiv:
1004 if (CI2->equalsInt(1))
1005 return C1; // X / 1 == X
1006 if (CI2->equalsInt(0))
1007 return UndefValue::get(CI2->getType()); // X / 0 == undef
1009 case Instruction::URem:
1010 case Instruction::SRem:
1011 if (CI2->equalsInt(1))
1012 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1013 if (CI2->equalsInt(0))
1014 return UndefValue::get(CI2->getType()); // X % 0 == undef
1016 case Instruction::And:
1017 if (CI2->isZero()) return C2; // X & 0 == 0
1018 if (CI2->isAllOnesValue())
1019 return C1; // X & -1 == X
1021 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1022 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1023 if (CE1->getOpcode() == Instruction::ZExt) {
1024 unsigned DstWidth = CI2->getType()->getBitWidth();
1026 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1027 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1028 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1032 // If and'ing the address of a global with a constant, fold it.
1033 if (CE1->getOpcode() == Instruction::PtrToInt &&
1034 isa<GlobalValue>(CE1->getOperand(0))) {
1035 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1037 // Functions are at least 4-byte aligned.
1038 unsigned GVAlign = GV->getAlignment();
1039 if (isa<Function>(GV))
1040 GVAlign = std::max(GVAlign, 4U);
1043 unsigned DstWidth = CI2->getType()->getBitWidth();
1044 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1045 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1047 // If checking bits we know are clear, return zero.
1048 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1049 return Constant::getNullValue(CI2->getType());
1054 case Instruction::Or:
1055 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1056 if (CI2->isAllOnesValue())
1057 return C2; // X | -1 == -1
1059 case Instruction::Xor:
1060 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1062 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1063 switch (CE1->getOpcode()) {
1065 case Instruction::ICmp:
1066 case Instruction::FCmp:
1067 // cmp pred ^ true -> cmp !pred
1068 assert(CI2->equalsInt(1));
1069 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1070 pred = CmpInst::getInversePredicate(pred);
1071 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1072 CE1->getOperand(1));
1076 case Instruction::AShr:
1077 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1078 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1079 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1080 return ConstantExpr::getLShr(C1, C2);
1085 // At this point we know neither constant is an UndefValue.
1086 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1087 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1088 using namespace APIntOps;
1089 const APInt &C1V = CI1->getValue();
1090 const APInt &C2V = CI2->getValue();
1094 case Instruction::Add:
1095 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1096 case Instruction::Sub:
1097 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1098 case Instruction::Mul:
1099 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1100 case Instruction::UDiv:
1101 assert(!CI2->isNullValue() && "Div by zero handled above");
1102 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1103 case Instruction::SDiv:
1104 assert(!CI2->isNullValue() && "Div by zero handled above");
1105 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1106 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1107 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1108 case Instruction::URem:
1109 assert(!CI2->isNullValue() && "Div by zero handled above");
1110 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1111 case Instruction::SRem:
1112 assert(!CI2->isNullValue() && "Div by zero handled above");
1113 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1114 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1115 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1116 case Instruction::And:
1117 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1118 case Instruction::Or:
1119 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1120 case Instruction::Xor:
1121 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1122 case Instruction::Shl: {
1123 uint32_t shiftAmt = C2V.getZExtValue();
1124 if (shiftAmt < C1V.getBitWidth())
1125 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1127 return UndefValue::get(C1->getType()); // too big shift is undef
1129 case Instruction::LShr: {
1130 uint32_t shiftAmt = C2V.getZExtValue();
1131 if (shiftAmt < C1V.getBitWidth())
1132 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1134 return UndefValue::get(C1->getType()); // too big shift is undef
1136 case Instruction::AShr: {
1137 uint32_t shiftAmt = C2V.getZExtValue();
1138 if (shiftAmt < C1V.getBitWidth())
1139 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1141 return UndefValue::get(C1->getType()); // too big shift is undef
1147 case Instruction::SDiv:
1148 case Instruction::UDiv:
1149 case Instruction::URem:
1150 case Instruction::SRem:
1151 case Instruction::LShr:
1152 case Instruction::AShr:
1153 case Instruction::Shl:
1154 if (CI1->equalsInt(0)) return C1;
1159 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1160 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1161 APFloat C1V = CFP1->getValueAPF();
1162 APFloat C2V = CFP2->getValueAPF();
1163 APFloat C3V = C1V; // copy for modification
1167 case Instruction::FAdd:
1168 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1169 return ConstantFP::get(C1->getContext(), C3V);
1170 case Instruction::FSub:
1171 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1172 return ConstantFP::get(C1->getContext(), C3V);
1173 case Instruction::FMul:
1174 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1175 return ConstantFP::get(C1->getContext(), C3V);
1176 case Instruction::FDiv:
1177 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1178 return ConstantFP::get(C1->getContext(), C3V);
1179 case Instruction::FRem:
1180 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1181 return ConstantFP::get(C1->getContext(), C3V);
1184 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1185 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1186 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1187 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1188 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1189 std::vector<Constant*> Res;
1190 const Type* EltTy = VTy->getElementType();
1196 case Instruction::Add:
1197 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1198 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1199 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1200 Res.push_back(ConstantExpr::getAdd(C1, C2));
1202 return ConstantVector::get(Res);
1203 case Instruction::FAdd:
1204 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1205 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1206 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1207 Res.push_back(ConstantExpr::getFAdd(C1, C2));
1209 return ConstantVector::get(Res);
1210 case Instruction::Sub:
1211 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1212 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1213 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1214 Res.push_back(ConstantExpr::getSub(C1, C2));
1216 return ConstantVector::get(Res);
1217 case Instruction::FSub:
1218 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1219 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1220 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1221 Res.push_back(ConstantExpr::getFSub(C1, C2));
1223 return ConstantVector::get(Res);
1224 case Instruction::Mul:
1225 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1226 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1227 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1228 Res.push_back(ConstantExpr::getMul(C1, C2));
1230 return ConstantVector::get(Res);
1231 case Instruction::FMul:
1232 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1233 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1234 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1235 Res.push_back(ConstantExpr::getFMul(C1, C2));
1237 return ConstantVector::get(Res);
1238 case Instruction::UDiv:
1239 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1240 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1241 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1242 Res.push_back(ConstantExpr::getUDiv(C1, C2));
1244 return ConstantVector::get(Res);
1245 case Instruction::SDiv:
1246 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1247 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1248 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1249 Res.push_back(ConstantExpr::getSDiv(C1, C2));
1251 return ConstantVector::get(Res);
1252 case Instruction::FDiv:
1253 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1254 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1255 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1256 Res.push_back(ConstantExpr::getFDiv(C1, C2));
1258 return ConstantVector::get(Res);
1259 case Instruction::URem:
1260 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1261 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1262 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1263 Res.push_back(ConstantExpr::getURem(C1, C2));
1265 return ConstantVector::get(Res);
1266 case Instruction::SRem:
1267 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1268 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1269 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1270 Res.push_back(ConstantExpr::getSRem(C1, C2));
1272 return ConstantVector::get(Res);
1273 case Instruction::FRem:
1274 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1275 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1276 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1277 Res.push_back(ConstantExpr::getFRem(C1, C2));
1279 return ConstantVector::get(Res);
1280 case Instruction::And:
1281 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1282 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1283 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1284 Res.push_back(ConstantExpr::getAnd(C1, C2));
1286 return ConstantVector::get(Res);
1287 case Instruction::Or:
1288 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1289 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1290 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1291 Res.push_back(ConstantExpr::getOr(C1, C2));
1293 return ConstantVector::get(Res);
1294 case Instruction::Xor:
1295 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1296 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1297 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1298 Res.push_back(ConstantExpr::getXor(C1, C2));
1300 return ConstantVector::get(Res);
1301 case Instruction::LShr:
1302 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1303 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1304 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1305 Res.push_back(ConstantExpr::getLShr(C1, C2));
1307 return ConstantVector::get(Res);
1308 case Instruction::AShr:
1309 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1310 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1311 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1312 Res.push_back(ConstantExpr::getAShr(C1, C2));
1314 return ConstantVector::get(Res);
1315 case Instruction::Shl:
1316 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1317 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1318 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1319 Res.push_back(ConstantExpr::getShl(C1, C2));
1321 return ConstantVector::get(Res);
1326 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1327 // There are many possible foldings we could do here. We should probably
1328 // at least fold add of a pointer with an integer into the appropriate
1329 // getelementptr. This will improve alias analysis a bit.
1331 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1333 if (Instruction::isAssociative(Opcode, C1->getType()) &&
1334 CE1->getOpcode() == Opcode) {
1335 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1336 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1337 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1339 } else if (isa<ConstantExpr>(C2)) {
1340 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1341 // other way if possible.
1343 case Instruction::Add:
1344 case Instruction::FAdd:
1345 case Instruction::Mul:
1346 case Instruction::FMul:
1347 case Instruction::And:
1348 case Instruction::Or:
1349 case Instruction::Xor:
1350 // No change of opcode required.
1351 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1353 case Instruction::Shl:
1354 case Instruction::LShr:
1355 case Instruction::AShr:
1356 case Instruction::Sub:
1357 case Instruction::FSub:
1358 case Instruction::SDiv:
1359 case Instruction::UDiv:
1360 case Instruction::FDiv:
1361 case Instruction::URem:
1362 case Instruction::SRem:
1363 case Instruction::FRem:
1364 default: // These instructions cannot be flopped around.
1369 // i1 can be simplified in many cases.
1370 if (C1->getType()->isInteger(1)) {
1372 case Instruction::Add:
1373 case Instruction::Sub:
1374 return ConstantExpr::getXor(C1, C2);
1375 case Instruction::Mul:
1376 return ConstantExpr::getAnd(C1, C2);
1377 case Instruction::Shl:
1378 case Instruction::LShr:
1379 case Instruction::AShr:
1380 // We can assume that C2 == 0. If it were one the result would be
1381 // undefined because the shift value is as large as the bitwidth.
1383 case Instruction::SDiv:
1384 case Instruction::UDiv:
1385 // We can assume that C2 == 1. If it were zero the result would be
1386 // undefined through division by zero.
1388 case Instruction::URem:
1389 case Instruction::SRem:
1390 // We can assume that C2 == 1. If it were zero the result would be
1391 // undefined through division by zero.
1392 return ConstantInt::getFalse(C1->getContext());
1398 // We don't know how to fold this.
1402 /// isZeroSizedType - This type is zero sized if its an array or structure of
1403 /// zero sized types. The only leaf zero sized type is an empty structure.
1404 static bool isMaybeZeroSizedType(const Type *Ty) {
1405 if (isa<OpaqueType>(Ty)) return true; // Can't say.
1406 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1408 // If all of elements have zero size, this does too.
1409 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1410 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1413 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1414 return isMaybeZeroSizedType(ATy->getElementType());
1419 /// IdxCompare - Compare the two constants as though they were getelementptr
1420 /// indices. This allows coersion of the types to be the same thing.
1422 /// If the two constants are the "same" (after coersion), return 0. If the
1423 /// first is less than the second, return -1, if the second is less than the
1424 /// first, return 1. If the constants are not integral, return -2.
1426 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
1427 if (C1 == C2) return 0;
1429 // Ok, we found a different index. If they are not ConstantInt, we can't do
1430 // anything with them.
1431 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1432 return -2; // don't know!
1434 // Ok, we have two differing integer indices. Sign extend them to be the same
1435 // type. Long is always big enough, so we use it.
1436 if (!C1->getType()->isInteger(64))
1437 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1439 if (!C2->getType()->isInteger(64))
1440 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1442 if (C1 == C2) return 0; // They are equal
1444 // If the type being indexed over is really just a zero sized type, there is
1445 // no pointer difference being made here.
1446 if (isMaybeZeroSizedType(ElTy))
1447 return -2; // dunno.
1449 // If they are really different, now that they are the same type, then we
1450 // found a difference!
1451 if (cast<ConstantInt>(C1)->getSExtValue() <
1452 cast<ConstantInt>(C2)->getSExtValue())
1458 /// evaluateFCmpRelation - This function determines if there is anything we can
1459 /// decide about the two constants provided. This doesn't need to handle simple
1460 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1461 /// If we can determine that the two constants have a particular relation to
1462 /// each other, we should return the corresponding FCmpInst predicate,
1463 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1464 /// ConstantFoldCompareInstruction.
1466 /// To simplify this code we canonicalize the relation so that the first
1467 /// operand is always the most "complex" of the two. We consider ConstantFP
1468 /// to be the simplest, and ConstantExprs to be the most complex.
1469 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1470 assert(V1->getType() == V2->getType() &&
1471 "Cannot compare values of different types!");
1473 // No compile-time operations on this type yet.
1474 if (V1->getType()->isPPC_FP128Ty())
1475 return FCmpInst::BAD_FCMP_PREDICATE;
1477 // Handle degenerate case quickly
1478 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1480 if (!isa<ConstantExpr>(V1)) {
1481 if (!isa<ConstantExpr>(V2)) {
1482 // We distilled thisUse the standard constant folder for a few cases
1484 R = dyn_cast<ConstantInt>(
1485 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1486 if (R && !R->isZero())
1487 return FCmpInst::FCMP_OEQ;
1488 R = dyn_cast<ConstantInt>(
1489 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1490 if (R && !R->isZero())
1491 return FCmpInst::FCMP_OLT;
1492 R = dyn_cast<ConstantInt>(
1493 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1494 if (R && !R->isZero())
1495 return FCmpInst::FCMP_OGT;
1497 // Nothing more we can do
1498 return FCmpInst::BAD_FCMP_PREDICATE;
1501 // If the first operand is simple and second is ConstantExpr, swap operands.
1502 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1503 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1504 return FCmpInst::getSwappedPredicate(SwappedRelation);
1506 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1507 // constantexpr or a simple constant.
1508 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1509 switch (CE1->getOpcode()) {
1510 case Instruction::FPTrunc:
1511 case Instruction::FPExt:
1512 case Instruction::UIToFP:
1513 case Instruction::SIToFP:
1514 // We might be able to do something with these but we don't right now.
1520 // There are MANY other foldings that we could perform here. They will
1521 // probably be added on demand, as they seem needed.
1522 return FCmpInst::BAD_FCMP_PREDICATE;
1525 /// evaluateICmpRelation - This function determines if there is anything we can
1526 /// decide about the two constants provided. This doesn't need to handle simple
1527 /// things like integer comparisons, but should instead handle ConstantExprs
1528 /// and GlobalValues. If we can determine that the two constants have a
1529 /// particular relation to each other, we should return the corresponding ICmp
1530 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1532 /// To simplify this code we canonicalize the relation so that the first
1533 /// operand is always the most "complex" of the two. We consider simple
1534 /// constants (like ConstantInt) to be the simplest, followed by
1535 /// GlobalValues, followed by ConstantExpr's (the most complex).
1537 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1539 assert(V1->getType() == V2->getType() &&
1540 "Cannot compare different types of values!");
1541 if (V1 == V2) return ICmpInst::ICMP_EQ;
1543 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1544 !isa<BlockAddress>(V1)) {
1545 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1546 !isa<BlockAddress>(V2)) {
1547 // We distilled this down to a simple case, use the standard constant
1550 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1551 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1552 if (R && !R->isZero())
1554 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1555 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1556 if (R && !R->isZero())
1558 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1559 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1560 if (R && !R->isZero())
1563 // If we couldn't figure it out, bail.
1564 return ICmpInst::BAD_ICMP_PREDICATE;
1567 // If the first operand is simple, swap operands.
1568 ICmpInst::Predicate SwappedRelation =
1569 evaluateICmpRelation(V2, V1, isSigned);
1570 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1571 return ICmpInst::getSwappedPredicate(SwappedRelation);
1573 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1574 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1575 ICmpInst::Predicate SwappedRelation =
1576 evaluateICmpRelation(V2, V1, isSigned);
1577 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1578 return ICmpInst::getSwappedPredicate(SwappedRelation);
1579 return ICmpInst::BAD_ICMP_PREDICATE;
1582 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1583 // constant (which, since the types must match, means that it's a
1584 // ConstantPointerNull).
1585 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1586 // Don't try to decide equality of aliases.
1587 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1588 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1589 return ICmpInst::ICMP_NE;
1590 } else if (isa<BlockAddress>(V2)) {
1591 return ICmpInst::ICMP_NE; // Globals never equal labels.
1593 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1594 // GlobalVals can never be null unless they have external weak linkage.
1595 // We don't try to evaluate aliases here.
1596 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1597 return ICmpInst::ICMP_NE;
1599 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1600 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1601 ICmpInst::Predicate SwappedRelation =
1602 evaluateICmpRelation(V2, V1, isSigned);
1603 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1604 return ICmpInst::getSwappedPredicate(SwappedRelation);
1605 return ICmpInst::BAD_ICMP_PREDICATE;
1608 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1609 // constant (which, since the types must match, means that it is a
1610 // ConstantPointerNull).
1611 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1612 // Block address in another function can't equal this one, but block
1613 // addresses in the current function might be the same if blocks are
1615 if (BA2->getFunction() != BA->getFunction())
1616 return ICmpInst::ICMP_NE;
1618 // Block addresses aren't null, don't equal the address of globals.
1619 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1620 "Canonicalization guarantee!");
1621 return ICmpInst::ICMP_NE;
1624 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1625 // constantexpr, a global, block address, or a simple constant.
1626 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1627 Constant *CE1Op0 = CE1->getOperand(0);
1629 switch (CE1->getOpcode()) {
1630 case Instruction::Trunc:
1631 case Instruction::FPTrunc:
1632 case Instruction::FPExt:
1633 case Instruction::FPToUI:
1634 case Instruction::FPToSI:
1635 break; // We can't evaluate floating point casts or truncations.
1637 case Instruction::UIToFP:
1638 case Instruction::SIToFP:
1639 case Instruction::BitCast:
1640 case Instruction::ZExt:
1641 case Instruction::SExt:
1642 // If the cast is not actually changing bits, and the second operand is a
1643 // null pointer, do the comparison with the pre-casted value.
1644 if (V2->isNullValue() &&
1645 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
1646 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1647 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1648 return evaluateICmpRelation(CE1Op0,
1649 Constant::getNullValue(CE1Op0->getType()),
1654 case Instruction::GetElementPtr:
1655 // Ok, since this is a getelementptr, we know that the constant has a
1656 // pointer type. Check the various cases.
1657 if (isa<ConstantPointerNull>(V2)) {
1658 // If we are comparing a GEP to a null pointer, check to see if the base
1659 // of the GEP equals the null pointer.
1660 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1661 if (GV->hasExternalWeakLinkage())
1662 // Weak linkage GVals could be zero or not. We're comparing that
1663 // to null pointer so its greater-or-equal
1664 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1666 // If its not weak linkage, the GVal must have a non-zero address
1667 // so the result is greater-than
1668 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1669 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1670 // If we are indexing from a null pointer, check to see if we have any
1671 // non-zero indices.
1672 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1673 if (!CE1->getOperand(i)->isNullValue())
1674 // Offsetting from null, must not be equal.
1675 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1676 // Only zero indexes from null, must still be zero.
1677 return ICmpInst::ICMP_EQ;
1679 // Otherwise, we can't really say if the first operand is null or not.
1680 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1681 if (isa<ConstantPointerNull>(CE1Op0)) {
1682 if (GV2->hasExternalWeakLinkage())
1683 // Weak linkage GVals could be zero or not. We're comparing it to
1684 // a null pointer, so its less-or-equal
1685 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1687 // If its not weak linkage, the GVal must have a non-zero address
1688 // so the result is less-than
1689 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1690 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1692 // If this is a getelementptr of the same global, then it must be
1693 // different. Because the types must match, the getelementptr could
1694 // only have at most one index, and because we fold getelementptr's
1695 // with a single zero index, it must be nonzero.
1696 assert(CE1->getNumOperands() == 2 &&
1697 !CE1->getOperand(1)->isNullValue() &&
1698 "Suprising getelementptr!");
1699 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1701 // If they are different globals, we don't know what the value is,
1702 // but they can't be equal.
1703 return ICmpInst::ICMP_NE;
1707 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1708 Constant *CE2Op0 = CE2->getOperand(0);
1710 // There are MANY other foldings that we could perform here. They will
1711 // probably be added on demand, as they seem needed.
1712 switch (CE2->getOpcode()) {
1714 case Instruction::GetElementPtr:
1715 // By far the most common case to handle is when the base pointers are
1716 // obviously to the same or different globals.
1717 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1718 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1719 return ICmpInst::ICMP_NE;
1720 // Ok, we know that both getelementptr instructions are based on the
1721 // same global. From this, we can precisely determine the relative
1722 // ordering of the resultant pointers.
1725 // The logic below assumes that the result of the comparison
1726 // can be determined by finding the first index that differs.
1727 // This doesn't work if there is over-indexing in any
1728 // subsequent indices, so check for that case first.
1729 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1730 !CE2->isGEPWithNoNotionalOverIndexing())
1731 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1733 // Compare all of the operands the GEP's have in common.
1734 gep_type_iterator GTI = gep_type_begin(CE1);
1735 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1737 switch (IdxCompare(CE1->getOperand(i),
1738 CE2->getOperand(i), GTI.getIndexedType())) {
1739 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1740 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1741 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1744 // Ok, we ran out of things they have in common. If any leftovers
1745 // are non-zero then we have a difference, otherwise we are equal.
1746 for (; i < CE1->getNumOperands(); ++i)
1747 if (!CE1->getOperand(i)->isNullValue()) {
1748 if (isa<ConstantInt>(CE1->getOperand(i)))
1749 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1751 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1754 for (; i < CE2->getNumOperands(); ++i)
1755 if (!CE2->getOperand(i)->isNullValue()) {
1756 if (isa<ConstantInt>(CE2->getOperand(i)))
1757 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1759 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1761 return ICmpInst::ICMP_EQ;
1770 return ICmpInst::BAD_ICMP_PREDICATE;
1773 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1774 Constant *C1, Constant *C2) {
1775 const Type *ResultTy;
1776 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1777 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1778 VT->getNumElements());
1780 ResultTy = Type::getInt1Ty(C1->getContext());
1782 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1783 if (pred == FCmpInst::FCMP_FALSE)
1784 return Constant::getNullValue(ResultTy);
1786 if (pred == FCmpInst::FCMP_TRUE)
1787 return Constant::getAllOnesValue(ResultTy);
1789 // Handle some degenerate cases first
1790 if (isa<UndefValue>(C1) || isa<UndefValue>(C2))
1791 return UndefValue::get(ResultTy);
1793 // No compile-time operations on this type yet.
1794 if (C1->getType()->isPPC_FP128Ty())
1797 // icmp eq/ne(null,GV) -> false/true
1798 if (C1->isNullValue()) {
1799 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1800 // Don't try to evaluate aliases. External weak GV can be null.
1801 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1802 if (pred == ICmpInst::ICMP_EQ)
1803 return ConstantInt::getFalse(C1->getContext());
1804 else if (pred == ICmpInst::ICMP_NE)
1805 return ConstantInt::getTrue(C1->getContext());
1807 // icmp eq/ne(GV,null) -> false/true
1808 } else if (C2->isNullValue()) {
1809 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1810 // Don't try to evaluate aliases. External weak GV can be null.
1811 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1812 if (pred == ICmpInst::ICMP_EQ)
1813 return ConstantInt::getFalse(C1->getContext());
1814 else if (pred == ICmpInst::ICMP_NE)
1815 return ConstantInt::getTrue(C1->getContext());
1819 // If the comparison is a comparison between two i1's, simplify it.
1820 if (C1->getType()->isInteger(1)) {
1822 case ICmpInst::ICMP_EQ:
1823 if (isa<ConstantInt>(C2))
1824 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1825 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1826 case ICmpInst::ICMP_NE:
1827 return ConstantExpr::getXor(C1, C2);
1833 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1834 APInt V1 = cast<ConstantInt>(C1)->getValue();
1835 APInt V2 = cast<ConstantInt>(C2)->getValue();
1837 default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1838 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1839 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1840 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1841 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1842 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1843 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1844 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1845 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1846 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1847 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1849 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1850 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1851 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1852 APFloat::cmpResult R = C1V.compare(C2V);
1854 default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1855 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1856 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1857 case FCmpInst::FCMP_UNO:
1858 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1859 case FCmpInst::FCMP_ORD:
1860 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1861 case FCmpInst::FCMP_UEQ:
1862 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1863 R==APFloat::cmpEqual);
1864 case FCmpInst::FCMP_OEQ:
1865 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1866 case FCmpInst::FCMP_UNE:
1867 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1868 case FCmpInst::FCMP_ONE:
1869 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1870 R==APFloat::cmpGreaterThan);
1871 case FCmpInst::FCMP_ULT:
1872 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1873 R==APFloat::cmpLessThan);
1874 case FCmpInst::FCMP_OLT:
1875 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1876 case FCmpInst::FCMP_UGT:
1877 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1878 R==APFloat::cmpGreaterThan);
1879 case FCmpInst::FCMP_OGT:
1880 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1881 case FCmpInst::FCMP_ULE:
1882 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1883 case FCmpInst::FCMP_OLE:
1884 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1885 R==APFloat::cmpEqual);
1886 case FCmpInst::FCMP_UGE:
1887 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1888 case FCmpInst::FCMP_OGE:
1889 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1890 R==APFloat::cmpEqual);
1892 } else if (isa<VectorType>(C1->getType())) {
1893 SmallVector<Constant*, 16> C1Elts, C2Elts;
1894 C1->getVectorElements(C1Elts);
1895 C2->getVectorElements(C2Elts);
1896 if (C1Elts.empty() || C2Elts.empty())
1899 // If we can constant fold the comparison of each element, constant fold
1900 // the whole vector comparison.
1901 SmallVector<Constant*, 4> ResElts;
1902 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1903 // Compare the elements, producing an i1 result or constant expr.
1904 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1906 return ConstantVector::get(&ResElts[0], ResElts.size());
1909 if (C1->getType()->isFloatingPoint()) {
1910 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1911 switch (evaluateFCmpRelation(C1, C2)) {
1912 default: llvm_unreachable("Unknown relation!");
1913 case FCmpInst::FCMP_UNO:
1914 case FCmpInst::FCMP_ORD:
1915 case FCmpInst::FCMP_UEQ:
1916 case FCmpInst::FCMP_UNE:
1917 case FCmpInst::FCMP_ULT:
1918 case FCmpInst::FCMP_UGT:
1919 case FCmpInst::FCMP_ULE:
1920 case FCmpInst::FCMP_UGE:
1921 case FCmpInst::FCMP_TRUE:
1922 case FCmpInst::FCMP_FALSE:
1923 case FCmpInst::BAD_FCMP_PREDICATE:
1924 break; // Couldn't determine anything about these constants.
1925 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1926 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1927 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1928 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1930 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1931 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1932 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1933 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1935 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1936 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1937 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1938 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1940 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1941 // We can only partially decide this relation.
1942 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1944 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1947 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1948 // We can only partially decide this relation.
1949 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1951 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1954 case ICmpInst::ICMP_NE: // We know that C1 != C2
1955 // We can only partially decide this relation.
1956 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1958 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1963 // If we evaluated the result, return it now.
1965 return ConstantInt::get(ResultTy, Result);
1968 // Evaluate the relation between the two constants, per the predicate.
1969 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1970 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1971 default: llvm_unreachable("Unknown relational!");
1972 case ICmpInst::BAD_ICMP_PREDICATE:
1973 break; // Couldn't determine anything about these constants.
1974 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1975 // If we know the constants are equal, we can decide the result of this
1976 // computation precisely.
1977 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1979 case ICmpInst::ICMP_ULT:
1981 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1983 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1987 case ICmpInst::ICMP_SLT:
1989 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1991 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1995 case ICmpInst::ICMP_UGT:
1997 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1999 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2003 case ICmpInst::ICMP_SGT:
2005 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2007 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2011 case ICmpInst::ICMP_ULE:
2012 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2013 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2015 case ICmpInst::ICMP_SLE:
2016 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2017 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2019 case ICmpInst::ICMP_UGE:
2020 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2021 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2023 case ICmpInst::ICMP_SGE:
2024 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2025 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2027 case ICmpInst::ICMP_NE:
2028 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2029 if (pred == ICmpInst::ICMP_NE) Result = 1;
2033 // If we evaluated the result, return it now.
2035 return ConstantInt::get(ResultTy, Result);
2037 // If the right hand side is a bitcast, try using its inverse to simplify
2038 // it by moving it to the left hand side. We can't do this if it would turn
2039 // a vector compare into a scalar compare or visa versa.
2040 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2041 Constant *CE2Op0 = CE2->getOperand(0);
2042 if (CE2->getOpcode() == Instruction::BitCast &&
2043 isa<VectorType>(CE2->getType())==isa<VectorType>(CE2Op0->getType())) {
2044 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2045 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2049 // If the left hand side is an extension, try eliminating it.
2050 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2051 if (CE1->getOpcode() == Instruction::SExt ||
2052 CE1->getOpcode() == Instruction::ZExt) {
2053 Constant *CE1Op0 = CE1->getOperand(0);
2054 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2055 if (CE1Inverse == CE1Op0) {
2056 // Check whether we can safely truncate the right hand side.
2057 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2058 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2059 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2065 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2066 (C1->isNullValue() && !C2->isNullValue())) {
2067 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2068 // other way if possible.
2069 // Also, if C1 is null and C2 isn't, flip them around.
2071 case ICmpInst::ICMP_EQ:
2072 case ICmpInst::ICMP_NE:
2073 // No change of predicate required.
2074 return ConstantExpr::getICmp(pred, C2, C1);
2076 case ICmpInst::ICMP_ULT:
2077 case ICmpInst::ICMP_SLT:
2078 case ICmpInst::ICMP_UGT:
2079 case ICmpInst::ICMP_SGT:
2080 case ICmpInst::ICMP_ULE:
2081 case ICmpInst::ICMP_SLE:
2082 case ICmpInst::ICMP_UGE:
2083 case ICmpInst::ICMP_SGE:
2084 // Change the predicate as necessary to swap the operands.
2085 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2086 return ConstantExpr::getICmp(pred, C2, C1);
2088 default: // These predicates cannot be flopped around.
2096 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2098 static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) {
2099 // No indices means nothing that could be out of bounds.
2100 if (NumIdx == 0) return true;
2102 // If the first index is zero, it's in bounds.
2103 if (Idxs[0]->isNullValue()) return true;
2105 // If the first index is one and all the rest are zero, it's in bounds,
2106 // by the one-past-the-end rule.
2107 if (!cast<ConstantInt>(Idxs[0])->isOne())
2109 for (unsigned i = 1, e = NumIdx; i != e; ++i)
2110 if (!Idxs[i]->isNullValue())
2115 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2117 Constant* const *Idxs,
2120 (NumIdx == 1 && Idxs[0]->isNullValue()))
2123 if (isa<UndefValue>(C)) {
2124 const PointerType *Ptr = cast<PointerType>(C->getType());
2125 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
2127 (Value **)Idxs+NumIdx);
2128 assert(Ty != 0 && "Invalid indices for GEP!");
2129 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2132 Constant *Idx0 = Idxs[0];
2133 if (C->isNullValue()) {
2135 for (unsigned i = 0, e = NumIdx; i != e; ++i)
2136 if (!Idxs[i]->isNullValue()) {
2141 const PointerType *Ptr = cast<PointerType>(C->getType());
2142 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr,
2144 (Value**)Idxs+NumIdx);
2145 assert(Ty != 0 && "Invalid indices for GEP!");
2146 return ConstantPointerNull::get(
2147 PointerType::get(Ty,Ptr->getAddressSpace()));
2151 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2152 // Combine Indices - If the source pointer to this getelementptr instruction
2153 // is a getelementptr instruction, combine the indices of the two
2154 // getelementptr instructions into a single instruction.
2156 if (CE->getOpcode() == Instruction::GetElementPtr) {
2157 const Type *LastTy = 0;
2158 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2162 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
2163 SmallVector<Value*, 16> NewIndices;
2164 NewIndices.reserve(NumIdx + CE->getNumOperands());
2165 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2166 NewIndices.push_back(CE->getOperand(i));
2168 // Add the last index of the source with the first index of the new GEP.
2169 // Make sure to handle the case when they are actually different types.
2170 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2171 // Otherwise it must be an array.
2172 if (!Idx0->isNullValue()) {
2173 const Type *IdxTy = Combined->getType();
2174 if (IdxTy != Idx0->getType()) {
2175 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2176 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2177 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2178 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2181 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2185 NewIndices.push_back(Combined);
2186 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
2187 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ?
2188 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0),
2190 NewIndices.size()) :
2191 ConstantExpr::getGetElementPtr(CE->getOperand(0),
2197 // Implement folding of:
2198 // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*),
2200 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
2202 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
2203 if (const PointerType *SPT =
2204 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2205 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2206 if (const ArrayType *CAT =
2207 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2208 if (CAT->getElementType() == SAT->getElementType())
2210 ConstantExpr::getInBoundsGetElementPtr(
2211 (Constant*)CE->getOperand(0), Idxs, NumIdx) :
2212 ConstantExpr::getGetElementPtr(
2213 (Constant*)CE->getOperand(0), Idxs, NumIdx);
2217 // Check to see if any array indices are not within the corresponding
2218 // notional array bounds. If so, try to determine if they can be factored
2219 // out into preceding dimensions.
2220 bool Unknown = false;
2221 SmallVector<Constant *, 8> NewIdxs;
2222 const Type *Ty = C->getType();
2223 const Type *Prev = 0;
2224 for (unsigned i = 0; i != NumIdx;
2225 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2226 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2227 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2228 if (ATy->getNumElements() <= INT64_MAX &&
2229 ATy->getNumElements() != 0 &&
2230 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2231 if (isa<SequentialType>(Prev)) {
2232 // It's out of range, but we can factor it into the prior
2234 NewIdxs.resize(NumIdx);
2235 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2236 ATy->getNumElements());
2237 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2239 Constant *PrevIdx = Idxs[i-1];
2240 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2242 // Before adding, extend both operands to i64 to avoid
2243 // overflow trouble.
2244 if (!PrevIdx->getType()->isInteger(64))
2245 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2246 Type::getInt64Ty(Div->getContext()));
2247 if (!Div->getType()->isInteger(64))
2248 Div = ConstantExpr::getSExt(Div,
2249 Type::getInt64Ty(Div->getContext()));
2251 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2253 // It's out of range, but the prior dimension is a struct
2254 // so we can't do anything about it.
2259 // We don't know if it's in range or not.
2264 // If we did any factoring, start over with the adjusted indices.
2265 if (!NewIdxs.empty()) {
2266 for (unsigned i = 0; i != NumIdx; ++i)
2267 if (!NewIdxs[i]) NewIdxs[i] = Idxs[i];
2269 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(),
2271 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size());
2274 // If all indices are known integers and normalized, we can do a simple
2275 // check for the "inbounds" property.
2276 if (!Unknown && !inBounds &&
2277 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx))
2278 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx);