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 DataLayout, and the pieces that do. This is to avoid
16 // a dependence in IR on Target.
18 //===----------------------------------------------------------------------===//
20 #include "ConstantFold.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GlobalAlias.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Operator.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/ManagedStatic.h"
33 #include "llvm/Support/MathExtras.h"
37 //===----------------------------------------------------------------------===//
38 // ConstantFold*Instruction Implementations
39 //===----------------------------------------------------------------------===//
41 /// BitCastConstantVector - Convert the specified vector Constant node to the
42 /// specified vector type. At this point, we know that the elements of the
43 /// input vector constant are all simple integer or FP values.
44 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
47 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49 // If this cast changes element count then we can't handle it here:
50 // doing so requires endianness information. This should be handled by
51 // Analysis/ConstantFolding.cpp
52 unsigned NumElts = DstTy->getNumElements();
53 if (NumElts != CV->getType()->getVectorNumElements())
56 Type *DstEltTy = DstTy->getElementType();
58 SmallVector<Constant*, 16> Result;
59 Type *Ty = IntegerType::get(CV->getContext(), 32);
60 for (unsigned i = 0; i != NumElts; ++i) {
62 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
63 C = ConstantExpr::getBitCast(C, DstEltTy);
67 return ConstantVector::get(Result);
70 /// This function determines which opcode to use to fold two constant cast
71 /// expressions together. It uses CastInst::isEliminableCastPair to determine
72 /// the opcode. Consequently its just a wrapper around that function.
73 /// @brief Determine if it is valid to fold a cast of a cast
76 unsigned opc, ///< opcode of the second cast constant expression
77 ConstantExpr *Op, ///< the first cast constant expression
78 Type *DstTy ///< destination type of the first cast
80 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
81 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
82 assert(CastInst::isCast(opc) && "Invalid cast opcode");
84 // The the types and opcodes for the two Cast constant expressions
85 Type *SrcTy = Op->getOperand(0)->getType();
86 Type *MidTy = Op->getType();
87 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
88 Instruction::CastOps secondOp = Instruction::CastOps(opc);
90 // Assume that pointers are never more than 64 bits wide, and only use this
91 // for the middle type. Otherwise we could end up folding away illegal
92 // bitcasts between address spaces with different sizes.
93 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
95 // Let CastInst::isEliminableCastPair do the heavy lifting.
96 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
100 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
101 Type *SrcTy = V->getType();
103 return V; // no-op cast
105 // Check to see if we are casting a pointer to an aggregate to a pointer to
106 // the first element. If so, return the appropriate GEP instruction.
107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
109 if (PTy->getAddressSpace() == DPTy->getAddressSpace()
110 && DPTy->getElementType()->isSized()) {
111 SmallVector<Value*, 8> IdxList;
113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
114 IdxList.push_back(Zero);
115 Type *ElTy = PTy->getElementType();
116 while (ElTy != DPTy->getElementType()) {
117 if (StructType *STy = dyn_cast<StructType>(ElTy)) {
118 if (STy->getNumElements() == 0) break;
119 ElTy = STy->getElementType(0);
120 IdxList.push_back(Zero);
121 } else if (SequentialType *STy =
122 dyn_cast<SequentialType>(ElTy)) {
123 if (ElTy->isPointerTy()) break; // Can't index into pointers!
124 ElTy = STy->getElementType();
125 IdxList.push_back(Zero);
131 if (ElTy == DPTy->getElementType())
132 // This GEP is inbounds because all indices are zero.
133 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList);
136 // Handle casts from one vector constant to another. We know that the src
137 // and dest type have the same size (otherwise its an illegal cast).
138 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
139 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
140 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
141 "Not cast between same sized vectors!");
143 // First, check for null. Undef is already handled.
144 if (isa<ConstantAggregateZero>(V))
145 return Constant::getNullValue(DestTy);
147 // Handle ConstantVector and ConstantAggregateVector.
148 return BitCastConstantVector(V, DestPTy);
151 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
152 // This allows for other simplifications (although some of them
153 // can only be handled by Analysis/ConstantFolding.cpp).
154 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
155 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
158 // Finally, implement bitcast folding now. The code below doesn't handle
160 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
161 return ConstantPointerNull::get(cast<PointerType>(DestTy));
163 // Handle integral constant input.
164 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
165 if (DestTy->isIntegerTy())
166 // Integral -> Integral. This is a no-op because the bit widths must
167 // be the same. Consequently, we just fold to V.
170 if (DestTy->isFloatingPointTy())
171 return ConstantFP::get(DestTy->getContext(),
172 APFloat(DestTy->getFltSemantics(),
175 // Otherwise, can't fold this (vector?)
179 // Handle ConstantFP input: FP -> Integral.
180 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
181 return ConstantInt::get(FP->getContext(),
182 FP->getValueAPF().bitcastToAPInt());
188 /// ExtractConstantBytes - V is an integer constant which only has a subset of
189 /// its bytes used. The bytes used are indicated by ByteStart (which is the
190 /// first byte used, counting from the least significant byte) and ByteSize,
191 /// which is the number of bytes used.
193 /// This function analyzes the specified constant to see if the specified byte
194 /// range can be returned as a simplified constant. If so, the constant is
195 /// returned, otherwise null is returned.
197 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
199 assert(C->getType()->isIntegerTy() &&
200 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
201 "Non-byte sized integer input");
202 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
203 assert(ByteSize && "Must be accessing some piece");
204 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
205 assert(ByteSize != CSize && "Should not extract everything");
207 // Constant Integers are simple.
208 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
209 APInt V = CI->getValue();
211 V = V.lshr(ByteStart*8);
212 V = V.trunc(ByteSize*8);
213 return ConstantInt::get(CI->getContext(), V);
216 // In the input is a constant expr, we might be able to recursively simplify.
217 // If not, we definitely can't do anything.
218 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
219 if (CE == 0) return 0;
221 switch (CE->getOpcode()) {
223 case Instruction::Or: {
224 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
229 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
230 if (RHSC->isAllOnesValue())
233 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
236 return ConstantExpr::getOr(LHS, RHS);
238 case Instruction::And: {
239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
244 if (RHS->isNullValue())
247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
250 return ConstantExpr::getAnd(LHS, RHS);
252 case Instruction::LShr: {
253 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
256 unsigned ShAmt = Amt->getZExtValue();
257 // Cannot analyze non-byte shifts.
258 if ((ShAmt & 7) != 0)
262 // If the extract is known to be all zeros, return zero.
263 if (ByteStart >= CSize-ShAmt)
264 return Constant::getNullValue(IntegerType::get(CE->getContext(),
266 // If the extract is known to be fully in the input, extract it.
267 if (ByteStart+ByteSize+ShAmt <= CSize)
268 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
270 // TODO: Handle the 'partially zero' case.
274 case Instruction::Shl: {
275 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
278 unsigned ShAmt = Amt->getZExtValue();
279 // Cannot analyze non-byte shifts.
280 if ((ShAmt & 7) != 0)
284 // If the extract is known to be all zeros, return zero.
285 if (ByteStart+ByteSize <= ShAmt)
286 return Constant::getNullValue(IntegerType::get(CE->getContext(),
288 // If the extract is known to be fully in the input, extract it.
289 if (ByteStart >= ShAmt)
290 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
292 // TODO: Handle the 'partially zero' case.
296 case Instruction::ZExt: {
297 unsigned SrcBitSize =
298 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
300 // If extracting something that is completely zero, return 0.
301 if (ByteStart*8 >= SrcBitSize)
302 return Constant::getNullValue(IntegerType::get(CE->getContext(),
305 // If exactly extracting the input, return it.
306 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
307 return CE->getOperand(0);
309 // If extracting something completely in the input, if if the input is a
310 // multiple of 8 bits, recurse.
311 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
312 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
314 // Otherwise, if extracting a subset of the input, which is not multiple of
315 // 8 bits, do a shift and trunc to get the bits.
316 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
317 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
318 Constant *Res = CE->getOperand(0);
320 Res = ConstantExpr::getLShr(Res,
321 ConstantInt::get(Res->getType(), ByteStart*8));
322 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
326 // TODO: Handle the 'partially zero' case.
332 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
333 /// on Ty, with any known factors factored out. If Folded is false,
334 /// return null if no factoring was possible, to avoid endlessly
335 /// bouncing an unfoldable expression back into the top-level folder.
337 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
339 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
340 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
341 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
342 return ConstantExpr::getNUWMul(E, N);
345 if (StructType *STy = dyn_cast<StructType>(Ty))
346 if (!STy->isPacked()) {
347 unsigned NumElems = STy->getNumElements();
348 // An empty struct has size zero.
350 return ConstantExpr::getNullValue(DestTy);
351 // Check for a struct with all members having the same size.
352 Constant *MemberSize =
353 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
355 for (unsigned i = 1; i != NumElems; ++i)
357 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
362 Constant *N = ConstantInt::get(DestTy, NumElems);
363 return ConstantExpr::getNUWMul(MemberSize, N);
367 // Pointer size doesn't depend on the pointee type, so canonicalize them
368 // to an arbitrary pointee.
369 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
370 if (!PTy->getElementType()->isIntegerTy(1))
372 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
373 PTy->getAddressSpace()),
376 // If there's no interesting folding happening, bail so that we don't create
377 // a constant that looks like it needs folding but really doesn't.
381 // Base case: Get a regular sizeof expression.
382 Constant *C = ConstantExpr::getSizeOf(Ty);
383 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
389 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
390 /// on Ty, with any known factors factored out. If Folded is false,
391 /// return null if no factoring was possible, to avoid endlessly
392 /// bouncing an unfoldable expression back into the top-level folder.
394 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
396 // The alignment of an array is equal to the alignment of the
397 // array element. Note that this is not always true for vectors.
398 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
399 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
400 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
407 if (StructType *STy = dyn_cast<StructType>(Ty)) {
408 // Packed structs always have an alignment of 1.
410 return ConstantInt::get(DestTy, 1);
412 // Otherwise, struct alignment is the maximum alignment of any member.
413 // Without target data, we can't compare much, but we can check to see
414 // if all the members have the same alignment.
415 unsigned NumElems = STy->getNumElements();
416 // An empty struct has minimal alignment.
418 return ConstantInt::get(DestTy, 1);
419 // Check for a struct with all members having the same alignment.
420 Constant *MemberAlign =
421 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
423 for (unsigned i = 1; i != NumElems; ++i)
424 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
432 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
433 // to an arbitrary pointee.
434 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
435 if (!PTy->getElementType()->isIntegerTy(1))
437 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
439 PTy->getAddressSpace()),
442 // If there's no interesting folding happening, bail so that we don't create
443 // a constant that looks like it needs folding but really doesn't.
447 // Base case: Get a regular alignof expression.
448 Constant *C = ConstantExpr::getAlignOf(Ty);
449 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
455 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
456 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
457 /// return null if no factoring was possible, to avoid endlessly
458 /// bouncing an unfoldable expression back into the top-level folder.
460 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
463 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
464 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
467 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
468 return ConstantExpr::getNUWMul(E, N);
471 if (StructType *STy = dyn_cast<StructType>(Ty))
472 if (!STy->isPacked()) {
473 unsigned NumElems = STy->getNumElements();
474 // An empty struct has no members.
477 // Check for a struct with all members having the same size.
478 Constant *MemberSize =
479 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
481 for (unsigned i = 1; i != NumElems; ++i)
483 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
488 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
493 return ConstantExpr::getNUWMul(MemberSize, N);
497 // If there's no interesting folding happening, bail so that we don't create
498 // a constant that looks like it needs folding but really doesn't.
502 // Base case: Get a regular offsetof expression.
503 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
504 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
510 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
512 if (isa<UndefValue>(V)) {
513 // zext(undef) = 0, because the top bits will be zero.
514 // sext(undef) = 0, because the top bits will all be the same.
515 // [us]itofp(undef) = 0, because the result value is bounded.
516 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
517 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
518 return Constant::getNullValue(DestTy);
519 return UndefValue::get(DestTy);
522 if (V->isNullValue() && !DestTy->isX86_MMXTy())
523 return Constant::getNullValue(DestTy);
525 // If the cast operand is a constant expression, there's a few things we can
526 // do to try to simplify it.
527 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
529 // Try hard to fold cast of cast because they are often eliminable.
530 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
531 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
532 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
533 // If all of the indexes in the GEP are null values, there is no pointer
534 // adjustment going on. We might as well cast the source pointer.
535 bool isAllNull = true;
536 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
537 if (!CE->getOperand(i)->isNullValue()) {
542 // This is casting one pointer type to another, always BitCast
543 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
547 // If the cast operand is a constant vector, perform the cast by
548 // operating on each element. In the cast of bitcasts, the element
549 // count may be mismatched; don't attempt to handle that here.
550 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
551 DestTy->isVectorTy() &&
552 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
553 SmallVector<Constant*, 16> res;
554 VectorType *DestVecTy = cast<VectorType>(DestTy);
555 Type *DstEltTy = DestVecTy->getElementType();
556 Type *Ty = IntegerType::get(V->getContext(), 32);
557 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
559 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
560 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
562 return ConstantVector::get(res);
565 // We actually have to do a cast now. Perform the cast according to the
569 llvm_unreachable("Failed to cast constant expression");
570 case Instruction::FPTrunc:
571 case Instruction::FPExt:
572 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
574 APFloat Val = FPC->getValueAPF();
575 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf :
576 DestTy->isFloatTy() ? APFloat::IEEEsingle :
577 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
578 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
579 DestTy->isFP128Ty() ? APFloat::IEEEquad :
580 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble :
582 APFloat::rmNearestTiesToEven, &ignored);
583 return ConstantFP::get(V->getContext(), Val);
585 return 0; // Can't fold.
586 case Instruction::FPToUI:
587 case Instruction::FPToSI:
588 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
589 const APFloat &V = FPC->getValueAPF();
592 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
593 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
594 APFloat::rmTowardZero, &ignored);
595 APInt Val(DestBitWidth, x);
596 return ConstantInt::get(FPC->getContext(), Val);
598 return 0; // Can't fold.
599 case Instruction::IntToPtr: //always treated as unsigned
600 if (V->isNullValue()) // Is it an integral null value?
601 return ConstantPointerNull::get(cast<PointerType>(DestTy));
602 return 0; // Other pointer types cannot be casted
603 case Instruction::PtrToInt: // always treated as unsigned
604 // Is it a null pointer value?
605 if (V->isNullValue())
606 return ConstantInt::get(DestTy, 0);
607 // If this is a sizeof-like expression, pull out multiplications by
608 // known factors to expose them to subsequent folding. If it's an
609 // alignof-like expression, factor out known factors.
610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
611 if (CE->getOpcode() == Instruction::GetElementPtr &&
612 CE->getOperand(0)->isNullValue()) {
614 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
615 if (CE->getNumOperands() == 2) {
616 // Handle a sizeof-like expression.
617 Constant *Idx = CE->getOperand(1);
618 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
619 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
620 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
623 return ConstantExpr::getMul(C, Idx);
625 } else if (CE->getNumOperands() == 3 &&
626 CE->getOperand(1)->isNullValue()) {
627 // Handle an alignof-like expression.
628 if (StructType *STy = dyn_cast<StructType>(Ty))
629 if (!STy->isPacked()) {
630 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
632 STy->getNumElements() == 2 &&
633 STy->getElementType(0)->isIntegerTy(1)) {
634 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
637 // Handle an offsetof-like expression.
638 if (Ty->isStructTy() || Ty->isArrayTy()) {
639 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
645 // Other pointer types cannot be casted
647 case Instruction::UIToFP:
648 case Instruction::SIToFP:
649 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
650 APInt api = CI->getValue();
651 APFloat apf(DestTy->getFltSemantics(),
652 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
653 (void)apf.convertFromAPInt(api,
654 opc==Instruction::SIToFP,
655 APFloat::rmNearestTiesToEven);
656 return ConstantFP::get(V->getContext(), apf);
659 case Instruction::ZExt:
660 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
661 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
662 return ConstantInt::get(V->getContext(),
663 CI->getValue().zext(BitWidth));
666 case Instruction::SExt:
667 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
668 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
669 return ConstantInt::get(V->getContext(),
670 CI->getValue().sext(BitWidth));
673 case Instruction::Trunc: {
674 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
675 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
676 return ConstantInt::get(V->getContext(),
677 CI->getValue().trunc(DestBitWidth));
680 // The input must be a constantexpr. See if we can simplify this based on
681 // the bytes we are demanding. Only do this if the source and dest are an
682 // even multiple of a byte.
683 if ((DestBitWidth & 7) == 0 &&
684 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
685 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
690 case Instruction::BitCast:
691 return FoldBitCast(V, DestTy);
695 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
696 Constant *V1, Constant *V2) {
697 // Check for i1 and vector true/false conditions.
698 if (Cond->isNullValue()) return V2;
699 if (Cond->isAllOnesValue()) return V1;
701 // If the condition is a vector constant, fold the result elementwise.
702 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
703 SmallVector<Constant*, 16> Result;
704 Type *Ty = IntegerType::get(CondV->getContext(), 32);
705 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
706 ConstantInt *Cond = dyn_cast<ConstantInt>(CondV->getOperand(i));
707 if (Cond == 0) break;
709 Constant *V = Cond->isNullValue() ? V2 : V1;
710 Constant *Res = ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
711 Result.push_back(Res);
714 // If we were able to build the vector, return it.
715 if (Result.size() == V1->getType()->getVectorNumElements())
716 return ConstantVector::get(Result);
719 if (isa<UndefValue>(Cond)) {
720 if (isa<UndefValue>(V1)) return V1;
723 if (isa<UndefValue>(V1)) return V2;
724 if (isa<UndefValue>(V2)) return V1;
725 if (V1 == V2) return V1;
727 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
728 if (TrueVal->getOpcode() == Instruction::Select)
729 if (TrueVal->getOperand(0) == Cond)
730 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
732 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
733 if (FalseVal->getOpcode() == Instruction::Select)
734 if (FalseVal->getOperand(0) == Cond)
735 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
741 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
743 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
744 return UndefValue::get(Val->getType()->getVectorElementType());
745 if (Val->isNullValue()) // ee(zero, x) -> zero
746 return Constant::getNullValue(Val->getType()->getVectorElementType());
747 // ee({w,x,y,z}, undef) -> undef
748 if (isa<UndefValue>(Idx))
749 return UndefValue::get(Val->getType()->getVectorElementType());
751 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
752 uint64_t Index = CIdx->getZExtValue();
753 // ee({w,x,y,z}, wrong_value) -> undef
754 if (Index >= Val->getType()->getVectorNumElements())
755 return UndefValue::get(Val->getType()->getVectorElementType());
756 return Val->getAggregateElement(Index);
761 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
764 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
766 const APInt &IdxVal = CIdx->getValue();
768 SmallVector<Constant*, 16> Result;
769 Type *Ty = IntegerType::get(Val->getContext(), 32);
770 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){
772 Result.push_back(Elt);
777 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
781 return ConstantVector::get(Result);
784 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
787 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
788 Type *EltTy = V1->getType()->getVectorElementType();
790 // Undefined shuffle mask -> undefined value.
791 if (isa<UndefValue>(Mask))
792 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
794 // Don't break the bitcode reader hack.
795 if (isa<ConstantExpr>(Mask)) return 0;
797 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
799 // Loop over the shuffle mask, evaluating each element.
800 SmallVector<Constant*, 32> Result;
801 for (unsigned i = 0; i != MaskNumElts; ++i) {
802 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
804 Result.push_back(UndefValue::get(EltTy));
808 if (unsigned(Elt) >= SrcNumElts*2)
809 InElt = UndefValue::get(EltTy);
810 else if (unsigned(Elt) >= SrcNumElts) {
811 Type *Ty = IntegerType::get(V2->getContext(), 32);
813 ConstantExpr::getExtractElement(V2,
814 ConstantInt::get(Ty, Elt - SrcNumElts));
816 Type *Ty = IntegerType::get(V1->getContext(), 32);
817 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
819 Result.push_back(InElt);
822 return ConstantVector::get(Result);
825 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
826 ArrayRef<unsigned> Idxs) {
827 // Base case: no indices, so return the entire value.
831 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
832 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
837 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
839 ArrayRef<unsigned> Idxs) {
840 // Base case: no indices, so replace the entire value.
845 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
846 NumElts = ST->getNumElements();
847 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
848 NumElts = AT->getNumElements();
850 NumElts = Agg->getType()->getVectorNumElements();
852 SmallVector<Constant*, 32> Result;
853 for (unsigned i = 0; i != NumElts; ++i) {
854 Constant *C = Agg->getAggregateElement(i);
855 if (C == 0) return 0;
858 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
863 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
864 return ConstantStruct::get(ST, Result);
865 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
866 return ConstantArray::get(AT, Result);
867 return ConstantVector::get(Result);
871 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
872 Constant *C1, Constant *C2) {
873 // Handle UndefValue up front.
874 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
876 case Instruction::Xor:
877 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
878 // Handle undef ^ undef -> 0 special case. This is a common
880 return Constant::getNullValue(C1->getType());
882 case Instruction::Add:
883 case Instruction::Sub:
884 return UndefValue::get(C1->getType());
885 case Instruction::And:
886 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
888 return Constant::getNullValue(C1->getType()); // undef & X -> 0
889 case Instruction::Mul: {
891 // X * undef -> undef if X is odd or undef
892 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
893 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
894 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
895 return UndefValue::get(C1->getType());
897 // X * undef -> 0 otherwise
898 return Constant::getNullValue(C1->getType());
900 case Instruction::UDiv:
901 case Instruction::SDiv:
902 // undef / 1 -> undef
903 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
904 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
908 case Instruction::URem:
909 case Instruction::SRem:
910 if (!isa<UndefValue>(C2)) // undef / X -> 0
911 return Constant::getNullValue(C1->getType());
912 return C2; // X / undef -> undef
913 case Instruction::Or: // X | undef -> -1
914 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
916 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
917 case Instruction::LShr:
918 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
919 return C1; // undef lshr undef -> undef
920 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
922 case Instruction::AShr:
923 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
924 return Constant::getAllOnesValue(C1->getType());
925 else if (isa<UndefValue>(C1))
926 return C1; // undef ashr undef -> undef
928 return C1; // X ashr undef --> X
929 case Instruction::Shl:
930 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
931 return C1; // undef shl undef -> undef
932 // undef << X -> 0 or X << undef -> 0
933 return Constant::getNullValue(C1->getType());
937 // Handle simplifications when the RHS is a constant int.
938 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
940 case Instruction::Add:
941 if (CI2->equalsInt(0)) return C1; // X + 0 == X
943 case Instruction::Sub:
944 if (CI2->equalsInt(0)) return C1; // X - 0 == X
946 case Instruction::Mul:
947 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
948 if (CI2->equalsInt(1))
949 return C1; // X * 1 == X
951 case Instruction::UDiv:
952 case Instruction::SDiv:
953 if (CI2->equalsInt(1))
954 return C1; // X / 1 == X
955 if (CI2->equalsInt(0))
956 return UndefValue::get(CI2->getType()); // X / 0 == undef
958 case Instruction::URem:
959 case Instruction::SRem:
960 if (CI2->equalsInt(1))
961 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
962 if (CI2->equalsInt(0))
963 return UndefValue::get(CI2->getType()); // X % 0 == undef
965 case Instruction::And:
966 if (CI2->isZero()) return C2; // X & 0 == 0
967 if (CI2->isAllOnesValue())
968 return C1; // X & -1 == X
970 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
971 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
972 if (CE1->getOpcode() == Instruction::ZExt) {
973 unsigned DstWidth = CI2->getType()->getBitWidth();
975 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
976 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
977 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
981 // If and'ing the address of a global with a constant, fold it.
982 if (CE1->getOpcode() == Instruction::PtrToInt &&
983 isa<GlobalValue>(CE1->getOperand(0))) {
984 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
986 // Functions are at least 4-byte aligned.
987 unsigned GVAlign = GV->getAlignment();
988 if (isa<Function>(GV))
989 GVAlign = std::max(GVAlign, 4U);
992 unsigned DstWidth = CI2->getType()->getBitWidth();
993 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
994 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
996 // If checking bits we know are clear, return zero.
997 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
998 return Constant::getNullValue(CI2->getType());
1003 case Instruction::Or:
1004 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1005 if (CI2->isAllOnesValue())
1006 return C2; // X | -1 == -1
1008 case Instruction::Xor:
1009 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1011 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1012 switch (CE1->getOpcode()) {
1014 case Instruction::ICmp:
1015 case Instruction::FCmp:
1016 // cmp pred ^ true -> cmp !pred
1017 assert(CI2->equalsInt(1));
1018 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1019 pred = CmpInst::getInversePredicate(pred);
1020 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1021 CE1->getOperand(1));
1025 case Instruction::AShr:
1026 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1027 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1028 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1029 return ConstantExpr::getLShr(C1, C2);
1032 } else if (isa<ConstantInt>(C1)) {
1033 // If C1 is a ConstantInt and C2 is not, swap the operands.
1034 if (Instruction::isCommutative(Opcode))
1035 return ConstantExpr::get(Opcode, C2, C1);
1038 // At this point we know neither constant is an UndefValue.
1039 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1040 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1041 const APInt &C1V = CI1->getValue();
1042 const APInt &C2V = CI2->getValue();
1046 case Instruction::Add:
1047 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1048 case Instruction::Sub:
1049 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1050 case Instruction::Mul:
1051 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1052 case Instruction::UDiv:
1053 assert(!CI2->isNullValue() && "Div by zero handled above");
1054 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1055 case Instruction::SDiv:
1056 assert(!CI2->isNullValue() && "Div by zero handled above");
1057 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1058 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1059 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1060 case Instruction::URem:
1061 assert(!CI2->isNullValue() && "Div by zero handled above");
1062 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1063 case Instruction::SRem:
1064 assert(!CI2->isNullValue() && "Div by zero handled above");
1065 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1066 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1067 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1068 case Instruction::And:
1069 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1070 case Instruction::Or:
1071 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1072 case Instruction::Xor:
1073 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1074 case Instruction::Shl: {
1075 uint32_t shiftAmt = C2V.getZExtValue();
1076 if (shiftAmt < C1V.getBitWidth())
1077 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1079 return UndefValue::get(C1->getType()); // too big shift is undef
1081 case Instruction::LShr: {
1082 uint32_t shiftAmt = C2V.getZExtValue();
1083 if (shiftAmt < C1V.getBitWidth())
1084 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1086 return UndefValue::get(C1->getType()); // too big shift is undef
1088 case Instruction::AShr: {
1089 uint32_t shiftAmt = C2V.getZExtValue();
1090 if (shiftAmt < C1V.getBitWidth())
1091 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1093 return UndefValue::get(C1->getType()); // too big shift is undef
1099 case Instruction::SDiv:
1100 case Instruction::UDiv:
1101 case Instruction::URem:
1102 case Instruction::SRem:
1103 case Instruction::LShr:
1104 case Instruction::AShr:
1105 case Instruction::Shl:
1106 if (CI1->equalsInt(0)) return C1;
1111 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1112 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1113 APFloat C1V = CFP1->getValueAPF();
1114 APFloat C2V = CFP2->getValueAPF();
1115 APFloat C3V = C1V; // copy for modification
1119 case Instruction::FAdd:
1120 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1121 return ConstantFP::get(C1->getContext(), C3V);
1122 case Instruction::FSub:
1123 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1124 return ConstantFP::get(C1->getContext(), C3V);
1125 case Instruction::FMul:
1126 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1127 return ConstantFP::get(C1->getContext(), C3V);
1128 case Instruction::FDiv:
1129 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1130 return ConstantFP::get(C1->getContext(), C3V);
1131 case Instruction::FRem:
1132 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1133 return ConstantFP::get(C1->getContext(), C3V);
1136 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1137 // Perform elementwise folding.
1138 SmallVector<Constant*, 16> Result;
1139 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1140 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1142 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1144 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1146 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1149 return ConstantVector::get(Result);
1152 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1153 // There are many possible foldings we could do here. We should probably
1154 // at least fold add of a pointer with an integer into the appropriate
1155 // getelementptr. This will improve alias analysis a bit.
1157 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1159 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1160 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1161 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1162 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1164 } else if (isa<ConstantExpr>(C2)) {
1165 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1166 // other way if possible.
1167 if (Instruction::isCommutative(Opcode))
1168 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1171 // i1 can be simplified in many cases.
1172 if (C1->getType()->isIntegerTy(1)) {
1174 case Instruction::Add:
1175 case Instruction::Sub:
1176 return ConstantExpr::getXor(C1, C2);
1177 case Instruction::Mul:
1178 return ConstantExpr::getAnd(C1, C2);
1179 case Instruction::Shl:
1180 case Instruction::LShr:
1181 case Instruction::AShr:
1182 // We can assume that C2 == 0. If it were one the result would be
1183 // undefined because the shift value is as large as the bitwidth.
1185 case Instruction::SDiv:
1186 case Instruction::UDiv:
1187 // We can assume that C2 == 1. If it were zero the result would be
1188 // undefined through division by zero.
1190 case Instruction::URem:
1191 case Instruction::SRem:
1192 // We can assume that C2 == 1. If it were zero the result would be
1193 // undefined through division by zero.
1194 return ConstantInt::getFalse(C1->getContext());
1200 // We don't know how to fold this.
1204 /// isZeroSizedType - This type is zero sized if its an array or structure of
1205 /// zero sized types. The only leaf zero sized type is an empty structure.
1206 static bool isMaybeZeroSizedType(Type *Ty) {
1207 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1208 if (STy->isOpaque()) return true; // Can't say.
1210 // If all of elements have zero size, this does too.
1211 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1212 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1215 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1216 return isMaybeZeroSizedType(ATy->getElementType());
1221 /// IdxCompare - Compare the two constants as though they were getelementptr
1222 /// indices. This allows coersion of the types to be the same thing.
1224 /// If the two constants are the "same" (after coersion), return 0. If the
1225 /// first is less than the second, return -1, if the second is less than the
1226 /// first, return 1. If the constants are not integral, return -2.
1228 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1229 if (C1 == C2) return 0;
1231 // Ok, we found a different index. If they are not ConstantInt, we can't do
1232 // anything with them.
1233 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1234 return -2; // don't know!
1236 // Ok, we have two differing integer indices. Sign extend them to be the same
1237 // type. Long is always big enough, so we use it.
1238 if (!C1->getType()->isIntegerTy(64))
1239 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1241 if (!C2->getType()->isIntegerTy(64))
1242 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1244 if (C1 == C2) return 0; // They are equal
1246 // If the type being indexed over is really just a zero sized type, there is
1247 // no pointer difference being made here.
1248 if (isMaybeZeroSizedType(ElTy))
1249 return -2; // dunno.
1251 // If they are really different, now that they are the same type, then we
1252 // found a difference!
1253 if (cast<ConstantInt>(C1)->getSExtValue() <
1254 cast<ConstantInt>(C2)->getSExtValue())
1260 /// evaluateFCmpRelation - This function determines if there is anything we can
1261 /// decide about the two constants provided. This doesn't need to handle simple
1262 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1263 /// If we can determine that the two constants have a particular relation to
1264 /// each other, we should return the corresponding FCmpInst predicate,
1265 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1266 /// ConstantFoldCompareInstruction.
1268 /// To simplify this code we canonicalize the relation so that the first
1269 /// operand is always the most "complex" of the two. We consider ConstantFP
1270 /// to be the simplest, and ConstantExprs to be the most complex.
1271 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1272 assert(V1->getType() == V2->getType() &&
1273 "Cannot compare values of different types!");
1275 // Handle degenerate case quickly
1276 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1278 if (!isa<ConstantExpr>(V1)) {
1279 if (!isa<ConstantExpr>(V2)) {
1280 // We distilled thisUse the standard constant folder for a few cases
1282 R = dyn_cast<ConstantInt>(
1283 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1284 if (R && !R->isZero())
1285 return FCmpInst::FCMP_OEQ;
1286 R = dyn_cast<ConstantInt>(
1287 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1288 if (R && !R->isZero())
1289 return FCmpInst::FCMP_OLT;
1290 R = dyn_cast<ConstantInt>(
1291 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1292 if (R && !R->isZero())
1293 return FCmpInst::FCMP_OGT;
1295 // Nothing more we can do
1296 return FCmpInst::BAD_FCMP_PREDICATE;
1299 // If the first operand is simple and second is ConstantExpr, swap operands.
1300 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1301 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1302 return FCmpInst::getSwappedPredicate(SwappedRelation);
1304 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1305 // constantexpr or a simple constant.
1306 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1307 switch (CE1->getOpcode()) {
1308 case Instruction::FPTrunc:
1309 case Instruction::FPExt:
1310 case Instruction::UIToFP:
1311 case Instruction::SIToFP:
1312 // We might be able to do something with these but we don't right now.
1318 // There are MANY other foldings that we could perform here. They will
1319 // probably be added on demand, as they seem needed.
1320 return FCmpInst::BAD_FCMP_PREDICATE;
1323 /// evaluateICmpRelation - This function determines if there is anything we can
1324 /// decide about the two constants provided. This doesn't need to handle simple
1325 /// things like integer comparisons, but should instead handle ConstantExprs
1326 /// and GlobalValues. If we can determine that the two constants have a
1327 /// particular relation to each other, we should return the corresponding ICmp
1328 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1330 /// To simplify this code we canonicalize the relation so that the first
1331 /// operand is always the most "complex" of the two. We consider simple
1332 /// constants (like ConstantInt) to be the simplest, followed by
1333 /// GlobalValues, followed by ConstantExpr's (the most complex).
1335 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1337 assert(V1->getType() == V2->getType() &&
1338 "Cannot compare different types of values!");
1339 if (V1 == V2) return ICmpInst::ICMP_EQ;
1341 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1342 !isa<BlockAddress>(V1)) {
1343 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1344 !isa<BlockAddress>(V2)) {
1345 // We distilled this down to a simple case, use the standard constant
1348 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1349 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1350 if (R && !R->isZero())
1352 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1353 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1354 if (R && !R->isZero())
1356 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1357 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1358 if (R && !R->isZero())
1361 // If we couldn't figure it out, bail.
1362 return ICmpInst::BAD_ICMP_PREDICATE;
1365 // If the first operand is simple, swap operands.
1366 ICmpInst::Predicate SwappedRelation =
1367 evaluateICmpRelation(V2, V1, isSigned);
1368 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1369 return ICmpInst::getSwappedPredicate(SwappedRelation);
1371 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1372 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1373 ICmpInst::Predicate SwappedRelation =
1374 evaluateICmpRelation(V2, V1, isSigned);
1375 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1376 return ICmpInst::getSwappedPredicate(SwappedRelation);
1377 return ICmpInst::BAD_ICMP_PREDICATE;
1380 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1381 // constant (which, since the types must match, means that it's a
1382 // ConstantPointerNull).
1383 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1384 // Don't try to decide equality of aliases.
1385 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1386 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1387 return ICmpInst::ICMP_NE;
1388 } else if (isa<BlockAddress>(V2)) {
1389 return ICmpInst::ICMP_NE; // Globals never equal labels.
1391 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1392 // GlobalVals can never be null unless they have external weak linkage.
1393 // We don't try to evaluate aliases here.
1394 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1395 return ICmpInst::ICMP_NE;
1397 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1398 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1399 ICmpInst::Predicate SwappedRelation =
1400 evaluateICmpRelation(V2, V1, isSigned);
1401 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1402 return ICmpInst::getSwappedPredicate(SwappedRelation);
1403 return ICmpInst::BAD_ICMP_PREDICATE;
1406 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1407 // constant (which, since the types must match, means that it is a
1408 // ConstantPointerNull).
1409 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1410 // Block address in another function can't equal this one, but block
1411 // addresses in the current function might be the same if blocks are
1413 if (BA2->getFunction() != BA->getFunction())
1414 return ICmpInst::ICMP_NE;
1416 // Block addresses aren't null, don't equal the address of globals.
1417 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1418 "Canonicalization guarantee!");
1419 return ICmpInst::ICMP_NE;
1422 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1423 // constantexpr, a global, block address, or a simple constant.
1424 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1425 Constant *CE1Op0 = CE1->getOperand(0);
1427 switch (CE1->getOpcode()) {
1428 case Instruction::Trunc:
1429 case Instruction::FPTrunc:
1430 case Instruction::FPExt:
1431 case Instruction::FPToUI:
1432 case Instruction::FPToSI:
1433 break; // We can't evaluate floating point casts or truncations.
1435 case Instruction::UIToFP:
1436 case Instruction::SIToFP:
1437 case Instruction::BitCast:
1438 case Instruction::ZExt:
1439 case Instruction::SExt:
1440 // If the cast is not actually changing bits, and the second operand is a
1441 // null pointer, do the comparison with the pre-casted value.
1442 if (V2->isNullValue() &&
1443 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1444 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1445 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1446 return evaluateICmpRelation(CE1Op0,
1447 Constant::getNullValue(CE1Op0->getType()),
1452 case Instruction::GetElementPtr:
1453 // Ok, since this is a getelementptr, we know that the constant has a
1454 // pointer type. Check the various cases.
1455 if (isa<ConstantPointerNull>(V2)) {
1456 // If we are comparing a GEP to a null pointer, check to see if the base
1457 // of the GEP equals the null pointer.
1458 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1459 if (GV->hasExternalWeakLinkage())
1460 // Weak linkage GVals could be zero or not. We're comparing that
1461 // to null pointer so its greater-or-equal
1462 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1464 // If its not weak linkage, the GVal must have a non-zero address
1465 // so the result is greater-than
1466 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1467 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1468 // If we are indexing from a null pointer, check to see if we have any
1469 // non-zero indices.
1470 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1471 if (!CE1->getOperand(i)->isNullValue())
1472 // Offsetting from null, must not be equal.
1473 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1474 // Only zero indexes from null, must still be zero.
1475 return ICmpInst::ICMP_EQ;
1477 // Otherwise, we can't really say if the first operand is null or not.
1478 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1479 if (isa<ConstantPointerNull>(CE1Op0)) {
1480 if (GV2->hasExternalWeakLinkage())
1481 // Weak linkage GVals could be zero or not. We're comparing it to
1482 // a null pointer, so its less-or-equal
1483 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1485 // If its not weak linkage, the GVal must have a non-zero address
1486 // so the result is less-than
1487 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1488 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1490 // If this is a getelementptr of the same global, then it must be
1491 // different. Because the types must match, the getelementptr could
1492 // only have at most one index, and because we fold getelementptr's
1493 // with a single zero index, it must be nonzero.
1494 assert(CE1->getNumOperands() == 2 &&
1495 !CE1->getOperand(1)->isNullValue() &&
1496 "Surprising getelementptr!");
1497 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1499 // If they are different globals, we don't know what the value is.
1500 return ICmpInst::BAD_ICMP_PREDICATE;
1504 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1505 Constant *CE2Op0 = CE2->getOperand(0);
1507 // There are MANY other foldings that we could perform here. They will
1508 // probably be added on demand, as they seem needed.
1509 switch (CE2->getOpcode()) {
1511 case Instruction::GetElementPtr:
1512 // By far the most common case to handle is when the base pointers are
1513 // obviously to the same global.
1514 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1515 if (CE1Op0 != CE2Op0) // Don't know relative ordering.
1516 return ICmpInst::BAD_ICMP_PREDICATE;
1517 // Ok, we know that both getelementptr instructions are based on the
1518 // same global. From this, we can precisely determine the relative
1519 // ordering of the resultant pointers.
1522 // The logic below assumes that the result of the comparison
1523 // can be determined by finding the first index that differs.
1524 // This doesn't work if there is over-indexing in any
1525 // subsequent indices, so check for that case first.
1526 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1527 !CE2->isGEPWithNoNotionalOverIndexing())
1528 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1530 // Compare all of the operands the GEP's have in common.
1531 gep_type_iterator GTI = gep_type_begin(CE1);
1532 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1534 switch (IdxCompare(CE1->getOperand(i),
1535 CE2->getOperand(i), GTI.getIndexedType())) {
1536 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1537 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1538 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1541 // Ok, we ran out of things they have in common. If any leftovers
1542 // are non-zero then we have a difference, otherwise we are equal.
1543 for (; i < CE1->getNumOperands(); ++i)
1544 if (!CE1->getOperand(i)->isNullValue()) {
1545 if (isa<ConstantInt>(CE1->getOperand(i)))
1546 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1548 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1551 for (; i < CE2->getNumOperands(); ++i)
1552 if (!CE2->getOperand(i)->isNullValue()) {
1553 if (isa<ConstantInt>(CE2->getOperand(i)))
1554 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1556 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1558 return ICmpInst::ICMP_EQ;
1567 return ICmpInst::BAD_ICMP_PREDICATE;
1570 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1571 Constant *C1, Constant *C2) {
1573 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1574 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1575 VT->getNumElements());
1577 ResultTy = Type::getInt1Ty(C1->getContext());
1579 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1580 if (pred == FCmpInst::FCMP_FALSE)
1581 return Constant::getNullValue(ResultTy);
1583 if (pred == FCmpInst::FCMP_TRUE)
1584 return Constant::getAllOnesValue(ResultTy);
1586 // Handle some degenerate cases first
1587 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1588 // For EQ and NE, we can always pick a value for the undef to make the
1589 // predicate pass or fail, so we can return undef.
1590 // Also, if both operands are undef, we can return undef.
1591 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1592 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1593 return UndefValue::get(ResultTy);
1594 // Otherwise, pick the same value as the non-undef operand, and fold
1595 // it to true or false.
1596 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1599 // icmp eq/ne(null,GV) -> false/true
1600 if (C1->isNullValue()) {
1601 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1602 // Don't try to evaluate aliases. External weak GV can be null.
1603 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1604 if (pred == ICmpInst::ICMP_EQ)
1605 return ConstantInt::getFalse(C1->getContext());
1606 else if (pred == ICmpInst::ICMP_NE)
1607 return ConstantInt::getTrue(C1->getContext());
1609 // icmp eq/ne(GV,null) -> false/true
1610 } else if (C2->isNullValue()) {
1611 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1612 // Don't try to evaluate aliases. External weak GV can be null.
1613 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1614 if (pred == ICmpInst::ICMP_EQ)
1615 return ConstantInt::getFalse(C1->getContext());
1616 else if (pred == ICmpInst::ICMP_NE)
1617 return ConstantInt::getTrue(C1->getContext());
1621 // If the comparison is a comparison between two i1's, simplify it.
1622 if (C1->getType()->isIntegerTy(1)) {
1624 case ICmpInst::ICMP_EQ:
1625 if (isa<ConstantInt>(C2))
1626 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1627 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1628 case ICmpInst::ICMP_NE:
1629 return ConstantExpr::getXor(C1, C2);
1635 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1636 APInt V1 = cast<ConstantInt>(C1)->getValue();
1637 APInt V2 = cast<ConstantInt>(C2)->getValue();
1639 default: llvm_unreachable("Invalid ICmp Predicate");
1640 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1641 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1642 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1643 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1644 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1645 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1646 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1647 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1648 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1649 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1651 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1652 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1653 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1654 APFloat::cmpResult R = C1V.compare(C2V);
1656 default: llvm_unreachable("Invalid FCmp Predicate");
1657 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1658 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1659 case FCmpInst::FCMP_UNO:
1660 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1661 case FCmpInst::FCMP_ORD:
1662 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1663 case FCmpInst::FCMP_UEQ:
1664 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1665 R==APFloat::cmpEqual);
1666 case FCmpInst::FCMP_OEQ:
1667 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1668 case FCmpInst::FCMP_UNE:
1669 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1670 case FCmpInst::FCMP_ONE:
1671 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1672 R==APFloat::cmpGreaterThan);
1673 case FCmpInst::FCMP_ULT:
1674 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1675 R==APFloat::cmpLessThan);
1676 case FCmpInst::FCMP_OLT:
1677 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1678 case FCmpInst::FCMP_UGT:
1679 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1680 R==APFloat::cmpGreaterThan);
1681 case FCmpInst::FCMP_OGT:
1682 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1683 case FCmpInst::FCMP_ULE:
1684 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1685 case FCmpInst::FCMP_OLE:
1686 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1687 R==APFloat::cmpEqual);
1688 case FCmpInst::FCMP_UGE:
1689 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1690 case FCmpInst::FCMP_OGE:
1691 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1692 R==APFloat::cmpEqual);
1694 } else if (C1->getType()->isVectorTy()) {
1695 // If we can constant fold the comparison of each element, constant fold
1696 // the whole vector comparison.
1697 SmallVector<Constant*, 4> ResElts;
1698 Type *Ty = IntegerType::get(C1->getContext(), 32);
1699 // Compare the elements, producing an i1 result or constant expr.
1700 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1702 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1704 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1706 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1709 return ConstantVector::get(ResElts);
1712 if (C1->getType()->isFloatingPointTy()) {
1713 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1714 switch (evaluateFCmpRelation(C1, C2)) {
1715 default: llvm_unreachable("Unknown relation!");
1716 case FCmpInst::FCMP_UNO:
1717 case FCmpInst::FCMP_ORD:
1718 case FCmpInst::FCMP_UEQ:
1719 case FCmpInst::FCMP_UNE:
1720 case FCmpInst::FCMP_ULT:
1721 case FCmpInst::FCMP_UGT:
1722 case FCmpInst::FCMP_ULE:
1723 case FCmpInst::FCMP_UGE:
1724 case FCmpInst::FCMP_TRUE:
1725 case FCmpInst::FCMP_FALSE:
1726 case FCmpInst::BAD_FCMP_PREDICATE:
1727 break; // Couldn't determine anything about these constants.
1728 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1729 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1730 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1731 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1733 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1734 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1735 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1736 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1738 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1739 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1740 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1741 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1743 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1744 // We can only partially decide this relation.
1745 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1747 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1750 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1751 // We can only partially decide this relation.
1752 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1754 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1757 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1758 // We can only partially decide this relation.
1759 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1761 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1766 // If we evaluated the result, return it now.
1768 return ConstantInt::get(ResultTy, Result);
1771 // Evaluate the relation between the two constants, per the predicate.
1772 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1773 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1774 default: llvm_unreachable("Unknown relational!");
1775 case ICmpInst::BAD_ICMP_PREDICATE:
1776 break; // Couldn't determine anything about these constants.
1777 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1778 // If we know the constants are equal, we can decide the result of this
1779 // computation precisely.
1780 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1782 case ICmpInst::ICMP_ULT:
1784 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1786 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1790 case ICmpInst::ICMP_SLT:
1792 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1794 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1798 case ICmpInst::ICMP_UGT:
1800 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1802 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1806 case ICmpInst::ICMP_SGT:
1808 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1810 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1814 case ICmpInst::ICMP_ULE:
1815 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1816 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1818 case ICmpInst::ICMP_SLE:
1819 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1820 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1822 case ICmpInst::ICMP_UGE:
1823 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1824 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1826 case ICmpInst::ICMP_SGE:
1827 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1828 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1830 case ICmpInst::ICMP_NE:
1831 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1832 if (pred == ICmpInst::ICMP_NE) Result = 1;
1836 // If we evaluated the result, return it now.
1838 return ConstantInt::get(ResultTy, Result);
1840 // If the right hand side is a bitcast, try using its inverse to simplify
1841 // it by moving it to the left hand side. We can't do this if it would turn
1842 // a vector compare into a scalar compare or visa versa.
1843 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1844 Constant *CE2Op0 = CE2->getOperand(0);
1845 if (CE2->getOpcode() == Instruction::BitCast &&
1846 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1847 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1848 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1852 // If the left hand side is an extension, try eliminating it.
1853 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1854 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
1855 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
1856 Constant *CE1Op0 = CE1->getOperand(0);
1857 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1858 if (CE1Inverse == CE1Op0) {
1859 // Check whether we can safely truncate the right hand side.
1860 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1861 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1862 C2->getType()) == C2)
1863 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1868 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1869 (C1->isNullValue() && !C2->isNullValue())) {
1870 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1871 // other way if possible.
1872 // Also, if C1 is null and C2 isn't, flip them around.
1873 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1874 return ConstantExpr::getICmp(pred, C2, C1);
1880 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1882 template<typename IndexTy>
1883 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1884 // No indices means nothing that could be out of bounds.
1885 if (Idxs.empty()) return true;
1887 // If the first index is zero, it's in bounds.
1888 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1890 // If the first index is one and all the rest are zero, it's in bounds,
1891 // by the one-past-the-end rule.
1892 if (!cast<ConstantInt>(Idxs[0])->isOne())
1894 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1895 if (!cast<Constant>(Idxs[i])->isNullValue())
1900 template<typename IndexTy>
1901 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
1903 ArrayRef<IndexTy> Idxs) {
1904 if (Idxs.empty()) return C;
1905 Constant *Idx0 = cast<Constant>(Idxs[0]);
1906 if ((Idxs.size() == 1 && Idx0->isNullValue()))
1909 if (isa<UndefValue>(C)) {
1910 PointerType *Ptr = cast<PointerType>(C->getType());
1911 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1912 assert(Ty != 0 && "Invalid indices for GEP!");
1913 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1916 if (C->isNullValue()) {
1918 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
1919 if (!cast<Constant>(Idxs[i])->isNullValue()) {
1924 PointerType *Ptr = cast<PointerType>(C->getType());
1925 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1926 assert(Ty != 0 && "Invalid indices for GEP!");
1927 return ConstantPointerNull::get(PointerType::get(Ty,
1928 Ptr->getAddressSpace()));
1932 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1933 // Combine Indices - If the source pointer to this getelementptr instruction
1934 // is a getelementptr instruction, combine the indices of the two
1935 // getelementptr instructions into a single instruction.
1937 if (CE->getOpcode() == Instruction::GetElementPtr) {
1939 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1943 if ((LastTy && isa<SequentialType>(LastTy)) || Idx0->isNullValue()) {
1944 SmallVector<Value*, 16> NewIndices;
1945 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
1946 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1947 NewIndices.push_back(CE->getOperand(i));
1949 // Add the last index of the source with the first index of the new GEP.
1950 // Make sure to handle the case when they are actually different types.
1951 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1952 // Otherwise it must be an array.
1953 if (!Idx0->isNullValue()) {
1954 Type *IdxTy = Combined->getType();
1955 if (IdxTy != Idx0->getType()) {
1956 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
1957 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
1958 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
1959 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1962 ConstantExpr::get(Instruction::Add, Idx0, Combined);
1966 NewIndices.push_back(Combined);
1967 NewIndices.append(Idxs.begin() + 1, Idxs.end());
1969 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
1971 cast<GEPOperator>(CE)->isInBounds());
1975 // Attempt to fold casts to the same type away. For example, folding:
1977 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
1981 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
1983 // Don't fold if the cast is changing address spaces.
1984 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
1985 PointerType *SrcPtrTy =
1986 dyn_cast<PointerType>(CE->getOperand(0)->getType());
1987 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
1988 if (SrcPtrTy && DstPtrTy) {
1989 ArrayType *SrcArrayTy =
1990 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
1991 ArrayType *DstArrayTy =
1992 dyn_cast<ArrayType>(DstPtrTy->getElementType());
1993 if (SrcArrayTy && DstArrayTy
1994 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
1995 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
1996 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
2002 // Check to see if any array indices are not within the corresponding
2003 // notional array bounds. If so, try to determine if they can be factored
2004 // out into preceding dimensions.
2005 bool Unknown = false;
2006 SmallVector<Constant *, 8> NewIdxs;
2007 Type *Ty = C->getType();
2009 for (unsigned i = 0, e = Idxs.size(); i != e;
2010 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2011 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2012 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2013 if (ATy->getNumElements() <= INT64_MAX &&
2014 ATy->getNumElements() != 0 &&
2015 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2016 if (isa<SequentialType>(Prev)) {
2017 // It's out of range, but we can factor it into the prior
2019 NewIdxs.resize(Idxs.size());
2020 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2021 ATy->getNumElements());
2022 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2024 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2025 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2027 // Before adding, extend both operands to i64 to avoid
2028 // overflow trouble.
2029 if (!PrevIdx->getType()->isIntegerTy(64))
2030 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2031 Type::getInt64Ty(Div->getContext()));
2032 if (!Div->getType()->isIntegerTy(64))
2033 Div = ConstantExpr::getSExt(Div,
2034 Type::getInt64Ty(Div->getContext()));
2036 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2038 // It's out of range, but the prior dimension is a struct
2039 // so we can't do anything about it.
2044 // We don't know if it's in range or not.
2049 // If we did any factoring, start over with the adjusted indices.
2050 if (!NewIdxs.empty()) {
2051 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2052 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2053 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2056 // If all indices are known integers and normalized, we can do a simple
2057 // check for the "inbounds" property.
2058 if (!Unknown && !inBounds &&
2059 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs))
2060 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2065 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2067 ArrayRef<Constant *> Idxs) {
2068 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2071 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2073 ArrayRef<Value *> Idxs) {
2074 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);