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 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/Operator.h"
28 #include "llvm/ADT/SmallVector.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 ///< desintation 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 // Let CastInst::isEliminableCastPair do the heavy lifting.
91 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
92 Type::getInt64Ty(DstTy->getContext()));
95 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
96 Type *SrcTy = V->getType();
98 return V; // no-op cast
100 // Check to see if we are casting a pointer to an aggregate to a pointer to
101 // the first element. If so, return the appropriate GEP instruction.
102 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
103 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
104 if (PTy->getAddressSpace() == DPTy->getAddressSpace()
105 && DPTy->getElementType()->isSized()) {
106 SmallVector<Value*, 8> IdxList;
108 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
109 IdxList.push_back(Zero);
110 Type *ElTy = PTy->getElementType();
111 while (ElTy != DPTy->getElementType()) {
112 if (StructType *STy = dyn_cast<StructType>(ElTy)) {
113 if (STy->getNumElements() == 0) break;
114 ElTy = STy->getElementType(0);
115 IdxList.push_back(Zero);
116 } else if (SequentialType *STy =
117 dyn_cast<SequentialType>(ElTy)) {
118 if (ElTy->isPointerTy()) break; // Can't index into pointers!
119 ElTy = STy->getElementType();
120 IdxList.push_back(Zero);
126 if (ElTy == DPTy->getElementType())
127 // This GEP is inbounds because all indices are zero.
128 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList);
131 // Handle casts from one vector constant to another. We know that the src
132 // and dest type have the same size (otherwise its an illegal cast).
133 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
134 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
135 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
136 "Not cast between same sized vectors!");
138 // First, check for null. Undef is already handled.
139 if (isa<ConstantAggregateZero>(V))
140 return Constant::getNullValue(DestTy);
142 // Handle ConstantVector and ConstantAggregateVector.
143 return BitCastConstantVector(V, DestPTy);
146 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
147 // This allows for other simplifications (although some of them
148 // can only be handled by Analysis/ConstantFolding.cpp).
149 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
150 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
153 // Finally, implement bitcast folding now. The code below doesn't handle
155 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
156 return ConstantPointerNull::get(cast<PointerType>(DestTy));
158 // Handle integral constant input.
159 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
160 if (DestTy->isIntegerTy())
161 // Integral -> Integral. This is a no-op because the bit widths must
162 // be the same. Consequently, we just fold to V.
165 if (DestTy->isFloatingPointTy())
166 return ConstantFP::get(DestTy->getContext(),
167 APFloat(CI->getValue(),
168 !DestTy->isPPC_FP128Ty()));
170 // Otherwise, can't fold this (vector?)
174 // Handle ConstantFP input: FP -> Integral.
175 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
176 return ConstantInt::get(FP->getContext(),
177 FP->getValueAPF().bitcastToAPInt());
183 /// ExtractConstantBytes - V is an integer constant which only has a subset of
184 /// its bytes used. The bytes used are indicated by ByteStart (which is the
185 /// first byte used, counting from the least significant byte) and ByteSize,
186 /// which is the number of bytes used.
188 /// This function analyzes the specified constant to see if the specified byte
189 /// range can be returned as a simplified constant. If so, the constant is
190 /// returned, otherwise null is returned.
192 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
194 assert(C->getType()->isIntegerTy() &&
195 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
196 "Non-byte sized integer input");
197 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
198 assert(ByteSize && "Must be accessing some piece");
199 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
200 assert(ByteSize != CSize && "Should not extract everything");
202 // Constant Integers are simple.
203 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
204 APInt V = CI->getValue();
206 V = V.lshr(ByteStart*8);
207 V = V.trunc(ByteSize*8);
208 return ConstantInt::get(CI->getContext(), V);
211 // In the input is a constant expr, we might be able to recursively simplify.
212 // If not, we definitely can't do anything.
213 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
214 if (CE == 0) return 0;
216 switch (CE->getOpcode()) {
218 case Instruction::Or: {
219 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
224 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
225 if (RHSC->isAllOnesValue())
228 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
231 return ConstantExpr::getOr(LHS, RHS);
233 case Instruction::And: {
234 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
239 if (RHS->isNullValue())
242 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
245 return ConstantExpr::getAnd(LHS, RHS);
247 case Instruction::LShr: {
248 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
251 unsigned ShAmt = Amt->getZExtValue();
252 // Cannot analyze non-byte shifts.
253 if ((ShAmt & 7) != 0)
257 // If the extract is known to be all zeros, return zero.
258 if (ByteStart >= CSize-ShAmt)
259 return Constant::getNullValue(IntegerType::get(CE->getContext(),
261 // If the extract is known to be fully in the input, extract it.
262 if (ByteStart+ByteSize+ShAmt <= CSize)
263 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
265 // TODO: Handle the 'partially zero' case.
269 case Instruction::Shl: {
270 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
273 unsigned ShAmt = Amt->getZExtValue();
274 // Cannot analyze non-byte shifts.
275 if ((ShAmt & 7) != 0)
279 // If the extract is known to be all zeros, return zero.
280 if (ByteStart+ByteSize <= ShAmt)
281 return Constant::getNullValue(IntegerType::get(CE->getContext(),
283 // If the extract is known to be fully in the input, extract it.
284 if (ByteStart >= ShAmt)
285 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
287 // TODO: Handle the 'partially zero' case.
291 case Instruction::ZExt: {
292 unsigned SrcBitSize =
293 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
295 // If extracting something that is completely zero, return 0.
296 if (ByteStart*8 >= SrcBitSize)
297 return Constant::getNullValue(IntegerType::get(CE->getContext(),
300 // If exactly extracting the input, return it.
301 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
302 return CE->getOperand(0);
304 // If extracting something completely in the input, if if the input is a
305 // multiple of 8 bits, recurse.
306 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
307 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
309 // Otherwise, if extracting a subset of the input, which is not multiple of
310 // 8 bits, do a shift and trunc to get the bits.
311 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
312 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
313 Constant *Res = CE->getOperand(0);
315 Res = ConstantExpr::getLShr(Res,
316 ConstantInt::get(Res->getType(), ByteStart*8));
317 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
321 // TODO: Handle the 'partially zero' case.
327 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
328 /// on Ty, with any known factors factored out. If Folded is false,
329 /// return null if no factoring was possible, to avoid endlessly
330 /// bouncing an unfoldable expression back into the top-level folder.
332 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
334 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
335 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
336 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
337 return ConstantExpr::getNUWMul(E, N);
340 if (StructType *STy = dyn_cast<StructType>(Ty))
341 if (!STy->isPacked()) {
342 unsigned NumElems = STy->getNumElements();
343 // An empty struct has size zero.
345 return ConstantExpr::getNullValue(DestTy);
346 // Check for a struct with all members having the same size.
347 Constant *MemberSize =
348 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
350 for (unsigned i = 1; i != NumElems; ++i)
352 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
357 Constant *N = ConstantInt::get(DestTy, NumElems);
358 return ConstantExpr::getNUWMul(MemberSize, N);
362 // Pointer size doesn't depend on the pointee type, so canonicalize them
363 // to an arbitrary pointee.
364 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
365 if (!PTy->getElementType()->isIntegerTy(1))
367 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
368 PTy->getAddressSpace()),
371 // If there's no interesting folding happening, bail so that we don't create
372 // a constant that looks like it needs folding but really doesn't.
376 // Base case: Get a regular sizeof expression.
377 Constant *C = ConstantExpr::getSizeOf(Ty);
378 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
384 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
385 /// on Ty, with any known factors factored out. If Folded is false,
386 /// return null if no factoring was possible, to avoid endlessly
387 /// bouncing an unfoldable expression back into the top-level folder.
389 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
391 // The alignment of an array is equal to the alignment of the
392 // array element. Note that this is not always true for vectors.
393 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
394 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
395 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
402 if (StructType *STy = dyn_cast<StructType>(Ty)) {
403 // Packed structs always have an alignment of 1.
405 return ConstantInt::get(DestTy, 1);
407 // Otherwise, struct alignment is the maximum alignment of any member.
408 // Without target data, we can't compare much, but we can check to see
409 // if all the members have the same alignment.
410 unsigned NumElems = STy->getNumElements();
411 // An empty struct has minimal alignment.
413 return ConstantInt::get(DestTy, 1);
414 // Check for a struct with all members having the same alignment.
415 Constant *MemberAlign =
416 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
418 for (unsigned i = 1; i != NumElems; ++i)
419 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
427 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
428 // to an arbitrary pointee.
429 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
430 if (!PTy->getElementType()->isIntegerTy(1))
432 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
434 PTy->getAddressSpace()),
437 // If there's no interesting folding happening, bail so that we don't create
438 // a constant that looks like it needs folding but really doesn't.
442 // Base case: Get a regular alignof expression.
443 Constant *C = ConstantExpr::getAlignOf(Ty);
444 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
450 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
451 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
452 /// return null if no factoring was possible, to avoid endlessly
453 /// bouncing an unfoldable expression back into the top-level folder.
455 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
458 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
459 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
462 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
463 return ConstantExpr::getNUWMul(E, N);
466 if (StructType *STy = dyn_cast<StructType>(Ty))
467 if (!STy->isPacked()) {
468 unsigned NumElems = STy->getNumElements();
469 // An empty struct has no members.
472 // Check for a struct with all members having the same size.
473 Constant *MemberSize =
474 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
476 for (unsigned i = 1; i != NumElems; ++i)
478 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
483 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
488 return ConstantExpr::getNUWMul(MemberSize, N);
492 // If there's no interesting folding happening, bail so that we don't create
493 // a constant that looks like it needs folding but really doesn't.
497 // Base case: Get a regular offsetof expression.
498 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
499 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
505 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
507 if (isa<UndefValue>(V)) {
508 // zext(undef) = 0, because the top bits will be zero.
509 // sext(undef) = 0, because the top bits will all be the same.
510 // [us]itofp(undef) = 0, because the result value is bounded.
511 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
512 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
513 return Constant::getNullValue(DestTy);
514 return UndefValue::get(DestTy);
517 if (V->isNullValue() && !DestTy->isX86_MMXTy())
518 return Constant::getNullValue(DestTy);
520 // If the cast operand is a constant expression, there's a few things we can
521 // do to try to simplify it.
522 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
524 // Try hard to fold cast of cast because they are often eliminable.
525 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
526 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
527 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
528 // If all of the indexes in the GEP are null values, there is no pointer
529 // adjustment going on. We might as well cast the source pointer.
530 bool isAllNull = true;
531 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
532 if (!CE->getOperand(i)->isNullValue()) {
537 // This is casting one pointer type to another, always BitCast
538 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
542 // If the cast operand is a constant vector, perform the cast by
543 // operating on each element. In the cast of bitcasts, the element
544 // count may be mismatched; don't attempt to handle that here.
545 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
546 DestTy->isVectorTy() &&
547 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
548 SmallVector<Constant*, 16> res;
549 VectorType *DestVecTy = cast<VectorType>(DestTy);
550 Type *DstEltTy = DestVecTy->getElementType();
551 Type *Ty = IntegerType::get(V->getContext(), 32);
552 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
554 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
555 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
557 return ConstantVector::get(res);
560 // We actually have to do a cast now. Perform the cast according to the
564 llvm_unreachable("Failed to cast constant expression");
565 case Instruction::FPTrunc:
566 case Instruction::FPExt:
567 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
569 APFloat Val = FPC->getValueAPF();
570 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf :
571 DestTy->isFloatTy() ? APFloat::IEEEsingle :
572 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
573 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
574 DestTy->isFP128Ty() ? APFloat::IEEEquad :
575 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble :
577 APFloat::rmNearestTiesToEven, &ignored);
578 return ConstantFP::get(V->getContext(), Val);
580 return 0; // Can't fold.
581 case Instruction::FPToUI:
582 case Instruction::FPToSI:
583 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
584 const APFloat &V = FPC->getValueAPF();
587 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
588 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
589 APFloat::rmTowardZero, &ignored);
590 APInt Val(DestBitWidth, x);
591 return ConstantInt::get(FPC->getContext(), Val);
593 return 0; // Can't fold.
594 case Instruction::IntToPtr: //always treated as unsigned
595 if (V->isNullValue()) // Is it an integral null value?
596 return ConstantPointerNull::get(cast<PointerType>(DestTy));
597 return 0; // Other pointer types cannot be casted
598 case Instruction::PtrToInt: // always treated as unsigned
599 // Is it a null pointer value?
600 if (V->isNullValue())
601 return ConstantInt::get(DestTy, 0);
602 // If this is a sizeof-like expression, pull out multiplications by
603 // known factors to expose them to subsequent folding. If it's an
604 // alignof-like expression, factor out known factors.
605 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
606 if (CE->getOpcode() == Instruction::GetElementPtr &&
607 CE->getOperand(0)->isNullValue()) {
609 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
610 if (CE->getNumOperands() == 2) {
611 // Handle a sizeof-like expression.
612 Constant *Idx = CE->getOperand(1);
613 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
614 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
615 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
618 return ConstantExpr::getMul(C, Idx);
620 } else if (CE->getNumOperands() == 3 &&
621 CE->getOperand(1)->isNullValue()) {
622 // Handle an alignof-like expression.
623 if (StructType *STy = dyn_cast<StructType>(Ty))
624 if (!STy->isPacked()) {
625 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
627 STy->getNumElements() == 2 &&
628 STy->getElementType(0)->isIntegerTy(1)) {
629 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
632 // Handle an offsetof-like expression.
633 if (Ty->isStructTy() || Ty->isArrayTy()) {
634 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
640 // Other pointer types cannot be casted
642 case Instruction::UIToFP:
643 case Instruction::SIToFP:
644 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
645 APInt api = CI->getValue();
646 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()),
647 !DestTy->isPPC_FP128Ty() /* isEEEE */);
648 (void)apf.convertFromAPInt(api,
649 opc==Instruction::SIToFP,
650 APFloat::rmNearestTiesToEven);
651 return ConstantFP::get(V->getContext(), apf);
654 case Instruction::ZExt:
655 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
656 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
657 return ConstantInt::get(V->getContext(),
658 CI->getValue().zext(BitWidth));
661 case Instruction::SExt:
662 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
663 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
664 return ConstantInt::get(V->getContext(),
665 CI->getValue().sext(BitWidth));
668 case Instruction::Trunc: {
669 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
670 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
671 return ConstantInt::get(V->getContext(),
672 CI->getValue().trunc(DestBitWidth));
675 // The input must be a constantexpr. See if we can simplify this based on
676 // the bytes we are demanding. Only do this if the source and dest are an
677 // even multiple of a byte.
678 if ((DestBitWidth & 7) == 0 &&
679 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
680 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
685 case Instruction::BitCast:
686 return FoldBitCast(V, DestTy);
690 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
691 Constant *V1, Constant *V2) {
692 // Check for i1 and vector true/false conditions.
693 if (Cond->isNullValue()) return V2;
694 if (Cond->isAllOnesValue()) return V1;
696 // If the condition is a vector constant, fold the result elementwise.
697 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
698 SmallVector<Constant*, 16> Result;
699 Type *Ty = IntegerType::get(CondV->getContext(), 32);
700 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
701 ConstantInt *Cond = dyn_cast<ConstantInt>(CondV->getOperand(i));
702 if (Cond == 0) break;
704 Constant *V = Cond->isNullValue() ? V2 : V1;
705 Constant *Res = ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
706 Result.push_back(Res);
709 // If we were able to build the vector, return it.
710 if (Result.size() == V1->getType()->getVectorNumElements())
711 return ConstantVector::get(Result);
714 if (isa<UndefValue>(Cond)) {
715 if (isa<UndefValue>(V1)) return V1;
718 if (isa<UndefValue>(V1)) return V2;
719 if (isa<UndefValue>(V2)) return V1;
720 if (V1 == V2) return V1;
722 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
723 if (TrueVal->getOpcode() == Instruction::Select)
724 if (TrueVal->getOperand(0) == Cond)
725 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
727 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
728 if (FalseVal->getOpcode() == Instruction::Select)
729 if (FalseVal->getOperand(0) == Cond)
730 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
736 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
738 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
739 return UndefValue::get(Val->getType()->getVectorElementType());
740 if (Val->isNullValue()) // ee(zero, x) -> zero
741 return Constant::getNullValue(Val->getType()->getVectorElementType());
742 // ee({w,x,y,z}, undef) -> undef
743 if (isa<UndefValue>(Idx))
744 return UndefValue::get(Val->getType()->getVectorElementType());
746 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
747 uint64_t Index = CIdx->getZExtValue();
748 // ee({w,x,y,z}, wrong_value) -> undef
749 if (Index >= Val->getType()->getVectorNumElements())
750 return UndefValue::get(Val->getType()->getVectorElementType());
751 return Val->getAggregateElement(Index);
756 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
759 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
761 const APInt &IdxVal = CIdx->getValue();
763 SmallVector<Constant*, 16> Result;
764 Type *Ty = IntegerType::get(Val->getContext(), 32);
765 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){
767 Result.push_back(Elt);
772 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
776 return ConstantVector::get(Result);
779 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
782 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
783 Type *EltTy = V1->getType()->getVectorElementType();
785 // Undefined shuffle mask -> undefined value.
786 if (isa<UndefValue>(Mask))
787 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
789 // Don't break the bitcode reader hack.
790 if (isa<ConstantExpr>(Mask)) return 0;
792 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
794 // Loop over the shuffle mask, evaluating each element.
795 SmallVector<Constant*, 32> Result;
796 for (unsigned i = 0; i != MaskNumElts; ++i) {
797 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
799 Result.push_back(UndefValue::get(EltTy));
803 if (unsigned(Elt) >= SrcNumElts*2)
804 InElt = UndefValue::get(EltTy);
805 else if (unsigned(Elt) >= SrcNumElts) {
806 Type *Ty = IntegerType::get(V2->getContext(), 32);
808 ConstantExpr::getExtractElement(V2,
809 ConstantInt::get(Ty, Elt - SrcNumElts));
811 Type *Ty = IntegerType::get(V1->getContext(), 32);
812 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
814 Result.push_back(InElt);
817 return ConstantVector::get(Result);
820 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
821 ArrayRef<unsigned> Idxs) {
822 // Base case: no indices, so return the entire value.
826 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
827 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
832 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
834 ArrayRef<unsigned> Idxs) {
835 // Base case: no indices, so replace the entire value.
840 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
841 NumElts = ST->getNumElements();
842 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
843 NumElts = AT->getNumElements();
845 NumElts = AT->getVectorNumElements();
847 SmallVector<Constant*, 32> Result;
848 for (unsigned i = 0; i != NumElts; ++i) {
849 Constant *C = Agg->getAggregateElement(i);
850 if (C == 0) return 0;
853 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
858 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
859 return ConstantStruct::get(ST, Result);
860 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
861 return ConstantArray::get(AT, Result);
862 return ConstantVector::get(Result);
866 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
867 Constant *C1, Constant *C2) {
868 // Handle UndefValue up front.
869 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
871 case Instruction::Xor:
872 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
873 // Handle undef ^ undef -> 0 special case. This is a common
875 return Constant::getNullValue(C1->getType());
877 case Instruction::Add:
878 case Instruction::Sub:
879 return UndefValue::get(C1->getType());
880 case Instruction::And:
881 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
883 return Constant::getNullValue(C1->getType()); // undef & X -> 0
884 case Instruction::Mul: {
886 // X * undef -> undef if X is odd or undef
887 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
888 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
889 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
890 return UndefValue::get(C1->getType());
892 // X * undef -> 0 otherwise
893 return Constant::getNullValue(C1->getType());
895 case Instruction::UDiv:
896 case Instruction::SDiv:
897 // undef / 1 -> undef
898 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
899 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
903 case Instruction::URem:
904 case Instruction::SRem:
905 if (!isa<UndefValue>(C2)) // undef / X -> 0
906 return Constant::getNullValue(C1->getType());
907 return C2; // X / undef -> undef
908 case Instruction::Or: // X | undef -> -1
909 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
911 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
912 case Instruction::LShr:
913 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
914 return C1; // undef lshr undef -> undef
915 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
917 case Instruction::AShr:
918 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
919 return Constant::getAllOnesValue(C1->getType());
920 else if (isa<UndefValue>(C1))
921 return C1; // undef ashr undef -> undef
923 return C1; // X ashr undef --> X
924 case Instruction::Shl:
925 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
926 return C1; // undef shl undef -> undef
927 // undef << X -> 0 or X << undef -> 0
928 return Constant::getNullValue(C1->getType());
932 // Handle simplifications when the RHS is a constant int.
933 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
935 case Instruction::Add:
936 if (CI2->equalsInt(0)) return C1; // X + 0 == X
938 case Instruction::Sub:
939 if (CI2->equalsInt(0)) return C1; // X - 0 == X
941 case Instruction::Mul:
942 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
943 if (CI2->equalsInt(1))
944 return C1; // X * 1 == X
946 case Instruction::UDiv:
947 case Instruction::SDiv:
948 if (CI2->equalsInt(1))
949 return C1; // X / 1 == X
950 if (CI2->equalsInt(0))
951 return UndefValue::get(CI2->getType()); // X / 0 == undef
953 case Instruction::URem:
954 case Instruction::SRem:
955 if (CI2->equalsInt(1))
956 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
957 if (CI2->equalsInt(0))
958 return UndefValue::get(CI2->getType()); // X % 0 == undef
960 case Instruction::And:
961 if (CI2->isZero()) return C2; // X & 0 == 0
962 if (CI2->isAllOnesValue())
963 return C1; // X & -1 == X
965 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
966 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
967 if (CE1->getOpcode() == Instruction::ZExt) {
968 unsigned DstWidth = CI2->getType()->getBitWidth();
970 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
971 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
972 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
976 // If and'ing the address of a global with a constant, fold it.
977 if (CE1->getOpcode() == Instruction::PtrToInt &&
978 isa<GlobalValue>(CE1->getOperand(0))) {
979 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
981 // Functions are at least 4-byte aligned.
982 unsigned GVAlign = GV->getAlignment();
983 if (isa<Function>(GV))
984 GVAlign = std::max(GVAlign, 4U);
987 unsigned DstWidth = CI2->getType()->getBitWidth();
988 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
989 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
991 // If checking bits we know are clear, return zero.
992 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
993 return Constant::getNullValue(CI2->getType());
998 case Instruction::Or:
999 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1000 if (CI2->isAllOnesValue())
1001 return C2; // X | -1 == -1
1003 case Instruction::Xor:
1004 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1006 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1007 switch (CE1->getOpcode()) {
1009 case Instruction::ICmp:
1010 case Instruction::FCmp:
1011 // cmp pred ^ true -> cmp !pred
1012 assert(CI2->equalsInt(1));
1013 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1014 pred = CmpInst::getInversePredicate(pred);
1015 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1016 CE1->getOperand(1));
1020 case Instruction::AShr:
1021 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1022 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1023 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1024 return ConstantExpr::getLShr(C1, C2);
1027 } else if (isa<ConstantInt>(C1)) {
1028 // If C1 is a ConstantInt and C2 is not, swap the operands.
1029 if (Instruction::isCommutative(Opcode))
1030 return ConstantExpr::get(Opcode, C2, C1);
1033 // At this point we know neither constant is an UndefValue.
1034 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1035 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1036 const APInt &C1V = CI1->getValue();
1037 const APInt &C2V = CI2->getValue();
1041 case Instruction::Add:
1042 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1043 case Instruction::Sub:
1044 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1045 case Instruction::Mul:
1046 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1047 case Instruction::UDiv:
1048 assert(!CI2->isNullValue() && "Div by zero handled above");
1049 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1050 case Instruction::SDiv:
1051 assert(!CI2->isNullValue() && "Div by zero handled above");
1052 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1053 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1054 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1055 case Instruction::URem:
1056 assert(!CI2->isNullValue() && "Div by zero handled above");
1057 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1058 case Instruction::SRem:
1059 assert(!CI2->isNullValue() && "Div by zero handled above");
1060 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1061 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1062 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1063 case Instruction::And:
1064 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1065 case Instruction::Or:
1066 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1067 case Instruction::Xor:
1068 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1069 case Instruction::Shl: {
1070 uint32_t shiftAmt = C2V.getZExtValue();
1071 if (shiftAmt < C1V.getBitWidth())
1072 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1074 return UndefValue::get(C1->getType()); // too big shift is undef
1076 case Instruction::LShr: {
1077 uint32_t shiftAmt = C2V.getZExtValue();
1078 if (shiftAmt < C1V.getBitWidth())
1079 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1081 return UndefValue::get(C1->getType()); // too big shift is undef
1083 case Instruction::AShr: {
1084 uint32_t shiftAmt = C2V.getZExtValue();
1085 if (shiftAmt < C1V.getBitWidth())
1086 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1088 return UndefValue::get(C1->getType()); // too big shift is undef
1094 case Instruction::SDiv:
1095 case Instruction::UDiv:
1096 case Instruction::URem:
1097 case Instruction::SRem:
1098 case Instruction::LShr:
1099 case Instruction::AShr:
1100 case Instruction::Shl:
1101 if (CI1->equalsInt(0)) return C1;
1106 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1107 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1108 APFloat C1V = CFP1->getValueAPF();
1109 APFloat C2V = CFP2->getValueAPF();
1110 APFloat C3V = C1V; // copy for modification
1114 case Instruction::FAdd:
1115 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1116 return ConstantFP::get(C1->getContext(), C3V);
1117 case Instruction::FSub:
1118 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1119 return ConstantFP::get(C1->getContext(), C3V);
1120 case Instruction::FMul:
1121 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1122 return ConstantFP::get(C1->getContext(), C3V);
1123 case Instruction::FDiv:
1124 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1125 return ConstantFP::get(C1->getContext(), C3V);
1126 case Instruction::FRem:
1127 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1128 return ConstantFP::get(C1->getContext(), C3V);
1131 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1132 // Perform elementwise folding.
1133 SmallVector<Constant*, 16> Result;
1134 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1135 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1137 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1139 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1141 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1144 return ConstantVector::get(Result);
1147 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1148 // There are many possible foldings we could do here. We should probably
1149 // at least fold add of a pointer with an integer into the appropriate
1150 // getelementptr. This will improve alias analysis a bit.
1152 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1154 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1155 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1156 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1157 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1159 } else if (isa<ConstantExpr>(C2)) {
1160 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1161 // other way if possible.
1162 if (Instruction::isCommutative(Opcode))
1163 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1166 // i1 can be simplified in many cases.
1167 if (C1->getType()->isIntegerTy(1)) {
1169 case Instruction::Add:
1170 case Instruction::Sub:
1171 return ConstantExpr::getXor(C1, C2);
1172 case Instruction::Mul:
1173 return ConstantExpr::getAnd(C1, C2);
1174 case Instruction::Shl:
1175 case Instruction::LShr:
1176 case Instruction::AShr:
1177 // We can assume that C2 == 0. If it were one the result would be
1178 // undefined because the shift value is as large as the bitwidth.
1180 case Instruction::SDiv:
1181 case Instruction::UDiv:
1182 // We can assume that C2 == 1. If it were zero the result would be
1183 // undefined through division by zero.
1185 case Instruction::URem:
1186 case Instruction::SRem:
1187 // We can assume that C2 == 1. If it were zero the result would be
1188 // undefined through division by zero.
1189 return ConstantInt::getFalse(C1->getContext());
1195 // We don't know how to fold this.
1199 /// isZeroSizedType - This type is zero sized if its an array or structure of
1200 /// zero sized types. The only leaf zero sized type is an empty structure.
1201 static bool isMaybeZeroSizedType(Type *Ty) {
1202 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1203 if (STy->isOpaque()) return true; // Can't say.
1205 // If all of elements have zero size, this does too.
1206 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1207 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1210 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1211 return isMaybeZeroSizedType(ATy->getElementType());
1216 /// IdxCompare - Compare the two constants as though they were getelementptr
1217 /// indices. This allows coersion of the types to be the same thing.
1219 /// If the two constants are the "same" (after coersion), return 0. If the
1220 /// first is less than the second, return -1, if the second is less than the
1221 /// first, return 1. If the constants are not integral, return -2.
1223 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1224 if (C1 == C2) return 0;
1226 // Ok, we found a different index. If they are not ConstantInt, we can't do
1227 // anything with them.
1228 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1229 return -2; // don't know!
1231 // Ok, we have two differing integer indices. Sign extend them to be the same
1232 // type. Long is always big enough, so we use it.
1233 if (!C1->getType()->isIntegerTy(64))
1234 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1236 if (!C2->getType()->isIntegerTy(64))
1237 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1239 if (C1 == C2) return 0; // They are equal
1241 // If the type being indexed over is really just a zero sized type, there is
1242 // no pointer difference being made here.
1243 if (isMaybeZeroSizedType(ElTy))
1244 return -2; // dunno.
1246 // If they are really different, now that they are the same type, then we
1247 // found a difference!
1248 if (cast<ConstantInt>(C1)->getSExtValue() <
1249 cast<ConstantInt>(C2)->getSExtValue())
1255 /// evaluateFCmpRelation - This function determines if there is anything we can
1256 /// decide about the two constants provided. This doesn't need to handle simple
1257 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1258 /// If we can determine that the two constants have a particular relation to
1259 /// each other, we should return the corresponding FCmpInst predicate,
1260 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1261 /// ConstantFoldCompareInstruction.
1263 /// To simplify this code we canonicalize the relation so that the first
1264 /// operand is always the most "complex" of the two. We consider ConstantFP
1265 /// to be the simplest, and ConstantExprs to be the most complex.
1266 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1267 assert(V1->getType() == V2->getType() &&
1268 "Cannot compare values of different types!");
1270 // Handle degenerate case quickly
1271 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1273 if (!isa<ConstantExpr>(V1)) {
1274 if (!isa<ConstantExpr>(V2)) {
1275 // We distilled thisUse the standard constant folder for a few cases
1277 R = dyn_cast<ConstantInt>(
1278 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1279 if (R && !R->isZero())
1280 return FCmpInst::FCMP_OEQ;
1281 R = dyn_cast<ConstantInt>(
1282 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1283 if (R && !R->isZero())
1284 return FCmpInst::FCMP_OLT;
1285 R = dyn_cast<ConstantInt>(
1286 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1287 if (R && !R->isZero())
1288 return FCmpInst::FCMP_OGT;
1290 // Nothing more we can do
1291 return FCmpInst::BAD_FCMP_PREDICATE;
1294 // If the first operand is simple and second is ConstantExpr, swap operands.
1295 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1296 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1297 return FCmpInst::getSwappedPredicate(SwappedRelation);
1299 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1300 // constantexpr or a simple constant.
1301 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1302 switch (CE1->getOpcode()) {
1303 case Instruction::FPTrunc:
1304 case Instruction::FPExt:
1305 case Instruction::UIToFP:
1306 case Instruction::SIToFP:
1307 // We might be able to do something with these but we don't right now.
1313 // There are MANY other foldings that we could perform here. They will
1314 // probably be added on demand, as they seem needed.
1315 return FCmpInst::BAD_FCMP_PREDICATE;
1318 /// evaluateICmpRelation - This function determines if there is anything we can
1319 /// decide about the two constants provided. This doesn't need to handle simple
1320 /// things like integer comparisons, but should instead handle ConstantExprs
1321 /// and GlobalValues. If we can determine that the two constants have a
1322 /// particular relation to each other, we should return the corresponding ICmp
1323 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1325 /// To simplify this code we canonicalize the relation so that the first
1326 /// operand is always the most "complex" of the two. We consider simple
1327 /// constants (like ConstantInt) to be the simplest, followed by
1328 /// GlobalValues, followed by ConstantExpr's (the most complex).
1330 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1332 assert(V1->getType() == V2->getType() &&
1333 "Cannot compare different types of values!");
1334 if (V1 == V2) return ICmpInst::ICMP_EQ;
1336 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1337 !isa<BlockAddress>(V1)) {
1338 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1339 !isa<BlockAddress>(V2)) {
1340 // We distilled this down to a simple case, use the standard constant
1343 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1344 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1345 if (R && !R->isZero())
1347 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1348 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1349 if (R && !R->isZero())
1351 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1352 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1353 if (R && !R->isZero())
1356 // If we couldn't figure it out, bail.
1357 return ICmpInst::BAD_ICMP_PREDICATE;
1360 // If the first operand is simple, swap operands.
1361 ICmpInst::Predicate SwappedRelation =
1362 evaluateICmpRelation(V2, V1, isSigned);
1363 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1364 return ICmpInst::getSwappedPredicate(SwappedRelation);
1366 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1367 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1368 ICmpInst::Predicate SwappedRelation =
1369 evaluateICmpRelation(V2, V1, isSigned);
1370 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1371 return ICmpInst::getSwappedPredicate(SwappedRelation);
1372 return ICmpInst::BAD_ICMP_PREDICATE;
1375 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1376 // constant (which, since the types must match, means that it's a
1377 // ConstantPointerNull).
1378 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1379 // Don't try to decide equality of aliases.
1380 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1381 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1382 return ICmpInst::ICMP_NE;
1383 } else if (isa<BlockAddress>(V2)) {
1384 return ICmpInst::ICMP_NE; // Globals never equal labels.
1386 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1387 // GlobalVals can never be null unless they have external weak linkage.
1388 // We don't try to evaluate aliases here.
1389 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1390 return ICmpInst::ICMP_NE;
1392 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1393 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1394 ICmpInst::Predicate SwappedRelation =
1395 evaluateICmpRelation(V2, V1, isSigned);
1396 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1397 return ICmpInst::getSwappedPredicate(SwappedRelation);
1398 return ICmpInst::BAD_ICMP_PREDICATE;
1401 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1402 // constant (which, since the types must match, means that it is a
1403 // ConstantPointerNull).
1404 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1405 // Block address in another function can't equal this one, but block
1406 // addresses in the current function might be the same if blocks are
1408 if (BA2->getFunction() != BA->getFunction())
1409 return ICmpInst::ICMP_NE;
1411 // Block addresses aren't null, don't equal the address of globals.
1412 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1413 "Canonicalization guarantee!");
1414 return ICmpInst::ICMP_NE;
1417 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1418 // constantexpr, a global, block address, or a simple constant.
1419 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1420 Constant *CE1Op0 = CE1->getOperand(0);
1422 switch (CE1->getOpcode()) {
1423 case Instruction::Trunc:
1424 case Instruction::FPTrunc:
1425 case Instruction::FPExt:
1426 case Instruction::FPToUI:
1427 case Instruction::FPToSI:
1428 break; // We can't evaluate floating point casts or truncations.
1430 case Instruction::UIToFP:
1431 case Instruction::SIToFP:
1432 case Instruction::BitCast:
1433 case Instruction::ZExt:
1434 case Instruction::SExt:
1435 // If the cast is not actually changing bits, and the second operand is a
1436 // null pointer, do the comparison with the pre-casted value.
1437 if (V2->isNullValue() &&
1438 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1439 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1440 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1441 return evaluateICmpRelation(CE1Op0,
1442 Constant::getNullValue(CE1Op0->getType()),
1447 case Instruction::GetElementPtr:
1448 // Ok, since this is a getelementptr, we know that the constant has a
1449 // pointer type. Check the various cases.
1450 if (isa<ConstantPointerNull>(V2)) {
1451 // If we are comparing a GEP to a null pointer, check to see if the base
1452 // of the GEP equals the null pointer.
1453 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1454 if (GV->hasExternalWeakLinkage())
1455 // Weak linkage GVals could be zero or not. We're comparing that
1456 // to null pointer so its greater-or-equal
1457 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1459 // If its not weak linkage, the GVal must have a non-zero address
1460 // so the result is greater-than
1461 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1462 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1463 // If we are indexing from a null pointer, check to see if we have any
1464 // non-zero indices.
1465 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1466 if (!CE1->getOperand(i)->isNullValue())
1467 // Offsetting from null, must not be equal.
1468 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1469 // Only zero indexes from null, must still be zero.
1470 return ICmpInst::ICMP_EQ;
1472 // Otherwise, we can't really say if the first operand is null or not.
1473 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1474 if (isa<ConstantPointerNull>(CE1Op0)) {
1475 if (GV2->hasExternalWeakLinkage())
1476 // Weak linkage GVals could be zero or not. We're comparing it to
1477 // a null pointer, so its less-or-equal
1478 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1480 // If its not weak linkage, the GVal must have a non-zero address
1481 // so the result is less-than
1482 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1483 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1485 // If this is a getelementptr of the same global, then it must be
1486 // different. Because the types must match, the getelementptr could
1487 // only have at most one index, and because we fold getelementptr's
1488 // with a single zero index, it must be nonzero.
1489 assert(CE1->getNumOperands() == 2 &&
1490 !CE1->getOperand(1)->isNullValue() &&
1491 "Surprising getelementptr!");
1492 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1494 // If they are different globals, we don't know what the value is,
1495 // but they can't be equal.
1496 return ICmpInst::ICMP_NE;
1500 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1501 Constant *CE2Op0 = CE2->getOperand(0);
1503 // There are MANY other foldings that we could perform here. They will
1504 // probably be added on demand, as they seem needed.
1505 switch (CE2->getOpcode()) {
1507 case Instruction::GetElementPtr:
1508 // By far the most common case to handle is when the base pointers are
1509 // obviously to the same or different globals.
1510 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1511 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1512 return ICmpInst::ICMP_NE;
1513 // Ok, we know that both getelementptr instructions are based on the
1514 // same global. From this, we can precisely determine the relative
1515 // ordering of the resultant pointers.
1518 // The logic below assumes that the result of the comparison
1519 // can be determined by finding the first index that differs.
1520 // This doesn't work if there is over-indexing in any
1521 // subsequent indices, so check for that case first.
1522 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1523 !CE2->isGEPWithNoNotionalOverIndexing())
1524 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1526 // Compare all of the operands the GEP's have in common.
1527 gep_type_iterator GTI = gep_type_begin(CE1);
1528 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1530 switch (IdxCompare(CE1->getOperand(i),
1531 CE2->getOperand(i), GTI.getIndexedType())) {
1532 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1533 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1534 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1537 // Ok, we ran out of things they have in common. If any leftovers
1538 // are non-zero then we have a difference, otherwise we are equal.
1539 for (; i < CE1->getNumOperands(); ++i)
1540 if (!CE1->getOperand(i)->isNullValue()) {
1541 if (isa<ConstantInt>(CE1->getOperand(i)))
1542 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1544 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1547 for (; i < CE2->getNumOperands(); ++i)
1548 if (!CE2->getOperand(i)->isNullValue()) {
1549 if (isa<ConstantInt>(CE2->getOperand(i)))
1550 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1552 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1554 return ICmpInst::ICMP_EQ;
1563 return ICmpInst::BAD_ICMP_PREDICATE;
1566 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1567 Constant *C1, Constant *C2) {
1569 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1570 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1571 VT->getNumElements());
1573 ResultTy = Type::getInt1Ty(C1->getContext());
1575 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1576 if (pred == FCmpInst::FCMP_FALSE)
1577 return Constant::getNullValue(ResultTy);
1579 if (pred == FCmpInst::FCMP_TRUE)
1580 return Constant::getAllOnesValue(ResultTy);
1582 // Handle some degenerate cases first
1583 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1584 // For EQ and NE, we can always pick a value for the undef to make the
1585 // predicate pass or fail, so we can return undef.
1586 // Also, if both operands are undef, we can return undef.
1587 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1588 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1589 return UndefValue::get(ResultTy);
1590 // Otherwise, pick the same value as the non-undef operand, and fold
1591 // it to true or false.
1592 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1595 // icmp eq/ne(null,GV) -> false/true
1596 if (C1->isNullValue()) {
1597 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1598 // Don't try to evaluate aliases. External weak GV can be null.
1599 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1600 if (pred == ICmpInst::ICMP_EQ)
1601 return ConstantInt::getFalse(C1->getContext());
1602 else if (pred == ICmpInst::ICMP_NE)
1603 return ConstantInt::getTrue(C1->getContext());
1605 // icmp eq/ne(GV,null) -> false/true
1606 } else if (C2->isNullValue()) {
1607 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1608 // Don't try to evaluate aliases. External weak GV can be null.
1609 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1610 if (pred == ICmpInst::ICMP_EQ)
1611 return ConstantInt::getFalse(C1->getContext());
1612 else if (pred == ICmpInst::ICMP_NE)
1613 return ConstantInt::getTrue(C1->getContext());
1617 // If the comparison is a comparison between two i1's, simplify it.
1618 if (C1->getType()->isIntegerTy(1)) {
1620 case ICmpInst::ICMP_EQ:
1621 if (isa<ConstantInt>(C2))
1622 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1623 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1624 case ICmpInst::ICMP_NE:
1625 return ConstantExpr::getXor(C1, C2);
1631 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1632 APInt V1 = cast<ConstantInt>(C1)->getValue();
1633 APInt V2 = cast<ConstantInt>(C2)->getValue();
1635 default: llvm_unreachable("Invalid ICmp Predicate");
1636 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1637 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1638 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1639 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1640 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1641 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1642 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1643 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1644 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1645 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1647 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1648 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1649 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1650 APFloat::cmpResult R = C1V.compare(C2V);
1652 default: llvm_unreachable("Invalid FCmp Predicate");
1653 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1654 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1655 case FCmpInst::FCMP_UNO:
1656 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1657 case FCmpInst::FCMP_ORD:
1658 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1659 case FCmpInst::FCMP_UEQ:
1660 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1661 R==APFloat::cmpEqual);
1662 case FCmpInst::FCMP_OEQ:
1663 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1664 case FCmpInst::FCMP_UNE:
1665 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1666 case FCmpInst::FCMP_ONE:
1667 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1668 R==APFloat::cmpGreaterThan);
1669 case FCmpInst::FCMP_ULT:
1670 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1671 R==APFloat::cmpLessThan);
1672 case FCmpInst::FCMP_OLT:
1673 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1674 case FCmpInst::FCMP_UGT:
1675 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1676 R==APFloat::cmpGreaterThan);
1677 case FCmpInst::FCMP_OGT:
1678 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1679 case FCmpInst::FCMP_ULE:
1680 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1681 case FCmpInst::FCMP_OLE:
1682 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1683 R==APFloat::cmpEqual);
1684 case FCmpInst::FCMP_UGE:
1685 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1686 case FCmpInst::FCMP_OGE:
1687 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1688 R==APFloat::cmpEqual);
1690 } else if (C1->getType()->isVectorTy()) {
1691 // If we can constant fold the comparison of each element, constant fold
1692 // the whole vector comparison.
1693 SmallVector<Constant*, 4> ResElts;
1694 Type *Ty = IntegerType::get(C1->getContext(), 32);
1695 // Compare the elements, producing an i1 result or constant expr.
1696 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1698 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1700 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1702 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1705 return ConstantVector::get(ResElts);
1708 if (C1->getType()->isFloatingPointTy()) {
1709 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1710 switch (evaluateFCmpRelation(C1, C2)) {
1711 default: llvm_unreachable("Unknown relation!");
1712 case FCmpInst::FCMP_UNO:
1713 case FCmpInst::FCMP_ORD:
1714 case FCmpInst::FCMP_UEQ:
1715 case FCmpInst::FCMP_UNE:
1716 case FCmpInst::FCMP_ULT:
1717 case FCmpInst::FCMP_UGT:
1718 case FCmpInst::FCMP_ULE:
1719 case FCmpInst::FCMP_UGE:
1720 case FCmpInst::FCMP_TRUE:
1721 case FCmpInst::FCMP_FALSE:
1722 case FCmpInst::BAD_FCMP_PREDICATE:
1723 break; // Couldn't determine anything about these constants.
1724 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1725 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1726 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1727 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1729 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1730 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1731 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1732 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1734 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1735 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1736 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1737 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1739 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1740 // We can only partially decide this relation.
1741 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1743 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1746 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1747 // We can only partially decide this relation.
1748 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1750 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1753 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1754 // We can only partially decide this relation.
1755 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1757 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1762 // If we evaluated the result, return it now.
1764 return ConstantInt::get(ResultTy, Result);
1767 // Evaluate the relation between the two constants, per the predicate.
1768 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1769 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1770 default: llvm_unreachable("Unknown relational!");
1771 case ICmpInst::BAD_ICMP_PREDICATE:
1772 break; // Couldn't determine anything about these constants.
1773 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1774 // If we know the constants are equal, we can decide the result of this
1775 // computation precisely.
1776 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1778 case ICmpInst::ICMP_ULT:
1780 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1782 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1786 case ICmpInst::ICMP_SLT:
1788 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1790 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1794 case ICmpInst::ICMP_UGT:
1796 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1798 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1802 case ICmpInst::ICMP_SGT:
1804 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1806 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1810 case ICmpInst::ICMP_ULE:
1811 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1812 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1814 case ICmpInst::ICMP_SLE:
1815 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1816 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1818 case ICmpInst::ICMP_UGE:
1819 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1820 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1822 case ICmpInst::ICMP_SGE:
1823 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1824 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1826 case ICmpInst::ICMP_NE:
1827 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1828 if (pred == ICmpInst::ICMP_NE) Result = 1;
1832 // If we evaluated the result, return it now.
1834 return ConstantInt::get(ResultTy, Result);
1836 // If the right hand side is a bitcast, try using its inverse to simplify
1837 // it by moving it to the left hand side. We can't do this if it would turn
1838 // a vector compare into a scalar compare or visa versa.
1839 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1840 Constant *CE2Op0 = CE2->getOperand(0);
1841 if (CE2->getOpcode() == Instruction::BitCast &&
1842 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1843 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1844 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1848 // If the left hand side is an extension, try eliminating it.
1849 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1850 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
1851 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
1852 Constant *CE1Op0 = CE1->getOperand(0);
1853 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1854 if (CE1Inverse == CE1Op0) {
1855 // Check whether we can safely truncate the right hand side.
1856 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1857 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
1858 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1864 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1865 (C1->isNullValue() && !C2->isNullValue())) {
1866 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1867 // other way if possible.
1868 // Also, if C1 is null and C2 isn't, flip them around.
1869 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1870 return ConstantExpr::getICmp(pred, C2, C1);
1876 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1878 template<typename IndexTy>
1879 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1880 // No indices means nothing that could be out of bounds.
1881 if (Idxs.empty()) return true;
1883 // If the first index is zero, it's in bounds.
1884 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1886 // If the first index is one and all the rest are zero, it's in bounds,
1887 // by the one-past-the-end rule.
1888 if (!cast<ConstantInt>(Idxs[0])->isOne())
1890 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1891 if (!cast<Constant>(Idxs[i])->isNullValue())
1896 template<typename IndexTy>
1897 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
1899 ArrayRef<IndexTy> Idxs) {
1900 if (Idxs.empty()) return C;
1901 Constant *Idx0 = cast<Constant>(Idxs[0]);
1902 if ((Idxs.size() == 1 && Idx0->isNullValue()))
1905 if (isa<UndefValue>(C)) {
1906 PointerType *Ptr = cast<PointerType>(C->getType());
1907 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1908 assert(Ty != 0 && "Invalid indices for GEP!");
1909 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1912 if (C->isNullValue()) {
1914 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
1915 if (!cast<Constant>(Idxs[i])->isNullValue()) {
1920 PointerType *Ptr = cast<PointerType>(C->getType());
1921 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1922 assert(Ty != 0 && "Invalid indices for GEP!");
1923 return ConstantPointerNull::get(PointerType::get(Ty,
1924 Ptr->getAddressSpace()));
1928 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1929 // Combine Indices - If the source pointer to this getelementptr instruction
1930 // is a getelementptr instruction, combine the indices of the two
1931 // getelementptr instructions into a single instruction.
1933 if (CE->getOpcode() == Instruction::GetElementPtr) {
1935 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1939 if ((LastTy && isa<SequentialType>(LastTy)) || Idx0->isNullValue()) {
1940 SmallVector<Value*, 16> NewIndices;
1941 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
1942 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1943 NewIndices.push_back(CE->getOperand(i));
1945 // Add the last index of the source with the first index of the new GEP.
1946 // Make sure to handle the case when they are actually different types.
1947 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1948 // Otherwise it must be an array.
1949 if (!Idx0->isNullValue()) {
1950 Type *IdxTy = Combined->getType();
1951 if (IdxTy != Idx0->getType()) {
1952 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
1953 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
1954 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
1955 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1958 ConstantExpr::get(Instruction::Add, Idx0, Combined);
1962 NewIndices.push_back(Combined);
1963 NewIndices.append(Idxs.begin() + 1, Idxs.end());
1965 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
1967 cast<GEPOperator>(CE)->isInBounds());
1971 // Implement folding of:
1972 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
1974 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
1976 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
1977 if (PointerType *SPT =
1978 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
1979 if (ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
1980 if (ArrayType *CAT =
1981 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
1982 if (CAT->getElementType() == SAT->getElementType())
1984 ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
1989 // Check to see if any array indices are not within the corresponding
1990 // notional array bounds. If so, try to determine if they can be factored
1991 // out into preceding dimensions.
1992 bool Unknown = false;
1993 SmallVector<Constant *, 8> NewIdxs;
1994 Type *Ty = C->getType();
1996 for (unsigned i = 0, e = Idxs.size(); i != e;
1997 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
1998 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
1999 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2000 if (ATy->getNumElements() <= INT64_MAX &&
2001 ATy->getNumElements() != 0 &&
2002 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2003 if (isa<SequentialType>(Prev)) {
2004 // It's out of range, but we can factor it into the prior
2006 NewIdxs.resize(Idxs.size());
2007 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2008 ATy->getNumElements());
2009 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2011 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2012 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2014 // Before adding, extend both operands to i64 to avoid
2015 // overflow trouble.
2016 if (!PrevIdx->getType()->isIntegerTy(64))
2017 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2018 Type::getInt64Ty(Div->getContext()));
2019 if (!Div->getType()->isIntegerTy(64))
2020 Div = ConstantExpr::getSExt(Div,
2021 Type::getInt64Ty(Div->getContext()));
2023 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2025 // It's out of range, but the prior dimension is a struct
2026 // so we can't do anything about it.
2031 // We don't know if it's in range or not.
2036 // If we did any factoring, start over with the adjusted indices.
2037 if (!NewIdxs.empty()) {
2038 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2039 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2040 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2043 // If all indices are known integers and normalized, we can do a simple
2044 // check for the "inbounds" property.
2045 if (!Unknown && !inBounds &&
2046 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs))
2047 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2052 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2054 ArrayRef<Constant *> Idxs) {
2055 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2058 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2060 ArrayRef<Value *> Idxs) {
2061 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);