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/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/Support/ErrorHandling.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,
97 nullptr, FakeIntPtrTy, nullptr);
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) return nullptr;
221 switch (CE->getOpcode()) {
222 default: return nullptr;
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 nullptr; // 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 nullptr; // 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 nullptr; // 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);
692 case Instruction::AddrSpaceCast:
697 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
698 Constant *V1, Constant *V2) {
699 // Check for i1 and vector true/false conditions.
700 if (Cond->isNullValue()) return V2;
701 if (Cond->isAllOnesValue()) return V1;
703 // If the condition is a vector constant, fold the result elementwise.
704 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
705 SmallVector<Constant*, 16> Result;
706 Type *Ty = IntegerType::get(CondV->getContext(), 32);
707 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
709 Constant *V1Element = ConstantExpr::getExtractElement(V1,
710 ConstantInt::get(Ty, i));
711 Constant *V2Element = ConstantExpr::getExtractElement(V2,
712 ConstantInt::get(Ty, i));
713 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
714 if (V1Element == V2Element) {
716 } else if (isa<UndefValue>(Cond)) {
717 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
719 if (!isa<ConstantInt>(Cond)) break;
720 V = Cond->isNullValue() ? V2Element : V1Element;
725 // If we were able to build the vector, return it.
726 if (Result.size() == V1->getType()->getVectorNumElements())
727 return ConstantVector::get(Result);
730 if (isa<UndefValue>(Cond)) {
731 if (isa<UndefValue>(V1)) return V1;
734 if (isa<UndefValue>(V1)) return V2;
735 if (isa<UndefValue>(V2)) return V1;
736 if (V1 == V2) return V1;
738 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
739 if (TrueVal->getOpcode() == Instruction::Select)
740 if (TrueVal->getOperand(0) == Cond)
741 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
743 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
744 if (FalseVal->getOpcode() == Instruction::Select)
745 if (FalseVal->getOperand(0) == Cond)
746 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
752 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
754 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
755 return UndefValue::get(Val->getType()->getVectorElementType());
756 if (Val->isNullValue()) // ee(zero, x) -> zero
757 return Constant::getNullValue(Val->getType()->getVectorElementType());
758 // ee({w,x,y,z}, undef) -> undef
759 if (isa<UndefValue>(Idx))
760 return UndefValue::get(Val->getType()->getVectorElementType());
762 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
763 uint64_t Index = CIdx->getZExtValue();
764 // ee({w,x,y,z}, wrong_value) -> undef
765 if (Index >= Val->getType()->getVectorNumElements())
766 return UndefValue::get(Val->getType()->getVectorElementType());
767 return Val->getAggregateElement(Index);
772 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
775 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
776 if (!CIdx) return nullptr;
777 const APInt &IdxVal = CIdx->getValue();
779 SmallVector<Constant*, 16> Result;
780 Type *Ty = IntegerType::get(Val->getContext(), 32);
781 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){
783 Result.push_back(Elt);
788 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
792 return ConstantVector::get(Result);
795 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
798 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
799 Type *EltTy = V1->getType()->getVectorElementType();
801 // Undefined shuffle mask -> undefined value.
802 if (isa<UndefValue>(Mask))
803 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
805 // Don't break the bitcode reader hack.
806 if (isa<ConstantExpr>(Mask)) return nullptr;
808 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
810 // Loop over the shuffle mask, evaluating each element.
811 SmallVector<Constant*, 32> Result;
812 for (unsigned i = 0; i != MaskNumElts; ++i) {
813 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
815 Result.push_back(UndefValue::get(EltTy));
819 if (unsigned(Elt) >= SrcNumElts*2)
820 InElt = UndefValue::get(EltTy);
821 else if (unsigned(Elt) >= SrcNumElts) {
822 Type *Ty = IntegerType::get(V2->getContext(), 32);
824 ConstantExpr::getExtractElement(V2,
825 ConstantInt::get(Ty, Elt - SrcNumElts));
827 Type *Ty = IntegerType::get(V1->getContext(), 32);
828 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
830 Result.push_back(InElt);
833 return ConstantVector::get(Result);
836 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
837 ArrayRef<unsigned> Idxs) {
838 // Base case: no indices, so return the entire value.
842 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
843 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
848 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
850 ArrayRef<unsigned> Idxs) {
851 // Base case: no indices, so replace the entire value.
856 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
857 NumElts = ST->getNumElements();
858 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
859 NumElts = AT->getNumElements();
861 NumElts = Agg->getType()->getVectorNumElements();
863 SmallVector<Constant*, 32> Result;
864 for (unsigned i = 0; i != NumElts; ++i) {
865 Constant *C = Agg->getAggregateElement(i);
866 if (!C) return nullptr;
869 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
874 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
875 return ConstantStruct::get(ST, Result);
876 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
877 return ConstantArray::get(AT, Result);
878 return ConstantVector::get(Result);
882 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
883 Constant *C1, Constant *C2) {
884 // Handle UndefValue up front.
885 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
887 case Instruction::Xor:
888 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
889 // Handle undef ^ undef -> 0 special case. This is a common
891 return Constant::getNullValue(C1->getType());
893 case Instruction::Add:
894 case Instruction::Sub:
895 return UndefValue::get(C1->getType());
896 case Instruction::And:
897 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
899 return Constant::getNullValue(C1->getType()); // undef & X -> 0
900 case Instruction::Mul: {
902 // X * undef -> undef if X is odd or undef
903 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
904 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
905 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
906 return UndefValue::get(C1->getType());
908 // X * undef -> 0 otherwise
909 return Constant::getNullValue(C1->getType());
911 case Instruction::UDiv:
912 case Instruction::SDiv:
913 // undef / 1 -> undef
914 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
915 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
919 case Instruction::URem:
920 case Instruction::SRem:
921 if (!isa<UndefValue>(C2)) // undef / X -> 0
922 return Constant::getNullValue(C1->getType());
923 return C2; // X / undef -> undef
924 case Instruction::Or: // X | undef -> -1
925 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
927 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
928 case Instruction::LShr:
929 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
930 return C1; // undef lshr undef -> undef
931 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
933 case Instruction::AShr:
934 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
935 return Constant::getAllOnesValue(C1->getType());
936 else if (isa<UndefValue>(C1))
937 return C1; // undef ashr undef -> undef
939 return C1; // X ashr undef --> X
940 case Instruction::Shl:
941 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
942 return C1; // undef shl undef -> undef
943 // undef << X -> 0 or X << undef -> 0
944 return Constant::getNullValue(C1->getType());
948 // Handle simplifications when the RHS is a constant int.
949 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
951 case Instruction::Add:
952 if (CI2->equalsInt(0)) return C1; // X + 0 == X
954 case Instruction::Sub:
955 if (CI2->equalsInt(0)) return C1; // X - 0 == X
957 case Instruction::Mul:
958 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
959 if (CI2->equalsInt(1))
960 return C1; // X * 1 == X
962 case Instruction::UDiv:
963 case Instruction::SDiv:
964 if (CI2->equalsInt(1))
965 return C1; // X / 1 == X
966 if (CI2->equalsInt(0))
967 return UndefValue::get(CI2->getType()); // X / 0 == undef
969 case Instruction::URem:
970 case Instruction::SRem:
971 if (CI2->equalsInt(1))
972 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
973 if (CI2->equalsInt(0))
974 return UndefValue::get(CI2->getType()); // X % 0 == undef
976 case Instruction::And:
977 if (CI2->isZero()) return C2; // X & 0 == 0
978 if (CI2->isAllOnesValue())
979 return C1; // X & -1 == X
981 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
982 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
983 if (CE1->getOpcode() == Instruction::ZExt) {
984 unsigned DstWidth = CI2->getType()->getBitWidth();
986 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
987 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
988 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
992 // If and'ing the address of a global with a constant, fold it.
993 if (CE1->getOpcode() == Instruction::PtrToInt &&
994 isa<GlobalValue>(CE1->getOperand(0))) {
995 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
997 // Functions are at least 4-byte aligned.
998 unsigned GVAlign = GV->getAlignment();
999 if (isa<Function>(GV))
1000 GVAlign = std::max(GVAlign, 4U);
1003 unsigned DstWidth = CI2->getType()->getBitWidth();
1004 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1005 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1007 // If checking bits we know are clear, return zero.
1008 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1009 return Constant::getNullValue(CI2->getType());
1014 case Instruction::Or:
1015 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1016 if (CI2->isAllOnesValue())
1017 return C2; // X | -1 == -1
1019 case Instruction::Xor:
1020 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1022 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1023 switch (CE1->getOpcode()) {
1025 case Instruction::ICmp:
1026 case Instruction::FCmp:
1027 // cmp pred ^ true -> cmp !pred
1028 assert(CI2->equalsInt(1));
1029 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1030 pred = CmpInst::getInversePredicate(pred);
1031 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1032 CE1->getOperand(1));
1036 case Instruction::AShr:
1037 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1038 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1039 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1040 return ConstantExpr::getLShr(C1, C2);
1043 } else if (isa<ConstantInt>(C1)) {
1044 // If C1 is a ConstantInt and C2 is not, swap the operands.
1045 if (Instruction::isCommutative(Opcode))
1046 return ConstantExpr::get(Opcode, C2, C1);
1049 // At this point we know neither constant is an UndefValue.
1050 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1051 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1052 const APInt &C1V = CI1->getValue();
1053 const APInt &C2V = CI2->getValue();
1057 case Instruction::Add:
1058 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1059 case Instruction::Sub:
1060 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1061 case Instruction::Mul:
1062 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1063 case Instruction::UDiv:
1064 assert(!CI2->isNullValue() && "Div by zero handled above");
1065 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1066 case Instruction::SDiv:
1067 assert(!CI2->isNullValue() && "Div by zero handled above");
1068 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1069 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1070 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1071 case Instruction::URem:
1072 assert(!CI2->isNullValue() && "Div by zero handled above");
1073 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1074 case Instruction::SRem:
1075 assert(!CI2->isNullValue() && "Div by zero handled above");
1076 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1077 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1078 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1079 case Instruction::And:
1080 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1081 case Instruction::Or:
1082 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1083 case Instruction::Xor:
1084 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1085 case Instruction::Shl: {
1086 uint32_t shiftAmt = C2V.getZExtValue();
1087 if (shiftAmt < C1V.getBitWidth())
1088 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1090 return UndefValue::get(C1->getType()); // too big shift is undef
1092 case Instruction::LShr: {
1093 uint32_t shiftAmt = C2V.getZExtValue();
1094 if (shiftAmt < C1V.getBitWidth())
1095 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1097 return UndefValue::get(C1->getType()); // too big shift is undef
1099 case Instruction::AShr: {
1100 uint32_t shiftAmt = C2V.getZExtValue();
1101 if (shiftAmt < C1V.getBitWidth())
1102 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1104 return UndefValue::get(C1->getType()); // too big shift is undef
1110 case Instruction::SDiv:
1111 case Instruction::UDiv:
1112 case Instruction::URem:
1113 case Instruction::SRem:
1114 case Instruction::LShr:
1115 case Instruction::AShr:
1116 case Instruction::Shl:
1117 if (CI1->equalsInt(0)) return C1;
1122 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1123 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1124 APFloat C1V = CFP1->getValueAPF();
1125 APFloat C2V = CFP2->getValueAPF();
1126 APFloat C3V = C1V; // copy for modification
1130 case Instruction::FAdd:
1131 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1132 return ConstantFP::get(C1->getContext(), C3V);
1133 case Instruction::FSub:
1134 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1135 return ConstantFP::get(C1->getContext(), C3V);
1136 case Instruction::FMul:
1137 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1138 return ConstantFP::get(C1->getContext(), C3V);
1139 case Instruction::FDiv:
1140 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1141 return ConstantFP::get(C1->getContext(), C3V);
1142 case Instruction::FRem:
1143 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1144 return ConstantFP::get(C1->getContext(), C3V);
1147 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1148 // Perform elementwise folding.
1149 SmallVector<Constant*, 16> Result;
1150 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1151 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1153 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1155 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1157 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1160 return ConstantVector::get(Result);
1163 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1164 // There are many possible foldings we could do here. We should probably
1165 // at least fold add of a pointer with an integer into the appropriate
1166 // getelementptr. This will improve alias analysis a bit.
1168 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1170 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1171 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1172 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1173 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1175 } else if (isa<ConstantExpr>(C2)) {
1176 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1177 // other way if possible.
1178 if (Instruction::isCommutative(Opcode))
1179 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1182 // i1 can be simplified in many cases.
1183 if (C1->getType()->isIntegerTy(1)) {
1185 case Instruction::Add:
1186 case Instruction::Sub:
1187 return ConstantExpr::getXor(C1, C2);
1188 case Instruction::Mul:
1189 return ConstantExpr::getAnd(C1, C2);
1190 case Instruction::Shl:
1191 case Instruction::LShr:
1192 case Instruction::AShr:
1193 // We can assume that C2 == 0. If it were one the result would be
1194 // undefined because the shift value is as large as the bitwidth.
1196 case Instruction::SDiv:
1197 case Instruction::UDiv:
1198 // We can assume that C2 == 1. If it were zero the result would be
1199 // undefined through division by zero.
1201 case Instruction::URem:
1202 case Instruction::SRem:
1203 // We can assume that C2 == 1. If it were zero the result would be
1204 // undefined through division by zero.
1205 return ConstantInt::getFalse(C1->getContext());
1211 // We don't know how to fold this.
1215 /// isZeroSizedType - This type is zero sized if its an array or structure of
1216 /// zero sized types. The only leaf zero sized type is an empty structure.
1217 static bool isMaybeZeroSizedType(Type *Ty) {
1218 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1219 if (STy->isOpaque()) return true; // Can't say.
1221 // If all of elements have zero size, this does too.
1222 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1223 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1226 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1227 return isMaybeZeroSizedType(ATy->getElementType());
1232 /// IdxCompare - Compare the two constants as though they were getelementptr
1233 /// indices. This allows coersion of the types to be the same thing.
1235 /// If the two constants are the "same" (after coersion), return 0. If the
1236 /// first is less than the second, return -1, if the second is less than the
1237 /// first, return 1. If the constants are not integral, return -2.
1239 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1240 if (C1 == C2) return 0;
1242 // Ok, we found a different index. If they are not ConstantInt, we can't do
1243 // anything with them.
1244 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1245 return -2; // don't know!
1247 // Ok, we have two differing integer indices. Sign extend them to be the same
1248 // type. Long is always big enough, so we use it.
1249 if (!C1->getType()->isIntegerTy(64))
1250 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1252 if (!C2->getType()->isIntegerTy(64))
1253 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1255 if (C1 == C2) return 0; // They are equal
1257 // If the type being indexed over is really just a zero sized type, there is
1258 // no pointer difference being made here.
1259 if (isMaybeZeroSizedType(ElTy))
1260 return -2; // dunno.
1262 // If they are really different, now that they are the same type, then we
1263 // found a difference!
1264 if (cast<ConstantInt>(C1)->getSExtValue() <
1265 cast<ConstantInt>(C2)->getSExtValue())
1271 /// evaluateFCmpRelation - This function determines if there is anything we can
1272 /// decide about the two constants provided. This doesn't need to handle simple
1273 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1274 /// If we can determine that the two constants have a particular relation to
1275 /// each other, we should return the corresponding FCmpInst predicate,
1276 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1277 /// ConstantFoldCompareInstruction.
1279 /// To simplify this code we canonicalize the relation so that the first
1280 /// operand is always the most "complex" of the two. We consider ConstantFP
1281 /// to be the simplest, and ConstantExprs to be the most complex.
1282 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1283 assert(V1->getType() == V2->getType() &&
1284 "Cannot compare values of different types!");
1286 // Handle degenerate case quickly
1287 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1289 if (!isa<ConstantExpr>(V1)) {
1290 if (!isa<ConstantExpr>(V2)) {
1291 // We distilled thisUse the standard constant folder for a few cases
1292 ConstantInt *R = nullptr;
1293 R = dyn_cast<ConstantInt>(
1294 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1295 if (R && !R->isZero())
1296 return FCmpInst::FCMP_OEQ;
1297 R = dyn_cast<ConstantInt>(
1298 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1299 if (R && !R->isZero())
1300 return FCmpInst::FCMP_OLT;
1301 R = dyn_cast<ConstantInt>(
1302 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1303 if (R && !R->isZero())
1304 return FCmpInst::FCMP_OGT;
1306 // Nothing more we can do
1307 return FCmpInst::BAD_FCMP_PREDICATE;
1310 // If the first operand is simple and second is ConstantExpr, swap operands.
1311 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1312 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1313 return FCmpInst::getSwappedPredicate(SwappedRelation);
1315 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1316 // constantexpr or a simple constant.
1317 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1318 switch (CE1->getOpcode()) {
1319 case Instruction::FPTrunc:
1320 case Instruction::FPExt:
1321 case Instruction::UIToFP:
1322 case Instruction::SIToFP:
1323 // We might be able to do something with these but we don't right now.
1329 // There are MANY other foldings that we could perform here. They will
1330 // probably be added on demand, as they seem needed.
1331 return FCmpInst::BAD_FCMP_PREDICATE;
1334 /// evaluateICmpRelation - This function determines if there is anything we can
1335 /// decide about the two constants provided. This doesn't need to handle simple
1336 /// things like integer comparisons, but should instead handle ConstantExprs
1337 /// and GlobalValues. If we can determine that the two constants have a
1338 /// particular relation to each other, we should return the corresponding ICmp
1339 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1341 /// To simplify this code we canonicalize the relation so that the first
1342 /// operand is always the most "complex" of the two. We consider simple
1343 /// constants (like ConstantInt) to be the simplest, followed by
1344 /// GlobalValues, followed by ConstantExpr's (the most complex).
1346 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1348 assert(V1->getType() == V2->getType() &&
1349 "Cannot compare different types of values!");
1350 if (V1 == V2) return ICmpInst::ICMP_EQ;
1352 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1353 !isa<BlockAddress>(V1)) {
1354 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1355 !isa<BlockAddress>(V2)) {
1356 // We distilled this down to a simple case, use the standard constant
1358 ConstantInt *R = nullptr;
1359 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1360 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1361 if (R && !R->isZero())
1363 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1364 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1365 if (R && !R->isZero())
1367 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1368 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1369 if (R && !R->isZero())
1372 // If we couldn't figure it out, bail.
1373 return ICmpInst::BAD_ICMP_PREDICATE;
1376 // If the first operand is simple, swap operands.
1377 ICmpInst::Predicate SwappedRelation =
1378 evaluateICmpRelation(V2, V1, isSigned);
1379 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1380 return ICmpInst::getSwappedPredicate(SwappedRelation);
1382 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1383 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1384 ICmpInst::Predicate SwappedRelation =
1385 evaluateICmpRelation(V2, V1, isSigned);
1386 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1387 return ICmpInst::getSwappedPredicate(SwappedRelation);
1388 return ICmpInst::BAD_ICMP_PREDICATE;
1391 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1392 // constant (which, since the types must match, means that it's a
1393 // ConstantPointerNull).
1394 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1395 // Don't try to decide equality of aliases.
1396 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1397 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1398 return ICmpInst::ICMP_NE;
1399 } else if (isa<BlockAddress>(V2)) {
1400 return ICmpInst::ICMP_NE; // Globals never equal labels.
1402 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1403 // GlobalVals can never be null unless they have external weak linkage.
1404 // We don't try to evaluate aliases here.
1405 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1406 return ICmpInst::ICMP_NE;
1408 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1409 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1410 ICmpInst::Predicate SwappedRelation =
1411 evaluateICmpRelation(V2, V1, isSigned);
1412 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1413 return ICmpInst::getSwappedPredicate(SwappedRelation);
1414 return ICmpInst::BAD_ICMP_PREDICATE;
1417 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1418 // constant (which, since the types must match, means that it is a
1419 // ConstantPointerNull).
1420 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1421 // Block address in another function can't equal this one, but block
1422 // addresses in the current function might be the same if blocks are
1424 if (BA2->getFunction() != BA->getFunction())
1425 return ICmpInst::ICMP_NE;
1427 // Block addresses aren't null, don't equal the address of globals.
1428 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1429 "Canonicalization guarantee!");
1430 return ICmpInst::ICMP_NE;
1433 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1434 // constantexpr, a global, block address, or a simple constant.
1435 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1436 Constant *CE1Op0 = CE1->getOperand(0);
1438 switch (CE1->getOpcode()) {
1439 case Instruction::Trunc:
1440 case Instruction::FPTrunc:
1441 case Instruction::FPExt:
1442 case Instruction::FPToUI:
1443 case Instruction::FPToSI:
1444 break; // We can't evaluate floating point casts or truncations.
1446 case Instruction::UIToFP:
1447 case Instruction::SIToFP:
1448 case Instruction::BitCast:
1449 case Instruction::ZExt:
1450 case Instruction::SExt:
1451 // If the cast is not actually changing bits, and the second operand is a
1452 // null pointer, do the comparison with the pre-casted value.
1453 if (V2->isNullValue() &&
1454 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1455 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1456 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1457 return evaluateICmpRelation(CE1Op0,
1458 Constant::getNullValue(CE1Op0->getType()),
1463 case Instruction::GetElementPtr:
1464 // Ok, since this is a getelementptr, we know that the constant has a
1465 // pointer type. Check the various cases.
1466 if (isa<ConstantPointerNull>(V2)) {
1467 // If we are comparing a GEP to a null pointer, check to see if the base
1468 // of the GEP equals the null pointer.
1469 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1470 if (GV->hasExternalWeakLinkage())
1471 // Weak linkage GVals could be zero or not. We're comparing that
1472 // to null pointer so its greater-or-equal
1473 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1475 // If its not weak linkage, the GVal must have a non-zero address
1476 // so the result is greater-than
1477 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1478 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1479 // If we are indexing from a null pointer, check to see if we have any
1480 // non-zero indices.
1481 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1482 if (!CE1->getOperand(i)->isNullValue())
1483 // Offsetting from null, must not be equal.
1484 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1485 // Only zero indexes from null, must still be zero.
1486 return ICmpInst::ICMP_EQ;
1488 // Otherwise, we can't really say if the first operand is null or not.
1489 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1490 if (isa<ConstantPointerNull>(CE1Op0)) {
1491 if (GV2->hasExternalWeakLinkage())
1492 // Weak linkage GVals could be zero or not. We're comparing it to
1493 // a null pointer, so its less-or-equal
1494 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1496 // If its not weak linkage, the GVal must have a non-zero address
1497 // so the result is less-than
1498 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1499 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1501 // If this is a getelementptr of the same global, then it must be
1502 // different. Because the types must match, the getelementptr could
1503 // only have at most one index, and because we fold getelementptr's
1504 // with a single zero index, it must be nonzero.
1505 assert(CE1->getNumOperands() == 2 &&
1506 !CE1->getOperand(1)->isNullValue() &&
1507 "Surprising getelementptr!");
1508 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1510 // If they are different globals, we don't know what the value is.
1511 return ICmpInst::BAD_ICMP_PREDICATE;
1515 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1516 Constant *CE2Op0 = CE2->getOperand(0);
1518 // There are MANY other foldings that we could perform here. They will
1519 // probably be added on demand, as they seem needed.
1520 switch (CE2->getOpcode()) {
1522 case Instruction::GetElementPtr:
1523 // By far the most common case to handle is when the base pointers are
1524 // obviously to the same global.
1525 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1526 if (CE1Op0 != CE2Op0) // Don't know relative ordering.
1527 return ICmpInst::BAD_ICMP_PREDICATE;
1528 // Ok, we know that both getelementptr instructions are based on the
1529 // same global. From this, we can precisely determine the relative
1530 // ordering of the resultant pointers.
1533 // The logic below assumes that the result of the comparison
1534 // can be determined by finding the first index that differs.
1535 // This doesn't work if there is over-indexing in any
1536 // subsequent indices, so check for that case first.
1537 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1538 !CE2->isGEPWithNoNotionalOverIndexing())
1539 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1541 // Compare all of the operands the GEP's have in common.
1542 gep_type_iterator GTI = gep_type_begin(CE1);
1543 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1545 switch (IdxCompare(CE1->getOperand(i),
1546 CE2->getOperand(i), GTI.getIndexedType())) {
1547 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1548 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1549 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1552 // Ok, we ran out of things they have in common. If any leftovers
1553 // are non-zero then we have a difference, otherwise we are equal.
1554 for (; i < CE1->getNumOperands(); ++i)
1555 if (!CE1->getOperand(i)->isNullValue()) {
1556 if (isa<ConstantInt>(CE1->getOperand(i)))
1557 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1559 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1562 for (; i < CE2->getNumOperands(); ++i)
1563 if (!CE2->getOperand(i)->isNullValue()) {
1564 if (isa<ConstantInt>(CE2->getOperand(i)))
1565 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1567 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1569 return ICmpInst::ICMP_EQ;
1578 return ICmpInst::BAD_ICMP_PREDICATE;
1581 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1582 Constant *C1, Constant *C2) {
1584 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1585 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1586 VT->getNumElements());
1588 ResultTy = Type::getInt1Ty(C1->getContext());
1590 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1591 if (pred == FCmpInst::FCMP_FALSE)
1592 return Constant::getNullValue(ResultTy);
1594 if (pred == FCmpInst::FCMP_TRUE)
1595 return Constant::getAllOnesValue(ResultTy);
1597 // Handle some degenerate cases first
1598 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1599 // For EQ and NE, we can always pick a value for the undef to make the
1600 // predicate pass or fail, so we can return undef.
1601 // Also, if both operands are undef, we can return undef.
1602 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1603 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1604 return UndefValue::get(ResultTy);
1605 // Otherwise, pick the same value as the non-undef operand, and fold
1606 // it to true or false.
1607 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1610 // icmp eq/ne(null,GV) -> false/true
1611 if (C1->isNullValue()) {
1612 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1613 // Don't try to evaluate aliases. External weak GV can be null.
1614 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1615 if (pred == ICmpInst::ICMP_EQ)
1616 return ConstantInt::getFalse(C1->getContext());
1617 else if (pred == ICmpInst::ICMP_NE)
1618 return ConstantInt::getTrue(C1->getContext());
1620 // icmp eq/ne(GV,null) -> false/true
1621 } else if (C2->isNullValue()) {
1622 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1623 // Don't try to evaluate aliases. External weak GV can be null.
1624 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1625 if (pred == ICmpInst::ICMP_EQ)
1626 return ConstantInt::getFalse(C1->getContext());
1627 else if (pred == ICmpInst::ICMP_NE)
1628 return ConstantInt::getTrue(C1->getContext());
1632 // If the comparison is a comparison between two i1's, simplify it.
1633 if (C1->getType()->isIntegerTy(1)) {
1635 case ICmpInst::ICMP_EQ:
1636 if (isa<ConstantInt>(C2))
1637 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1638 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1639 case ICmpInst::ICMP_NE:
1640 return ConstantExpr::getXor(C1, C2);
1646 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1647 APInt V1 = cast<ConstantInt>(C1)->getValue();
1648 APInt V2 = cast<ConstantInt>(C2)->getValue();
1650 default: llvm_unreachable("Invalid ICmp Predicate");
1651 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1652 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1653 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1654 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1655 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1656 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1657 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1658 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1659 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1660 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1662 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1663 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1664 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1665 APFloat::cmpResult R = C1V.compare(C2V);
1667 default: llvm_unreachable("Invalid FCmp Predicate");
1668 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1669 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1670 case FCmpInst::FCMP_UNO:
1671 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1672 case FCmpInst::FCMP_ORD:
1673 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1674 case FCmpInst::FCMP_UEQ:
1675 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1676 R==APFloat::cmpEqual);
1677 case FCmpInst::FCMP_OEQ:
1678 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1679 case FCmpInst::FCMP_UNE:
1680 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1681 case FCmpInst::FCMP_ONE:
1682 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1683 R==APFloat::cmpGreaterThan);
1684 case FCmpInst::FCMP_ULT:
1685 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1686 R==APFloat::cmpLessThan);
1687 case FCmpInst::FCMP_OLT:
1688 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1689 case FCmpInst::FCMP_UGT:
1690 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1691 R==APFloat::cmpGreaterThan);
1692 case FCmpInst::FCMP_OGT:
1693 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1694 case FCmpInst::FCMP_ULE:
1695 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1696 case FCmpInst::FCMP_OLE:
1697 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1698 R==APFloat::cmpEqual);
1699 case FCmpInst::FCMP_UGE:
1700 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1701 case FCmpInst::FCMP_OGE:
1702 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1703 R==APFloat::cmpEqual);
1705 } else if (C1->getType()->isVectorTy()) {
1706 // If we can constant fold the comparison of each element, constant fold
1707 // the whole vector comparison.
1708 SmallVector<Constant*, 4> ResElts;
1709 Type *Ty = IntegerType::get(C1->getContext(), 32);
1710 // Compare the elements, producing an i1 result or constant expr.
1711 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1713 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1715 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1717 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1720 return ConstantVector::get(ResElts);
1723 if (C1->getType()->isFloatingPointTy()) {
1724 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1725 switch (evaluateFCmpRelation(C1, C2)) {
1726 default: llvm_unreachable("Unknown relation!");
1727 case FCmpInst::FCMP_UNO:
1728 case FCmpInst::FCMP_ORD:
1729 case FCmpInst::FCMP_UEQ:
1730 case FCmpInst::FCMP_UNE:
1731 case FCmpInst::FCMP_ULT:
1732 case FCmpInst::FCMP_UGT:
1733 case FCmpInst::FCMP_ULE:
1734 case FCmpInst::FCMP_UGE:
1735 case FCmpInst::FCMP_TRUE:
1736 case FCmpInst::FCMP_FALSE:
1737 case FCmpInst::BAD_FCMP_PREDICATE:
1738 break; // Couldn't determine anything about these constants.
1739 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1740 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1741 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1742 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1744 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1745 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1746 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1747 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1749 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1750 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1751 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1752 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1754 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1755 // We can only partially decide this relation.
1756 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1758 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1761 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1762 // We can only partially decide this relation.
1763 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1765 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1768 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1769 // We can only partially decide this relation.
1770 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1772 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1777 // If we evaluated the result, return it now.
1779 return ConstantInt::get(ResultTy, Result);
1782 // Evaluate the relation between the two constants, per the predicate.
1783 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1784 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1785 default: llvm_unreachable("Unknown relational!");
1786 case ICmpInst::BAD_ICMP_PREDICATE:
1787 break; // Couldn't determine anything about these constants.
1788 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1789 // If we know the constants are equal, we can decide the result of this
1790 // computation precisely.
1791 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1793 case ICmpInst::ICMP_ULT:
1795 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1797 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1801 case ICmpInst::ICMP_SLT:
1803 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1805 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1809 case ICmpInst::ICMP_UGT:
1811 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1813 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1817 case ICmpInst::ICMP_SGT:
1819 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1821 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1825 case ICmpInst::ICMP_ULE:
1826 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1827 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1829 case ICmpInst::ICMP_SLE:
1830 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1831 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1833 case ICmpInst::ICMP_UGE:
1834 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1835 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1837 case ICmpInst::ICMP_SGE:
1838 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1839 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1841 case ICmpInst::ICMP_NE:
1842 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1843 if (pred == ICmpInst::ICMP_NE) Result = 1;
1847 // If we evaluated the result, return it now.
1849 return ConstantInt::get(ResultTy, Result);
1851 // If the right hand side is a bitcast, try using its inverse to simplify
1852 // it by moving it to the left hand side. We can't do this if it would turn
1853 // a vector compare into a scalar compare or visa versa.
1854 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1855 Constant *CE2Op0 = CE2->getOperand(0);
1856 if (CE2->getOpcode() == Instruction::BitCast &&
1857 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1858 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1859 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1863 // If the left hand side is an extension, try eliminating it.
1864 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1865 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
1866 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
1867 Constant *CE1Op0 = CE1->getOperand(0);
1868 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1869 if (CE1Inverse == CE1Op0) {
1870 // Check whether we can safely truncate the right hand side.
1871 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1872 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1873 C2->getType()) == C2)
1874 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1879 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1880 (C1->isNullValue() && !C2->isNullValue())) {
1881 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1882 // other way if possible.
1883 // Also, if C1 is null and C2 isn't, flip them around.
1884 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1885 return ConstantExpr::getICmp(pred, C2, C1);
1891 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1893 template<typename IndexTy>
1894 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1895 // No indices means nothing that could be out of bounds.
1896 if (Idxs.empty()) return true;
1898 // If the first index is zero, it's in bounds.
1899 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1901 // If the first index is one and all the rest are zero, it's in bounds,
1902 // by the one-past-the-end rule.
1903 if (!cast<ConstantInt>(Idxs[0])->isOne())
1905 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1906 if (!cast<Constant>(Idxs[i])->isNullValue())
1911 /// \brief Test whether a given ConstantInt is in-range for a SequentialType.
1912 static bool isIndexInRangeOfSequentialType(const SequentialType *STy,
1913 const ConstantInt *CI) {
1914 if (const PointerType *PTy = dyn_cast<PointerType>(STy))
1915 // Only handle pointers to sized types, not pointers to functions.
1916 return PTy->getElementType()->isSized();
1918 uint64_t NumElements = 0;
1919 // Determine the number of elements in our sequential type.
1920 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
1921 NumElements = ATy->getNumElements();
1922 else if (const VectorType *VTy = dyn_cast<VectorType>(STy))
1923 NumElements = VTy->getNumElements();
1925 assert((isa<ArrayType>(STy) || NumElements > 0) &&
1926 "didn't expect non-array type to have zero elements!");
1928 // We cannot bounds check the index if it doesn't fit in an int64_t.
1929 if (CI->getValue().getActiveBits() > 64)
1932 // A negative index or an index past the end of our sequential type is
1933 // considered out-of-range.
1934 int64_t IndexVal = CI->getSExtValue();
1935 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
1938 // Otherwise, it is in-range.
1942 template<typename IndexTy>
1943 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
1945 ArrayRef<IndexTy> Idxs) {
1946 if (Idxs.empty()) return C;
1947 Constant *Idx0 = cast<Constant>(Idxs[0]);
1948 if ((Idxs.size() == 1 && Idx0->isNullValue()))
1951 if (isa<UndefValue>(C)) {
1952 PointerType *Ptr = cast<PointerType>(C->getType());
1953 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1954 assert(Ty != 0 && "Invalid indices for GEP!");
1955 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
1958 if (C->isNullValue()) {
1960 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
1961 if (!cast<Constant>(Idxs[i])->isNullValue()) {
1966 PointerType *Ptr = cast<PointerType>(C->getType());
1967 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
1968 assert(Ty != 0 && "Invalid indices for GEP!");
1969 return ConstantPointerNull::get(PointerType::get(Ty,
1970 Ptr->getAddressSpace()));
1974 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1975 // Combine Indices - If the source pointer to this getelementptr instruction
1976 // is a getelementptr instruction, combine the indices of the two
1977 // getelementptr instructions into a single instruction.
1979 if (CE->getOpcode() == Instruction::GetElementPtr) {
1980 Type *LastTy = nullptr;
1981 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1985 // We cannot combine indices if doing so would take us outside of an
1986 // array or vector. Doing otherwise could trick us if we evaluated such a
1987 // GEP as part of a load.
1989 // e.g. Consider if the original GEP was:
1990 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
1991 // i32 0, i32 0, i64 0)
1993 // If we then tried to offset it by '8' to get to the third element,
1994 // an i8, we should *not* get:
1995 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
1996 // i32 0, i32 0, i64 8)
1998 // This GEP tries to index array element '8 which runs out-of-bounds.
1999 // Subsequent evaluation would get confused and produce erroneous results.
2001 // The following prohibits such a GEP from being formed by checking to see
2002 // if the index is in-range with respect to an array or vector.
2003 bool PerformFold = false;
2004 if (Idx0->isNullValue())
2006 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy))
2007 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2008 PerformFold = isIndexInRangeOfSequentialType(STy, CI);
2011 SmallVector<Value*, 16> NewIndices;
2012 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2013 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2014 NewIndices.push_back(CE->getOperand(i));
2016 // Add the last index of the source with the first index of the new GEP.
2017 // Make sure to handle the case when they are actually different types.
2018 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2019 // Otherwise it must be an array.
2020 if (!Idx0->isNullValue()) {
2021 Type *IdxTy = Combined->getType();
2022 if (IdxTy != Idx0->getType()) {
2023 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2024 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2025 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2026 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2029 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2033 NewIndices.push_back(Combined);
2034 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2036 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
2038 cast<GEPOperator>(CE)->isInBounds());
2042 // Attempt to fold casts to the same type away. For example, folding:
2044 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2048 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2050 // Don't fold if the cast is changing address spaces.
2051 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2052 PointerType *SrcPtrTy =
2053 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2054 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2055 if (SrcPtrTy && DstPtrTy) {
2056 ArrayType *SrcArrayTy =
2057 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2058 ArrayType *DstArrayTy =
2059 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2060 if (SrcArrayTy && DstArrayTy
2061 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2062 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2063 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
2069 // Check to see if any array indices are not within the corresponding
2070 // notional array or vector bounds. If so, try to determine if they can be
2071 // factored out into preceding dimensions.
2072 bool Unknown = false;
2073 SmallVector<Constant *, 8> NewIdxs;
2074 Type *Ty = C->getType();
2075 Type *Prev = nullptr;
2076 for (unsigned i = 0, e = Idxs.size(); i != e;
2077 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2078 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2079 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty))
2080 if (CI->getSExtValue() > 0 &&
2081 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) {
2082 if (isa<SequentialType>(Prev)) {
2083 // It's out of range, but we can factor it into the prior
2085 NewIdxs.resize(Idxs.size());
2086 uint64_t NumElements = 0;
2087 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2088 NumElements = ATy->getNumElements();
2090 NumElements = cast<VectorType>(Ty)->getNumElements();
2092 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements);
2093 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2095 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2096 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2098 // Before adding, extend both operands to i64 to avoid
2099 // overflow trouble.
2100 if (!PrevIdx->getType()->isIntegerTy(64))
2101 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2102 Type::getInt64Ty(Div->getContext()));
2103 if (!Div->getType()->isIntegerTy(64))
2104 Div = ConstantExpr::getSExt(Div,
2105 Type::getInt64Ty(Div->getContext()));
2107 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2109 // It's out of range, but the prior dimension is a struct
2110 // so we can't do anything about it.
2115 // We don't know if it's in range or not.
2120 // If we did any factoring, start over with the adjusted indices.
2121 if (!NewIdxs.empty()) {
2122 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2123 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2124 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2127 // If all indices are known integers and normalized, we can do a simple
2128 // check for the "inbounds" property.
2129 if (!Unknown && !inBounds &&
2130 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs))
2131 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2136 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2138 ArrayRef<Constant *> Idxs) {
2139 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2142 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2144 ArrayRef<Value *> Idxs) {
2145 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);