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/IR/PatternMatch.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ManagedStatic.h"
34 #include "llvm/Support/MathExtras.h"
37 using namespace llvm::PatternMatch;
39 //===----------------------------------------------------------------------===//
40 // ConstantFold*Instruction Implementations
41 //===----------------------------------------------------------------------===//
43 /// BitCastConstantVector - Convert the specified vector Constant node to the
44 /// specified vector type. At this point, we know that the elements of the
45 /// input vector constant are all simple integer or FP values.
46 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
48 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
49 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
51 // If this cast changes element count then we can't handle it here:
52 // doing so requires endianness information. This should be handled by
53 // Analysis/ConstantFolding.cpp
54 unsigned NumElts = DstTy->getNumElements();
55 if (NumElts != CV->getType()->getVectorNumElements())
58 Type *DstEltTy = DstTy->getElementType();
60 SmallVector<Constant*, 16> Result;
61 Type *Ty = IntegerType::get(CV->getContext(), 32);
62 for (unsigned i = 0; i != NumElts; ++i) {
64 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
65 C = ConstantExpr::getBitCast(C, DstEltTy);
69 return ConstantVector::get(Result);
72 /// This function determines which opcode to use to fold two constant cast
73 /// expressions together. It uses CastInst::isEliminableCastPair to determine
74 /// the opcode. Consequently its just a wrapper around that function.
75 /// @brief Determine if it is valid to fold a cast of a cast
78 unsigned opc, ///< opcode of the second cast constant expression
79 ConstantExpr *Op, ///< the first cast constant expression
80 Type *DstTy ///< destination type of the first cast
82 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
83 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
84 assert(CastInst::isCast(opc) && "Invalid cast opcode");
86 // The the types and opcodes for the two Cast constant expressions
87 Type *SrcTy = Op->getOperand(0)->getType();
88 Type *MidTy = Op->getType();
89 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
90 Instruction::CastOps secondOp = Instruction::CastOps(opc);
92 // Assume that pointers are never more than 64 bits wide, and only use this
93 // for the middle type. Otherwise we could end up folding away illegal
94 // bitcasts between address spaces with different sizes.
95 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
97 // Let CastInst::isEliminableCastPair do the heavy lifting.
98 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
99 nullptr, FakeIntPtrTy, nullptr);
102 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
103 Type *SrcTy = V->getType();
105 return V; // no-op cast
107 // Check to see if we are casting a pointer to an aggregate to a pointer to
108 // the first element. If so, return the appropriate GEP instruction.
109 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
110 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
111 if (PTy->getAddressSpace() == DPTy->getAddressSpace()
112 && DPTy->getElementType()->isSized()) {
113 SmallVector<Value*, 8> IdxList;
115 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
116 IdxList.push_back(Zero);
117 Type *ElTy = PTy->getElementType();
118 while (ElTy != DPTy->getElementType()) {
119 if (StructType *STy = dyn_cast<StructType>(ElTy)) {
120 if (STy->getNumElements() == 0) break;
121 ElTy = STy->getElementType(0);
122 IdxList.push_back(Zero);
123 } else if (SequentialType *STy =
124 dyn_cast<SequentialType>(ElTy)) {
125 if (ElTy->isPointerTy()) break; // Can't index into pointers!
126 ElTy = STy->getElementType();
127 IdxList.push_back(Zero);
133 if (ElTy == DPTy->getElementType())
134 // This GEP is inbounds because all indices are zero.
135 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList);
138 // Handle casts from one vector constant to another. We know that the src
139 // and dest type have the same size (otherwise its an illegal cast).
140 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
141 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
142 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
143 "Not cast between same sized vectors!");
145 // First, check for null. Undef is already handled.
146 if (isa<ConstantAggregateZero>(V))
147 return Constant::getNullValue(DestTy);
149 // Handle ConstantVector and ConstantAggregateVector.
150 return BitCastConstantVector(V, DestPTy);
153 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
154 // This allows for other simplifications (although some of them
155 // can only be handled by Analysis/ConstantFolding.cpp).
156 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
157 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
160 // Finally, implement bitcast folding now. The code below doesn't handle
162 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
163 return ConstantPointerNull::get(cast<PointerType>(DestTy));
165 // Handle integral constant input.
166 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
167 if (DestTy->isIntegerTy())
168 // Integral -> Integral. This is a no-op because the bit widths must
169 // be the same. Consequently, we just fold to V.
172 // See note below regarding the PPC_FP128 restriction.
173 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
174 return ConstantFP::get(DestTy->getContext(),
175 APFloat(DestTy->getFltSemantics(),
178 // Otherwise, can't fold this (vector?)
182 // Handle ConstantFP input: FP -> Integral.
183 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
184 // PPC_FP128 is really the sum of two consecutive doubles, where the first
185 // double is always stored first in memory, regardless of the target
186 // endianness. The memory layout of i128, however, depends on the target
187 // endianness, and so we can't fold this without target endianness
188 // information. This should instead be handled by
189 // Analysis/ConstantFolding.cpp
190 if (FP->getType()->isPPC_FP128Ty())
193 return ConstantInt::get(FP->getContext(),
194 FP->getValueAPF().bitcastToAPInt());
201 /// ExtractConstantBytes - V is an integer constant which only has a subset of
202 /// its bytes used. The bytes used are indicated by ByteStart (which is the
203 /// first byte used, counting from the least significant byte) and ByteSize,
204 /// which is the number of bytes used.
206 /// This function analyzes the specified constant to see if the specified byte
207 /// range can be returned as a simplified constant. If so, the constant is
208 /// returned, otherwise null is returned.
210 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
212 assert(C->getType()->isIntegerTy() &&
213 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
214 "Non-byte sized integer input");
215 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
216 assert(ByteSize && "Must be accessing some piece");
217 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
218 assert(ByteSize != CSize && "Should not extract everything");
220 // Constant Integers are simple.
221 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
222 APInt V = CI->getValue();
224 V = V.lshr(ByteStart*8);
225 V = V.trunc(ByteSize*8);
226 return ConstantInt::get(CI->getContext(), V);
229 // In the input is a constant expr, we might be able to recursively simplify.
230 // If not, we definitely can't do anything.
231 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
232 if (!CE) return nullptr;
234 switch (CE->getOpcode()) {
235 default: return nullptr;
236 case Instruction::Or: {
237 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
242 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
243 if (RHSC->isAllOnesValue())
246 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
249 return ConstantExpr::getOr(LHS, RHS);
251 case Instruction::And: {
252 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
257 if (RHS->isNullValue())
260 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
263 return ConstantExpr::getAnd(LHS, RHS);
265 case Instruction::LShr: {
266 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
269 unsigned ShAmt = Amt->getZExtValue();
270 // Cannot analyze non-byte shifts.
271 if ((ShAmt & 7) != 0)
275 // If the extract is known to be all zeros, return zero.
276 if (ByteStart >= CSize-ShAmt)
277 return Constant::getNullValue(IntegerType::get(CE->getContext(),
279 // If the extract is known to be fully in the input, extract it.
280 if (ByteStart+ByteSize+ShAmt <= CSize)
281 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
283 // TODO: Handle the 'partially zero' case.
287 case Instruction::Shl: {
288 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
291 unsigned ShAmt = Amt->getZExtValue();
292 // Cannot analyze non-byte shifts.
293 if ((ShAmt & 7) != 0)
297 // If the extract is known to be all zeros, return zero.
298 if (ByteStart+ByteSize <= ShAmt)
299 return Constant::getNullValue(IntegerType::get(CE->getContext(),
301 // If the extract is known to be fully in the input, extract it.
302 if (ByteStart >= ShAmt)
303 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
305 // TODO: Handle the 'partially zero' case.
309 case Instruction::ZExt: {
310 unsigned SrcBitSize =
311 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
313 // If extracting something that is completely zero, return 0.
314 if (ByteStart*8 >= SrcBitSize)
315 return Constant::getNullValue(IntegerType::get(CE->getContext(),
318 // If exactly extracting the input, return it.
319 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
320 return CE->getOperand(0);
322 // If extracting something completely in the input, if if the input is a
323 // multiple of 8 bits, recurse.
324 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
325 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
327 // Otherwise, if extracting a subset of the input, which is not multiple of
328 // 8 bits, do a shift and trunc to get the bits.
329 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
330 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
331 Constant *Res = CE->getOperand(0);
333 Res = ConstantExpr::getLShr(Res,
334 ConstantInt::get(Res->getType(), ByteStart*8));
335 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
339 // TODO: Handle the 'partially zero' case.
345 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
346 /// on Ty, with any known factors factored out. If Folded is false,
347 /// return null if no factoring was possible, to avoid endlessly
348 /// bouncing an unfoldable expression back into the top-level folder.
350 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
352 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
353 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
354 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
355 return ConstantExpr::getNUWMul(E, N);
358 if (StructType *STy = dyn_cast<StructType>(Ty))
359 if (!STy->isPacked()) {
360 unsigned NumElems = STy->getNumElements();
361 // An empty struct has size zero.
363 return ConstantExpr::getNullValue(DestTy);
364 // Check for a struct with all members having the same size.
365 Constant *MemberSize =
366 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
368 for (unsigned i = 1; i != NumElems; ++i)
370 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
375 Constant *N = ConstantInt::get(DestTy, NumElems);
376 return ConstantExpr::getNUWMul(MemberSize, N);
380 // Pointer size doesn't depend on the pointee type, so canonicalize them
381 // to an arbitrary pointee.
382 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
383 if (!PTy->getElementType()->isIntegerTy(1))
385 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
386 PTy->getAddressSpace()),
389 // If there's no interesting folding happening, bail so that we don't create
390 // a constant that looks like it needs folding but really doesn't.
394 // Base case: Get a regular sizeof expression.
395 Constant *C = ConstantExpr::getSizeOf(Ty);
396 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
402 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
403 /// on Ty, with any known factors factored out. If Folded is false,
404 /// return null if no factoring was possible, to avoid endlessly
405 /// bouncing an unfoldable expression back into the top-level folder.
407 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
409 // The alignment of an array is equal to the alignment of the
410 // array element. Note that this is not always true for vectors.
411 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
412 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
413 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
420 if (StructType *STy = dyn_cast<StructType>(Ty)) {
421 // Packed structs always have an alignment of 1.
423 return ConstantInt::get(DestTy, 1);
425 // Otherwise, struct alignment is the maximum alignment of any member.
426 // Without target data, we can't compare much, but we can check to see
427 // if all the members have the same alignment.
428 unsigned NumElems = STy->getNumElements();
429 // An empty struct has minimal alignment.
431 return ConstantInt::get(DestTy, 1);
432 // Check for a struct with all members having the same alignment.
433 Constant *MemberAlign =
434 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
436 for (unsigned i = 1; i != NumElems; ++i)
437 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
445 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
446 // to an arbitrary pointee.
447 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
448 if (!PTy->getElementType()->isIntegerTy(1))
450 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
452 PTy->getAddressSpace()),
455 // If there's no interesting folding happening, bail so that we don't create
456 // a constant that looks like it needs folding but really doesn't.
460 // Base case: Get a regular alignof expression.
461 Constant *C = ConstantExpr::getAlignOf(Ty);
462 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
468 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
469 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
470 /// return null if no factoring was possible, to avoid endlessly
471 /// bouncing an unfoldable expression back into the top-level folder.
473 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo,
476 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
477 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
480 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
481 return ConstantExpr::getNUWMul(E, N);
484 if (StructType *STy = dyn_cast<StructType>(Ty))
485 if (!STy->isPacked()) {
486 unsigned NumElems = STy->getNumElements();
487 // An empty struct has no members.
490 // Check for a struct with all members having the same size.
491 Constant *MemberSize =
492 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
494 for (unsigned i = 1; i != NumElems; ++i)
496 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
501 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
506 return ConstantExpr::getNUWMul(MemberSize, N);
510 // If there's no interesting folding happening, bail so that we don't create
511 // a constant that looks like it needs folding but really doesn't.
515 // Base case: Get a regular offsetof expression.
516 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
517 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
523 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
525 if (isa<UndefValue>(V)) {
526 // zext(undef) = 0, because the top bits will be zero.
527 // sext(undef) = 0, because the top bits will all be the same.
528 // [us]itofp(undef) = 0, because the result value is bounded.
529 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
530 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
531 return Constant::getNullValue(DestTy);
532 return UndefValue::get(DestTy);
535 if (V->isNullValue() && !DestTy->isX86_MMXTy())
536 return Constant::getNullValue(DestTy);
538 // If the cast operand is a constant expression, there's a few things we can
539 // do to try to simplify it.
540 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
542 // Try hard to fold cast of cast because they are often eliminable.
543 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
544 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
545 } else if (CE->getOpcode() == Instruction::GetElementPtr &&
546 // Do not fold addrspacecast (gep 0, .., 0). It might make the
547 // addrspacecast uncanonicalized.
548 opc != Instruction::AddrSpaceCast) {
549 // If all of the indexes in the GEP are null values, there is no pointer
550 // adjustment going on. We might as well cast the source pointer.
551 bool isAllNull = true;
552 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
553 if (!CE->getOperand(i)->isNullValue()) {
558 // This is casting one pointer type to another, always BitCast
559 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
563 // If the cast operand is a constant vector, perform the cast by
564 // operating on each element. In the cast of bitcasts, the element
565 // count may be mismatched; don't attempt to handle that here.
566 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
567 DestTy->isVectorTy() &&
568 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
569 SmallVector<Constant*, 16> res;
570 VectorType *DestVecTy = cast<VectorType>(DestTy);
571 Type *DstEltTy = DestVecTy->getElementType();
572 Type *Ty = IntegerType::get(V->getContext(), 32);
573 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
575 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
576 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
578 return ConstantVector::get(res);
581 // We actually have to do a cast now. Perform the cast according to the
585 llvm_unreachable("Failed to cast constant expression");
586 case Instruction::FPTrunc:
587 case Instruction::FPExt:
588 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
590 APFloat Val = FPC->getValueAPF();
591 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf :
592 DestTy->isFloatTy() ? APFloat::IEEEsingle :
593 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
594 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
595 DestTy->isFP128Ty() ? APFloat::IEEEquad :
596 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble :
598 APFloat::rmNearestTiesToEven, &ignored);
599 return ConstantFP::get(V->getContext(), Val);
601 return nullptr; // Can't fold.
602 case Instruction::FPToUI:
603 case Instruction::FPToSI:
604 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
605 const APFloat &V = FPC->getValueAPF();
608 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
609 if (APFloat::opInvalidOp ==
610 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
611 APFloat::rmTowardZero, &ignored)) {
612 // Undefined behavior invoked - the destination type can't represent
613 // the input constant.
614 return UndefValue::get(DestTy);
616 APInt Val(DestBitWidth, x);
617 return ConstantInt::get(FPC->getContext(), Val);
619 return nullptr; // Can't fold.
620 case Instruction::IntToPtr: //always treated as unsigned
621 if (V->isNullValue()) // Is it an integral null value?
622 return ConstantPointerNull::get(cast<PointerType>(DestTy));
623 return nullptr; // Other pointer types cannot be casted
624 case Instruction::PtrToInt: // always treated as unsigned
625 // Is it a null pointer value?
626 if (V->isNullValue())
627 return ConstantInt::get(DestTy, 0);
628 // If this is a sizeof-like expression, pull out multiplications by
629 // known factors to expose them to subsequent folding. If it's an
630 // alignof-like expression, factor out known factors.
631 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
632 if (CE->getOpcode() == Instruction::GetElementPtr &&
633 CE->getOperand(0)->isNullValue()) {
635 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
636 if (CE->getNumOperands() == 2) {
637 // Handle a sizeof-like expression.
638 Constant *Idx = CE->getOperand(1);
639 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
640 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
641 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
644 return ConstantExpr::getMul(C, Idx);
646 } else if (CE->getNumOperands() == 3 &&
647 CE->getOperand(1)->isNullValue()) {
648 // Handle an alignof-like expression.
649 if (StructType *STy = dyn_cast<StructType>(Ty))
650 if (!STy->isPacked()) {
651 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
653 STy->getNumElements() == 2 &&
654 STy->getElementType(0)->isIntegerTy(1)) {
655 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
658 // Handle an offsetof-like expression.
659 if (Ty->isStructTy() || Ty->isArrayTy()) {
660 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
666 // Other pointer types cannot be casted
668 case Instruction::UIToFP:
669 case Instruction::SIToFP:
670 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
671 APInt api = CI->getValue();
672 APFloat apf(DestTy->getFltSemantics(),
673 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
674 if (APFloat::opOverflow &
675 apf.convertFromAPInt(api, opc==Instruction::SIToFP,
676 APFloat::rmNearestTiesToEven)) {
677 // Undefined behavior invoked - the destination type can't represent
678 // the input constant.
679 return UndefValue::get(DestTy);
681 return ConstantFP::get(V->getContext(), apf);
684 case Instruction::ZExt:
685 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
686 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
687 return ConstantInt::get(V->getContext(),
688 CI->getValue().zext(BitWidth));
691 case Instruction::SExt:
692 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
693 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
694 return ConstantInt::get(V->getContext(),
695 CI->getValue().sext(BitWidth));
698 case Instruction::Trunc: {
699 if (V->getType()->isVectorTy())
702 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
703 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
704 return ConstantInt::get(V->getContext(),
705 CI->getValue().trunc(DestBitWidth));
708 // The input must be a constantexpr. See if we can simplify this based on
709 // the bytes we are demanding. Only do this if the source and dest are an
710 // even multiple of a byte.
711 if ((DestBitWidth & 7) == 0 &&
712 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
713 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
718 case Instruction::BitCast:
719 return FoldBitCast(V, DestTy);
720 case Instruction::AddrSpaceCast:
725 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
726 Constant *V1, Constant *V2) {
727 // Check for i1 and vector true/false conditions.
728 if (Cond->isNullValue()) return V2;
729 if (Cond->isAllOnesValue()) return V1;
731 // If the condition is a vector constant, fold the result elementwise.
732 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
733 SmallVector<Constant*, 16> Result;
734 Type *Ty = IntegerType::get(CondV->getContext(), 32);
735 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
737 Constant *V1Element = ConstantExpr::getExtractElement(V1,
738 ConstantInt::get(Ty, i));
739 Constant *V2Element = ConstantExpr::getExtractElement(V2,
740 ConstantInt::get(Ty, i));
741 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
742 if (V1Element == V2Element) {
744 } else if (isa<UndefValue>(Cond)) {
745 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
747 if (!isa<ConstantInt>(Cond)) break;
748 V = Cond->isNullValue() ? V2Element : V1Element;
753 // If we were able to build the vector, return it.
754 if (Result.size() == V1->getType()->getVectorNumElements())
755 return ConstantVector::get(Result);
758 if (isa<UndefValue>(Cond)) {
759 if (isa<UndefValue>(V1)) return V1;
762 if (isa<UndefValue>(V1)) return V2;
763 if (isa<UndefValue>(V2)) return V1;
764 if (V1 == V2) return V1;
766 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
767 if (TrueVal->getOpcode() == Instruction::Select)
768 if (TrueVal->getOperand(0) == Cond)
769 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
771 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
772 if (FalseVal->getOpcode() == Instruction::Select)
773 if (FalseVal->getOperand(0) == Cond)
774 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
780 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
782 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
783 return UndefValue::get(Val->getType()->getVectorElementType());
784 if (Val->isNullValue()) // ee(zero, x) -> zero
785 return Constant::getNullValue(Val->getType()->getVectorElementType());
786 // ee({w,x,y,z}, undef) -> undef
787 if (isa<UndefValue>(Idx))
788 return UndefValue::get(Val->getType()->getVectorElementType());
790 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
791 uint64_t Index = CIdx->getZExtValue();
792 // ee({w,x,y,z}, wrong_value) -> undef
793 if (Index >= Val->getType()->getVectorNumElements())
794 return UndefValue::get(Val->getType()->getVectorElementType());
795 return Val->getAggregateElement(Index);
800 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
803 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
804 if (!CIdx) return nullptr;
805 const APInt &IdxVal = CIdx->getValue();
807 SmallVector<Constant*, 16> Result;
808 Type *Ty = IntegerType::get(Val->getContext(), 32);
809 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){
811 Result.push_back(Elt);
816 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
820 return ConstantVector::get(Result);
823 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
826 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
827 Type *EltTy = V1->getType()->getVectorElementType();
829 // Undefined shuffle mask -> undefined value.
830 if (isa<UndefValue>(Mask))
831 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
833 // Don't break the bitcode reader hack.
834 if (isa<ConstantExpr>(Mask)) return nullptr;
836 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
838 // Loop over the shuffle mask, evaluating each element.
839 SmallVector<Constant*, 32> Result;
840 for (unsigned i = 0; i != MaskNumElts; ++i) {
841 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
843 Result.push_back(UndefValue::get(EltTy));
847 if (unsigned(Elt) >= SrcNumElts*2)
848 InElt = UndefValue::get(EltTy);
849 else if (unsigned(Elt) >= SrcNumElts) {
850 Type *Ty = IntegerType::get(V2->getContext(), 32);
852 ConstantExpr::getExtractElement(V2,
853 ConstantInt::get(Ty, Elt - SrcNumElts));
855 Type *Ty = IntegerType::get(V1->getContext(), 32);
856 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
858 Result.push_back(InElt);
861 return ConstantVector::get(Result);
864 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
865 ArrayRef<unsigned> Idxs) {
866 // Base case: no indices, so return the entire value.
870 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
871 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
876 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
878 ArrayRef<unsigned> Idxs) {
879 // Base case: no indices, so replace the entire value.
884 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
885 NumElts = ST->getNumElements();
886 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
887 NumElts = AT->getNumElements();
889 NumElts = Agg->getType()->getVectorNumElements();
891 SmallVector<Constant*, 32> Result;
892 for (unsigned i = 0; i != NumElts; ++i) {
893 Constant *C = Agg->getAggregateElement(i);
894 if (!C) return nullptr;
897 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
902 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
903 return ConstantStruct::get(ST, Result);
904 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
905 return ConstantArray::get(AT, Result);
906 return ConstantVector::get(Result);
910 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
911 Constant *C1, Constant *C2) {
912 // Handle UndefValue up front.
913 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
915 case Instruction::Xor:
916 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
917 // Handle undef ^ undef -> 0 special case. This is a common
919 return Constant::getNullValue(C1->getType());
921 case Instruction::Add:
922 case Instruction::Sub:
923 return UndefValue::get(C1->getType());
924 case Instruction::And:
925 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
927 return Constant::getNullValue(C1->getType()); // undef & X -> 0
928 case Instruction::Mul: {
929 // undef * undef -> undef
930 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
933 // X * undef -> undef if X is odd
934 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
936 return UndefValue::get(C1->getType());
938 // X * undef -> 0 otherwise
939 return Constant::getNullValue(C1->getType());
941 case Instruction::SDiv:
942 case Instruction::UDiv:
943 // X / undef -> undef
944 if (match(C1, m_Zero()))
946 // undef / 0 -> undef
947 // undef / 1 -> undef
948 if (match(C2, m_Zero()) || match(C2, m_One()))
950 // undef / X -> 0 otherwise
951 return Constant::getNullValue(C1->getType());
952 case Instruction::URem:
953 case Instruction::SRem:
954 // X % undef -> undef
955 if (match(C2, m_Undef()))
957 // undef % 0 -> undef
958 if (match(C2, m_Zero()))
960 // undef % X -> 0 otherwise
961 return Constant::getNullValue(C1->getType());
962 case Instruction::Or: // X | undef -> -1
963 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
965 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
966 case Instruction::LShr:
967 // X >>l undef -> undef
968 if (isa<UndefValue>(C2))
970 // undef >>l 0 -> undef
971 if (match(C2, m_Zero()))
974 return Constant::getNullValue(C1->getType());
975 case Instruction::AShr:
976 // X >>a undef -> undef
977 if (isa<UndefValue>(C2))
979 // undef >>a 0 -> undef
980 if (match(C2, m_Zero()))
982 // TODO: undef >>a X -> undef if the shift is exact
984 return Constant::getNullValue(C1->getType());
985 case Instruction::Shl:
986 // X << undef -> undef
987 if (isa<UndefValue>(C2))
989 // undef << 0 -> undef
990 if (match(C2, m_Zero()))
993 return Constant::getNullValue(C1->getType());
997 // Handle simplifications when the RHS is a constant int.
998 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1000 case Instruction::Add:
1001 if (CI2->equalsInt(0)) return C1; // X + 0 == X
1003 case Instruction::Sub:
1004 if (CI2->equalsInt(0)) return C1; // X - 0 == X
1006 case Instruction::Mul:
1007 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
1008 if (CI2->equalsInt(1))
1009 return C1; // X * 1 == X
1011 case Instruction::UDiv:
1012 case Instruction::SDiv:
1013 if (CI2->equalsInt(1))
1014 return C1; // X / 1 == X
1015 if (CI2->equalsInt(0))
1016 return UndefValue::get(CI2->getType()); // X / 0 == undef
1018 case Instruction::URem:
1019 case Instruction::SRem:
1020 if (CI2->equalsInt(1))
1021 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1022 if (CI2->equalsInt(0))
1023 return UndefValue::get(CI2->getType()); // X % 0 == undef
1025 case Instruction::And:
1026 if (CI2->isZero()) return C2; // X & 0 == 0
1027 if (CI2->isAllOnesValue())
1028 return C1; // X & -1 == X
1030 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1031 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1032 if (CE1->getOpcode() == Instruction::ZExt) {
1033 unsigned DstWidth = CI2->getType()->getBitWidth();
1035 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1036 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1037 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1041 // If and'ing the address of a global with a constant, fold it.
1042 if (CE1->getOpcode() == Instruction::PtrToInt &&
1043 isa<GlobalValue>(CE1->getOperand(0))) {
1044 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1046 // Functions are at least 4-byte aligned.
1047 unsigned GVAlign = GV->getAlignment();
1048 if (isa<Function>(GV))
1049 GVAlign = std::max(GVAlign, 4U);
1052 unsigned DstWidth = CI2->getType()->getBitWidth();
1053 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1054 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1056 // If checking bits we know are clear, return zero.
1057 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1058 return Constant::getNullValue(CI2->getType());
1063 case Instruction::Or:
1064 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1065 if (CI2->isAllOnesValue())
1066 return C2; // X | -1 == -1
1068 case Instruction::Xor:
1069 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1071 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1072 switch (CE1->getOpcode()) {
1074 case Instruction::ICmp:
1075 case Instruction::FCmp:
1076 // cmp pred ^ true -> cmp !pred
1077 assert(CI2->equalsInt(1));
1078 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1079 pred = CmpInst::getInversePredicate(pred);
1080 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1081 CE1->getOperand(1));
1085 case Instruction::AShr:
1086 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1087 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1088 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1089 return ConstantExpr::getLShr(C1, C2);
1092 } else if (isa<ConstantInt>(C1)) {
1093 // If C1 is a ConstantInt and C2 is not, swap the operands.
1094 if (Instruction::isCommutative(Opcode))
1095 return ConstantExpr::get(Opcode, C2, C1);
1098 // At this point we know neither constant is an UndefValue.
1099 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1100 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1101 const APInt &C1V = CI1->getValue();
1102 const APInt &C2V = CI2->getValue();
1106 case Instruction::Add:
1107 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1108 case Instruction::Sub:
1109 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1110 case Instruction::Mul:
1111 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1112 case Instruction::UDiv:
1113 assert(!CI2->isNullValue() && "Div by zero handled above");
1114 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1115 case Instruction::SDiv:
1116 assert(!CI2->isNullValue() && "Div by zero handled above");
1117 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1118 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1119 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1120 case Instruction::URem:
1121 assert(!CI2->isNullValue() && "Div by zero handled above");
1122 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1123 case Instruction::SRem:
1124 assert(!CI2->isNullValue() && "Div by zero handled above");
1125 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1126 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1127 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1128 case Instruction::And:
1129 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1130 case Instruction::Or:
1131 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1132 case Instruction::Xor:
1133 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1134 case Instruction::Shl:
1135 if (C2V.ult(C1V.getBitWidth()))
1136 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1137 return UndefValue::get(C1->getType()); // too big shift is undef
1138 case Instruction::LShr:
1139 if (C2V.ult(C1V.getBitWidth()))
1140 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1141 return UndefValue::get(C1->getType()); // too big shift is undef
1142 case Instruction::AShr:
1143 if (C2V.ult(C1V.getBitWidth()))
1144 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1145 return UndefValue::get(C1->getType()); // too big shift is undef
1150 case Instruction::SDiv:
1151 case Instruction::UDiv:
1152 case Instruction::URem:
1153 case Instruction::SRem:
1154 case Instruction::LShr:
1155 case Instruction::AShr:
1156 case Instruction::Shl:
1157 if (CI1->equalsInt(0)) return C1;
1162 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1163 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1164 APFloat C1V = CFP1->getValueAPF();
1165 APFloat C2V = CFP2->getValueAPF();
1166 APFloat C3V = C1V; // copy for modification
1170 case Instruction::FAdd:
1171 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1172 return ConstantFP::get(C1->getContext(), C3V);
1173 case Instruction::FSub:
1174 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1175 return ConstantFP::get(C1->getContext(), C3V);
1176 case Instruction::FMul:
1177 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1178 return ConstantFP::get(C1->getContext(), C3V);
1179 case Instruction::FDiv:
1180 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1181 return ConstantFP::get(C1->getContext(), C3V);
1182 case Instruction::FRem:
1183 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1184 return ConstantFP::get(C1->getContext(), C3V);
1187 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1188 // Perform elementwise folding.
1189 SmallVector<Constant*, 16> Result;
1190 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1191 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1193 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1195 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1197 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1200 return ConstantVector::get(Result);
1203 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1204 // There are many possible foldings we could do here. We should probably
1205 // at least fold add of a pointer with an integer into the appropriate
1206 // getelementptr. This will improve alias analysis a bit.
1208 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1210 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1211 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1212 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1213 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1215 } else if (isa<ConstantExpr>(C2)) {
1216 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1217 // other way if possible.
1218 if (Instruction::isCommutative(Opcode))
1219 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1222 // i1 can be simplified in many cases.
1223 if (C1->getType()->isIntegerTy(1)) {
1225 case Instruction::Add:
1226 case Instruction::Sub:
1227 return ConstantExpr::getXor(C1, C2);
1228 case Instruction::Mul:
1229 return ConstantExpr::getAnd(C1, C2);
1230 case Instruction::Shl:
1231 case Instruction::LShr:
1232 case Instruction::AShr:
1233 // We can assume that C2 == 0. If it were one the result would be
1234 // undefined because the shift value is as large as the bitwidth.
1236 case Instruction::SDiv:
1237 case Instruction::UDiv:
1238 // We can assume that C2 == 1. If it were zero the result would be
1239 // undefined through division by zero.
1241 case Instruction::URem:
1242 case Instruction::SRem:
1243 // We can assume that C2 == 1. If it were zero the result would be
1244 // undefined through division by zero.
1245 return ConstantInt::getFalse(C1->getContext());
1251 // We don't know how to fold this.
1255 /// isZeroSizedType - This type is zero sized if its an array or structure of
1256 /// zero sized types. The only leaf zero sized type is an empty structure.
1257 static bool isMaybeZeroSizedType(Type *Ty) {
1258 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1259 if (STy->isOpaque()) return true; // Can't say.
1261 // If all of elements have zero size, this does too.
1262 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1263 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1266 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1267 return isMaybeZeroSizedType(ATy->getElementType());
1272 /// IdxCompare - Compare the two constants as though they were getelementptr
1273 /// indices. This allows coersion of the types to be the same thing.
1275 /// If the two constants are the "same" (after coersion), return 0. If the
1276 /// first is less than the second, return -1, if the second is less than the
1277 /// first, return 1. If the constants are not integral, return -2.
1279 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1280 if (C1 == C2) return 0;
1282 // Ok, we found a different index. If they are not ConstantInt, we can't do
1283 // anything with them.
1284 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1285 return -2; // don't know!
1287 // We cannot compare the indices if they don't fit in an int64_t.
1288 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1289 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1290 return -2; // don't know!
1292 // Ok, we have two differing integer indices. Sign extend them to be the same
1294 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1295 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1297 if (C1Val == C2Val) return 0; // They are equal
1299 // If the type being indexed over is really just a zero sized type, there is
1300 // no pointer difference being made here.
1301 if (isMaybeZeroSizedType(ElTy))
1302 return -2; // dunno.
1304 // If they are really different, now that they are the same type, then we
1305 // found a difference!
1312 /// evaluateFCmpRelation - This function determines if there is anything we can
1313 /// decide about the two constants provided. This doesn't need to handle simple
1314 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1315 /// If we can determine that the two constants have a particular relation to
1316 /// each other, we should return the corresponding FCmpInst predicate,
1317 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1318 /// ConstantFoldCompareInstruction.
1320 /// To simplify this code we canonicalize the relation so that the first
1321 /// operand is always the most "complex" of the two. We consider ConstantFP
1322 /// to be the simplest, and ConstantExprs to be the most complex.
1323 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1324 assert(V1->getType() == V2->getType() &&
1325 "Cannot compare values of different types!");
1327 // Handle degenerate case quickly
1328 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1330 if (!isa<ConstantExpr>(V1)) {
1331 if (!isa<ConstantExpr>(V2)) {
1332 // Simple case, use the standard constant folder.
1333 ConstantInt *R = nullptr;
1334 R = dyn_cast<ConstantInt>(
1335 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1336 if (R && !R->isZero())
1337 return FCmpInst::FCMP_OEQ;
1338 R = dyn_cast<ConstantInt>(
1339 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1340 if (R && !R->isZero())
1341 return FCmpInst::FCMP_OLT;
1342 R = dyn_cast<ConstantInt>(
1343 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1344 if (R && !R->isZero())
1345 return FCmpInst::FCMP_OGT;
1347 // Nothing more we can do
1348 return FCmpInst::BAD_FCMP_PREDICATE;
1351 // If the first operand is simple and second is ConstantExpr, swap operands.
1352 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1353 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1354 return FCmpInst::getSwappedPredicate(SwappedRelation);
1356 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1357 // constantexpr or a simple constant.
1358 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1359 switch (CE1->getOpcode()) {
1360 case Instruction::FPTrunc:
1361 case Instruction::FPExt:
1362 case Instruction::UIToFP:
1363 case Instruction::SIToFP:
1364 // We might be able to do something with these but we don't right now.
1370 // There are MANY other foldings that we could perform here. They will
1371 // probably be added on demand, as they seem needed.
1372 return FCmpInst::BAD_FCMP_PREDICATE;
1375 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1376 const GlobalValue *GV2) {
1377 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1378 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1380 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1381 Type *Ty = GVar->getType()->getPointerElementType();
1382 // A global with opaque type might end up being zero sized.
1385 // A global with an empty type might lie at the address of any other
1387 if (Ty->isEmptyTy())
1392 // Don't try to decide equality of aliases.
1393 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1394 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1395 return ICmpInst::ICMP_NE;
1396 return ICmpInst::BAD_ICMP_PREDICATE;
1399 /// evaluateICmpRelation - This function determines if there is anything we can
1400 /// decide about the two constants provided. This doesn't need to handle simple
1401 /// things like integer comparisons, but should instead handle ConstantExprs
1402 /// and GlobalValues. If we can determine that the two constants have a
1403 /// particular relation to each other, we should return the corresponding ICmp
1404 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1406 /// To simplify this code we canonicalize the relation so that the first
1407 /// operand is always the most "complex" of the two. We consider simple
1408 /// constants (like ConstantInt) to be the simplest, followed by
1409 /// GlobalValues, followed by ConstantExpr's (the most complex).
1411 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1413 assert(V1->getType() == V2->getType() &&
1414 "Cannot compare different types of values!");
1415 if (V1 == V2) return ICmpInst::ICMP_EQ;
1417 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1418 !isa<BlockAddress>(V1)) {
1419 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1420 !isa<BlockAddress>(V2)) {
1421 // We distilled this down to a simple case, use the standard constant
1423 ConstantInt *R = nullptr;
1424 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1425 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1426 if (R && !R->isZero())
1428 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1429 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1430 if (R && !R->isZero())
1432 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1433 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1434 if (R && !R->isZero())
1437 // If we couldn't figure it out, bail.
1438 return ICmpInst::BAD_ICMP_PREDICATE;
1441 // If the first operand is simple, swap operands.
1442 ICmpInst::Predicate SwappedRelation =
1443 evaluateICmpRelation(V2, V1, isSigned);
1444 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1445 return ICmpInst::getSwappedPredicate(SwappedRelation);
1447 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1448 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1449 ICmpInst::Predicate SwappedRelation =
1450 evaluateICmpRelation(V2, V1, isSigned);
1451 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1452 return ICmpInst::getSwappedPredicate(SwappedRelation);
1453 return ICmpInst::BAD_ICMP_PREDICATE;
1456 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1457 // constant (which, since the types must match, means that it's a
1458 // ConstantPointerNull).
1459 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1460 return areGlobalsPotentiallyEqual(GV, GV2);
1461 } else if (isa<BlockAddress>(V2)) {
1462 return ICmpInst::ICMP_NE; // Globals never equal labels.
1464 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1465 // GlobalVals can never be null unless they have external weak linkage.
1466 // We don't try to evaluate aliases here.
1467 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1468 return ICmpInst::ICMP_NE;
1470 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1471 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1472 ICmpInst::Predicate SwappedRelation =
1473 evaluateICmpRelation(V2, V1, isSigned);
1474 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1475 return ICmpInst::getSwappedPredicate(SwappedRelation);
1476 return ICmpInst::BAD_ICMP_PREDICATE;
1479 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1480 // constant (which, since the types must match, means that it is a
1481 // ConstantPointerNull).
1482 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1483 // Block address in another function can't equal this one, but block
1484 // addresses in the current function might be the same if blocks are
1486 if (BA2->getFunction() != BA->getFunction())
1487 return ICmpInst::ICMP_NE;
1489 // Block addresses aren't null, don't equal the address of globals.
1490 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1491 "Canonicalization guarantee!");
1492 return ICmpInst::ICMP_NE;
1495 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1496 // constantexpr, a global, block address, or a simple constant.
1497 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1498 Constant *CE1Op0 = CE1->getOperand(0);
1500 switch (CE1->getOpcode()) {
1501 case Instruction::Trunc:
1502 case Instruction::FPTrunc:
1503 case Instruction::FPExt:
1504 case Instruction::FPToUI:
1505 case Instruction::FPToSI:
1506 break; // We can't evaluate floating point casts or truncations.
1508 case Instruction::UIToFP:
1509 case Instruction::SIToFP:
1510 case Instruction::BitCast:
1511 case Instruction::ZExt:
1512 case Instruction::SExt:
1513 // If the cast is not actually changing bits, and the second operand is a
1514 // null pointer, do the comparison with the pre-casted value.
1515 if (V2->isNullValue() &&
1516 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1517 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1518 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1519 return evaluateICmpRelation(CE1Op0,
1520 Constant::getNullValue(CE1Op0->getType()),
1525 case Instruction::GetElementPtr: {
1526 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1527 // Ok, since this is a getelementptr, we know that the constant has a
1528 // pointer type. Check the various cases.
1529 if (isa<ConstantPointerNull>(V2)) {
1530 // If we are comparing a GEP to a null pointer, check to see if the base
1531 // of the GEP equals the null pointer.
1532 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1533 if (GV->hasExternalWeakLinkage())
1534 // Weak linkage GVals could be zero or not. We're comparing that
1535 // to null pointer so its greater-or-equal
1536 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1538 // If its not weak linkage, the GVal must have a non-zero address
1539 // so the result is greater-than
1540 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1541 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1542 // If we are indexing from a null pointer, check to see if we have any
1543 // non-zero indices.
1544 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1545 if (!CE1->getOperand(i)->isNullValue())
1546 // Offsetting from null, must not be equal.
1547 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1548 // Only zero indexes from null, must still be zero.
1549 return ICmpInst::ICMP_EQ;
1551 // Otherwise, we can't really say if the first operand is null or not.
1552 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1553 if (isa<ConstantPointerNull>(CE1Op0)) {
1554 if (GV2->hasExternalWeakLinkage())
1555 // Weak linkage GVals could be zero or not. We're comparing it to
1556 // a null pointer, so its less-or-equal
1557 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1559 // If its not weak linkage, the GVal must have a non-zero address
1560 // so the result is less-than
1561 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1562 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1564 // If this is a getelementptr of the same global, then it must be
1565 // different. Because the types must match, the getelementptr could
1566 // only have at most one index, and because we fold getelementptr's
1567 // with a single zero index, it must be nonzero.
1568 assert(CE1->getNumOperands() == 2 &&
1569 !CE1->getOperand(1)->isNullValue() &&
1570 "Surprising getelementptr!");
1571 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1573 if (CE1GEP->hasAllZeroIndices())
1574 return areGlobalsPotentiallyEqual(GV, GV2);
1575 return ICmpInst::BAD_ICMP_PREDICATE;
1579 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1580 Constant *CE2Op0 = CE2->getOperand(0);
1582 // There are MANY other foldings that we could perform here. They will
1583 // probably be added on demand, as they seem needed.
1584 switch (CE2->getOpcode()) {
1586 case Instruction::GetElementPtr:
1587 // By far the most common case to handle is when the base pointers are
1588 // obviously to the same global.
1589 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1590 // Don't know relative ordering, but check for inequality.
1591 if (CE1Op0 != CE2Op0) {
1592 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1593 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1594 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1595 cast<GlobalValue>(CE2Op0));
1596 return ICmpInst::BAD_ICMP_PREDICATE;
1598 // Ok, we know that both getelementptr instructions are based on the
1599 // same global. From this, we can precisely determine the relative
1600 // ordering of the resultant pointers.
1603 // The logic below assumes that the result of the comparison
1604 // can be determined by finding the first index that differs.
1605 // This doesn't work if there is over-indexing in any
1606 // subsequent indices, so check for that case first.
1607 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1608 !CE2->isGEPWithNoNotionalOverIndexing())
1609 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1611 // Compare all of the operands the GEP's have in common.
1612 gep_type_iterator GTI = gep_type_begin(CE1);
1613 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1615 switch (IdxCompare(CE1->getOperand(i),
1616 CE2->getOperand(i), GTI.getIndexedType())) {
1617 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1618 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1619 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1622 // Ok, we ran out of things they have in common. If any leftovers
1623 // are non-zero then we have a difference, otherwise we are equal.
1624 for (; i < CE1->getNumOperands(); ++i)
1625 if (!CE1->getOperand(i)->isNullValue()) {
1626 if (isa<ConstantInt>(CE1->getOperand(i)))
1627 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1629 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1632 for (; i < CE2->getNumOperands(); ++i)
1633 if (!CE2->getOperand(i)->isNullValue()) {
1634 if (isa<ConstantInt>(CE2->getOperand(i)))
1635 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1637 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1639 return ICmpInst::ICMP_EQ;
1649 return ICmpInst::BAD_ICMP_PREDICATE;
1652 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1653 Constant *C1, Constant *C2) {
1655 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1656 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1657 VT->getNumElements());
1659 ResultTy = Type::getInt1Ty(C1->getContext());
1661 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1662 if (pred == FCmpInst::FCMP_FALSE)
1663 return Constant::getNullValue(ResultTy);
1665 if (pred == FCmpInst::FCMP_TRUE)
1666 return Constant::getAllOnesValue(ResultTy);
1668 // Handle some degenerate cases first
1669 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1670 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1671 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1672 // For EQ and NE, we can always pick a value for the undef to make the
1673 // predicate pass or fail, so we can return undef.
1674 // Also, if both operands are undef, we can return undef for int comparison.
1675 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1676 return UndefValue::get(ResultTy);
1678 // Otherwise, for integer compare, pick the same value as the non-undef
1679 // operand, and fold it to true or false.
1680 if (isIntegerPredicate)
1681 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1683 // Choosing NaN for the undef will always make unordered comparison succeed
1684 // and ordered comparison fails.
1685 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1688 // icmp eq/ne(null,GV) -> false/true
1689 if (C1->isNullValue()) {
1690 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1691 // Don't try to evaluate aliases. External weak GV can be null.
1692 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1693 if (pred == ICmpInst::ICMP_EQ)
1694 return ConstantInt::getFalse(C1->getContext());
1695 else if (pred == ICmpInst::ICMP_NE)
1696 return ConstantInt::getTrue(C1->getContext());
1698 // icmp eq/ne(GV,null) -> false/true
1699 } else if (C2->isNullValue()) {
1700 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1701 // Don't try to evaluate aliases. External weak GV can be null.
1702 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1703 if (pred == ICmpInst::ICMP_EQ)
1704 return ConstantInt::getFalse(C1->getContext());
1705 else if (pred == ICmpInst::ICMP_NE)
1706 return ConstantInt::getTrue(C1->getContext());
1710 // If the comparison is a comparison between two i1's, simplify it.
1711 if (C1->getType()->isIntegerTy(1)) {
1713 case ICmpInst::ICMP_EQ:
1714 if (isa<ConstantInt>(C2))
1715 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1716 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1717 case ICmpInst::ICMP_NE:
1718 return ConstantExpr::getXor(C1, C2);
1724 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1725 APInt V1 = cast<ConstantInt>(C1)->getValue();
1726 APInt V2 = cast<ConstantInt>(C2)->getValue();
1728 default: llvm_unreachable("Invalid ICmp Predicate");
1729 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1730 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1731 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1732 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1733 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1734 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1735 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1736 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1737 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1738 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1740 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1741 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1742 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1743 APFloat::cmpResult R = C1V.compare(C2V);
1745 default: llvm_unreachable("Invalid FCmp Predicate");
1746 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1747 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1748 case FCmpInst::FCMP_UNO:
1749 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1750 case FCmpInst::FCMP_ORD:
1751 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1752 case FCmpInst::FCMP_UEQ:
1753 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1754 R==APFloat::cmpEqual);
1755 case FCmpInst::FCMP_OEQ:
1756 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1757 case FCmpInst::FCMP_UNE:
1758 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1759 case FCmpInst::FCMP_ONE:
1760 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1761 R==APFloat::cmpGreaterThan);
1762 case FCmpInst::FCMP_ULT:
1763 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1764 R==APFloat::cmpLessThan);
1765 case FCmpInst::FCMP_OLT:
1766 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1767 case FCmpInst::FCMP_UGT:
1768 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1769 R==APFloat::cmpGreaterThan);
1770 case FCmpInst::FCMP_OGT:
1771 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1772 case FCmpInst::FCMP_ULE:
1773 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1774 case FCmpInst::FCMP_OLE:
1775 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1776 R==APFloat::cmpEqual);
1777 case FCmpInst::FCMP_UGE:
1778 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1779 case FCmpInst::FCMP_OGE:
1780 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1781 R==APFloat::cmpEqual);
1783 } else if (C1->getType()->isVectorTy()) {
1784 // If we can constant fold the comparison of each element, constant fold
1785 // the whole vector comparison.
1786 SmallVector<Constant*, 4> ResElts;
1787 Type *Ty = IntegerType::get(C1->getContext(), 32);
1788 // Compare the elements, producing an i1 result or constant expr.
1789 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1791 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1793 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1795 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1798 return ConstantVector::get(ResElts);
1801 if (C1->getType()->isFloatingPointTy() &&
1802 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1803 // infinite recursive loop
1804 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1805 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1806 switch (evaluateFCmpRelation(C1, C2)) {
1807 default: llvm_unreachable("Unknown relation!");
1808 case FCmpInst::FCMP_UNO:
1809 case FCmpInst::FCMP_ORD:
1810 case FCmpInst::FCMP_UEQ:
1811 case FCmpInst::FCMP_UNE:
1812 case FCmpInst::FCMP_ULT:
1813 case FCmpInst::FCMP_UGT:
1814 case FCmpInst::FCMP_ULE:
1815 case FCmpInst::FCMP_UGE:
1816 case FCmpInst::FCMP_TRUE:
1817 case FCmpInst::FCMP_FALSE:
1818 case FCmpInst::BAD_FCMP_PREDICATE:
1819 break; // Couldn't determine anything about these constants.
1820 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1821 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1822 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1823 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1825 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1826 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1827 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1828 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1830 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1831 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1832 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1833 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1835 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1836 // We can only partially decide this relation.
1837 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1839 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1842 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1843 // We can only partially decide this relation.
1844 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1846 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1849 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1850 // We can only partially decide this relation.
1851 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1853 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1858 // If we evaluated the result, return it now.
1860 return ConstantInt::get(ResultTy, Result);
1863 // Evaluate the relation between the two constants, per the predicate.
1864 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1865 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1866 default: llvm_unreachable("Unknown relational!");
1867 case ICmpInst::BAD_ICMP_PREDICATE:
1868 break; // Couldn't determine anything about these constants.
1869 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1870 // If we know the constants are equal, we can decide the result of this
1871 // computation precisely.
1872 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1874 case ICmpInst::ICMP_ULT:
1876 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1878 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1882 case ICmpInst::ICMP_SLT:
1884 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1886 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1890 case ICmpInst::ICMP_UGT:
1892 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1894 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1898 case ICmpInst::ICMP_SGT:
1900 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1902 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1906 case ICmpInst::ICMP_ULE:
1907 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1908 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1910 case ICmpInst::ICMP_SLE:
1911 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1912 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1914 case ICmpInst::ICMP_UGE:
1915 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1916 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1918 case ICmpInst::ICMP_SGE:
1919 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1920 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1922 case ICmpInst::ICMP_NE:
1923 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1924 if (pred == ICmpInst::ICMP_NE) Result = 1;
1928 // If we evaluated the result, return it now.
1930 return ConstantInt::get(ResultTy, Result);
1932 // If the right hand side is a bitcast, try using its inverse to simplify
1933 // it by moving it to the left hand side. We can't do this if it would turn
1934 // a vector compare into a scalar compare or visa versa.
1935 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1936 Constant *CE2Op0 = CE2->getOperand(0);
1937 if (CE2->getOpcode() == Instruction::BitCast &&
1938 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1939 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1940 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1944 // If the left hand side is an extension, try eliminating it.
1945 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1946 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
1947 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
1948 Constant *CE1Op0 = CE1->getOperand(0);
1949 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1950 if (CE1Inverse == CE1Op0) {
1951 // Check whether we can safely truncate the right hand side.
1952 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1953 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1954 C2->getType()) == C2)
1955 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1960 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1961 (C1->isNullValue() && !C2->isNullValue())) {
1962 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1963 // other way if possible.
1964 // Also, if C1 is null and C2 isn't, flip them around.
1965 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1966 return ConstantExpr::getICmp(pred, C2, C1);
1972 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
1974 template<typename IndexTy>
1975 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
1976 // No indices means nothing that could be out of bounds.
1977 if (Idxs.empty()) return true;
1979 // If the first index is zero, it's in bounds.
1980 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
1982 // If the first index is one and all the rest are zero, it's in bounds,
1983 // by the one-past-the-end rule.
1984 if (!cast<ConstantInt>(Idxs[0])->isOne())
1986 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
1987 if (!cast<Constant>(Idxs[i])->isNullValue())
1992 /// \brief Test whether a given ConstantInt is in-range for a SequentialType.
1993 static bool isIndexInRangeOfSequentialType(const SequentialType *STy,
1994 const ConstantInt *CI) {
1995 if (const PointerType *PTy = dyn_cast<PointerType>(STy))
1996 // Only handle pointers to sized types, not pointers to functions.
1997 return PTy->getElementType()->isSized();
1999 uint64_t NumElements = 0;
2000 // Determine the number of elements in our sequential type.
2001 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
2002 NumElements = ATy->getNumElements();
2003 else if (const VectorType *VTy = dyn_cast<VectorType>(STy))
2004 NumElements = VTy->getNumElements();
2006 assert((isa<ArrayType>(STy) || NumElements > 0) &&
2007 "didn't expect non-array type to have zero elements!");
2009 // We cannot bounds check the index if it doesn't fit in an int64_t.
2010 if (CI->getValue().getActiveBits() > 64)
2013 // A negative index or an index past the end of our sequential type is
2014 // considered out-of-range.
2015 int64_t IndexVal = CI->getSExtValue();
2016 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2019 // Otherwise, it is in-range.
2023 template<typename IndexTy>
2024 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
2026 ArrayRef<IndexTy> Idxs) {
2027 if (Idxs.empty()) return C;
2028 Constant *Idx0 = cast<Constant>(Idxs[0]);
2029 if ((Idxs.size() == 1 && Idx0->isNullValue()))
2032 if (isa<UndefValue>(C)) {
2033 PointerType *Ptr = cast<PointerType>(C->getType());
2034 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
2035 assert(Ty && "Invalid indices for GEP!");
2036 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2039 if (C->isNullValue()) {
2041 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2042 if (!cast<Constant>(Idxs[i])->isNullValue()) {
2047 PointerType *Ptr = cast<PointerType>(C->getType());
2048 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs);
2049 assert(Ty && "Invalid indices for GEP!");
2050 return ConstantPointerNull::get(PointerType::get(Ty,
2051 Ptr->getAddressSpace()));
2055 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2056 // Combine Indices - If the source pointer to this getelementptr instruction
2057 // is a getelementptr instruction, combine the indices of the two
2058 // getelementptr instructions into a single instruction.
2060 if (CE->getOpcode() == Instruction::GetElementPtr) {
2061 Type *LastTy = nullptr;
2062 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2066 // We cannot combine indices if doing so would take us outside of an
2067 // array or vector. Doing otherwise could trick us if we evaluated such a
2068 // GEP as part of a load.
2070 // e.g. Consider if the original GEP was:
2071 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2072 // i32 0, i32 0, i64 0)
2074 // If we then tried to offset it by '8' to get to the third element,
2075 // an i8, we should *not* get:
2076 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2077 // i32 0, i32 0, i64 8)
2079 // This GEP tries to index array element '8 which runs out-of-bounds.
2080 // Subsequent evaluation would get confused and produce erroneous results.
2082 // The following prohibits such a GEP from being formed by checking to see
2083 // if the index is in-range with respect to an array or vector.
2084 bool PerformFold = false;
2085 if (Idx0->isNullValue())
2087 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy))
2088 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2089 PerformFold = isIndexInRangeOfSequentialType(STy, CI);
2092 SmallVector<Value*, 16> NewIndices;
2093 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2094 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2096 // Add the last index of the source with the first index of the new GEP.
2097 // Make sure to handle the case when they are actually different types.
2098 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2099 // Otherwise it must be an array.
2100 if (!Idx0->isNullValue()) {
2101 Type *IdxTy = Combined->getType();
2102 if (IdxTy != Idx0->getType()) {
2103 unsigned CommonExtendedWidth =
2104 std::max(IdxTy->getIntegerBitWidth(),
2105 Idx0->getType()->getIntegerBitWidth());
2106 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2109 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2110 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2111 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2112 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2115 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2119 NewIndices.push_back(Combined);
2120 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2122 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices,
2124 cast<GEPOperator>(CE)->isInBounds());
2128 // Attempt to fold casts to the same type away. For example, folding:
2130 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2134 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2136 // Don't fold if the cast is changing address spaces.
2137 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2138 PointerType *SrcPtrTy =
2139 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2140 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2141 if (SrcPtrTy && DstPtrTy) {
2142 ArrayType *SrcArrayTy =
2143 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2144 ArrayType *DstArrayTy =
2145 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2146 if (SrcArrayTy && DstArrayTy
2147 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2148 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2149 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0),
2155 // Check to see if any array indices are not within the corresponding
2156 // notional array or vector bounds. If so, try to determine if they can be
2157 // factored out into preceding dimensions.
2158 bool Unknown = false;
2159 SmallVector<Constant *, 8> NewIdxs;
2160 Type *Ty = C->getType();
2161 Type *Prev = nullptr;
2162 for (unsigned i = 0, e = Idxs.size(); i != e;
2163 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2164 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2165 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty))
2166 if (CI->getSExtValue() > 0 &&
2167 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) {
2168 if (isa<SequentialType>(Prev)) {
2169 // It's out of range, but we can factor it into the prior
2171 NewIdxs.resize(Idxs.size());
2172 uint64_t NumElements = 0;
2173 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2174 NumElements = ATy->getNumElements();
2176 NumElements = cast<VectorType>(Ty)->getNumElements();
2178 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements);
2179 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2181 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2182 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2184 unsigned CommonExtendedWidth =
2185 std::max(PrevIdx->getType()->getIntegerBitWidth(),
2186 Div->getType()->getIntegerBitWidth());
2187 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2189 // Before adding, extend both operands to i64 to avoid
2190 // overflow trouble.
2191 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth))
2192 PrevIdx = ConstantExpr::getSExt(
2194 Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2195 if (!Div->getType()->isIntegerTy(CommonExtendedWidth))
2196 Div = ConstantExpr::getSExt(
2197 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2199 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2201 // It's out of range, but the prior dimension is a struct
2202 // so we can't do anything about it.
2207 // We don't know if it's in range or not.
2212 // If we did any factoring, start over with the adjusted indices.
2213 if (!NewIdxs.empty()) {
2214 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2215 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2216 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds);
2219 // If all indices are known integers and normalized, we can do a simple
2220 // check for the "inbounds" property.
2221 if (!Unknown && !inBounds)
2222 if (auto *GV = dyn_cast<GlobalVariable>(C))
2223 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2224 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs);
2229 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2231 ArrayRef<Constant *> Idxs) {
2232 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);
2235 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2237 ArrayRef<Value *> Idxs) {
2238 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs);