[Alignment][NFC] migrate DataLayout internal struct to llvm::Align
[llvm-complete.git] / lib / IR / ConstantFold.cpp
blob835fbb3443b8fb8be02c870cef634f5a340cca1b
1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements folding of constants for LLVM. This implements the
10 // (internal) ConstantFold.h interface, which is used by the
11 // ConstantExpr::get* methods to automatically fold constants when possible.
13 // The current constant folding implementation is implemented in two pieces: the
14 // pieces that don't need DataLayout, and the pieces that do. This is to avoid
15 // a dependence in IR on Target.
17 //===----------------------------------------------------------------------===//
19 #include "ConstantFold.h"
20 #include "llvm/ADT/APSInt.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/Module.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ManagedStatic.h"
34 #include "llvm/Support/MathExtras.h"
35 using namespace llvm;
36 using namespace llvm::PatternMatch;
38 //===----------------------------------------------------------------------===//
39 // ConstantFold*Instruction Implementations
40 //===----------------------------------------------------------------------===//
42 /// Convert the specified vector Constant node to the specified vector type.
43 /// At this point, we know that the elements of the input vector constant are
44 /// all simple integer or FP values.
45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
50 // If this cast changes element count then we can't handle it here:
51 // doing so requires endianness information. This should be handled by
52 // Analysis/ConstantFolding.cpp
53 unsigned NumElts = DstTy->getNumElements();
54 if (NumElts != CV->getType()->getVectorNumElements())
55 return nullptr;
57 Type *DstEltTy = DstTy->getElementType();
59 SmallVector<Constant*, 16> Result;
60 Type *Ty = IntegerType::get(CV->getContext(), 32);
61 for (unsigned i = 0; i != NumElts; ++i) {
62 Constant *C =
63 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
64 C = ConstantExpr::getBitCast(C, DstEltTy);
65 Result.push_back(C);
68 return ConstantVector::get(Result);
71 /// This function determines which opcode to use to fold two constant cast
72 /// expressions together. It uses CastInst::isEliminableCastPair to determine
73 /// the opcode. Consequently its just a wrapper around that function.
74 /// Determine if it is valid to fold a cast of a cast
75 static unsigned
76 foldConstantCastPair(
77 unsigned opc, ///< opcode of the second cast constant expression
78 ConstantExpr *Op, ///< the first cast constant expression
79 Type *DstTy ///< destination type of the first cast
80 ) {
81 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
82 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
83 assert(CastInst::isCast(opc) && "Invalid cast opcode");
85 // The types and opcodes for the two Cast constant expressions
86 Type *SrcTy = Op->getOperand(0)->getType();
87 Type *MidTy = Op->getType();
88 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
89 Instruction::CastOps secondOp = Instruction::CastOps(opc);
91 // Assume that pointers are never more than 64 bits wide, and only use this
92 // for the middle type. Otherwise we could end up folding away illegal
93 // bitcasts between address spaces with different sizes.
94 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
96 // Let CastInst::isEliminableCastPair do the heavy lifting.
97 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
98 nullptr, FakeIntPtrTy, nullptr);
101 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
102 Type *SrcTy = V->getType();
103 if (SrcTy == DestTy)
104 return V; // no-op cast
106 // Check to see if we are casting a pointer to an aggregate to a pointer to
107 // the first element. If so, return the appropriate GEP instruction.
108 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
109 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
110 if (PTy->getAddressSpace() == DPTy->getAddressSpace()
111 && PTy->getElementType()->isSized()) {
112 SmallVector<Value*, 8> IdxList;
113 Value *Zero =
114 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
115 IdxList.push_back(Zero);
116 Type *ElTy = PTy->getElementType();
117 while (ElTy != DPTy->getElementType()) {
118 if (StructType *STy = dyn_cast<StructType>(ElTy)) {
119 if (STy->getNumElements() == 0) break;
120 ElTy = STy->getElementType(0);
121 IdxList.push_back(Zero);
122 } else if (SequentialType *STy =
123 dyn_cast<SequentialType>(ElTy)) {
124 ElTy = STy->getElementType();
125 IdxList.push_back(Zero);
126 } else {
127 break;
131 if (ElTy == DPTy->getElementType())
132 // This GEP is inbounds because all indices are zero.
133 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
134 V, IdxList);
137 // Handle casts from one vector constant to another. We know that the src
138 // and dest type have the same size (otherwise its an illegal cast).
139 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
140 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
141 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
142 "Not cast between same sized vectors!");
143 SrcTy = nullptr;
144 // First, check for null. Undef is already handled.
145 if (isa<ConstantAggregateZero>(V))
146 return Constant::getNullValue(DestTy);
148 // Handle ConstantVector and ConstantAggregateVector.
149 return BitCastConstantVector(V, DestPTy);
152 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
153 // This allows for other simplifications (although some of them
154 // can only be handled by Analysis/ConstantFolding.cpp).
155 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
156 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
159 // Finally, implement bitcast folding now. The code below doesn't handle
160 // bitcast right.
161 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
162 return ConstantPointerNull::get(cast<PointerType>(DestTy));
164 // Handle integral constant input.
165 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
166 if (DestTy->isIntegerTy())
167 // Integral -> Integral. This is a no-op because the bit widths must
168 // be the same. Consequently, we just fold to V.
169 return V;
171 // See note below regarding the PPC_FP128 restriction.
172 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
173 return ConstantFP::get(DestTy->getContext(),
174 APFloat(DestTy->getFltSemantics(),
175 CI->getValue()));
177 // Otherwise, can't fold this (vector?)
178 return nullptr;
181 // Handle ConstantFP input: FP -> Integral.
182 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
183 // PPC_FP128 is really the sum of two consecutive doubles, where the first
184 // double is always stored first in memory, regardless of the target
185 // endianness. The memory layout of i128, however, depends on the target
186 // endianness, and so we can't fold this without target endianness
187 // information. This should instead be handled by
188 // Analysis/ConstantFolding.cpp
189 if (FP->getType()->isPPC_FP128Ty())
190 return nullptr;
192 // Make sure dest type is compatible with the folded integer constant.
193 if (!DestTy->isIntegerTy())
194 return nullptr;
196 return ConstantInt::get(FP->getContext(),
197 FP->getValueAPF().bitcastToAPInt());
200 return nullptr;
204 /// V is an integer constant which only has a subset of its bytes used.
205 /// The bytes used are indicated by ByteStart (which is the first byte used,
206 /// counting from the least significant byte) and ByteSize, which is the number
207 /// of bytes used.
209 /// This function analyzes the specified constant to see if the specified byte
210 /// range can be returned as a simplified constant. If so, the constant is
211 /// returned, otherwise null is returned.
212 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
213 unsigned ByteSize) {
214 assert(C->getType()->isIntegerTy() &&
215 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
216 "Non-byte sized integer input");
217 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
218 assert(ByteSize && "Must be accessing some piece");
219 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
220 assert(ByteSize != CSize && "Should not extract everything");
222 // Constant Integers are simple.
223 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
224 APInt V = CI->getValue();
225 if (ByteStart)
226 V.lshrInPlace(ByteStart*8);
227 V = V.trunc(ByteSize*8);
228 return ConstantInt::get(CI->getContext(), V);
231 // In the input is a constant expr, we might be able to recursively simplify.
232 // If not, we definitely can't do anything.
233 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
234 if (!CE) return nullptr;
236 switch (CE->getOpcode()) {
237 default: return nullptr;
238 case Instruction::Or: {
239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
240 if (!RHS)
241 return nullptr;
243 // X | -1 -> -1.
244 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
245 if (RHSC->isMinusOne())
246 return RHSC;
248 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
249 if (!LHS)
250 return nullptr;
251 return ConstantExpr::getOr(LHS, RHS);
253 case Instruction::And: {
254 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
255 if (!RHS)
256 return nullptr;
258 // X & 0 -> 0.
259 if (RHS->isNullValue())
260 return RHS;
262 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
263 if (!LHS)
264 return nullptr;
265 return ConstantExpr::getAnd(LHS, RHS);
267 case Instruction::LShr: {
268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
269 if (!Amt)
270 return nullptr;
271 APInt ShAmt = Amt->getValue();
272 // Cannot analyze non-byte shifts.
273 if ((ShAmt & 7) != 0)
274 return nullptr;
275 ShAmt.lshrInPlace(3);
277 // If the extract is known to be all zeros, return zero.
278 if (ShAmt.uge(CSize - ByteStart))
279 return Constant::getNullValue(
280 IntegerType::get(CE->getContext(), ByteSize * 8));
281 // If the extract is known to be fully in the input, extract it.
282 if (ShAmt.ule(CSize - (ByteStart + ByteSize)))
283 return ExtractConstantBytes(CE->getOperand(0),
284 ByteStart + ShAmt.getZExtValue(), ByteSize);
286 // TODO: Handle the 'partially zero' case.
287 return nullptr;
290 case Instruction::Shl: {
291 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
292 if (!Amt)
293 return nullptr;
294 APInt ShAmt = Amt->getValue();
295 // Cannot analyze non-byte shifts.
296 if ((ShAmt & 7) != 0)
297 return nullptr;
298 ShAmt.lshrInPlace(3);
300 // If the extract is known to be all zeros, return zero.
301 if (ShAmt.uge(ByteStart + ByteSize))
302 return Constant::getNullValue(
303 IntegerType::get(CE->getContext(), ByteSize * 8));
304 // If the extract is known to be fully in the input, extract it.
305 if (ShAmt.ule(ByteStart))
306 return ExtractConstantBytes(CE->getOperand(0),
307 ByteStart - ShAmt.getZExtValue(), ByteSize);
309 // TODO: Handle the 'partially zero' case.
310 return nullptr;
313 case Instruction::ZExt: {
314 unsigned SrcBitSize =
315 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
317 // If extracting something that is completely zero, return 0.
318 if (ByteStart*8 >= SrcBitSize)
319 return Constant::getNullValue(IntegerType::get(CE->getContext(),
320 ByteSize*8));
322 // If exactly extracting the input, return it.
323 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
324 return CE->getOperand(0);
326 // If extracting something completely in the input, if the input is a
327 // multiple of 8 bits, recurse.
328 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
329 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
331 // Otherwise, if extracting a subset of the input, which is not multiple of
332 // 8 bits, do a shift and trunc to get the bits.
333 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
334 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
335 Constant *Res = CE->getOperand(0);
336 if (ByteStart)
337 Res = ConstantExpr::getLShr(Res,
338 ConstantInt::get(Res->getType(), ByteStart*8));
339 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
340 ByteSize*8));
343 // TODO: Handle the 'partially zero' case.
344 return nullptr;
349 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known
350 /// factors factored out. If Folded is false, return null if no factoring was
351 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
352 /// top-level folder.
353 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded) {
354 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
355 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
356 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
357 return ConstantExpr::getNUWMul(E, N);
360 if (StructType *STy = dyn_cast<StructType>(Ty))
361 if (!STy->isPacked()) {
362 unsigned NumElems = STy->getNumElements();
363 // An empty struct has size zero.
364 if (NumElems == 0)
365 return ConstantExpr::getNullValue(DestTy);
366 // Check for a struct with all members having the same size.
367 Constant *MemberSize =
368 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
369 bool AllSame = true;
370 for (unsigned i = 1; i != NumElems; ++i)
371 if (MemberSize !=
372 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
373 AllSame = false;
374 break;
376 if (AllSame) {
377 Constant *N = ConstantInt::get(DestTy, NumElems);
378 return ConstantExpr::getNUWMul(MemberSize, N);
382 // Pointer size doesn't depend on the pointee type, so canonicalize them
383 // to an arbitrary pointee.
384 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
385 if (!PTy->getElementType()->isIntegerTy(1))
386 return
387 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
388 PTy->getAddressSpace()),
389 DestTy, true);
391 // If there's no interesting folding happening, bail so that we don't create
392 // a constant that looks like it needs folding but really doesn't.
393 if (!Folded)
394 return nullptr;
396 // Base case: Get a regular sizeof expression.
397 Constant *C = ConstantExpr::getSizeOf(Ty);
398 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
399 DestTy, false),
400 C, DestTy);
401 return C;
404 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known
405 /// factors factored out. If Folded is false, return null if no factoring was
406 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
407 /// top-level folder.
408 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) {
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,
414 DestTy,
415 false),
416 C, DestTy);
417 return C;
420 if (StructType *STy = dyn_cast<StructType>(Ty)) {
421 // Packed structs always have an alignment of 1.
422 if (STy->isPacked())
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.
430 if (NumElems == 0)
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);
435 bool AllSame = true;
436 for (unsigned i = 1; i != NumElems; ++i)
437 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
438 AllSame = false;
439 break;
441 if (AllSame)
442 return MemberAlign;
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))
449 return
450 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
452 PTy->getAddressSpace()),
453 DestTy, true);
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.
457 if (!Folded)
458 return nullptr;
460 // Base case: Get a regular alignof expression.
461 Constant *C = ConstantExpr::getAlignOf(Ty);
462 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
463 DestTy, false),
464 C, DestTy);
465 return C;
468 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with
469 /// any known factors factored out. If Folded is false, return null if no
470 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression
471 /// back into the top-level folder.
472 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, Type *DestTy,
473 bool Folded) {
474 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
475 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
476 DestTy, false),
477 FieldNo, DestTy);
478 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
479 return ConstantExpr::getNUWMul(E, N);
482 if (StructType *STy = dyn_cast<StructType>(Ty))
483 if (!STy->isPacked()) {
484 unsigned NumElems = STy->getNumElements();
485 // An empty struct has no members.
486 if (NumElems == 0)
487 return nullptr;
488 // Check for a struct with all members having the same size.
489 Constant *MemberSize =
490 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
491 bool AllSame = true;
492 for (unsigned i = 1; i != NumElems; ++i)
493 if (MemberSize !=
494 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
495 AllSame = false;
496 break;
498 if (AllSame) {
499 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
500 false,
501 DestTy,
502 false),
503 FieldNo, DestTy);
504 return ConstantExpr::getNUWMul(MemberSize, N);
508 // If there's no interesting folding happening, bail so that we don't create
509 // a constant that looks like it needs folding but really doesn't.
510 if (!Folded)
511 return nullptr;
513 // Base case: Get a regular offsetof expression.
514 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
515 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
516 DestTy, false),
517 C, DestTy);
518 return C;
521 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
522 Type *DestTy) {
523 if (isa<UndefValue>(V)) {
524 // zext(undef) = 0, because the top bits will be zero.
525 // sext(undef) = 0, because the top bits will all be the same.
526 // [us]itofp(undef) = 0, because the result value is bounded.
527 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
528 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
529 return Constant::getNullValue(DestTy);
530 return UndefValue::get(DestTy);
533 if (V->isNullValue() && !DestTy->isX86_MMXTy() &&
534 opc != Instruction::AddrSpaceCast)
535 return Constant::getNullValue(DestTy);
537 // If the cast operand is a constant expression, there's a few things we can
538 // do to try to simplify it.
539 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
540 if (CE->isCast()) {
541 // Try hard to fold cast of cast because they are often eliminable.
542 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
543 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
544 } else if (CE->getOpcode() == Instruction::GetElementPtr &&
545 // Do not fold addrspacecast (gep 0, .., 0). It might make the
546 // addrspacecast uncanonicalized.
547 opc != Instruction::AddrSpaceCast &&
548 // Do not fold bitcast (gep) with inrange index, as this loses
549 // information.
550 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() &&
551 // Do not fold if the gep type is a vector, as bitcasting
552 // operand 0 of a vector gep will result in a bitcast between
553 // different sizes.
554 !CE->getType()->isVectorTy()) {
555 // If all of the indexes in the GEP are null values, there is no pointer
556 // adjustment going on. We might as well cast the source pointer.
557 bool isAllNull = true;
558 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
559 if (!CE->getOperand(i)->isNullValue()) {
560 isAllNull = false;
561 break;
563 if (isAllNull)
564 // This is casting one pointer type to another, always BitCast
565 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
569 // If the cast operand is a constant vector, perform the cast by
570 // operating on each element. In the cast of bitcasts, the element
571 // count may be mismatched; don't attempt to handle that here.
572 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
573 DestTy->isVectorTy() &&
574 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
575 SmallVector<Constant*, 16> res;
576 VectorType *DestVecTy = cast<VectorType>(DestTy);
577 Type *DstEltTy = DestVecTy->getElementType();
578 Type *Ty = IntegerType::get(V->getContext(), 32);
579 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
580 Constant *C =
581 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
582 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
584 return ConstantVector::get(res);
587 // We actually have to do a cast now. Perform the cast according to the
588 // opcode specified.
589 switch (opc) {
590 default:
591 llvm_unreachable("Failed to cast constant expression");
592 case Instruction::FPTrunc:
593 case Instruction::FPExt:
594 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
595 bool ignored;
596 APFloat Val = FPC->getValueAPF();
597 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
598 DestTy->isFloatTy() ? APFloat::IEEEsingle() :
599 DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
600 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
601 DestTy->isFP128Ty() ? APFloat::IEEEquad() :
602 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
603 APFloat::Bogus(),
604 APFloat::rmNearestTiesToEven, &ignored);
605 return ConstantFP::get(V->getContext(), Val);
607 return nullptr; // Can't fold.
608 case Instruction::FPToUI:
609 case Instruction::FPToSI:
610 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
611 const APFloat &V = FPC->getValueAPF();
612 bool ignored;
613 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
614 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
615 if (APFloat::opInvalidOp ==
616 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
617 // Undefined behavior invoked - the destination type can't represent
618 // the input constant.
619 return UndefValue::get(DestTy);
621 return ConstantInt::get(FPC->getContext(), IntVal);
623 return nullptr; // Can't fold.
624 case Instruction::IntToPtr: //always treated as unsigned
625 if (V->isNullValue()) // Is it an integral null value?
626 return ConstantPointerNull::get(cast<PointerType>(DestTy));
627 return nullptr; // Other pointer types cannot be casted
628 case Instruction::PtrToInt: // always treated as unsigned
629 // Is it a null pointer value?
630 if (V->isNullValue())
631 return ConstantInt::get(DestTy, 0);
632 // If this is a sizeof-like expression, pull out multiplications by
633 // known factors to expose them to subsequent folding. If it's an
634 // alignof-like expression, factor out known factors.
635 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
636 if (CE->getOpcode() == Instruction::GetElementPtr &&
637 CE->getOperand(0)->isNullValue()) {
638 // FIXME: Looks like getFoldedSizeOf(), getFoldedOffsetOf() and
639 // getFoldedAlignOf() don't handle the case when DestTy is a vector of
640 // pointers yet. We end up in asserts in CastInst::getCastOpcode (see
641 // test/Analysis/ConstantFolding/cast-vector.ll). I've only seen this
642 // happen in one "real" C-code test case, so it does not seem to be an
643 // important optimization to handle vectors here. For now, simply bail
644 // out.
645 if (DestTy->isVectorTy())
646 return nullptr;
647 GEPOperator *GEPO = cast<GEPOperator>(CE);
648 Type *Ty = GEPO->getSourceElementType();
649 if (CE->getNumOperands() == 2) {
650 // Handle a sizeof-like expression.
651 Constant *Idx = CE->getOperand(1);
652 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
653 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
654 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
655 DestTy, false),
656 Idx, DestTy);
657 return ConstantExpr::getMul(C, Idx);
659 } else if (CE->getNumOperands() == 3 &&
660 CE->getOperand(1)->isNullValue()) {
661 // Handle an alignof-like expression.
662 if (StructType *STy = dyn_cast<StructType>(Ty))
663 if (!STy->isPacked()) {
664 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
665 if (CI->isOne() &&
666 STy->getNumElements() == 2 &&
667 STy->getElementType(0)->isIntegerTy(1)) {
668 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
671 // Handle an offsetof-like expression.
672 if (Ty->isStructTy() || Ty->isArrayTy()) {
673 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
674 DestTy, false))
675 return C;
679 // Other pointer types cannot be casted
680 return nullptr;
681 case Instruction::UIToFP:
682 case Instruction::SIToFP:
683 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
684 const APInt &api = CI->getValue();
685 APFloat apf(DestTy->getFltSemantics(),
686 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
687 apf.convertFromAPInt(api, opc==Instruction::SIToFP,
688 APFloat::rmNearestTiesToEven);
689 return ConstantFP::get(V->getContext(), apf);
691 return nullptr;
692 case Instruction::ZExt:
693 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
694 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
695 return ConstantInt::get(V->getContext(),
696 CI->getValue().zext(BitWidth));
698 return nullptr;
699 case Instruction::SExt:
700 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
701 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
702 return ConstantInt::get(V->getContext(),
703 CI->getValue().sext(BitWidth));
705 return nullptr;
706 case Instruction::Trunc: {
707 if (V->getType()->isVectorTy())
708 return nullptr;
710 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
711 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
712 return ConstantInt::get(V->getContext(),
713 CI->getValue().trunc(DestBitWidth));
716 // The input must be a constantexpr. See if we can simplify this based on
717 // the bytes we are demanding. Only do this if the source and dest are an
718 // even multiple of a byte.
719 if ((DestBitWidth & 7) == 0 &&
720 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
721 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
722 return Res;
724 return nullptr;
726 case Instruction::BitCast:
727 return FoldBitCast(V, DestTy);
728 case Instruction::AddrSpaceCast:
729 return nullptr;
733 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
734 Constant *V1, Constant *V2) {
735 // Check for i1 and vector true/false conditions.
736 if (Cond->isNullValue()) return V2;
737 if (Cond->isAllOnesValue()) return V1;
739 // If the condition is a vector constant, fold the result elementwise.
740 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
741 SmallVector<Constant*, 16> Result;
742 Type *Ty = IntegerType::get(CondV->getContext(), 32);
743 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
744 Constant *V;
745 Constant *V1Element = ConstantExpr::getExtractElement(V1,
746 ConstantInt::get(Ty, i));
747 Constant *V2Element = ConstantExpr::getExtractElement(V2,
748 ConstantInt::get(Ty, i));
749 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
750 if (V1Element == V2Element) {
751 V = V1Element;
752 } else if (isa<UndefValue>(Cond)) {
753 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
754 } else {
755 if (!isa<ConstantInt>(Cond)) break;
756 V = Cond->isNullValue() ? V2Element : V1Element;
758 Result.push_back(V);
761 // If we were able to build the vector, return it.
762 if (Result.size() == V1->getType()->getVectorNumElements())
763 return ConstantVector::get(Result);
766 if (isa<UndefValue>(Cond)) {
767 if (isa<UndefValue>(V1)) return V1;
768 return V2;
770 if (isa<UndefValue>(V1)) return V2;
771 if (isa<UndefValue>(V2)) return V1;
772 if (V1 == V2) return V1;
774 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
775 if (TrueVal->getOpcode() == Instruction::Select)
776 if (TrueVal->getOperand(0) == Cond)
777 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
779 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
780 if (FalseVal->getOpcode() == Instruction::Select)
781 if (FalseVal->getOperand(0) == Cond)
782 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
785 return nullptr;
788 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
789 Constant *Idx) {
790 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
791 return UndefValue::get(Val->getType()->getVectorElementType());
792 if (Val->isNullValue()) // ee(zero, x) -> zero
793 return Constant::getNullValue(Val->getType()->getVectorElementType());
794 // ee({w,x,y,z}, undef) -> undef
795 if (isa<UndefValue>(Idx))
796 return UndefValue::get(Val->getType()->getVectorElementType());
798 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
799 // ee({w,x,y,z}, wrong_value) -> undef
800 if (CIdx->uge(Val->getType()->getVectorNumElements()))
801 return UndefValue::get(Val->getType()->getVectorElementType());
802 return Val->getAggregateElement(CIdx->getZExtValue());
804 return nullptr;
807 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
808 Constant *Elt,
809 Constant *Idx) {
810 if (isa<UndefValue>(Idx))
811 return UndefValue::get(Val->getType());
813 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
814 if (!CIdx) return nullptr;
816 unsigned NumElts = Val->getType()->getVectorNumElements();
817 if (CIdx->uge(NumElts))
818 return UndefValue::get(Val->getType());
820 SmallVector<Constant*, 16> Result;
821 Result.reserve(NumElts);
822 auto *Ty = Type::getInt32Ty(Val->getContext());
823 uint64_t IdxVal = CIdx->getZExtValue();
824 for (unsigned i = 0; i != NumElts; ++i) {
825 if (i == IdxVal) {
826 Result.push_back(Elt);
827 continue;
830 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
831 Result.push_back(C);
834 return ConstantVector::get(Result);
837 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
838 Constant *V2,
839 Constant *Mask) {
840 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
841 Type *EltTy = V1->getType()->getVectorElementType();
843 // Undefined shuffle mask -> undefined value.
844 if (isa<UndefValue>(Mask))
845 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
847 // Don't break the bitcode reader hack.
848 if (isa<ConstantExpr>(Mask)) return nullptr;
850 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
852 // Loop over the shuffle mask, evaluating each element.
853 SmallVector<Constant*, 32> Result;
854 for (unsigned i = 0; i != MaskNumElts; ++i) {
855 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
856 if (Elt == -1) {
857 Result.push_back(UndefValue::get(EltTy));
858 continue;
860 Constant *InElt;
861 if (unsigned(Elt) >= SrcNumElts*2)
862 InElt = UndefValue::get(EltTy);
863 else if (unsigned(Elt) >= SrcNumElts) {
864 Type *Ty = IntegerType::get(V2->getContext(), 32);
865 InElt =
866 ConstantExpr::getExtractElement(V2,
867 ConstantInt::get(Ty, Elt - SrcNumElts));
868 } else {
869 Type *Ty = IntegerType::get(V1->getContext(), 32);
870 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
872 Result.push_back(InElt);
875 return ConstantVector::get(Result);
878 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
879 ArrayRef<unsigned> Idxs) {
880 // Base case: no indices, so return the entire value.
881 if (Idxs.empty())
882 return Agg;
884 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
885 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
887 return nullptr;
890 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
891 Constant *Val,
892 ArrayRef<unsigned> Idxs) {
893 // Base case: no indices, so replace the entire value.
894 if (Idxs.empty())
895 return Val;
897 unsigned NumElts;
898 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
899 NumElts = ST->getNumElements();
900 else
901 NumElts = cast<SequentialType>(Agg->getType())->getNumElements();
903 SmallVector<Constant*, 32> Result;
904 for (unsigned i = 0; i != NumElts; ++i) {
905 Constant *C = Agg->getAggregateElement(i);
906 if (!C) return nullptr;
908 if (Idxs[0] == i)
909 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
911 Result.push_back(C);
914 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
915 return ConstantStruct::get(ST, Result);
916 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
917 return ConstantArray::get(AT, Result);
918 return ConstantVector::get(Result);
921 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) {
922 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected");
924 // Handle scalar UndefValue. Vectors are always evaluated per element.
925 bool HasScalarUndef = !C->getType()->isVectorTy() && isa<UndefValue>(C);
927 if (HasScalarUndef) {
928 switch (static_cast<Instruction::UnaryOps>(Opcode)) {
929 case Instruction::FNeg:
930 return C; // -undef -> undef
931 case Instruction::UnaryOpsEnd:
932 llvm_unreachable("Invalid UnaryOp");
936 // Constant should not be UndefValue, unless these are vector constants.
937 assert(!HasScalarUndef && "Unexpected UndefValue");
938 // We only have FP UnaryOps right now.
939 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp");
941 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
942 const APFloat &CV = CFP->getValueAPF();
943 switch (Opcode) {
944 default:
945 break;
946 case Instruction::FNeg:
947 return ConstantFP::get(C->getContext(), neg(CV));
949 } else if (VectorType *VTy = dyn_cast<VectorType>(C->getType())) {
950 // Fold each element and create a vector constant from those constants.
951 SmallVector<Constant*, 16> Result;
952 Type *Ty = IntegerType::get(VTy->getContext(), 32);
953 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
954 Constant *ExtractIdx = ConstantInt::get(Ty, i);
955 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx);
957 Result.push_back(ConstantExpr::get(Opcode, Elt));
960 return ConstantVector::get(Result);
963 // We don't know how to fold this.
964 return nullptr;
967 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
968 Constant *C2) {
969 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
971 // Handle scalar UndefValue. Vectors are always evaluated per element.
972 bool HasScalarUndef = !C1->getType()->isVectorTy() &&
973 (isa<UndefValue>(C1) || isa<UndefValue>(C2));
974 if (HasScalarUndef) {
975 switch (static_cast<Instruction::BinaryOps>(Opcode)) {
976 case Instruction::Xor:
977 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
978 // Handle undef ^ undef -> 0 special case. This is a common
979 // idiom (misuse).
980 return Constant::getNullValue(C1->getType());
981 LLVM_FALLTHROUGH;
982 case Instruction::Add:
983 case Instruction::Sub:
984 return UndefValue::get(C1->getType());
985 case Instruction::And:
986 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
987 return C1;
988 return Constant::getNullValue(C1->getType()); // undef & X -> 0
989 case Instruction::Mul: {
990 // undef * undef -> undef
991 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
992 return C1;
993 const APInt *CV;
994 // X * undef -> undef if X is odd
995 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
996 if ((*CV)[0])
997 return UndefValue::get(C1->getType());
999 // X * undef -> 0 otherwise
1000 return Constant::getNullValue(C1->getType());
1002 case Instruction::SDiv:
1003 case Instruction::UDiv:
1004 // X / undef -> undef
1005 if (isa<UndefValue>(C2))
1006 return C2;
1007 // undef / 0 -> undef
1008 // undef / 1 -> undef
1009 if (match(C2, m_Zero()) || match(C2, m_One()))
1010 return C1;
1011 // undef / X -> 0 otherwise
1012 return Constant::getNullValue(C1->getType());
1013 case Instruction::URem:
1014 case Instruction::SRem:
1015 // X % undef -> undef
1016 if (match(C2, m_Undef()))
1017 return C2;
1018 // undef % 0 -> undef
1019 if (match(C2, m_Zero()))
1020 return C1;
1021 // undef % X -> 0 otherwise
1022 return Constant::getNullValue(C1->getType());
1023 case Instruction::Or: // X | undef -> -1
1024 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
1025 return C1;
1026 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
1027 case Instruction::LShr:
1028 // X >>l undef -> undef
1029 if (isa<UndefValue>(C2))
1030 return C2;
1031 // undef >>l 0 -> undef
1032 if (match(C2, m_Zero()))
1033 return C1;
1034 // undef >>l X -> 0
1035 return Constant::getNullValue(C1->getType());
1036 case Instruction::AShr:
1037 // X >>a undef -> undef
1038 if (isa<UndefValue>(C2))
1039 return C2;
1040 // undef >>a 0 -> undef
1041 if (match(C2, m_Zero()))
1042 return C1;
1043 // TODO: undef >>a X -> undef if the shift is exact
1044 // undef >>a X -> 0
1045 return Constant::getNullValue(C1->getType());
1046 case Instruction::Shl:
1047 // X << undef -> undef
1048 if (isa<UndefValue>(C2))
1049 return C2;
1050 // undef << 0 -> undef
1051 if (match(C2, m_Zero()))
1052 return C1;
1053 // undef << X -> 0
1054 return Constant::getNullValue(C1->getType());
1055 case Instruction::FAdd:
1056 case Instruction::FSub:
1057 case Instruction::FMul:
1058 case Instruction::FDiv:
1059 case Instruction::FRem:
1060 // [any flop] undef, undef -> undef
1061 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1062 return C1;
1063 // [any flop] C, undef -> NaN
1064 // [any flop] undef, C -> NaN
1065 // We could potentially specialize NaN/Inf constants vs. 'normal'
1066 // constants (possibly differently depending on opcode and operand). This
1067 // would allow returning undef sometimes. But it is always safe to fold to
1068 // NaN because we can choose the undef operand as NaN, and any FP opcode
1069 // with a NaN operand will propagate NaN.
1070 return ConstantFP::getNaN(C1->getType());
1071 case Instruction::BinaryOpsEnd:
1072 llvm_unreachable("Invalid BinaryOp");
1076 // Neither constant should be UndefValue, unless these are vector constants.
1077 assert(!HasScalarUndef && "Unexpected UndefValue");
1079 // Handle simplifications when the RHS is a constant int.
1080 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1081 switch (Opcode) {
1082 case Instruction::Add:
1083 if (CI2->isZero()) return C1; // X + 0 == X
1084 break;
1085 case Instruction::Sub:
1086 if (CI2->isZero()) return C1; // X - 0 == X
1087 break;
1088 case Instruction::Mul:
1089 if (CI2->isZero()) return C2; // X * 0 == 0
1090 if (CI2->isOne())
1091 return C1; // X * 1 == X
1092 break;
1093 case Instruction::UDiv:
1094 case Instruction::SDiv:
1095 if (CI2->isOne())
1096 return C1; // X / 1 == X
1097 if (CI2->isZero())
1098 return UndefValue::get(CI2->getType()); // X / 0 == undef
1099 break;
1100 case Instruction::URem:
1101 case Instruction::SRem:
1102 if (CI2->isOne())
1103 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1104 if (CI2->isZero())
1105 return UndefValue::get(CI2->getType()); // X % 0 == undef
1106 break;
1107 case Instruction::And:
1108 if (CI2->isZero()) return C2; // X & 0 == 0
1109 if (CI2->isMinusOne())
1110 return C1; // X & -1 == X
1112 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1113 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1114 if (CE1->getOpcode() == Instruction::ZExt) {
1115 unsigned DstWidth = CI2->getType()->getBitWidth();
1116 unsigned SrcWidth =
1117 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1118 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1119 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1120 return C1;
1123 // If and'ing the address of a global with a constant, fold it.
1124 if (CE1->getOpcode() == Instruction::PtrToInt &&
1125 isa<GlobalValue>(CE1->getOperand(0))) {
1126 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1128 unsigned GVAlign;
1130 if (Module *TheModule = GV->getParent()) {
1131 GVAlign = GV->getPointerAlignment(TheModule->getDataLayout());
1133 // If the function alignment is not specified then assume that it
1134 // is 4.
1135 // This is dangerous; on x86, the alignment of the pointer
1136 // corresponds to the alignment of the function, but might be less
1137 // than 4 if it isn't explicitly specified.
1138 // However, a fix for this behaviour was reverted because it
1139 // increased code size (see https://reviews.llvm.org/D55115)
1140 // FIXME: This code should be deleted once existing targets have
1141 // appropriate defaults
1142 if (GVAlign == 0U && isa<Function>(GV))
1143 GVAlign = 4U;
1144 } else if (isa<Function>(GV)) {
1145 // Without a datalayout we have to assume the worst case: that the
1146 // function pointer isn't aligned at all.
1147 GVAlign = 0U;
1148 } else {
1149 GVAlign = GV->getAlignment();
1152 if (GVAlign > 1) {
1153 unsigned DstWidth = CI2->getType()->getBitWidth();
1154 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1155 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1157 // If checking bits we know are clear, return zero.
1158 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1159 return Constant::getNullValue(CI2->getType());
1163 break;
1164 case Instruction::Or:
1165 if (CI2->isZero()) return C1; // X | 0 == X
1166 if (CI2->isMinusOne())
1167 return C2; // X | -1 == -1
1168 break;
1169 case Instruction::Xor:
1170 if (CI2->isZero()) return C1; // X ^ 0 == X
1172 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1173 switch (CE1->getOpcode()) {
1174 default: break;
1175 case Instruction::ICmp:
1176 case Instruction::FCmp:
1177 // cmp pred ^ true -> cmp !pred
1178 assert(CI2->isOne());
1179 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1180 pred = CmpInst::getInversePredicate(pred);
1181 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1182 CE1->getOperand(1));
1185 break;
1186 case Instruction::AShr:
1187 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1188 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1189 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1190 return ConstantExpr::getLShr(C1, C2);
1191 break;
1193 } else if (isa<ConstantInt>(C1)) {
1194 // If C1 is a ConstantInt and C2 is not, swap the operands.
1195 if (Instruction::isCommutative(Opcode))
1196 return ConstantExpr::get(Opcode, C2, C1);
1199 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1200 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1201 const APInt &C1V = CI1->getValue();
1202 const APInt &C2V = CI2->getValue();
1203 switch (Opcode) {
1204 default:
1205 break;
1206 case Instruction::Add:
1207 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1208 case Instruction::Sub:
1209 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1210 case Instruction::Mul:
1211 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1212 case Instruction::UDiv:
1213 assert(!CI2->isZero() && "Div by zero handled above");
1214 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1215 case Instruction::SDiv:
1216 assert(!CI2->isZero() && "Div by zero handled above");
1217 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1218 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1219 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1220 case Instruction::URem:
1221 assert(!CI2->isZero() && "Div by zero handled above");
1222 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1223 case Instruction::SRem:
1224 assert(!CI2->isZero() && "Div by zero handled above");
1225 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1226 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1227 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1228 case Instruction::And:
1229 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1230 case Instruction::Or:
1231 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1232 case Instruction::Xor:
1233 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1234 case Instruction::Shl:
1235 if (C2V.ult(C1V.getBitWidth()))
1236 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1237 return UndefValue::get(C1->getType()); // too big shift is undef
1238 case Instruction::LShr:
1239 if (C2V.ult(C1V.getBitWidth()))
1240 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1241 return UndefValue::get(C1->getType()); // too big shift is undef
1242 case Instruction::AShr:
1243 if (C2V.ult(C1V.getBitWidth()))
1244 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1245 return UndefValue::get(C1->getType()); // too big shift is undef
1249 switch (Opcode) {
1250 case Instruction::SDiv:
1251 case Instruction::UDiv:
1252 case Instruction::URem:
1253 case Instruction::SRem:
1254 case Instruction::LShr:
1255 case Instruction::AShr:
1256 case Instruction::Shl:
1257 if (CI1->isZero()) return C1;
1258 break;
1259 default:
1260 break;
1262 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1263 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1264 const APFloat &C1V = CFP1->getValueAPF();
1265 const APFloat &C2V = CFP2->getValueAPF();
1266 APFloat C3V = C1V; // copy for modification
1267 switch (Opcode) {
1268 default:
1269 break;
1270 case Instruction::FAdd:
1271 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1272 return ConstantFP::get(C1->getContext(), C3V);
1273 case Instruction::FSub:
1274 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1275 return ConstantFP::get(C1->getContext(), C3V);
1276 case Instruction::FMul:
1277 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1278 return ConstantFP::get(C1->getContext(), C3V);
1279 case Instruction::FDiv:
1280 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1281 return ConstantFP::get(C1->getContext(), C3V);
1282 case Instruction::FRem:
1283 (void)C3V.mod(C2V);
1284 return ConstantFP::get(C1->getContext(), C3V);
1287 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1288 // Fold each element and create a vector constant from those constants.
1289 SmallVector<Constant*, 16> Result;
1290 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1291 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1292 Constant *ExtractIdx = ConstantInt::get(Ty, i);
1293 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1294 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1296 // If any element of a divisor vector is zero, the whole op is undef.
1297 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1298 return UndefValue::get(VTy);
1300 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1303 return ConstantVector::get(Result);
1306 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1307 // There are many possible foldings we could do here. We should probably
1308 // at least fold add of a pointer with an integer into the appropriate
1309 // getelementptr. This will improve alias analysis a bit.
1311 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1312 // (a + (b + c)).
1313 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1314 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1315 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1316 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1318 } else if (isa<ConstantExpr>(C2)) {
1319 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1320 // other way if possible.
1321 if (Instruction::isCommutative(Opcode))
1322 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1325 // i1 can be simplified in many cases.
1326 if (C1->getType()->isIntegerTy(1)) {
1327 switch (Opcode) {
1328 case Instruction::Add:
1329 case Instruction::Sub:
1330 return ConstantExpr::getXor(C1, C2);
1331 case Instruction::Mul:
1332 return ConstantExpr::getAnd(C1, C2);
1333 case Instruction::Shl:
1334 case Instruction::LShr:
1335 case Instruction::AShr:
1336 // We can assume that C2 == 0. If it were one the result would be
1337 // undefined because the shift value is as large as the bitwidth.
1338 return C1;
1339 case Instruction::SDiv:
1340 case Instruction::UDiv:
1341 // We can assume that C2 == 1. If it were zero the result would be
1342 // undefined through division by zero.
1343 return C1;
1344 case Instruction::URem:
1345 case Instruction::SRem:
1346 // We can assume that C2 == 1. If it were zero the result would be
1347 // undefined through division by zero.
1348 return ConstantInt::getFalse(C1->getContext());
1349 default:
1350 break;
1354 // We don't know how to fold this.
1355 return nullptr;
1358 /// This type is zero-sized if it's an array or structure of zero-sized types.
1359 /// The only leaf zero-sized type is an empty structure.
1360 static bool isMaybeZeroSizedType(Type *Ty) {
1361 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1362 if (STy->isOpaque()) return true; // Can't say.
1364 // If all of elements have zero size, this does too.
1365 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1366 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1367 return true;
1369 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1370 return isMaybeZeroSizedType(ATy->getElementType());
1372 return false;
1375 /// Compare the two constants as though they were getelementptr indices.
1376 /// This allows coercion of the types to be the same thing.
1378 /// If the two constants are the "same" (after coercion), return 0. If the
1379 /// first is less than the second, return -1, if the second is less than the
1380 /// first, return 1. If the constants are not integral, return -2.
1382 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1383 if (C1 == C2) return 0;
1385 // Ok, we found a different index. If they are not ConstantInt, we can't do
1386 // anything with them.
1387 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1388 return -2; // don't know!
1390 // We cannot compare the indices if they don't fit in an int64_t.
1391 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1392 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1393 return -2; // don't know!
1395 // Ok, we have two differing integer indices. Sign extend them to be the same
1396 // type.
1397 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1398 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1400 if (C1Val == C2Val) return 0; // They are equal
1402 // If the type being indexed over is really just a zero sized type, there is
1403 // no pointer difference being made here.
1404 if (isMaybeZeroSizedType(ElTy))
1405 return -2; // dunno.
1407 // If they are really different, now that they are the same type, then we
1408 // found a difference!
1409 if (C1Val < C2Val)
1410 return -1;
1411 else
1412 return 1;
1415 /// This function determines if there is anything we can decide about the two
1416 /// constants provided. This doesn't need to handle simple things like
1417 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1418 /// If we can determine that the two constants have a particular relation to
1419 /// each other, we should return the corresponding FCmpInst predicate,
1420 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1421 /// ConstantFoldCompareInstruction.
1423 /// To simplify this code we canonicalize the relation so that the first
1424 /// operand is always the most "complex" of the two. We consider ConstantFP
1425 /// to be the simplest, and ConstantExprs to be the most complex.
1426 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1427 assert(V1->getType() == V2->getType() &&
1428 "Cannot compare values of different types!");
1430 // We do not know if a constant expression will evaluate to a number or NaN.
1431 // Therefore, we can only say that the relation is unordered or equal.
1432 if (V1 == V2) return FCmpInst::FCMP_UEQ;
1434 if (!isa<ConstantExpr>(V1)) {
1435 if (!isa<ConstantExpr>(V2)) {
1436 // Simple case, use the standard constant folder.
1437 ConstantInt *R = nullptr;
1438 R = dyn_cast<ConstantInt>(
1439 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1440 if (R && !R->isZero())
1441 return FCmpInst::FCMP_OEQ;
1442 R = dyn_cast<ConstantInt>(
1443 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1444 if (R && !R->isZero())
1445 return FCmpInst::FCMP_OLT;
1446 R = dyn_cast<ConstantInt>(
1447 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1448 if (R && !R->isZero())
1449 return FCmpInst::FCMP_OGT;
1451 // Nothing more we can do
1452 return FCmpInst::BAD_FCMP_PREDICATE;
1455 // If the first operand is simple and second is ConstantExpr, swap operands.
1456 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1457 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1458 return FCmpInst::getSwappedPredicate(SwappedRelation);
1459 } else {
1460 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1461 // constantexpr or a simple constant.
1462 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1463 switch (CE1->getOpcode()) {
1464 case Instruction::FPTrunc:
1465 case Instruction::FPExt:
1466 case Instruction::UIToFP:
1467 case Instruction::SIToFP:
1468 // We might be able to do something with these but we don't right now.
1469 break;
1470 default:
1471 break;
1474 // There are MANY other foldings that we could perform here. They will
1475 // probably be added on demand, as they seem needed.
1476 return FCmpInst::BAD_FCMP_PREDICATE;
1479 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1480 const GlobalValue *GV2) {
1481 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1482 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1483 return true;
1484 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1485 Type *Ty = GVar->getValueType();
1486 // A global with opaque type might end up being zero sized.
1487 if (!Ty->isSized())
1488 return true;
1489 // A global with an empty type might lie at the address of any other
1490 // global.
1491 if (Ty->isEmptyTy())
1492 return true;
1494 return false;
1496 // Don't try to decide equality of aliases.
1497 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1498 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1499 return ICmpInst::ICMP_NE;
1500 return ICmpInst::BAD_ICMP_PREDICATE;
1503 /// This function determines if there is anything we can decide about the two
1504 /// constants provided. This doesn't need to handle simple things like integer
1505 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1506 /// If we can determine that the two constants have a particular relation to
1507 /// each other, we should return the corresponding ICmp predicate, otherwise
1508 /// return ICmpInst::BAD_ICMP_PREDICATE.
1510 /// To simplify this code we canonicalize the relation so that the first
1511 /// operand is always the most "complex" of the two. We consider simple
1512 /// constants (like ConstantInt) to be the simplest, followed by
1513 /// GlobalValues, followed by ConstantExpr's (the most complex).
1515 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1516 bool isSigned) {
1517 assert(V1->getType() == V2->getType() &&
1518 "Cannot compare different types of values!");
1519 if (V1 == V2) return ICmpInst::ICMP_EQ;
1521 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1522 !isa<BlockAddress>(V1)) {
1523 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1524 !isa<BlockAddress>(V2)) {
1525 // We distilled this down to a simple case, use the standard constant
1526 // folder.
1527 ConstantInt *R = nullptr;
1528 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1529 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1530 if (R && !R->isZero())
1531 return pred;
1532 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1533 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1534 if (R && !R->isZero())
1535 return pred;
1536 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1537 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1538 if (R && !R->isZero())
1539 return pred;
1541 // If we couldn't figure it out, bail.
1542 return ICmpInst::BAD_ICMP_PREDICATE;
1545 // If the first operand is simple, swap operands.
1546 ICmpInst::Predicate SwappedRelation =
1547 evaluateICmpRelation(V2, V1, isSigned);
1548 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1549 return ICmpInst::getSwappedPredicate(SwappedRelation);
1551 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1552 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1553 ICmpInst::Predicate SwappedRelation =
1554 evaluateICmpRelation(V2, V1, isSigned);
1555 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1556 return ICmpInst::getSwappedPredicate(SwappedRelation);
1557 return ICmpInst::BAD_ICMP_PREDICATE;
1560 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1561 // constant (which, since the types must match, means that it's a
1562 // ConstantPointerNull).
1563 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1564 return areGlobalsPotentiallyEqual(GV, GV2);
1565 } else if (isa<BlockAddress>(V2)) {
1566 return ICmpInst::ICMP_NE; // Globals never equal labels.
1567 } else {
1568 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1569 // GlobalVals can never be null unless they have external weak linkage.
1570 // We don't try to evaluate aliases here.
1571 // NOTE: We should not be doing this constant folding if null pointer
1572 // is considered valid for the function. But currently there is no way to
1573 // query it from the Constant type.
1574 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1575 !NullPointerIsDefined(nullptr /* F */,
1576 GV->getType()->getAddressSpace()))
1577 return ICmpInst::ICMP_NE;
1579 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1580 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1581 ICmpInst::Predicate SwappedRelation =
1582 evaluateICmpRelation(V2, V1, isSigned);
1583 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1584 return ICmpInst::getSwappedPredicate(SwappedRelation);
1585 return ICmpInst::BAD_ICMP_PREDICATE;
1588 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1589 // constant (which, since the types must match, means that it is a
1590 // ConstantPointerNull).
1591 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1592 // Block address in another function can't equal this one, but block
1593 // addresses in the current function might be the same if blocks are
1594 // empty.
1595 if (BA2->getFunction() != BA->getFunction())
1596 return ICmpInst::ICMP_NE;
1597 } else {
1598 // Block addresses aren't null, don't equal the address of globals.
1599 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1600 "Canonicalization guarantee!");
1601 return ICmpInst::ICMP_NE;
1603 } else {
1604 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1605 // constantexpr, a global, block address, or a simple constant.
1606 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1607 Constant *CE1Op0 = CE1->getOperand(0);
1609 switch (CE1->getOpcode()) {
1610 case Instruction::Trunc:
1611 case Instruction::FPTrunc:
1612 case Instruction::FPExt:
1613 case Instruction::FPToUI:
1614 case Instruction::FPToSI:
1615 break; // We can't evaluate floating point casts or truncations.
1617 case Instruction::UIToFP:
1618 case Instruction::SIToFP:
1619 case Instruction::BitCast:
1620 case Instruction::ZExt:
1621 case Instruction::SExt:
1622 // We can't evaluate floating point casts or truncations.
1623 if (CE1Op0->getType()->isFPOrFPVectorTy())
1624 break;
1626 // If the cast is not actually changing bits, and the second operand is a
1627 // null pointer, do the comparison with the pre-casted value.
1628 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1629 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1630 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1631 return evaluateICmpRelation(CE1Op0,
1632 Constant::getNullValue(CE1Op0->getType()),
1633 isSigned);
1635 break;
1637 case Instruction::GetElementPtr: {
1638 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1639 // Ok, since this is a getelementptr, we know that the constant has a
1640 // pointer type. Check the various cases.
1641 if (isa<ConstantPointerNull>(V2)) {
1642 // If we are comparing a GEP to a null pointer, check to see if the base
1643 // of the GEP equals the null pointer.
1644 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1645 if (GV->hasExternalWeakLinkage())
1646 // Weak linkage GVals could be zero or not. We're comparing that
1647 // to null pointer so its greater-or-equal
1648 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1649 else
1650 // If its not weak linkage, the GVal must have a non-zero address
1651 // so the result is greater-than
1652 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1653 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1654 // If we are indexing from a null pointer, check to see if we have any
1655 // non-zero indices.
1656 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1657 if (!CE1->getOperand(i)->isNullValue())
1658 // Offsetting from null, must not be equal.
1659 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1660 // Only zero indexes from null, must still be zero.
1661 return ICmpInst::ICMP_EQ;
1663 // Otherwise, we can't really say if the first operand is null or not.
1664 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1665 if (isa<ConstantPointerNull>(CE1Op0)) {
1666 if (GV2->hasExternalWeakLinkage())
1667 // Weak linkage GVals could be zero or not. We're comparing it to
1668 // a null pointer, so its less-or-equal
1669 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1670 else
1671 // If its not weak linkage, the GVal must have a non-zero address
1672 // so the result is less-than
1673 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1674 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1675 if (GV == GV2) {
1676 // If this is a getelementptr of the same global, then it must be
1677 // different. Because the types must match, the getelementptr could
1678 // only have at most one index, and because we fold getelementptr's
1679 // with a single zero index, it must be nonzero.
1680 assert(CE1->getNumOperands() == 2 &&
1681 !CE1->getOperand(1)->isNullValue() &&
1682 "Surprising getelementptr!");
1683 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1684 } else {
1685 if (CE1GEP->hasAllZeroIndices())
1686 return areGlobalsPotentiallyEqual(GV, GV2);
1687 return ICmpInst::BAD_ICMP_PREDICATE;
1690 } else {
1691 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1692 Constant *CE2Op0 = CE2->getOperand(0);
1694 // There are MANY other foldings that we could perform here. They will
1695 // probably be added on demand, as they seem needed.
1696 switch (CE2->getOpcode()) {
1697 default: break;
1698 case Instruction::GetElementPtr:
1699 // By far the most common case to handle is when the base pointers are
1700 // obviously to the same global.
1701 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1702 // Don't know relative ordering, but check for inequality.
1703 if (CE1Op0 != CE2Op0) {
1704 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1705 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1706 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1707 cast<GlobalValue>(CE2Op0));
1708 return ICmpInst::BAD_ICMP_PREDICATE;
1710 // Ok, we know that both getelementptr instructions are based on the
1711 // same global. From this, we can precisely determine the relative
1712 // ordering of the resultant pointers.
1713 unsigned i = 1;
1715 // The logic below assumes that the result of the comparison
1716 // can be determined by finding the first index that differs.
1717 // This doesn't work if there is over-indexing in any
1718 // subsequent indices, so check for that case first.
1719 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1720 !CE2->isGEPWithNoNotionalOverIndexing())
1721 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1723 // Compare all of the operands the GEP's have in common.
1724 gep_type_iterator GTI = gep_type_begin(CE1);
1725 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1726 ++i, ++GTI)
1727 switch (IdxCompare(CE1->getOperand(i),
1728 CE2->getOperand(i), GTI.getIndexedType())) {
1729 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1730 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1731 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1734 // Ok, we ran out of things they have in common. If any leftovers
1735 // are non-zero then we have a difference, otherwise we are equal.
1736 for (; i < CE1->getNumOperands(); ++i)
1737 if (!CE1->getOperand(i)->isNullValue()) {
1738 if (isa<ConstantInt>(CE1->getOperand(i)))
1739 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1740 else
1741 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1744 for (; i < CE2->getNumOperands(); ++i)
1745 if (!CE2->getOperand(i)->isNullValue()) {
1746 if (isa<ConstantInt>(CE2->getOperand(i)))
1747 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1748 else
1749 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1751 return ICmpInst::ICMP_EQ;
1755 break;
1757 default:
1758 break;
1762 return ICmpInst::BAD_ICMP_PREDICATE;
1765 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1766 Constant *C1, Constant *C2) {
1767 Type *ResultTy;
1768 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1769 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1770 VT->getNumElements());
1771 else
1772 ResultTy = Type::getInt1Ty(C1->getContext());
1774 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1775 if (pred == FCmpInst::FCMP_FALSE)
1776 return Constant::getNullValue(ResultTy);
1778 if (pred == FCmpInst::FCMP_TRUE)
1779 return Constant::getAllOnesValue(ResultTy);
1781 // Handle some degenerate cases first
1782 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1783 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1784 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1785 // For EQ and NE, we can always pick a value for the undef to make the
1786 // predicate pass or fail, so we can return undef.
1787 // Also, if both operands are undef, we can return undef for int comparison.
1788 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1789 return UndefValue::get(ResultTy);
1791 // Otherwise, for integer compare, pick the same value as the non-undef
1792 // operand, and fold it to true or false.
1793 if (isIntegerPredicate)
1794 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1796 // Choosing NaN for the undef will always make unordered comparison succeed
1797 // and ordered comparison fails.
1798 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1801 // icmp eq/ne(null,GV) -> false/true
1802 if (C1->isNullValue()) {
1803 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1804 // Don't try to evaluate aliases. External weak GV can be null.
1805 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1806 !NullPointerIsDefined(nullptr /* F */,
1807 GV->getType()->getAddressSpace())) {
1808 if (pred == ICmpInst::ICMP_EQ)
1809 return ConstantInt::getFalse(C1->getContext());
1810 else if (pred == ICmpInst::ICMP_NE)
1811 return ConstantInt::getTrue(C1->getContext());
1813 // icmp eq/ne(GV,null) -> false/true
1814 } else if (C2->isNullValue()) {
1815 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1816 // Don't try to evaluate aliases. External weak GV can be null.
1817 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1818 !NullPointerIsDefined(nullptr /* F */,
1819 GV->getType()->getAddressSpace())) {
1820 if (pred == ICmpInst::ICMP_EQ)
1821 return ConstantInt::getFalse(C1->getContext());
1822 else if (pred == ICmpInst::ICMP_NE)
1823 return ConstantInt::getTrue(C1->getContext());
1827 // If the comparison is a comparison between two i1's, simplify it.
1828 if (C1->getType()->isIntegerTy(1)) {
1829 switch(pred) {
1830 case ICmpInst::ICMP_EQ:
1831 if (isa<ConstantInt>(C2))
1832 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1833 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1834 case ICmpInst::ICMP_NE:
1835 return ConstantExpr::getXor(C1, C2);
1836 default:
1837 break;
1841 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1842 const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1843 const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1844 switch (pred) {
1845 default: llvm_unreachable("Invalid ICmp Predicate");
1846 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1847 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1848 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1849 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1850 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1851 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1852 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1853 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1854 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1855 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1857 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1858 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1859 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1860 APFloat::cmpResult R = C1V.compare(C2V);
1861 switch (pred) {
1862 default: llvm_unreachable("Invalid FCmp Predicate");
1863 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1864 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1865 case FCmpInst::FCMP_UNO:
1866 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1867 case FCmpInst::FCMP_ORD:
1868 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1869 case FCmpInst::FCMP_UEQ:
1870 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1871 R==APFloat::cmpEqual);
1872 case FCmpInst::FCMP_OEQ:
1873 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1874 case FCmpInst::FCMP_UNE:
1875 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1876 case FCmpInst::FCMP_ONE:
1877 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1878 R==APFloat::cmpGreaterThan);
1879 case FCmpInst::FCMP_ULT:
1880 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1881 R==APFloat::cmpLessThan);
1882 case FCmpInst::FCMP_OLT:
1883 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1884 case FCmpInst::FCMP_UGT:
1885 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1886 R==APFloat::cmpGreaterThan);
1887 case FCmpInst::FCMP_OGT:
1888 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1889 case FCmpInst::FCMP_ULE:
1890 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1891 case FCmpInst::FCMP_OLE:
1892 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1893 R==APFloat::cmpEqual);
1894 case FCmpInst::FCMP_UGE:
1895 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1896 case FCmpInst::FCMP_OGE:
1897 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1898 R==APFloat::cmpEqual);
1900 } else if (C1->getType()->isVectorTy()) {
1901 // If we can constant fold the comparison of each element, constant fold
1902 // the whole vector comparison.
1903 SmallVector<Constant*, 4> ResElts;
1904 Type *Ty = IntegerType::get(C1->getContext(), 32);
1905 // Compare the elements, producing an i1 result or constant expr.
1906 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1907 Constant *C1E =
1908 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1909 Constant *C2E =
1910 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1912 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1915 return ConstantVector::get(ResElts);
1918 if (C1->getType()->isFloatingPointTy() &&
1919 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1920 // infinite recursive loop
1921 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1922 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1923 switch (evaluateFCmpRelation(C1, C2)) {
1924 default: llvm_unreachable("Unknown relation!");
1925 case FCmpInst::FCMP_UNO:
1926 case FCmpInst::FCMP_ORD:
1927 case FCmpInst::FCMP_UNE:
1928 case FCmpInst::FCMP_ULT:
1929 case FCmpInst::FCMP_UGT:
1930 case FCmpInst::FCMP_ULE:
1931 case FCmpInst::FCMP_UGE:
1932 case FCmpInst::FCMP_TRUE:
1933 case FCmpInst::FCMP_FALSE:
1934 case FCmpInst::BAD_FCMP_PREDICATE:
1935 break; // Couldn't determine anything about these constants.
1936 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1937 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1938 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1939 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1940 break;
1941 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1942 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1943 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1944 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1945 break;
1946 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1947 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1948 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1949 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1950 break;
1951 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1952 // We can only partially decide this relation.
1953 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1954 Result = 0;
1955 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1956 Result = 1;
1957 break;
1958 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1959 // We can only partially decide this relation.
1960 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1961 Result = 0;
1962 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1963 Result = 1;
1964 break;
1965 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1966 // We can only partially decide this relation.
1967 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1968 Result = 0;
1969 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1970 Result = 1;
1971 break;
1972 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
1973 // We can only partially decide this relation.
1974 if (pred == FCmpInst::FCMP_ONE)
1975 Result = 0;
1976 else if (pred == FCmpInst::FCMP_UEQ)
1977 Result = 1;
1978 break;
1981 // If we evaluated the result, return it now.
1982 if (Result != -1)
1983 return ConstantInt::get(ResultTy, Result);
1985 } else {
1986 // Evaluate the relation between the two constants, per the predicate.
1987 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1988 switch (evaluateICmpRelation(C1, C2,
1989 CmpInst::isSigned((CmpInst::Predicate)pred))) {
1990 default: llvm_unreachable("Unknown relational!");
1991 case ICmpInst::BAD_ICMP_PREDICATE:
1992 break; // Couldn't determine anything about these constants.
1993 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1994 // If we know the constants are equal, we can decide the result of this
1995 // computation precisely.
1996 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1997 break;
1998 case ICmpInst::ICMP_ULT:
1999 switch (pred) {
2000 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
2001 Result = 1; break;
2002 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
2003 Result = 0; break;
2005 break;
2006 case ICmpInst::ICMP_SLT:
2007 switch (pred) {
2008 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
2009 Result = 1; break;
2010 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
2011 Result = 0; break;
2013 break;
2014 case ICmpInst::ICMP_UGT:
2015 switch (pred) {
2016 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
2017 Result = 1; break;
2018 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2019 Result = 0; break;
2021 break;
2022 case ICmpInst::ICMP_SGT:
2023 switch (pred) {
2024 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2025 Result = 1; break;
2026 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2027 Result = 0; break;
2029 break;
2030 case ICmpInst::ICMP_ULE:
2031 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2032 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2033 break;
2034 case ICmpInst::ICMP_SLE:
2035 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2036 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2037 break;
2038 case ICmpInst::ICMP_UGE:
2039 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2040 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2041 break;
2042 case ICmpInst::ICMP_SGE:
2043 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2044 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2045 break;
2046 case ICmpInst::ICMP_NE:
2047 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2048 if (pred == ICmpInst::ICMP_NE) Result = 1;
2049 break;
2052 // If we evaluated the result, return it now.
2053 if (Result != -1)
2054 return ConstantInt::get(ResultTy, Result);
2056 // If the right hand side is a bitcast, try using its inverse to simplify
2057 // it by moving it to the left hand side. We can't do this if it would turn
2058 // a vector compare into a scalar compare or visa versa, or if it would turn
2059 // the operands into FP values.
2060 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2061 Constant *CE2Op0 = CE2->getOperand(0);
2062 if (CE2->getOpcode() == Instruction::BitCast &&
2063 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
2064 !CE2Op0->getType()->isFPOrFPVectorTy()) {
2065 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2066 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2070 // If the left hand side is an extension, try eliminating it.
2071 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2072 if ((CE1->getOpcode() == Instruction::SExt &&
2073 ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2074 (CE1->getOpcode() == Instruction::ZExt &&
2075 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2076 Constant *CE1Op0 = CE1->getOperand(0);
2077 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2078 if (CE1Inverse == CE1Op0) {
2079 // Check whether we can safely truncate the right hand side.
2080 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2081 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2082 C2->getType()) == C2)
2083 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2088 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2089 (C1->isNullValue() && !C2->isNullValue())) {
2090 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2091 // other way if possible.
2092 // Also, if C1 is null and C2 isn't, flip them around.
2093 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2094 return ConstantExpr::getICmp(pred, C2, C1);
2097 return nullptr;
2100 /// Test whether the given sequence of *normalized* indices is "inbounds".
2101 template<typename IndexTy>
2102 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2103 // No indices means nothing that could be out of bounds.
2104 if (Idxs.empty()) return true;
2106 // If the first index is zero, it's in bounds.
2107 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2109 // If the first index is one and all the rest are zero, it's in bounds,
2110 // by the one-past-the-end rule.
2111 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2112 if (!CI->isOne())
2113 return false;
2114 } else {
2115 auto *CV = cast<ConstantDataVector>(Idxs[0]);
2116 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2117 if (!CI || !CI->isOne())
2118 return false;
2121 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2122 if (!cast<Constant>(Idxs[i])->isNullValue())
2123 return false;
2124 return true;
2127 /// Test whether a given ConstantInt is in-range for a SequentialType.
2128 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2129 const ConstantInt *CI) {
2130 // We cannot bounds check the index if it doesn't fit in an int64_t.
2131 if (CI->getValue().getMinSignedBits() > 64)
2132 return false;
2134 // A negative index or an index past the end of our sequential type is
2135 // considered out-of-range.
2136 int64_t IndexVal = CI->getSExtValue();
2137 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2138 return false;
2140 // Otherwise, it is in-range.
2141 return true;
2144 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2145 bool InBounds,
2146 Optional<unsigned> InRangeIndex,
2147 ArrayRef<Value *> Idxs) {
2148 if (Idxs.empty()) return C;
2150 Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2151 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2153 if (isa<UndefValue>(C))
2154 return UndefValue::get(GEPTy);
2156 Constant *Idx0 = cast<Constant>(Idxs[0]);
2157 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2158 return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2159 ? ConstantVector::getSplat(
2160 cast<VectorType>(GEPTy)->getNumElements(), C)
2161 : C;
2163 if (C->isNullValue()) {
2164 bool isNull = true;
2165 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2166 if (!isa<UndefValue>(Idxs[i]) &&
2167 !cast<Constant>(Idxs[i])->isNullValue()) {
2168 isNull = false;
2169 break;
2171 if (isNull) {
2172 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2173 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2175 assert(Ty && "Invalid indices for GEP!");
2176 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2177 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2178 if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2179 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2181 // The GEP returns a vector of pointers when one of more of
2182 // its arguments is a vector.
2183 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2184 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2185 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2186 break;
2190 return Constant::getNullValue(GEPTy);
2194 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2195 // Combine Indices - If the source pointer to this getelementptr instruction
2196 // is a getelementptr instruction, combine the indices of the two
2197 // getelementptr instructions into a single instruction.
2199 if (CE->getOpcode() == Instruction::GetElementPtr) {
2200 gep_type_iterator LastI = gep_type_end(CE);
2201 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2202 I != E; ++I)
2203 LastI = I;
2205 // We cannot combine indices if doing so would take us outside of an
2206 // array or vector. Doing otherwise could trick us if we evaluated such a
2207 // GEP as part of a load.
2209 // e.g. Consider if the original GEP was:
2210 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2211 // i32 0, i32 0, i64 0)
2213 // If we then tried to offset it by '8' to get to the third element,
2214 // an i8, we should *not* get:
2215 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2216 // i32 0, i32 0, i64 8)
2218 // This GEP tries to index array element '8 which runs out-of-bounds.
2219 // Subsequent evaluation would get confused and produce erroneous results.
2221 // The following prohibits such a GEP from being formed by checking to see
2222 // if the index is in-range with respect to an array.
2223 // TODO: This code may be extended to handle vectors as well.
2224 bool PerformFold = false;
2225 if (Idx0->isNullValue())
2226 PerformFold = true;
2227 else if (LastI.isSequential())
2228 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2229 PerformFold = (!LastI.isBoundedSequential() ||
2230 isIndexInRangeOfArrayType(
2231 LastI.getSequentialNumElements(), CI)) &&
2232 !CE->getOperand(CE->getNumOperands() - 1)
2233 ->getType()
2234 ->isVectorTy();
2236 if (PerformFold) {
2237 SmallVector<Value*, 16> NewIndices;
2238 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2239 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2241 // Add the last index of the source with the first index of the new GEP.
2242 // Make sure to handle the case when they are actually different types.
2243 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2244 // Otherwise it must be an array.
2245 if (!Idx0->isNullValue()) {
2246 Type *IdxTy = Combined->getType();
2247 if (IdxTy != Idx0->getType()) {
2248 unsigned CommonExtendedWidth =
2249 std::max(IdxTy->getIntegerBitWidth(),
2250 Idx0->getType()->getIntegerBitWidth());
2251 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2253 Type *CommonTy =
2254 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2255 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2256 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2257 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2258 } else {
2259 Combined =
2260 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2264 NewIndices.push_back(Combined);
2265 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2267 // The combined GEP normally inherits its index inrange attribute from
2268 // the inner GEP, but if the inner GEP's last index was adjusted by the
2269 // outer GEP, any inbounds attribute on that index is invalidated.
2270 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex();
2271 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue())
2272 IRIndex = None;
2274 return ConstantExpr::getGetElementPtr(
2275 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2276 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(),
2277 IRIndex);
2281 // Attempt to fold casts to the same type away. For example, folding:
2283 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2284 // i64 0, i64 0)
2285 // into:
2287 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2289 // Don't fold if the cast is changing address spaces.
2290 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2291 PointerType *SrcPtrTy =
2292 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2293 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2294 if (SrcPtrTy && DstPtrTy) {
2295 ArrayType *SrcArrayTy =
2296 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2297 ArrayType *DstArrayTy =
2298 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2299 if (SrcArrayTy && DstArrayTy
2300 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2301 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2302 return ConstantExpr::getGetElementPtr(SrcArrayTy,
2303 (Constant *)CE->getOperand(0),
2304 Idxs, InBounds, InRangeIndex);
2309 // Check to see if any array indices are not within the corresponding
2310 // notional array or vector bounds. If so, try to determine if they can be
2311 // factored out into preceding dimensions.
2312 SmallVector<Constant *, 8> NewIdxs;
2313 Type *Ty = PointeeTy;
2314 Type *Prev = C->getType();
2315 bool Unknown =
2316 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2317 for (unsigned i = 1, e = Idxs.size(); i != e;
2318 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2319 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2320 // We don't know if it's in range or not.
2321 Unknown = true;
2322 continue;
2324 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2325 // Skip if the type of the previous index is not supported.
2326 continue;
2327 if (InRangeIndex && i == *InRangeIndex + 1) {
2328 // If an index is marked inrange, we cannot apply this canonicalization to
2329 // the following index, as that will cause the inrange index to point to
2330 // the wrong element.
2331 continue;
2333 if (isa<StructType>(Ty)) {
2334 // The verify makes sure that GEPs into a struct are in range.
2335 continue;
2337 auto *STy = cast<SequentialType>(Ty);
2338 if (isa<VectorType>(STy)) {
2339 // There can be awkward padding in after a non-power of two vector.
2340 Unknown = true;
2341 continue;
2343 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2344 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2345 // It's in range, skip to the next index.
2346 continue;
2347 if (CI->getSExtValue() < 0) {
2348 // It's out of range and negative, don't try to factor it.
2349 Unknown = true;
2350 continue;
2352 } else {
2353 auto *CV = cast<ConstantDataVector>(Idxs[i]);
2354 bool InRange = true;
2355 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2356 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2357 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2358 if (CI->getSExtValue() < 0) {
2359 Unknown = true;
2360 break;
2363 if (InRange || Unknown)
2364 // It's in range, skip to the next index.
2365 // It's out of range and negative, don't try to factor it.
2366 continue;
2368 if (isa<StructType>(Prev)) {
2369 // It's out of range, but the prior dimension is a struct
2370 // so we can't do anything about it.
2371 Unknown = true;
2372 continue;
2374 // It's out of range, but we can factor it into the prior
2375 // dimension.
2376 NewIdxs.resize(Idxs.size());
2377 // Determine the number of elements in our sequential type.
2378 uint64_t NumElements = STy->getArrayNumElements();
2380 // Expand the current index or the previous index to a vector from a scalar
2381 // if necessary.
2382 Constant *CurrIdx = cast<Constant>(Idxs[i]);
2383 auto *PrevIdx =
2384 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2385 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2386 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2387 bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2389 if (!IsCurrIdxVector && IsPrevIdxVector)
2390 CurrIdx = ConstantDataVector::getSplat(
2391 PrevIdx->getType()->getVectorNumElements(), CurrIdx);
2393 if (!IsPrevIdxVector && IsCurrIdxVector)
2394 PrevIdx = ConstantDataVector::getSplat(
2395 CurrIdx->getType()->getVectorNumElements(), PrevIdx);
2397 Constant *Factor =
2398 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2399 if (UseVector)
2400 Factor = ConstantDataVector::getSplat(
2401 IsPrevIdxVector ? PrevIdx->getType()->getVectorNumElements()
2402 : CurrIdx->getType()->getVectorNumElements(),
2403 Factor);
2405 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2407 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2409 unsigned CommonExtendedWidth =
2410 std::max(PrevIdx->getType()->getScalarSizeInBits(),
2411 Div->getType()->getScalarSizeInBits());
2412 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2414 // Before adding, extend both operands to i64 to avoid
2415 // overflow trouble.
2416 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2417 if (UseVector)
2418 ExtendedTy = VectorType::get(
2419 ExtendedTy, IsPrevIdxVector
2420 ? PrevIdx->getType()->getVectorNumElements()
2421 : CurrIdx->getType()->getVectorNumElements());
2423 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2424 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2426 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2427 Div = ConstantExpr::getSExt(Div, ExtendedTy);
2429 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2432 // If we did any factoring, start over with the adjusted indices.
2433 if (!NewIdxs.empty()) {
2434 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2435 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2436 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2437 InRangeIndex);
2440 // If all indices are known integers and normalized, we can do a simple
2441 // check for the "inbounds" property.
2442 if (!Unknown && !InBounds)
2443 if (auto *GV = dyn_cast<GlobalVariable>(C))
2444 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2445 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2446 /*InBounds=*/true, InRangeIndex);
2448 return nullptr;