llvm-config: Include -stdlib= in --cxxflags
[llvm-project.git] / llvm / lib / IR / ConstantFold.cpp
blob037f120f47db99620667eefb1b9f18a5f789e306
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/Operator.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/ManagedStatic.h"
33 #include "llvm/Support/MathExtras.h"
34 using namespace llvm;
35 using namespace llvm::PatternMatch;
37 //===----------------------------------------------------------------------===//
38 // ConstantFold*Instruction Implementations
39 //===----------------------------------------------------------------------===//
41 /// Convert the specified vector Constant node to the specified vector type.
42 /// At this point, we know that the elements of the input vector constant are
43 /// all simple integer or FP values.
44 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
47 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49 // If this cast changes element count then we can't handle it here:
50 // doing so requires endianness information. This should be handled by
51 // Analysis/ConstantFolding.cpp
52 unsigned NumElts = DstTy->getNumElements();
53 if (NumElts != CV->getType()->getVectorNumElements())
54 return nullptr;
56 Type *DstEltTy = DstTy->getElementType();
58 SmallVector<Constant*, 16> Result;
59 Type *Ty = IntegerType::get(CV->getContext(), 32);
60 for (unsigned i = 0; i != NumElts; ++i) {
61 Constant *C =
62 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
63 C = ConstantExpr::getBitCast(C, DstEltTy);
64 Result.push_back(C);
67 return ConstantVector::get(Result);
70 /// This function determines which opcode to use to fold two constant cast
71 /// expressions together. It uses CastInst::isEliminableCastPair to determine
72 /// the opcode. Consequently its just a wrapper around that function.
73 /// Determine if it is valid to fold a cast of a cast
74 static unsigned
75 foldConstantCastPair(
76 unsigned opc, ///< opcode of the second cast constant expression
77 ConstantExpr *Op, ///< the first cast constant expression
78 Type *DstTy ///< destination type of the first cast
79 ) {
80 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
81 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
82 assert(CastInst::isCast(opc) && "Invalid cast opcode");
84 // The types and opcodes for the two Cast constant expressions
85 Type *SrcTy = Op->getOperand(0)->getType();
86 Type *MidTy = Op->getType();
87 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
88 Instruction::CastOps secondOp = Instruction::CastOps(opc);
90 // Assume that pointers are never more than 64 bits wide, and only use this
91 // for the middle type. Otherwise we could end up folding away illegal
92 // bitcasts between address spaces with different sizes.
93 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
95 // Let CastInst::isEliminableCastPair do the heavy lifting.
96 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
97 nullptr, FakeIntPtrTy, nullptr);
100 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
101 Type *SrcTy = V->getType();
102 if (SrcTy == DestTy)
103 return V; // no-op cast
105 // Check to see if we are casting a pointer to an aggregate to a pointer to
106 // the first element. If so, return the appropriate GEP instruction.
107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
109 if (PTy->getAddressSpace() == DPTy->getAddressSpace()
110 && PTy->getElementType()->isSized()) {
111 SmallVector<Value*, 8> IdxList;
112 Value *Zero =
113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
114 IdxList.push_back(Zero);
115 Type *ElTy = PTy->getElementType();
116 while (ElTy != DPTy->getElementType()) {
117 if (StructType *STy = dyn_cast<StructType>(ElTy)) {
118 if (STy->getNumElements() == 0) break;
119 ElTy = STy->getElementType(0);
120 IdxList.push_back(Zero);
121 } else if (SequentialType *STy =
122 dyn_cast<SequentialType>(ElTy)) {
123 ElTy = STy->getElementType();
124 IdxList.push_back(Zero);
125 } else {
126 break;
130 if (ElTy == DPTy->getElementType())
131 // This GEP is inbounds because all indices are zero.
132 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
133 V, IdxList);
136 // Handle casts from one vector constant to another. We know that the src
137 // and dest type have the same size (otherwise its an illegal cast).
138 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
139 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
140 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
141 "Not cast between same sized vectors!");
142 SrcTy = nullptr;
143 // First, check for null. Undef is already handled.
144 if (isa<ConstantAggregateZero>(V))
145 return Constant::getNullValue(DestTy);
147 // Handle ConstantVector and ConstantAggregateVector.
148 return BitCastConstantVector(V, DestPTy);
151 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
152 // This allows for other simplifications (although some of them
153 // can only be handled by Analysis/ConstantFolding.cpp).
154 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
155 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
158 // Finally, implement bitcast folding now. The code below doesn't handle
159 // bitcast right.
160 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
161 return ConstantPointerNull::get(cast<PointerType>(DestTy));
163 // Handle integral constant input.
164 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
165 if (DestTy->isIntegerTy())
166 // Integral -> Integral. This is a no-op because the bit widths must
167 // be the same. Consequently, we just fold to V.
168 return V;
170 // See note below regarding the PPC_FP128 restriction.
171 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
172 return ConstantFP::get(DestTy->getContext(),
173 APFloat(DestTy->getFltSemantics(),
174 CI->getValue()));
176 // Otherwise, can't fold this (vector?)
177 return nullptr;
180 // Handle ConstantFP input: FP -> Integral.
181 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
182 // PPC_FP128 is really the sum of two consecutive doubles, where the first
183 // double is always stored first in memory, regardless of the target
184 // endianness. The memory layout of i128, however, depends on the target
185 // endianness, and so we can't fold this without target endianness
186 // information. This should instead be handled by
187 // Analysis/ConstantFolding.cpp
188 if (FP->getType()->isPPC_FP128Ty())
189 return nullptr;
191 // Make sure dest type is compatible with the folded integer constant.
192 if (!DestTy->isIntegerTy())
193 return nullptr;
195 return ConstantInt::get(FP->getContext(),
196 FP->getValueAPF().bitcastToAPInt());
199 return nullptr;
203 /// V is an integer constant which only has a subset of its bytes used.
204 /// The bytes used are indicated by ByteStart (which is the first byte used,
205 /// counting from the least significant byte) and ByteSize, which is the number
206 /// of bytes used.
208 /// This function analyzes the specified constant to see if the specified byte
209 /// range can be returned as a simplified constant. If so, the constant is
210 /// returned, otherwise null is returned.
211 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
212 unsigned ByteSize) {
213 assert(C->getType()->isIntegerTy() &&
214 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
215 "Non-byte sized integer input");
216 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
217 assert(ByteSize && "Must be accessing some piece");
218 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
219 assert(ByteSize != CSize && "Should not extract everything");
221 // Constant Integers are simple.
222 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
223 APInt V = CI->getValue();
224 if (ByteStart)
225 V.lshrInPlace(ByteStart*8);
226 V = V.trunc(ByteSize*8);
227 return ConstantInt::get(CI->getContext(), V);
230 // In the input is a constant expr, we might be able to recursively simplify.
231 // If not, we definitely can't do anything.
232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
233 if (!CE) return nullptr;
235 switch (CE->getOpcode()) {
236 default: return nullptr;
237 case Instruction::Or: {
238 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
239 if (!RHS)
240 return nullptr;
242 // X | -1 -> -1.
243 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
244 if (RHSC->isMinusOne())
245 return RHSC;
247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
248 if (!LHS)
249 return nullptr;
250 return ConstantExpr::getOr(LHS, RHS);
252 case Instruction::And: {
253 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
254 if (!RHS)
255 return nullptr;
257 // X & 0 -> 0.
258 if (RHS->isNullValue())
259 return RHS;
261 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
262 if (!LHS)
263 return nullptr;
264 return ConstantExpr::getAnd(LHS, RHS);
266 case Instruction::LShr: {
267 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
268 if (!Amt)
269 return nullptr;
270 unsigned ShAmt = Amt->getZExtValue();
271 // Cannot analyze non-byte shifts.
272 if ((ShAmt & 7) != 0)
273 return nullptr;
274 ShAmt >>= 3;
276 // If the extract is known to be all zeros, return zero.
277 if (ByteStart >= CSize-ShAmt)
278 return Constant::getNullValue(IntegerType::get(CE->getContext(),
279 ByteSize*8));
280 // If the extract is known to be fully in the input, extract it.
281 if (ByteStart+ByteSize+ShAmt <= CSize)
282 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
284 // TODO: Handle the 'partially zero' case.
285 return nullptr;
288 case Instruction::Shl: {
289 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
290 if (!Amt)
291 return nullptr;
292 unsigned ShAmt = Amt->getZExtValue();
293 // Cannot analyze non-byte shifts.
294 if ((ShAmt & 7) != 0)
295 return nullptr;
296 ShAmt >>= 3;
298 // If the extract is known to be all zeros, return zero.
299 if (ByteStart+ByteSize <= ShAmt)
300 return Constant::getNullValue(IntegerType::get(CE->getContext(),
301 ByteSize*8));
302 // If the extract is known to be fully in the input, extract it.
303 if (ByteStart >= ShAmt)
304 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
306 // TODO: Handle the 'partially zero' case.
307 return nullptr;
310 case Instruction::ZExt: {
311 unsigned SrcBitSize =
312 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
314 // If extracting something that is completely zero, return 0.
315 if (ByteStart*8 >= SrcBitSize)
316 return Constant::getNullValue(IntegerType::get(CE->getContext(),
317 ByteSize*8));
319 // If exactly extracting the input, return it.
320 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
321 return CE->getOperand(0);
323 // If extracting something completely in the input, if the input is a
324 // multiple of 8 bits, recurse.
325 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
326 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
328 // Otherwise, if extracting a subset of the input, which is not multiple of
329 // 8 bits, do a shift and trunc to get the bits.
330 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
331 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
332 Constant *Res = CE->getOperand(0);
333 if (ByteStart)
334 Res = ConstantExpr::getLShr(Res,
335 ConstantInt::get(Res->getType(), ByteStart*8));
336 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
337 ByteSize*8));
340 // TODO: Handle the 'partially zero' case.
341 return nullptr;
346 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known
347 /// factors factored out. If Folded is false, return null if no factoring was
348 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
349 /// top-level folder.
350 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded) {
351 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
352 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
353 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
354 return ConstantExpr::getNUWMul(E, N);
357 if (StructType *STy = dyn_cast<StructType>(Ty))
358 if (!STy->isPacked()) {
359 unsigned NumElems = STy->getNumElements();
360 // An empty struct has size zero.
361 if (NumElems == 0)
362 return ConstantExpr::getNullValue(DestTy);
363 // Check for a struct with all members having the same size.
364 Constant *MemberSize =
365 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
366 bool AllSame = true;
367 for (unsigned i = 1; i != NumElems; ++i)
368 if (MemberSize !=
369 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
370 AllSame = false;
371 break;
373 if (AllSame) {
374 Constant *N = ConstantInt::get(DestTy, NumElems);
375 return ConstantExpr::getNUWMul(MemberSize, N);
379 // Pointer size doesn't depend on the pointee type, so canonicalize them
380 // to an arbitrary pointee.
381 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
382 if (!PTy->getElementType()->isIntegerTy(1))
383 return
384 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
385 PTy->getAddressSpace()),
386 DestTy, true);
388 // If there's no interesting folding happening, bail so that we don't create
389 // a constant that looks like it needs folding but really doesn't.
390 if (!Folded)
391 return nullptr;
393 // Base case: Get a regular sizeof expression.
394 Constant *C = ConstantExpr::getSizeOf(Ty);
395 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
396 DestTy, false),
397 C, DestTy);
398 return C;
401 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known
402 /// factors factored out. If Folded is false, return null if no factoring was
403 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
404 /// top-level folder.
405 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) {
406 // The alignment of an array is equal to the alignment of the
407 // array element. Note that this is not always true for vectors.
408 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
409 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
410 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
411 DestTy,
412 false),
413 C, DestTy);
414 return C;
417 if (StructType *STy = dyn_cast<StructType>(Ty)) {
418 // Packed structs always have an alignment of 1.
419 if (STy->isPacked())
420 return ConstantInt::get(DestTy, 1);
422 // Otherwise, struct alignment is the maximum alignment of any member.
423 // Without target data, we can't compare much, but we can check to see
424 // if all the members have the same alignment.
425 unsigned NumElems = STy->getNumElements();
426 // An empty struct has minimal alignment.
427 if (NumElems == 0)
428 return ConstantInt::get(DestTy, 1);
429 // Check for a struct with all members having the same alignment.
430 Constant *MemberAlign =
431 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
432 bool AllSame = true;
433 for (unsigned i = 1; i != NumElems; ++i)
434 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
435 AllSame = false;
436 break;
438 if (AllSame)
439 return MemberAlign;
442 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
443 // to an arbitrary pointee.
444 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
445 if (!PTy->getElementType()->isIntegerTy(1))
446 return
447 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
449 PTy->getAddressSpace()),
450 DestTy, true);
452 // If there's no interesting folding happening, bail so that we don't create
453 // a constant that looks like it needs folding but really doesn't.
454 if (!Folded)
455 return nullptr;
457 // Base case: Get a regular alignof expression.
458 Constant *C = ConstantExpr::getAlignOf(Ty);
459 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
460 DestTy, false),
461 C, DestTy);
462 return C;
465 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with
466 /// any known factors factored out. If Folded is false, return null if no
467 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression
468 /// back into the top-level folder.
469 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, Type *DestTy,
470 bool Folded) {
471 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
472 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
473 DestTy, false),
474 FieldNo, DestTy);
475 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
476 return ConstantExpr::getNUWMul(E, N);
479 if (StructType *STy = dyn_cast<StructType>(Ty))
480 if (!STy->isPacked()) {
481 unsigned NumElems = STy->getNumElements();
482 // An empty struct has no members.
483 if (NumElems == 0)
484 return nullptr;
485 // Check for a struct with all members having the same size.
486 Constant *MemberSize =
487 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
488 bool AllSame = true;
489 for (unsigned i = 1; i != NumElems; ++i)
490 if (MemberSize !=
491 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
492 AllSame = false;
493 break;
495 if (AllSame) {
496 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
497 false,
498 DestTy,
499 false),
500 FieldNo, DestTy);
501 return ConstantExpr::getNUWMul(MemberSize, N);
505 // If there's no interesting folding happening, bail so that we don't create
506 // a constant that looks like it needs folding but really doesn't.
507 if (!Folded)
508 return nullptr;
510 // Base case: Get a regular offsetof expression.
511 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
512 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
513 DestTy, false),
514 C, DestTy);
515 return C;
518 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
519 Type *DestTy) {
520 if (isa<UndefValue>(V)) {
521 // zext(undef) = 0, because the top bits will be zero.
522 // sext(undef) = 0, because the top bits will all be the same.
523 // [us]itofp(undef) = 0, because the result value is bounded.
524 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
525 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
526 return Constant::getNullValue(DestTy);
527 return UndefValue::get(DestTy);
530 if (V->isNullValue() && !DestTy->isX86_MMXTy() &&
531 opc != Instruction::AddrSpaceCast)
532 return Constant::getNullValue(DestTy);
534 // If the cast operand is a constant expression, there's a few things we can
535 // do to try to simplify it.
536 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
537 if (CE->isCast()) {
538 // Try hard to fold cast of cast because they are often eliminable.
539 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
540 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
541 } else if (CE->getOpcode() == Instruction::GetElementPtr &&
542 // Do not fold addrspacecast (gep 0, .., 0). It might make the
543 // addrspacecast uncanonicalized.
544 opc != Instruction::AddrSpaceCast &&
545 // Do not fold bitcast (gep) with inrange index, as this loses
546 // information.
547 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() &&
548 // Do not fold if the gep type is a vector, as bitcasting
549 // operand 0 of a vector gep will result in a bitcast between
550 // different sizes.
551 !CE->getType()->isVectorTy()) {
552 // If all of the indexes in the GEP are null values, there is no pointer
553 // adjustment going on. We might as well cast the source pointer.
554 bool isAllNull = true;
555 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
556 if (!CE->getOperand(i)->isNullValue()) {
557 isAllNull = false;
558 break;
560 if (isAllNull)
561 // This is casting one pointer type to another, always BitCast
562 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
566 // If the cast operand is a constant vector, perform the cast by
567 // operating on each element. In the cast of bitcasts, the element
568 // count may be mismatched; don't attempt to handle that here.
569 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
570 DestTy->isVectorTy() &&
571 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
572 SmallVector<Constant*, 16> res;
573 VectorType *DestVecTy = cast<VectorType>(DestTy);
574 Type *DstEltTy = DestVecTy->getElementType();
575 Type *Ty = IntegerType::get(V->getContext(), 32);
576 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
577 Constant *C =
578 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
579 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
581 return ConstantVector::get(res);
584 // We actually have to do a cast now. Perform the cast according to the
585 // opcode specified.
586 switch (opc) {
587 default:
588 llvm_unreachable("Failed to cast constant expression");
589 case Instruction::FPTrunc:
590 case Instruction::FPExt:
591 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
592 bool ignored;
593 APFloat Val = FPC->getValueAPF();
594 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
595 DestTy->isFloatTy() ? APFloat::IEEEsingle() :
596 DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
597 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
598 DestTy->isFP128Ty() ? APFloat::IEEEquad() :
599 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
600 APFloat::Bogus(),
601 APFloat::rmNearestTiesToEven, &ignored);
602 return ConstantFP::get(V->getContext(), Val);
604 return nullptr; // Can't fold.
605 case Instruction::FPToUI:
606 case Instruction::FPToSI:
607 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
608 const APFloat &V = FPC->getValueAPF();
609 bool ignored;
610 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
611 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
612 if (APFloat::opInvalidOp ==
613 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
614 // Undefined behavior invoked - the destination type can't represent
615 // the input constant.
616 return UndefValue::get(DestTy);
618 return ConstantInt::get(FPC->getContext(), IntVal);
620 return nullptr; // Can't fold.
621 case Instruction::IntToPtr: //always treated as unsigned
622 if (V->isNullValue()) // Is it an integral null value?
623 return ConstantPointerNull::get(cast<PointerType>(DestTy));
624 return nullptr; // Other pointer types cannot be casted
625 case Instruction::PtrToInt: // always treated as unsigned
626 // Is it a null pointer value?
627 if (V->isNullValue())
628 return ConstantInt::get(DestTy, 0);
629 // If this is a sizeof-like expression, pull out multiplications by
630 // known factors to expose them to subsequent folding. If it's an
631 // alignof-like expression, factor out known factors.
632 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
633 if (CE->getOpcode() == Instruction::GetElementPtr &&
634 CE->getOperand(0)->isNullValue()) {
635 // FIXME: Looks like getFoldedSizeOf(), getFoldedOffsetOf() and
636 // getFoldedAlignOf() don't handle the case when DestTy is a vector of
637 // pointers yet. We end up in asserts in CastInst::getCastOpcode (see
638 // test/Analysis/ConstantFolding/cast-vector.ll). I've only seen this
639 // happen in one "real" C-code test case, so it does not seem to be an
640 // important optimization to handle vectors here. For now, simply bail
641 // out.
642 if (DestTy->isVectorTy())
643 return nullptr;
644 GEPOperator *GEPO = cast<GEPOperator>(CE);
645 Type *Ty = GEPO->getSourceElementType();
646 if (CE->getNumOperands() == 2) {
647 // Handle a sizeof-like expression.
648 Constant *Idx = CE->getOperand(1);
649 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
650 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
651 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
652 DestTy, false),
653 Idx, DestTy);
654 return ConstantExpr::getMul(C, Idx);
656 } else if (CE->getNumOperands() == 3 &&
657 CE->getOperand(1)->isNullValue()) {
658 // Handle an alignof-like expression.
659 if (StructType *STy = dyn_cast<StructType>(Ty))
660 if (!STy->isPacked()) {
661 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
662 if (CI->isOne() &&
663 STy->getNumElements() == 2 &&
664 STy->getElementType(0)->isIntegerTy(1)) {
665 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
668 // Handle an offsetof-like expression.
669 if (Ty->isStructTy() || Ty->isArrayTy()) {
670 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
671 DestTy, false))
672 return C;
676 // Other pointer types cannot be casted
677 return nullptr;
678 case Instruction::UIToFP:
679 case Instruction::SIToFP:
680 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
681 const APInt &api = CI->getValue();
682 APFloat apf(DestTy->getFltSemantics(),
683 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
684 apf.convertFromAPInt(api, opc==Instruction::SIToFP,
685 APFloat::rmNearestTiesToEven);
686 return ConstantFP::get(V->getContext(), apf);
688 return nullptr;
689 case Instruction::ZExt:
690 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
691 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
692 return ConstantInt::get(V->getContext(),
693 CI->getValue().zext(BitWidth));
695 return nullptr;
696 case Instruction::SExt:
697 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
698 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
699 return ConstantInt::get(V->getContext(),
700 CI->getValue().sext(BitWidth));
702 return nullptr;
703 case Instruction::Trunc: {
704 if (V->getType()->isVectorTy())
705 return nullptr;
707 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
708 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
709 return ConstantInt::get(V->getContext(),
710 CI->getValue().trunc(DestBitWidth));
713 // The input must be a constantexpr. See if we can simplify this based on
714 // the bytes we are demanding. Only do this if the source and dest are an
715 // even multiple of a byte.
716 if ((DestBitWidth & 7) == 0 &&
717 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
718 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
719 return Res;
721 return nullptr;
723 case Instruction::BitCast:
724 return FoldBitCast(V, DestTy);
725 case Instruction::AddrSpaceCast:
726 return nullptr;
730 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
731 Constant *V1, Constant *V2) {
732 // Check for i1 and vector true/false conditions.
733 if (Cond->isNullValue()) return V2;
734 if (Cond->isAllOnesValue()) return V1;
736 // If the condition is a vector constant, fold the result elementwise.
737 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
738 SmallVector<Constant*, 16> Result;
739 Type *Ty = IntegerType::get(CondV->getContext(), 32);
740 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
741 Constant *V;
742 Constant *V1Element = ConstantExpr::getExtractElement(V1,
743 ConstantInt::get(Ty, i));
744 Constant *V2Element = ConstantExpr::getExtractElement(V2,
745 ConstantInt::get(Ty, i));
746 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
747 if (V1Element == V2Element) {
748 V = V1Element;
749 } else if (isa<UndefValue>(Cond)) {
750 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
751 } else {
752 if (!isa<ConstantInt>(Cond)) break;
753 V = Cond->isNullValue() ? V2Element : V1Element;
755 Result.push_back(V);
758 // If we were able to build the vector, return it.
759 if (Result.size() == V1->getType()->getVectorNumElements())
760 return ConstantVector::get(Result);
763 if (isa<UndefValue>(Cond)) {
764 if (isa<UndefValue>(V1)) return V1;
765 return V2;
767 if (isa<UndefValue>(V1)) return V2;
768 if (isa<UndefValue>(V2)) return V1;
769 if (V1 == V2) return V1;
771 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
772 if (TrueVal->getOpcode() == Instruction::Select)
773 if (TrueVal->getOperand(0) == Cond)
774 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
776 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
777 if (FalseVal->getOpcode() == Instruction::Select)
778 if (FalseVal->getOperand(0) == Cond)
779 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
782 return nullptr;
785 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
786 Constant *Idx) {
787 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
788 return UndefValue::get(Val->getType()->getVectorElementType());
789 if (Val->isNullValue()) // ee(zero, x) -> zero
790 return Constant::getNullValue(Val->getType()->getVectorElementType());
791 // ee({w,x,y,z}, undef) -> undef
792 if (isa<UndefValue>(Idx))
793 return UndefValue::get(Val->getType()->getVectorElementType());
795 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
796 // ee({w,x,y,z}, wrong_value) -> undef
797 if (CIdx->uge(Val->getType()->getVectorNumElements()))
798 return UndefValue::get(Val->getType()->getVectorElementType());
799 return Val->getAggregateElement(CIdx->getZExtValue());
801 return nullptr;
804 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
805 Constant *Elt,
806 Constant *Idx) {
807 if (isa<UndefValue>(Idx))
808 return UndefValue::get(Val->getType());
810 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
811 if (!CIdx) return nullptr;
813 unsigned NumElts = Val->getType()->getVectorNumElements();
814 if (CIdx->uge(NumElts))
815 return UndefValue::get(Val->getType());
817 SmallVector<Constant*, 16> Result;
818 Result.reserve(NumElts);
819 auto *Ty = Type::getInt32Ty(Val->getContext());
820 uint64_t IdxVal = CIdx->getZExtValue();
821 for (unsigned i = 0; i != NumElts; ++i) {
822 if (i == IdxVal) {
823 Result.push_back(Elt);
824 continue;
827 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
828 Result.push_back(C);
831 return ConstantVector::get(Result);
834 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
835 Constant *V2,
836 Constant *Mask) {
837 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
838 Type *EltTy = V1->getType()->getVectorElementType();
840 // Undefined shuffle mask -> undefined value.
841 if (isa<UndefValue>(Mask))
842 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
844 // Don't break the bitcode reader hack.
845 if (isa<ConstantExpr>(Mask)) return nullptr;
847 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
849 // Loop over the shuffle mask, evaluating each element.
850 SmallVector<Constant*, 32> Result;
851 for (unsigned i = 0; i != MaskNumElts; ++i) {
852 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
853 if (Elt == -1) {
854 Result.push_back(UndefValue::get(EltTy));
855 continue;
857 Constant *InElt;
858 if (unsigned(Elt) >= SrcNumElts*2)
859 InElt = UndefValue::get(EltTy);
860 else if (unsigned(Elt) >= SrcNumElts) {
861 Type *Ty = IntegerType::get(V2->getContext(), 32);
862 InElt =
863 ConstantExpr::getExtractElement(V2,
864 ConstantInt::get(Ty, Elt - SrcNumElts));
865 } else {
866 Type *Ty = IntegerType::get(V1->getContext(), 32);
867 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
869 Result.push_back(InElt);
872 return ConstantVector::get(Result);
875 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
876 ArrayRef<unsigned> Idxs) {
877 // Base case: no indices, so return the entire value.
878 if (Idxs.empty())
879 return Agg;
881 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
882 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
884 return nullptr;
887 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
888 Constant *Val,
889 ArrayRef<unsigned> Idxs) {
890 // Base case: no indices, so replace the entire value.
891 if (Idxs.empty())
892 return Val;
894 unsigned NumElts;
895 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
896 NumElts = ST->getNumElements();
897 else
898 NumElts = cast<SequentialType>(Agg->getType())->getNumElements();
900 SmallVector<Constant*, 32> Result;
901 for (unsigned i = 0; i != NumElts; ++i) {
902 Constant *C = Agg->getAggregateElement(i);
903 if (!C) return nullptr;
905 if (Idxs[0] == i)
906 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
908 Result.push_back(C);
911 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
912 return ConstantStruct::get(ST, Result);
913 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
914 return ConstantArray::get(AT, Result);
915 return ConstantVector::get(Result);
918 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
919 Constant *C2) {
920 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
922 // Handle scalar UndefValue. Vectors are always evaluated per element.
923 bool HasScalarUndef = !C1->getType()->isVectorTy() &&
924 (isa<UndefValue>(C1) || isa<UndefValue>(C2));
925 if (HasScalarUndef) {
926 switch (static_cast<Instruction::BinaryOps>(Opcode)) {
927 case Instruction::Xor:
928 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
929 // Handle undef ^ undef -> 0 special case. This is a common
930 // idiom (misuse).
931 return Constant::getNullValue(C1->getType());
932 LLVM_FALLTHROUGH;
933 case Instruction::Add:
934 case Instruction::Sub:
935 return UndefValue::get(C1->getType());
936 case Instruction::And:
937 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
938 return C1;
939 return Constant::getNullValue(C1->getType()); // undef & X -> 0
940 case Instruction::Mul: {
941 // undef * undef -> undef
942 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
943 return C1;
944 const APInt *CV;
945 // X * undef -> undef if X is odd
946 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
947 if ((*CV)[0])
948 return UndefValue::get(C1->getType());
950 // X * undef -> 0 otherwise
951 return Constant::getNullValue(C1->getType());
953 case Instruction::SDiv:
954 case Instruction::UDiv:
955 // X / undef -> undef
956 if (isa<UndefValue>(C2))
957 return C2;
958 // undef / 0 -> undef
959 // undef / 1 -> undef
960 if (match(C2, m_Zero()) || match(C2, m_One()))
961 return C1;
962 // undef / X -> 0 otherwise
963 return Constant::getNullValue(C1->getType());
964 case Instruction::URem:
965 case Instruction::SRem:
966 // X % undef -> undef
967 if (match(C2, m_Undef()))
968 return C2;
969 // undef % 0 -> undef
970 if (match(C2, m_Zero()))
971 return C1;
972 // undef % X -> 0 otherwise
973 return Constant::getNullValue(C1->getType());
974 case Instruction::Or: // X | undef -> -1
975 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
976 return C1;
977 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
978 case Instruction::LShr:
979 // X >>l undef -> undef
980 if (isa<UndefValue>(C2))
981 return C2;
982 // undef >>l 0 -> undef
983 if (match(C2, m_Zero()))
984 return C1;
985 // undef >>l X -> 0
986 return Constant::getNullValue(C1->getType());
987 case Instruction::AShr:
988 // X >>a undef -> undef
989 if (isa<UndefValue>(C2))
990 return C2;
991 // undef >>a 0 -> undef
992 if (match(C2, m_Zero()))
993 return C1;
994 // TODO: undef >>a X -> undef if the shift is exact
995 // undef >>a X -> 0
996 return Constant::getNullValue(C1->getType());
997 case Instruction::Shl:
998 // X << undef -> undef
999 if (isa<UndefValue>(C2))
1000 return C2;
1001 // undef << 0 -> undef
1002 if (match(C2, m_Zero()))
1003 return C1;
1004 // undef << X -> 0
1005 return Constant::getNullValue(C1->getType());
1006 case Instruction::FAdd:
1007 case Instruction::FSub:
1008 case Instruction::FMul:
1009 case Instruction::FDiv:
1010 case Instruction::FRem:
1011 // [any flop] undef, undef -> undef
1012 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1013 return C1;
1014 // [any flop] C, undef -> NaN
1015 // [any flop] undef, C -> NaN
1016 // We could potentially specialize NaN/Inf constants vs. 'normal'
1017 // constants (possibly differently depending on opcode and operand). This
1018 // would allow returning undef sometimes. But it is always safe to fold to
1019 // NaN because we can choose the undef operand as NaN, and any FP opcode
1020 // with a NaN operand will propagate NaN.
1021 return ConstantFP::getNaN(C1->getType());
1022 case Instruction::BinaryOpsEnd:
1023 llvm_unreachable("Invalid BinaryOp");
1027 // Neither constant should be UndefValue, unless these are vector constants.
1028 assert(!HasScalarUndef && "Unexpected UndefValue");
1030 // Handle simplifications when the RHS is a constant int.
1031 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1032 switch (Opcode) {
1033 case Instruction::Add:
1034 if (CI2->isZero()) return C1; // X + 0 == X
1035 break;
1036 case Instruction::Sub:
1037 if (CI2->isZero()) return C1; // X - 0 == X
1038 break;
1039 case Instruction::Mul:
1040 if (CI2->isZero()) return C2; // X * 0 == 0
1041 if (CI2->isOne())
1042 return C1; // X * 1 == X
1043 break;
1044 case Instruction::UDiv:
1045 case Instruction::SDiv:
1046 if (CI2->isOne())
1047 return C1; // X / 1 == X
1048 if (CI2->isZero())
1049 return UndefValue::get(CI2->getType()); // X / 0 == undef
1050 break;
1051 case Instruction::URem:
1052 case Instruction::SRem:
1053 if (CI2->isOne())
1054 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1055 if (CI2->isZero())
1056 return UndefValue::get(CI2->getType()); // X % 0 == undef
1057 break;
1058 case Instruction::And:
1059 if (CI2->isZero()) return C2; // X & 0 == 0
1060 if (CI2->isMinusOne())
1061 return C1; // X & -1 == X
1063 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1064 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1065 if (CE1->getOpcode() == Instruction::ZExt) {
1066 unsigned DstWidth = CI2->getType()->getBitWidth();
1067 unsigned SrcWidth =
1068 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1069 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1070 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1071 return C1;
1074 // If and'ing the address of a global with a constant, fold it.
1075 if (CE1->getOpcode() == Instruction::PtrToInt &&
1076 isa<GlobalValue>(CE1->getOperand(0))) {
1077 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1079 // Functions are at least 4-byte aligned.
1080 unsigned GVAlign = GV->getAlignment();
1081 if (isa<Function>(GV))
1082 GVAlign = std::max(GVAlign, 4U);
1084 if (GVAlign > 1) {
1085 unsigned DstWidth = CI2->getType()->getBitWidth();
1086 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1087 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1089 // If checking bits we know are clear, return zero.
1090 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1091 return Constant::getNullValue(CI2->getType());
1095 break;
1096 case Instruction::Or:
1097 if (CI2->isZero()) return C1; // X | 0 == X
1098 if (CI2->isMinusOne())
1099 return C2; // X | -1 == -1
1100 break;
1101 case Instruction::Xor:
1102 if (CI2->isZero()) return C1; // X ^ 0 == X
1104 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1105 switch (CE1->getOpcode()) {
1106 default: break;
1107 case Instruction::ICmp:
1108 case Instruction::FCmp:
1109 // cmp pred ^ true -> cmp !pred
1110 assert(CI2->isOne());
1111 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1112 pred = CmpInst::getInversePredicate(pred);
1113 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1114 CE1->getOperand(1));
1117 break;
1118 case Instruction::AShr:
1119 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1120 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1121 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1122 return ConstantExpr::getLShr(C1, C2);
1123 break;
1125 } else if (isa<ConstantInt>(C1)) {
1126 // If C1 is a ConstantInt and C2 is not, swap the operands.
1127 if (Instruction::isCommutative(Opcode))
1128 return ConstantExpr::get(Opcode, C2, C1);
1131 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1132 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1133 const APInt &C1V = CI1->getValue();
1134 const APInt &C2V = CI2->getValue();
1135 switch (Opcode) {
1136 default:
1137 break;
1138 case Instruction::Add:
1139 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1140 case Instruction::Sub:
1141 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1142 case Instruction::Mul:
1143 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1144 case Instruction::UDiv:
1145 assert(!CI2->isZero() && "Div by zero handled above");
1146 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1147 case Instruction::SDiv:
1148 assert(!CI2->isZero() && "Div by zero handled above");
1149 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1150 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1151 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1152 case Instruction::URem:
1153 assert(!CI2->isZero() && "Div by zero handled above");
1154 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1155 case Instruction::SRem:
1156 assert(!CI2->isZero() && "Div by zero handled above");
1157 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1158 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1159 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1160 case Instruction::And:
1161 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1162 case Instruction::Or:
1163 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1164 case Instruction::Xor:
1165 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1166 case Instruction::Shl:
1167 if (C2V.ult(C1V.getBitWidth()))
1168 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1169 return UndefValue::get(C1->getType()); // too big shift is undef
1170 case Instruction::LShr:
1171 if (C2V.ult(C1V.getBitWidth()))
1172 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1173 return UndefValue::get(C1->getType()); // too big shift is undef
1174 case Instruction::AShr:
1175 if (C2V.ult(C1V.getBitWidth()))
1176 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1177 return UndefValue::get(C1->getType()); // too big shift is undef
1181 switch (Opcode) {
1182 case Instruction::SDiv:
1183 case Instruction::UDiv:
1184 case Instruction::URem:
1185 case Instruction::SRem:
1186 case Instruction::LShr:
1187 case Instruction::AShr:
1188 case Instruction::Shl:
1189 if (CI1->isZero()) return C1;
1190 break;
1191 default:
1192 break;
1194 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1195 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1196 const APFloat &C1V = CFP1->getValueAPF();
1197 const APFloat &C2V = CFP2->getValueAPF();
1198 APFloat C3V = C1V; // copy for modification
1199 switch (Opcode) {
1200 default:
1201 break;
1202 case Instruction::FAdd:
1203 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1204 return ConstantFP::get(C1->getContext(), C3V);
1205 case Instruction::FSub:
1206 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1207 return ConstantFP::get(C1->getContext(), C3V);
1208 case Instruction::FMul:
1209 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1210 return ConstantFP::get(C1->getContext(), C3V);
1211 case Instruction::FDiv:
1212 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1213 return ConstantFP::get(C1->getContext(), C3V);
1214 case Instruction::FRem:
1215 (void)C3V.mod(C2V);
1216 return ConstantFP::get(C1->getContext(), C3V);
1219 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1220 // Fold each element and create a vector constant from those constants.
1221 SmallVector<Constant*, 16> Result;
1222 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1223 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1224 Constant *ExtractIdx = ConstantInt::get(Ty, i);
1225 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1226 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1228 // If any element of a divisor vector is zero, the whole op is undef.
1229 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1230 return UndefValue::get(VTy);
1232 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1235 return ConstantVector::get(Result);
1238 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1239 // There are many possible foldings we could do here. We should probably
1240 // at least fold add of a pointer with an integer into the appropriate
1241 // getelementptr. This will improve alias analysis a bit.
1243 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1244 // (a + (b + c)).
1245 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1246 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1247 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1248 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1250 } else if (isa<ConstantExpr>(C2)) {
1251 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1252 // other way if possible.
1253 if (Instruction::isCommutative(Opcode))
1254 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1257 // i1 can be simplified in many cases.
1258 if (C1->getType()->isIntegerTy(1)) {
1259 switch (Opcode) {
1260 case Instruction::Add:
1261 case Instruction::Sub:
1262 return ConstantExpr::getXor(C1, C2);
1263 case Instruction::Mul:
1264 return ConstantExpr::getAnd(C1, C2);
1265 case Instruction::Shl:
1266 case Instruction::LShr:
1267 case Instruction::AShr:
1268 // We can assume that C2 == 0. If it were one the result would be
1269 // undefined because the shift value is as large as the bitwidth.
1270 return C1;
1271 case Instruction::SDiv:
1272 case Instruction::UDiv:
1273 // We can assume that C2 == 1. If it were zero the result would be
1274 // undefined through division by zero.
1275 return C1;
1276 case Instruction::URem:
1277 case Instruction::SRem:
1278 // We can assume that C2 == 1. If it were zero the result would be
1279 // undefined through division by zero.
1280 return ConstantInt::getFalse(C1->getContext());
1281 default:
1282 break;
1286 // We don't know how to fold this.
1287 return nullptr;
1290 /// This type is zero-sized if it's an array or structure of zero-sized types.
1291 /// The only leaf zero-sized type is an empty structure.
1292 static bool isMaybeZeroSizedType(Type *Ty) {
1293 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1294 if (STy->isOpaque()) return true; // Can't say.
1296 // If all of elements have zero size, this does too.
1297 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1298 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1299 return true;
1301 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1302 return isMaybeZeroSizedType(ATy->getElementType());
1304 return false;
1307 /// Compare the two constants as though they were getelementptr indices.
1308 /// This allows coercion of the types to be the same thing.
1310 /// If the two constants are the "same" (after coercion), return 0. If the
1311 /// first is less than the second, return -1, if the second is less than the
1312 /// first, return 1. If the constants are not integral, return -2.
1314 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1315 if (C1 == C2) return 0;
1317 // Ok, we found a different index. If they are not ConstantInt, we can't do
1318 // anything with them.
1319 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1320 return -2; // don't know!
1322 // We cannot compare the indices if they don't fit in an int64_t.
1323 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1324 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1325 return -2; // don't know!
1327 // Ok, we have two differing integer indices. Sign extend them to be the same
1328 // type.
1329 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1330 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1332 if (C1Val == C2Val) return 0; // They are equal
1334 // If the type being indexed over is really just a zero sized type, there is
1335 // no pointer difference being made here.
1336 if (isMaybeZeroSizedType(ElTy))
1337 return -2; // dunno.
1339 // If they are really different, now that they are the same type, then we
1340 // found a difference!
1341 if (C1Val < C2Val)
1342 return -1;
1343 else
1344 return 1;
1347 /// This function determines if there is anything we can decide about the two
1348 /// constants provided. This doesn't need to handle simple things like
1349 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1350 /// If we can determine that the two constants have a particular relation to
1351 /// each other, we should return the corresponding FCmpInst predicate,
1352 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1353 /// ConstantFoldCompareInstruction.
1355 /// To simplify this code we canonicalize the relation so that the first
1356 /// operand is always the most "complex" of the two. We consider ConstantFP
1357 /// to be the simplest, and ConstantExprs to be the most complex.
1358 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1359 assert(V1->getType() == V2->getType() &&
1360 "Cannot compare values of different types!");
1362 // We do not know if a constant expression will evaluate to a number or NaN.
1363 // Therefore, we can only say that the relation is unordered or equal.
1364 if (V1 == V2) return FCmpInst::FCMP_UEQ;
1366 if (!isa<ConstantExpr>(V1)) {
1367 if (!isa<ConstantExpr>(V2)) {
1368 // Simple case, use the standard constant folder.
1369 ConstantInt *R = nullptr;
1370 R = dyn_cast<ConstantInt>(
1371 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1372 if (R && !R->isZero())
1373 return FCmpInst::FCMP_OEQ;
1374 R = dyn_cast<ConstantInt>(
1375 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1376 if (R && !R->isZero())
1377 return FCmpInst::FCMP_OLT;
1378 R = dyn_cast<ConstantInt>(
1379 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1380 if (R && !R->isZero())
1381 return FCmpInst::FCMP_OGT;
1383 // Nothing more we can do
1384 return FCmpInst::BAD_FCMP_PREDICATE;
1387 // If the first operand is simple and second is ConstantExpr, swap operands.
1388 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1389 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1390 return FCmpInst::getSwappedPredicate(SwappedRelation);
1391 } else {
1392 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1393 // constantexpr or a simple constant.
1394 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1395 switch (CE1->getOpcode()) {
1396 case Instruction::FPTrunc:
1397 case Instruction::FPExt:
1398 case Instruction::UIToFP:
1399 case Instruction::SIToFP:
1400 // We might be able to do something with these but we don't right now.
1401 break;
1402 default:
1403 break;
1406 // There are MANY other foldings that we could perform here. They will
1407 // probably be added on demand, as they seem needed.
1408 return FCmpInst::BAD_FCMP_PREDICATE;
1411 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1412 const GlobalValue *GV2) {
1413 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1414 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1415 return true;
1416 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1417 Type *Ty = GVar->getValueType();
1418 // A global with opaque type might end up being zero sized.
1419 if (!Ty->isSized())
1420 return true;
1421 // A global with an empty type might lie at the address of any other
1422 // global.
1423 if (Ty->isEmptyTy())
1424 return true;
1426 return false;
1428 // Don't try to decide equality of aliases.
1429 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1430 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1431 return ICmpInst::ICMP_NE;
1432 return ICmpInst::BAD_ICMP_PREDICATE;
1435 /// This function determines if there is anything we can decide about the two
1436 /// constants provided. This doesn't need to handle simple things like integer
1437 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1438 /// If we can determine that the two constants have a particular relation to
1439 /// each other, we should return the corresponding ICmp predicate, otherwise
1440 /// return ICmpInst::BAD_ICMP_PREDICATE.
1442 /// To simplify this code we canonicalize the relation so that the first
1443 /// operand is always the most "complex" of the two. We consider simple
1444 /// constants (like ConstantInt) to be the simplest, followed by
1445 /// GlobalValues, followed by ConstantExpr's (the most complex).
1447 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1448 bool isSigned) {
1449 assert(V1->getType() == V2->getType() &&
1450 "Cannot compare different types of values!");
1451 if (V1 == V2) return ICmpInst::ICMP_EQ;
1453 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1454 !isa<BlockAddress>(V1)) {
1455 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1456 !isa<BlockAddress>(V2)) {
1457 // We distilled this down to a simple case, use the standard constant
1458 // folder.
1459 ConstantInt *R = nullptr;
1460 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1461 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1462 if (R && !R->isZero())
1463 return pred;
1464 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1465 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1466 if (R && !R->isZero())
1467 return pred;
1468 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1469 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1470 if (R && !R->isZero())
1471 return pred;
1473 // If we couldn't figure it out, bail.
1474 return ICmpInst::BAD_ICMP_PREDICATE;
1477 // If the first operand is simple, swap operands.
1478 ICmpInst::Predicate SwappedRelation =
1479 evaluateICmpRelation(V2, V1, isSigned);
1480 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1481 return ICmpInst::getSwappedPredicate(SwappedRelation);
1483 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1484 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1485 ICmpInst::Predicate SwappedRelation =
1486 evaluateICmpRelation(V2, V1, isSigned);
1487 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1488 return ICmpInst::getSwappedPredicate(SwappedRelation);
1489 return ICmpInst::BAD_ICMP_PREDICATE;
1492 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1493 // constant (which, since the types must match, means that it's a
1494 // ConstantPointerNull).
1495 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1496 return areGlobalsPotentiallyEqual(GV, GV2);
1497 } else if (isa<BlockAddress>(V2)) {
1498 return ICmpInst::ICMP_NE; // Globals never equal labels.
1499 } else {
1500 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1501 // GlobalVals can never be null unless they have external weak linkage.
1502 // We don't try to evaluate aliases here.
1503 // NOTE: We should not be doing this constant folding if null pointer
1504 // is considered valid for the function. But currently there is no way to
1505 // query it from the Constant type.
1506 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1507 !NullPointerIsDefined(nullptr /* F */,
1508 GV->getType()->getAddressSpace()))
1509 return ICmpInst::ICMP_NE;
1511 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1512 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1513 ICmpInst::Predicate SwappedRelation =
1514 evaluateICmpRelation(V2, V1, isSigned);
1515 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1516 return ICmpInst::getSwappedPredicate(SwappedRelation);
1517 return ICmpInst::BAD_ICMP_PREDICATE;
1520 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1521 // constant (which, since the types must match, means that it is a
1522 // ConstantPointerNull).
1523 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1524 // Block address in another function can't equal this one, but block
1525 // addresses in the current function might be the same if blocks are
1526 // empty.
1527 if (BA2->getFunction() != BA->getFunction())
1528 return ICmpInst::ICMP_NE;
1529 } else {
1530 // Block addresses aren't null, don't equal the address of globals.
1531 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1532 "Canonicalization guarantee!");
1533 return ICmpInst::ICMP_NE;
1535 } else {
1536 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1537 // constantexpr, a global, block address, or a simple constant.
1538 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1539 Constant *CE1Op0 = CE1->getOperand(0);
1541 switch (CE1->getOpcode()) {
1542 case Instruction::Trunc:
1543 case Instruction::FPTrunc:
1544 case Instruction::FPExt:
1545 case Instruction::FPToUI:
1546 case Instruction::FPToSI:
1547 break; // We can't evaluate floating point casts or truncations.
1549 case Instruction::UIToFP:
1550 case Instruction::SIToFP:
1551 case Instruction::BitCast:
1552 case Instruction::ZExt:
1553 case Instruction::SExt:
1554 // We can't evaluate floating point casts or truncations.
1555 if (CE1Op0->getType()->isFloatingPointTy())
1556 break;
1558 // If the cast is not actually changing bits, and the second operand is a
1559 // null pointer, do the comparison with the pre-casted value.
1560 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1561 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1562 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1563 return evaluateICmpRelation(CE1Op0,
1564 Constant::getNullValue(CE1Op0->getType()),
1565 isSigned);
1567 break;
1569 case Instruction::GetElementPtr: {
1570 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1571 // Ok, since this is a getelementptr, we know that the constant has a
1572 // pointer type. Check the various cases.
1573 if (isa<ConstantPointerNull>(V2)) {
1574 // If we are comparing a GEP to a null pointer, check to see if the base
1575 // of the GEP equals the null pointer.
1576 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1577 if (GV->hasExternalWeakLinkage())
1578 // Weak linkage GVals could be zero or not. We're comparing that
1579 // to null pointer so its greater-or-equal
1580 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1581 else
1582 // If its not weak linkage, the GVal must have a non-zero address
1583 // so the result is greater-than
1584 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1585 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1586 // If we are indexing from a null pointer, check to see if we have any
1587 // non-zero indices.
1588 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1589 if (!CE1->getOperand(i)->isNullValue())
1590 // Offsetting from null, must not be equal.
1591 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1592 // Only zero indexes from null, must still be zero.
1593 return ICmpInst::ICMP_EQ;
1595 // Otherwise, we can't really say if the first operand is null or not.
1596 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1597 if (isa<ConstantPointerNull>(CE1Op0)) {
1598 if (GV2->hasExternalWeakLinkage())
1599 // Weak linkage GVals could be zero or not. We're comparing it to
1600 // a null pointer, so its less-or-equal
1601 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1602 else
1603 // If its not weak linkage, the GVal must have a non-zero address
1604 // so the result is less-than
1605 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1606 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1607 if (GV == GV2) {
1608 // If this is a getelementptr of the same global, then it must be
1609 // different. Because the types must match, the getelementptr could
1610 // only have at most one index, and because we fold getelementptr's
1611 // with a single zero index, it must be nonzero.
1612 assert(CE1->getNumOperands() == 2 &&
1613 !CE1->getOperand(1)->isNullValue() &&
1614 "Surprising getelementptr!");
1615 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1616 } else {
1617 if (CE1GEP->hasAllZeroIndices())
1618 return areGlobalsPotentiallyEqual(GV, GV2);
1619 return ICmpInst::BAD_ICMP_PREDICATE;
1622 } else {
1623 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1624 Constant *CE2Op0 = CE2->getOperand(0);
1626 // There are MANY other foldings that we could perform here. They will
1627 // probably be added on demand, as they seem needed.
1628 switch (CE2->getOpcode()) {
1629 default: break;
1630 case Instruction::GetElementPtr:
1631 // By far the most common case to handle is when the base pointers are
1632 // obviously to the same global.
1633 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1634 // Don't know relative ordering, but check for inequality.
1635 if (CE1Op0 != CE2Op0) {
1636 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1637 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1638 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1639 cast<GlobalValue>(CE2Op0));
1640 return ICmpInst::BAD_ICMP_PREDICATE;
1642 // Ok, we know that both getelementptr instructions are based on the
1643 // same global. From this, we can precisely determine the relative
1644 // ordering of the resultant pointers.
1645 unsigned i = 1;
1647 // The logic below assumes that the result of the comparison
1648 // can be determined by finding the first index that differs.
1649 // This doesn't work if there is over-indexing in any
1650 // subsequent indices, so check for that case first.
1651 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1652 !CE2->isGEPWithNoNotionalOverIndexing())
1653 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1655 // Compare all of the operands the GEP's have in common.
1656 gep_type_iterator GTI = gep_type_begin(CE1);
1657 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1658 ++i, ++GTI)
1659 switch (IdxCompare(CE1->getOperand(i),
1660 CE2->getOperand(i), GTI.getIndexedType())) {
1661 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1662 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1663 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1666 // Ok, we ran out of things they have in common. If any leftovers
1667 // are non-zero then we have a difference, otherwise we are equal.
1668 for (; i < CE1->getNumOperands(); ++i)
1669 if (!CE1->getOperand(i)->isNullValue()) {
1670 if (isa<ConstantInt>(CE1->getOperand(i)))
1671 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1672 else
1673 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1676 for (; i < CE2->getNumOperands(); ++i)
1677 if (!CE2->getOperand(i)->isNullValue()) {
1678 if (isa<ConstantInt>(CE2->getOperand(i)))
1679 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1680 else
1681 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1683 return ICmpInst::ICMP_EQ;
1687 break;
1689 default:
1690 break;
1694 return ICmpInst::BAD_ICMP_PREDICATE;
1697 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1698 Constant *C1, Constant *C2) {
1699 Type *ResultTy;
1700 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1701 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1702 VT->getNumElements());
1703 else
1704 ResultTy = Type::getInt1Ty(C1->getContext());
1706 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1707 if (pred == FCmpInst::FCMP_FALSE)
1708 return Constant::getNullValue(ResultTy);
1710 if (pred == FCmpInst::FCMP_TRUE)
1711 return Constant::getAllOnesValue(ResultTy);
1713 // Handle some degenerate cases first
1714 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1715 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1716 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1717 // For EQ and NE, we can always pick a value for the undef to make the
1718 // predicate pass or fail, so we can return undef.
1719 // Also, if both operands are undef, we can return undef for int comparison.
1720 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1721 return UndefValue::get(ResultTy);
1723 // Otherwise, for integer compare, pick the same value as the non-undef
1724 // operand, and fold it to true or false.
1725 if (isIntegerPredicate)
1726 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1728 // Choosing NaN for the undef will always make unordered comparison succeed
1729 // and ordered comparison fails.
1730 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1733 // icmp eq/ne(null,GV) -> false/true
1734 if (C1->isNullValue()) {
1735 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1736 // Don't try to evaluate aliases. External weak GV can be null.
1737 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1738 !NullPointerIsDefined(nullptr /* F */,
1739 GV->getType()->getAddressSpace())) {
1740 if (pred == ICmpInst::ICMP_EQ)
1741 return ConstantInt::getFalse(C1->getContext());
1742 else if (pred == ICmpInst::ICMP_NE)
1743 return ConstantInt::getTrue(C1->getContext());
1745 // icmp eq/ne(GV,null) -> false/true
1746 } else if (C2->isNullValue()) {
1747 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1748 // Don't try to evaluate aliases. External weak GV can be null.
1749 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1750 !NullPointerIsDefined(nullptr /* F */,
1751 GV->getType()->getAddressSpace())) {
1752 if (pred == ICmpInst::ICMP_EQ)
1753 return ConstantInt::getFalse(C1->getContext());
1754 else if (pred == ICmpInst::ICMP_NE)
1755 return ConstantInt::getTrue(C1->getContext());
1759 // If the comparison is a comparison between two i1's, simplify it.
1760 if (C1->getType()->isIntegerTy(1)) {
1761 switch(pred) {
1762 case ICmpInst::ICMP_EQ:
1763 if (isa<ConstantInt>(C2))
1764 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1765 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1766 case ICmpInst::ICMP_NE:
1767 return ConstantExpr::getXor(C1, C2);
1768 default:
1769 break;
1773 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1774 const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1775 const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1776 switch (pred) {
1777 default: llvm_unreachable("Invalid ICmp Predicate");
1778 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1779 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1780 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1781 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1782 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1783 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1784 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1785 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1786 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1787 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1789 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1790 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1791 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1792 APFloat::cmpResult R = C1V.compare(C2V);
1793 switch (pred) {
1794 default: llvm_unreachable("Invalid FCmp Predicate");
1795 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1796 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1797 case FCmpInst::FCMP_UNO:
1798 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1799 case FCmpInst::FCMP_ORD:
1800 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1801 case FCmpInst::FCMP_UEQ:
1802 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1803 R==APFloat::cmpEqual);
1804 case FCmpInst::FCMP_OEQ:
1805 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1806 case FCmpInst::FCMP_UNE:
1807 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1808 case FCmpInst::FCMP_ONE:
1809 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1810 R==APFloat::cmpGreaterThan);
1811 case FCmpInst::FCMP_ULT:
1812 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1813 R==APFloat::cmpLessThan);
1814 case FCmpInst::FCMP_OLT:
1815 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1816 case FCmpInst::FCMP_UGT:
1817 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1818 R==APFloat::cmpGreaterThan);
1819 case FCmpInst::FCMP_OGT:
1820 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1821 case FCmpInst::FCMP_ULE:
1822 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1823 case FCmpInst::FCMP_OLE:
1824 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1825 R==APFloat::cmpEqual);
1826 case FCmpInst::FCMP_UGE:
1827 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1828 case FCmpInst::FCMP_OGE:
1829 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1830 R==APFloat::cmpEqual);
1832 } else if (C1->getType()->isVectorTy()) {
1833 // If we can constant fold the comparison of each element, constant fold
1834 // the whole vector comparison.
1835 SmallVector<Constant*, 4> ResElts;
1836 Type *Ty = IntegerType::get(C1->getContext(), 32);
1837 // Compare the elements, producing an i1 result or constant expr.
1838 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1839 Constant *C1E =
1840 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1841 Constant *C2E =
1842 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1844 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1847 return ConstantVector::get(ResElts);
1850 if (C1->getType()->isFloatingPointTy() &&
1851 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1852 // infinite recursive loop
1853 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1854 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1855 switch (evaluateFCmpRelation(C1, C2)) {
1856 default: llvm_unreachable("Unknown relation!");
1857 case FCmpInst::FCMP_UNO:
1858 case FCmpInst::FCMP_ORD:
1859 case FCmpInst::FCMP_UNE:
1860 case FCmpInst::FCMP_ULT:
1861 case FCmpInst::FCMP_UGT:
1862 case FCmpInst::FCMP_ULE:
1863 case FCmpInst::FCMP_UGE:
1864 case FCmpInst::FCMP_TRUE:
1865 case FCmpInst::FCMP_FALSE:
1866 case FCmpInst::BAD_FCMP_PREDICATE:
1867 break; // Couldn't determine anything about these constants.
1868 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1869 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1870 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1871 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1872 break;
1873 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1874 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1875 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1876 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1877 break;
1878 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1879 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1880 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1881 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1882 break;
1883 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1884 // We can only partially decide this relation.
1885 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1886 Result = 0;
1887 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1888 Result = 1;
1889 break;
1890 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1891 // We can only partially decide this relation.
1892 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1893 Result = 0;
1894 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1895 Result = 1;
1896 break;
1897 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1898 // We can only partially decide this relation.
1899 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1900 Result = 0;
1901 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1902 Result = 1;
1903 break;
1904 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
1905 // We can only partially decide this relation.
1906 if (pred == FCmpInst::FCMP_ONE)
1907 Result = 0;
1908 else if (pred == FCmpInst::FCMP_UEQ)
1909 Result = 1;
1910 break;
1913 // If we evaluated the result, return it now.
1914 if (Result != -1)
1915 return ConstantInt::get(ResultTy, Result);
1917 } else {
1918 // Evaluate the relation between the two constants, per the predicate.
1919 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1920 switch (evaluateICmpRelation(C1, C2,
1921 CmpInst::isSigned((CmpInst::Predicate)pred))) {
1922 default: llvm_unreachable("Unknown relational!");
1923 case ICmpInst::BAD_ICMP_PREDICATE:
1924 break; // Couldn't determine anything about these constants.
1925 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1926 // If we know the constants are equal, we can decide the result of this
1927 // computation precisely.
1928 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1929 break;
1930 case ICmpInst::ICMP_ULT:
1931 switch (pred) {
1932 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1933 Result = 1; break;
1934 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1935 Result = 0; break;
1937 break;
1938 case ICmpInst::ICMP_SLT:
1939 switch (pred) {
1940 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1941 Result = 1; break;
1942 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1943 Result = 0; break;
1945 break;
1946 case ICmpInst::ICMP_UGT:
1947 switch (pred) {
1948 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1949 Result = 1; break;
1950 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1951 Result = 0; break;
1953 break;
1954 case ICmpInst::ICMP_SGT:
1955 switch (pred) {
1956 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1957 Result = 1; break;
1958 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1959 Result = 0; break;
1961 break;
1962 case ICmpInst::ICMP_ULE:
1963 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1964 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1965 break;
1966 case ICmpInst::ICMP_SLE:
1967 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1968 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1969 break;
1970 case ICmpInst::ICMP_UGE:
1971 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1972 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1973 break;
1974 case ICmpInst::ICMP_SGE:
1975 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1976 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1977 break;
1978 case ICmpInst::ICMP_NE:
1979 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1980 if (pred == ICmpInst::ICMP_NE) Result = 1;
1981 break;
1984 // If we evaluated the result, return it now.
1985 if (Result != -1)
1986 return ConstantInt::get(ResultTy, Result);
1988 // If the right hand side is a bitcast, try using its inverse to simplify
1989 // it by moving it to the left hand side. We can't do this if it would turn
1990 // a vector compare into a scalar compare or visa versa, or if it would turn
1991 // the operands into FP values.
1992 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1993 Constant *CE2Op0 = CE2->getOperand(0);
1994 if (CE2->getOpcode() == Instruction::BitCast &&
1995 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
1996 !CE2Op0->getType()->isFPOrFPVectorTy()) {
1997 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1998 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2002 // If the left hand side is an extension, try eliminating it.
2003 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2004 if ((CE1->getOpcode() == Instruction::SExt &&
2005 ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2006 (CE1->getOpcode() == Instruction::ZExt &&
2007 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2008 Constant *CE1Op0 = CE1->getOperand(0);
2009 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2010 if (CE1Inverse == CE1Op0) {
2011 // Check whether we can safely truncate the right hand side.
2012 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2013 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2014 C2->getType()) == C2)
2015 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2020 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2021 (C1->isNullValue() && !C2->isNullValue())) {
2022 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2023 // other way if possible.
2024 // Also, if C1 is null and C2 isn't, flip them around.
2025 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2026 return ConstantExpr::getICmp(pred, C2, C1);
2029 return nullptr;
2032 /// Test whether the given sequence of *normalized* indices is "inbounds".
2033 template<typename IndexTy>
2034 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2035 // No indices means nothing that could be out of bounds.
2036 if (Idxs.empty()) return true;
2038 // If the first index is zero, it's in bounds.
2039 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2041 // If the first index is one and all the rest are zero, it's in bounds,
2042 // by the one-past-the-end rule.
2043 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2044 if (!CI->isOne())
2045 return false;
2046 } else {
2047 auto *CV = cast<ConstantDataVector>(Idxs[0]);
2048 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2049 if (!CI || !CI->isOne())
2050 return false;
2053 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2054 if (!cast<Constant>(Idxs[i])->isNullValue())
2055 return false;
2056 return true;
2059 /// Test whether a given ConstantInt is in-range for a SequentialType.
2060 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2061 const ConstantInt *CI) {
2062 // We cannot bounds check the index if it doesn't fit in an int64_t.
2063 if (CI->getValue().getMinSignedBits() > 64)
2064 return false;
2066 // A negative index or an index past the end of our sequential type is
2067 // considered out-of-range.
2068 int64_t IndexVal = CI->getSExtValue();
2069 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2070 return false;
2072 // Otherwise, it is in-range.
2073 return true;
2076 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2077 bool InBounds,
2078 Optional<unsigned> InRangeIndex,
2079 ArrayRef<Value *> Idxs) {
2080 if (Idxs.empty()) return C;
2082 Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2083 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2085 if (isa<UndefValue>(C))
2086 return UndefValue::get(GEPTy);
2088 Constant *Idx0 = cast<Constant>(Idxs[0]);
2089 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2090 return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2091 ? ConstantVector::getSplat(
2092 cast<VectorType>(GEPTy)->getNumElements(), C)
2093 : C;
2095 if (C->isNullValue()) {
2096 bool isNull = true;
2097 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2098 if (!isa<UndefValue>(Idxs[i]) &&
2099 !cast<Constant>(Idxs[i])->isNullValue()) {
2100 isNull = false;
2101 break;
2103 if (isNull) {
2104 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2105 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2107 assert(Ty && "Invalid indices for GEP!");
2108 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2109 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2110 if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2111 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2113 // The GEP returns a vector of pointers when one of more of
2114 // its arguments is a vector.
2115 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2116 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2117 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2118 break;
2122 return Constant::getNullValue(GEPTy);
2126 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2127 // Combine Indices - If the source pointer to this getelementptr instruction
2128 // is a getelementptr instruction, combine the indices of the two
2129 // getelementptr instructions into a single instruction.
2131 if (CE->getOpcode() == Instruction::GetElementPtr) {
2132 gep_type_iterator LastI = gep_type_end(CE);
2133 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2134 I != E; ++I)
2135 LastI = I;
2137 // We cannot combine indices if doing so would take us outside of an
2138 // array or vector. Doing otherwise could trick us if we evaluated such a
2139 // GEP as part of a load.
2141 // e.g. Consider if the original GEP was:
2142 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2143 // i32 0, i32 0, i64 0)
2145 // If we then tried to offset it by '8' to get to the third element,
2146 // an i8, we should *not* get:
2147 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2148 // i32 0, i32 0, i64 8)
2150 // This GEP tries to index array element '8 which runs out-of-bounds.
2151 // Subsequent evaluation would get confused and produce erroneous results.
2153 // The following prohibits such a GEP from being formed by checking to see
2154 // if the index is in-range with respect to an array.
2155 // TODO: This code may be extended to handle vectors as well.
2156 bool PerformFold = false;
2157 if (Idx0->isNullValue())
2158 PerformFold = true;
2159 else if (LastI.isSequential())
2160 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2161 PerformFold = (!LastI.isBoundedSequential() ||
2162 isIndexInRangeOfArrayType(
2163 LastI.getSequentialNumElements(), CI)) &&
2164 !CE->getOperand(CE->getNumOperands() - 1)
2165 ->getType()
2166 ->isVectorTy();
2168 if (PerformFold) {
2169 SmallVector<Value*, 16> NewIndices;
2170 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2171 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2173 // Add the last index of the source with the first index of the new GEP.
2174 // Make sure to handle the case when they are actually different types.
2175 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2176 // Otherwise it must be an array.
2177 if (!Idx0->isNullValue()) {
2178 Type *IdxTy = Combined->getType();
2179 if (IdxTy != Idx0->getType()) {
2180 unsigned CommonExtendedWidth =
2181 std::max(IdxTy->getIntegerBitWidth(),
2182 Idx0->getType()->getIntegerBitWidth());
2183 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2185 Type *CommonTy =
2186 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2187 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2188 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2189 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2190 } else {
2191 Combined =
2192 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2196 NewIndices.push_back(Combined);
2197 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2199 // The combined GEP normally inherits its index inrange attribute from
2200 // the inner GEP, but if the inner GEP's last index was adjusted by the
2201 // outer GEP, any inbounds attribute on that index is invalidated.
2202 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex();
2203 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue())
2204 IRIndex = None;
2206 return ConstantExpr::getGetElementPtr(
2207 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2208 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(),
2209 IRIndex);
2213 // Attempt to fold casts to the same type away. For example, folding:
2215 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2216 // i64 0, i64 0)
2217 // into:
2219 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2221 // Don't fold if the cast is changing address spaces.
2222 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2223 PointerType *SrcPtrTy =
2224 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2225 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2226 if (SrcPtrTy && DstPtrTy) {
2227 ArrayType *SrcArrayTy =
2228 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2229 ArrayType *DstArrayTy =
2230 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2231 if (SrcArrayTy && DstArrayTy
2232 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2233 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2234 return ConstantExpr::getGetElementPtr(SrcArrayTy,
2235 (Constant *)CE->getOperand(0),
2236 Idxs, InBounds, InRangeIndex);
2241 // Check to see if any array indices are not within the corresponding
2242 // notional array or vector bounds. If so, try to determine if they can be
2243 // factored out into preceding dimensions.
2244 SmallVector<Constant *, 8> NewIdxs;
2245 Type *Ty = PointeeTy;
2246 Type *Prev = C->getType();
2247 bool Unknown =
2248 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2249 for (unsigned i = 1, e = Idxs.size(); i != e;
2250 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2251 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2252 // We don't know if it's in range or not.
2253 Unknown = true;
2254 continue;
2256 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2257 // Skip if the type of the previous index is not supported.
2258 continue;
2259 if (InRangeIndex && i == *InRangeIndex + 1) {
2260 // If an index is marked inrange, we cannot apply this canonicalization to
2261 // the following index, as that will cause the inrange index to point to
2262 // the wrong element.
2263 continue;
2265 if (isa<StructType>(Ty)) {
2266 // The verify makes sure that GEPs into a struct are in range.
2267 continue;
2269 auto *STy = cast<SequentialType>(Ty);
2270 if (isa<VectorType>(STy)) {
2271 // There can be awkward padding in after a non-power of two vector.
2272 Unknown = true;
2273 continue;
2275 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2276 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2277 // It's in range, skip to the next index.
2278 continue;
2279 if (CI->getSExtValue() < 0) {
2280 // It's out of range and negative, don't try to factor it.
2281 Unknown = true;
2282 continue;
2284 } else {
2285 auto *CV = cast<ConstantDataVector>(Idxs[i]);
2286 bool InRange = true;
2287 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2288 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2289 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2290 if (CI->getSExtValue() < 0) {
2291 Unknown = true;
2292 break;
2295 if (InRange || Unknown)
2296 // It's in range, skip to the next index.
2297 // It's out of range and negative, don't try to factor it.
2298 continue;
2300 if (isa<StructType>(Prev)) {
2301 // It's out of range, but the prior dimension is a struct
2302 // so we can't do anything about it.
2303 Unknown = true;
2304 continue;
2306 // It's out of range, but we can factor it into the prior
2307 // dimension.
2308 NewIdxs.resize(Idxs.size());
2309 // Determine the number of elements in our sequential type.
2310 uint64_t NumElements = STy->getArrayNumElements();
2312 // Expand the current index or the previous index to a vector from a scalar
2313 // if necessary.
2314 Constant *CurrIdx = cast<Constant>(Idxs[i]);
2315 auto *PrevIdx =
2316 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2317 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2318 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2319 bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2321 if (!IsCurrIdxVector && IsPrevIdxVector)
2322 CurrIdx = ConstantDataVector::getSplat(
2323 PrevIdx->getType()->getVectorNumElements(), CurrIdx);
2325 if (!IsPrevIdxVector && IsCurrIdxVector)
2326 PrevIdx = ConstantDataVector::getSplat(
2327 CurrIdx->getType()->getVectorNumElements(), PrevIdx);
2329 Constant *Factor =
2330 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2331 if (UseVector)
2332 Factor = ConstantDataVector::getSplat(
2333 IsPrevIdxVector ? PrevIdx->getType()->getVectorNumElements()
2334 : CurrIdx->getType()->getVectorNumElements(),
2335 Factor);
2337 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2339 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2341 unsigned CommonExtendedWidth =
2342 std::max(PrevIdx->getType()->getScalarSizeInBits(),
2343 Div->getType()->getScalarSizeInBits());
2344 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2346 // Before adding, extend both operands to i64 to avoid
2347 // overflow trouble.
2348 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2349 if (UseVector)
2350 ExtendedTy = VectorType::get(
2351 ExtendedTy, IsPrevIdxVector
2352 ? PrevIdx->getType()->getVectorNumElements()
2353 : CurrIdx->getType()->getVectorNumElements());
2355 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2356 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2358 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2359 Div = ConstantExpr::getSExt(Div, ExtendedTy);
2361 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2364 // If we did any factoring, start over with the adjusted indices.
2365 if (!NewIdxs.empty()) {
2366 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2367 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2368 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2369 InRangeIndex);
2372 // If all indices are known integers and normalized, we can do a simple
2373 // check for the "inbounds" property.
2374 if (!Unknown && !InBounds)
2375 if (auto *GV = dyn_cast<GlobalVariable>(C))
2376 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2377 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2378 /*InBounds=*/true, InRangeIndex);
2380 return nullptr;