[yaml2obj/obj2yaml] - Add support for .stack_sizes sections.
[llvm-complete.git] / lib / IR / ConstantFold.cpp
blobee0ad322f4edc67bf4f374524f65bc2afb998b59
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 (!Cond) break;
751 if (V1Element == V2Element) {
752 V = V1Element;
753 } else if (isa<UndefValue>(Cond)) {
754 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
755 } else {
756 if (!isa<ConstantInt>(Cond)) break;
757 V = Cond->isNullValue() ? V2Element : V1Element;
759 Result.push_back(V);
762 // If we were able to build the vector, return it.
763 if (Result.size() == V1->getType()->getVectorNumElements())
764 return ConstantVector::get(Result);
767 if (isa<UndefValue>(Cond)) {
768 if (isa<UndefValue>(V1)) return V1;
769 return V2;
771 if (isa<UndefValue>(V1)) return V2;
772 if (isa<UndefValue>(V2)) return V1;
773 if (V1 == V2) return V1;
775 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
776 if (TrueVal->getOpcode() == Instruction::Select)
777 if (TrueVal->getOperand(0) == Cond)
778 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
780 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
781 if (FalseVal->getOpcode() == Instruction::Select)
782 if (FalseVal->getOperand(0) == Cond)
783 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
786 return nullptr;
789 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
790 Constant *Idx) {
791 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
792 return UndefValue::get(Val->getType()->getVectorElementType());
793 if (Val->isNullValue()) // ee(zero, x) -> zero
794 return Constant::getNullValue(Val->getType()->getVectorElementType());
795 // ee({w,x,y,z}, undef) -> undef
796 if (isa<UndefValue>(Idx))
797 return UndefValue::get(Val->getType()->getVectorElementType());
799 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
800 // ee({w,x,y,z}, wrong_value) -> undef
801 if (CIdx->uge(Val->getType()->getVectorNumElements()))
802 return UndefValue::get(Val->getType()->getVectorElementType());
803 return Val->getAggregateElement(CIdx->getZExtValue());
805 return nullptr;
808 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
809 Constant *Elt,
810 Constant *Idx) {
811 if (isa<UndefValue>(Idx))
812 return UndefValue::get(Val->getType());
814 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
815 if (!CIdx) return nullptr;
817 unsigned NumElts = Val->getType()->getVectorNumElements();
818 if (CIdx->uge(NumElts))
819 return UndefValue::get(Val->getType());
821 SmallVector<Constant*, 16> Result;
822 Result.reserve(NumElts);
823 auto *Ty = Type::getInt32Ty(Val->getContext());
824 uint64_t IdxVal = CIdx->getZExtValue();
825 for (unsigned i = 0; i != NumElts; ++i) {
826 if (i == IdxVal) {
827 Result.push_back(Elt);
828 continue;
831 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
832 Result.push_back(C);
835 return ConstantVector::get(Result);
838 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
839 Constant *V2,
840 Constant *Mask) {
841 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
842 Type *EltTy = V1->getType()->getVectorElementType();
844 // Undefined shuffle mask -> undefined value.
845 if (isa<UndefValue>(Mask))
846 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
848 // Don't break the bitcode reader hack.
849 if (isa<ConstantExpr>(Mask)) return nullptr;
851 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
853 // Loop over the shuffle mask, evaluating each element.
854 SmallVector<Constant*, 32> Result;
855 for (unsigned i = 0; i != MaskNumElts; ++i) {
856 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
857 if (Elt == -1) {
858 Result.push_back(UndefValue::get(EltTy));
859 continue;
861 Constant *InElt;
862 if (unsigned(Elt) >= SrcNumElts*2)
863 InElt = UndefValue::get(EltTy);
864 else if (unsigned(Elt) >= SrcNumElts) {
865 Type *Ty = IntegerType::get(V2->getContext(), 32);
866 InElt =
867 ConstantExpr::getExtractElement(V2,
868 ConstantInt::get(Ty, Elt - SrcNumElts));
869 } else {
870 Type *Ty = IntegerType::get(V1->getContext(), 32);
871 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
873 Result.push_back(InElt);
876 return ConstantVector::get(Result);
879 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
880 ArrayRef<unsigned> Idxs) {
881 // Base case: no indices, so return the entire value.
882 if (Idxs.empty())
883 return Agg;
885 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
886 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
888 return nullptr;
891 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
892 Constant *Val,
893 ArrayRef<unsigned> Idxs) {
894 // Base case: no indices, so replace the entire value.
895 if (Idxs.empty())
896 return Val;
898 unsigned NumElts;
899 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
900 NumElts = ST->getNumElements();
901 else
902 NumElts = cast<SequentialType>(Agg->getType())->getNumElements();
904 SmallVector<Constant*, 32> Result;
905 for (unsigned i = 0; i != NumElts; ++i) {
906 Constant *C = Agg->getAggregateElement(i);
907 if (!C) return nullptr;
909 if (Idxs[0] == i)
910 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
912 Result.push_back(C);
915 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
916 return ConstantStruct::get(ST, Result);
917 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
918 return ConstantArray::get(AT, Result);
919 return ConstantVector::get(Result);
922 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) {
923 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected");
925 // Handle scalar UndefValue. Vectors are always evaluated per element.
926 bool HasScalarUndef = !C->getType()->isVectorTy() && isa<UndefValue>(C);
928 if (HasScalarUndef) {
929 switch (static_cast<Instruction::UnaryOps>(Opcode)) {
930 case Instruction::FNeg:
931 return C; // -undef -> undef
932 case Instruction::UnaryOpsEnd:
933 llvm_unreachable("Invalid UnaryOp");
937 // Constant should not be UndefValue, unless these are vector constants.
938 assert(!HasScalarUndef && "Unexpected UndefValue");
939 // We only have FP UnaryOps right now.
940 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp");
942 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
943 const APFloat &CV = CFP->getValueAPF();
944 switch (Opcode) {
945 default:
946 break;
947 case Instruction::FNeg:
948 return ConstantFP::get(C->getContext(), neg(CV));
950 } else if (VectorType *VTy = dyn_cast<VectorType>(C->getType())) {
951 // Fold each element and create a vector constant from those constants.
952 SmallVector<Constant*, 16> Result;
953 Type *Ty = IntegerType::get(VTy->getContext(), 32);
954 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
955 Constant *ExtractIdx = ConstantInt::get(Ty, i);
956 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx);
958 Result.push_back(ConstantExpr::get(Opcode, Elt));
961 return ConstantVector::get(Result);
964 // We don't know how to fold this.
965 return nullptr;
968 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
969 Constant *C2) {
970 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
972 // Handle scalar UndefValue. Vectors are always evaluated per element.
973 bool HasScalarUndef = !C1->getType()->isVectorTy() &&
974 (isa<UndefValue>(C1) || isa<UndefValue>(C2));
975 if (HasScalarUndef) {
976 switch (static_cast<Instruction::BinaryOps>(Opcode)) {
977 case Instruction::Xor:
978 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
979 // Handle undef ^ undef -> 0 special case. This is a common
980 // idiom (misuse).
981 return Constant::getNullValue(C1->getType());
982 LLVM_FALLTHROUGH;
983 case Instruction::Add:
984 case Instruction::Sub:
985 return UndefValue::get(C1->getType());
986 case Instruction::And:
987 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
988 return C1;
989 return Constant::getNullValue(C1->getType()); // undef & X -> 0
990 case Instruction::Mul: {
991 // undef * undef -> undef
992 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
993 return C1;
994 const APInt *CV;
995 // X * undef -> undef if X is odd
996 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
997 if ((*CV)[0])
998 return UndefValue::get(C1->getType());
1000 // X * undef -> 0 otherwise
1001 return Constant::getNullValue(C1->getType());
1003 case Instruction::SDiv:
1004 case Instruction::UDiv:
1005 // X / undef -> undef
1006 if (isa<UndefValue>(C2))
1007 return C2;
1008 // undef / 0 -> undef
1009 // undef / 1 -> undef
1010 if (match(C2, m_Zero()) || match(C2, m_One()))
1011 return C1;
1012 // undef / X -> 0 otherwise
1013 return Constant::getNullValue(C1->getType());
1014 case Instruction::URem:
1015 case Instruction::SRem:
1016 // X % undef -> undef
1017 if (match(C2, m_Undef()))
1018 return C2;
1019 // undef % 0 -> undef
1020 if (match(C2, m_Zero()))
1021 return C1;
1022 // undef % X -> 0 otherwise
1023 return Constant::getNullValue(C1->getType());
1024 case Instruction::Or: // X | undef -> -1
1025 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
1026 return C1;
1027 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
1028 case Instruction::LShr:
1029 // X >>l undef -> undef
1030 if (isa<UndefValue>(C2))
1031 return C2;
1032 // undef >>l 0 -> undef
1033 if (match(C2, m_Zero()))
1034 return C1;
1035 // undef >>l X -> 0
1036 return Constant::getNullValue(C1->getType());
1037 case Instruction::AShr:
1038 // X >>a undef -> undef
1039 if (isa<UndefValue>(C2))
1040 return C2;
1041 // undef >>a 0 -> undef
1042 if (match(C2, m_Zero()))
1043 return C1;
1044 // TODO: undef >>a X -> undef if the shift is exact
1045 // undef >>a X -> 0
1046 return Constant::getNullValue(C1->getType());
1047 case Instruction::Shl:
1048 // X << undef -> undef
1049 if (isa<UndefValue>(C2))
1050 return C2;
1051 // undef << 0 -> undef
1052 if (match(C2, m_Zero()))
1053 return C1;
1054 // undef << X -> 0
1055 return Constant::getNullValue(C1->getType());
1056 case Instruction::FAdd:
1057 case Instruction::FSub:
1058 case Instruction::FMul:
1059 case Instruction::FDiv:
1060 case Instruction::FRem:
1061 // [any flop] undef, undef -> undef
1062 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1063 return C1;
1064 // [any flop] C, undef -> NaN
1065 // [any flop] undef, C -> NaN
1066 // We could potentially specialize NaN/Inf constants vs. 'normal'
1067 // constants (possibly differently depending on opcode and operand). This
1068 // would allow returning undef sometimes. But it is always safe to fold to
1069 // NaN because we can choose the undef operand as NaN, and any FP opcode
1070 // with a NaN operand will propagate NaN.
1071 return ConstantFP::getNaN(C1->getType());
1072 case Instruction::BinaryOpsEnd:
1073 llvm_unreachable("Invalid BinaryOp");
1077 // Neither constant should be UndefValue, unless these are vector constants.
1078 assert(!HasScalarUndef && "Unexpected UndefValue");
1080 // Handle simplifications when the RHS is a constant int.
1081 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1082 switch (Opcode) {
1083 case Instruction::Add:
1084 if (CI2->isZero()) return C1; // X + 0 == X
1085 break;
1086 case Instruction::Sub:
1087 if (CI2->isZero()) return C1; // X - 0 == X
1088 break;
1089 case Instruction::Mul:
1090 if (CI2->isZero()) return C2; // X * 0 == 0
1091 if (CI2->isOne())
1092 return C1; // X * 1 == X
1093 break;
1094 case Instruction::UDiv:
1095 case Instruction::SDiv:
1096 if (CI2->isOne())
1097 return C1; // X / 1 == X
1098 if (CI2->isZero())
1099 return UndefValue::get(CI2->getType()); // X / 0 == undef
1100 break;
1101 case Instruction::URem:
1102 case Instruction::SRem:
1103 if (CI2->isOne())
1104 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1105 if (CI2->isZero())
1106 return UndefValue::get(CI2->getType()); // X % 0 == undef
1107 break;
1108 case Instruction::And:
1109 if (CI2->isZero()) return C2; // X & 0 == 0
1110 if (CI2->isMinusOne())
1111 return C1; // X & -1 == X
1113 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1114 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1115 if (CE1->getOpcode() == Instruction::ZExt) {
1116 unsigned DstWidth = CI2->getType()->getBitWidth();
1117 unsigned SrcWidth =
1118 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1119 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1120 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1121 return C1;
1124 // If and'ing the address of a global with a constant, fold it.
1125 if (CE1->getOpcode() == Instruction::PtrToInt &&
1126 isa<GlobalValue>(CE1->getOperand(0))) {
1127 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1129 unsigned GVAlign;
1131 if (Module *TheModule = GV->getParent()) {
1132 GVAlign = GV->getPointerAlignment(TheModule->getDataLayout());
1134 // If the function alignment is not specified then assume that it
1135 // is 4.
1136 // This is dangerous; on x86, the alignment of the pointer
1137 // corresponds to the alignment of the function, but might be less
1138 // than 4 if it isn't explicitly specified.
1139 // However, a fix for this behaviour was reverted because it
1140 // increased code size (see https://reviews.llvm.org/D55115)
1141 // FIXME: This code should be deleted once existing targets have
1142 // appropriate defaults
1143 if (GVAlign == 0U && isa<Function>(GV))
1144 GVAlign = 4U;
1145 } else if (isa<Function>(GV)) {
1146 // Without a datalayout we have to assume the worst case: that the
1147 // function pointer isn't aligned at all.
1148 GVAlign = 0U;
1149 } else {
1150 GVAlign = GV->getAlignment();
1153 if (GVAlign > 1) {
1154 unsigned DstWidth = CI2->getType()->getBitWidth();
1155 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1156 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1158 // If checking bits we know are clear, return zero.
1159 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1160 return Constant::getNullValue(CI2->getType());
1164 break;
1165 case Instruction::Or:
1166 if (CI2->isZero()) return C1; // X | 0 == X
1167 if (CI2->isMinusOne())
1168 return C2; // X | -1 == -1
1169 break;
1170 case Instruction::Xor:
1171 if (CI2->isZero()) return C1; // X ^ 0 == X
1173 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1174 switch (CE1->getOpcode()) {
1175 default: break;
1176 case Instruction::ICmp:
1177 case Instruction::FCmp:
1178 // cmp pred ^ true -> cmp !pred
1179 assert(CI2->isOne());
1180 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1181 pred = CmpInst::getInversePredicate(pred);
1182 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1183 CE1->getOperand(1));
1186 break;
1187 case Instruction::AShr:
1188 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1189 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1190 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1191 return ConstantExpr::getLShr(C1, C2);
1192 break;
1194 } else if (isa<ConstantInt>(C1)) {
1195 // If C1 is a ConstantInt and C2 is not, swap the operands.
1196 if (Instruction::isCommutative(Opcode))
1197 return ConstantExpr::get(Opcode, C2, C1);
1200 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1201 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1202 const APInt &C1V = CI1->getValue();
1203 const APInt &C2V = CI2->getValue();
1204 switch (Opcode) {
1205 default:
1206 break;
1207 case Instruction::Add:
1208 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1209 case Instruction::Sub:
1210 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1211 case Instruction::Mul:
1212 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1213 case Instruction::UDiv:
1214 assert(!CI2->isZero() && "Div by zero handled above");
1215 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1216 case Instruction::SDiv:
1217 assert(!CI2->isZero() && "Div by zero handled above");
1218 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1219 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1220 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1221 case Instruction::URem:
1222 assert(!CI2->isZero() && "Div by zero handled above");
1223 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1224 case Instruction::SRem:
1225 assert(!CI2->isZero() && "Div by zero handled above");
1226 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1227 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1228 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1229 case Instruction::And:
1230 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1231 case Instruction::Or:
1232 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1233 case Instruction::Xor:
1234 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1235 case Instruction::Shl:
1236 if (C2V.ult(C1V.getBitWidth()))
1237 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1238 return UndefValue::get(C1->getType()); // too big shift is undef
1239 case Instruction::LShr:
1240 if (C2V.ult(C1V.getBitWidth()))
1241 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1242 return UndefValue::get(C1->getType()); // too big shift is undef
1243 case Instruction::AShr:
1244 if (C2V.ult(C1V.getBitWidth()))
1245 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1246 return UndefValue::get(C1->getType()); // too big shift is undef
1250 switch (Opcode) {
1251 case Instruction::SDiv:
1252 case Instruction::UDiv:
1253 case Instruction::URem:
1254 case Instruction::SRem:
1255 case Instruction::LShr:
1256 case Instruction::AShr:
1257 case Instruction::Shl:
1258 if (CI1->isZero()) return C1;
1259 break;
1260 default:
1261 break;
1263 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1264 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1265 const APFloat &C1V = CFP1->getValueAPF();
1266 const APFloat &C2V = CFP2->getValueAPF();
1267 APFloat C3V = C1V; // copy for modification
1268 switch (Opcode) {
1269 default:
1270 break;
1271 case Instruction::FAdd:
1272 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1273 return ConstantFP::get(C1->getContext(), C3V);
1274 case Instruction::FSub:
1275 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1276 return ConstantFP::get(C1->getContext(), C3V);
1277 case Instruction::FMul:
1278 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1279 return ConstantFP::get(C1->getContext(), C3V);
1280 case Instruction::FDiv:
1281 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1282 return ConstantFP::get(C1->getContext(), C3V);
1283 case Instruction::FRem:
1284 (void)C3V.mod(C2V);
1285 return ConstantFP::get(C1->getContext(), C3V);
1288 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1289 // Fold each element and create a vector constant from those constants.
1290 SmallVector<Constant*, 16> Result;
1291 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1292 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1293 Constant *ExtractIdx = ConstantInt::get(Ty, i);
1294 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1295 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1297 // If any element of a divisor vector is zero, the whole op is undef.
1298 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1299 return UndefValue::get(VTy);
1301 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1304 return ConstantVector::get(Result);
1307 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1308 // There are many possible foldings we could do here. We should probably
1309 // at least fold add of a pointer with an integer into the appropriate
1310 // getelementptr. This will improve alias analysis a bit.
1312 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1313 // (a + (b + c)).
1314 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1315 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1316 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1317 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1319 } else if (isa<ConstantExpr>(C2)) {
1320 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1321 // other way if possible.
1322 if (Instruction::isCommutative(Opcode))
1323 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1326 // i1 can be simplified in many cases.
1327 if (C1->getType()->isIntegerTy(1)) {
1328 switch (Opcode) {
1329 case Instruction::Add:
1330 case Instruction::Sub:
1331 return ConstantExpr::getXor(C1, C2);
1332 case Instruction::Mul:
1333 return ConstantExpr::getAnd(C1, C2);
1334 case Instruction::Shl:
1335 case Instruction::LShr:
1336 case Instruction::AShr:
1337 // We can assume that C2 == 0. If it were one the result would be
1338 // undefined because the shift value is as large as the bitwidth.
1339 return C1;
1340 case Instruction::SDiv:
1341 case Instruction::UDiv:
1342 // We can assume that C2 == 1. If it were zero the result would be
1343 // undefined through division by zero.
1344 return C1;
1345 case Instruction::URem:
1346 case Instruction::SRem:
1347 // We can assume that C2 == 1. If it were zero the result would be
1348 // undefined through division by zero.
1349 return ConstantInt::getFalse(C1->getContext());
1350 default:
1351 break;
1355 // We don't know how to fold this.
1356 return nullptr;
1359 /// This type is zero-sized if it's an array or structure of zero-sized types.
1360 /// The only leaf zero-sized type is an empty structure.
1361 static bool isMaybeZeroSizedType(Type *Ty) {
1362 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1363 if (STy->isOpaque()) return true; // Can't say.
1365 // If all of elements have zero size, this does too.
1366 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1367 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1368 return true;
1370 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1371 return isMaybeZeroSizedType(ATy->getElementType());
1373 return false;
1376 /// Compare the two constants as though they were getelementptr indices.
1377 /// This allows coercion of the types to be the same thing.
1379 /// If the two constants are the "same" (after coercion), return 0. If the
1380 /// first is less than the second, return -1, if the second is less than the
1381 /// first, return 1. If the constants are not integral, return -2.
1383 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1384 if (C1 == C2) return 0;
1386 // Ok, we found a different index. If they are not ConstantInt, we can't do
1387 // anything with them.
1388 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1389 return -2; // don't know!
1391 // We cannot compare the indices if they don't fit in an int64_t.
1392 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1393 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1394 return -2; // don't know!
1396 // Ok, we have two differing integer indices. Sign extend them to be the same
1397 // type.
1398 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1399 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1401 if (C1Val == C2Val) return 0; // They are equal
1403 // If the type being indexed over is really just a zero sized type, there is
1404 // no pointer difference being made here.
1405 if (isMaybeZeroSizedType(ElTy))
1406 return -2; // dunno.
1408 // If they are really different, now that they are the same type, then we
1409 // found a difference!
1410 if (C1Val < C2Val)
1411 return -1;
1412 else
1413 return 1;
1416 /// This function determines if there is anything we can decide about the two
1417 /// constants provided. This doesn't need to handle simple things like
1418 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1419 /// If we can determine that the two constants have a particular relation to
1420 /// each other, we should return the corresponding FCmpInst predicate,
1421 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1422 /// ConstantFoldCompareInstruction.
1424 /// To simplify this code we canonicalize the relation so that the first
1425 /// operand is always the most "complex" of the two. We consider ConstantFP
1426 /// to be the simplest, and ConstantExprs to be the most complex.
1427 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1428 assert(V1->getType() == V2->getType() &&
1429 "Cannot compare values of different types!");
1431 // We do not know if a constant expression will evaluate to a number or NaN.
1432 // Therefore, we can only say that the relation is unordered or equal.
1433 if (V1 == V2) return FCmpInst::FCMP_UEQ;
1435 if (!isa<ConstantExpr>(V1)) {
1436 if (!isa<ConstantExpr>(V2)) {
1437 // Simple case, use the standard constant folder.
1438 ConstantInt *R = nullptr;
1439 R = dyn_cast<ConstantInt>(
1440 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1441 if (R && !R->isZero())
1442 return FCmpInst::FCMP_OEQ;
1443 R = dyn_cast<ConstantInt>(
1444 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1445 if (R && !R->isZero())
1446 return FCmpInst::FCMP_OLT;
1447 R = dyn_cast<ConstantInt>(
1448 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1449 if (R && !R->isZero())
1450 return FCmpInst::FCMP_OGT;
1452 // Nothing more we can do
1453 return FCmpInst::BAD_FCMP_PREDICATE;
1456 // If the first operand is simple and second is ConstantExpr, swap operands.
1457 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1458 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1459 return FCmpInst::getSwappedPredicate(SwappedRelation);
1460 } else {
1461 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1462 // constantexpr or a simple constant.
1463 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1464 switch (CE1->getOpcode()) {
1465 case Instruction::FPTrunc:
1466 case Instruction::FPExt:
1467 case Instruction::UIToFP:
1468 case Instruction::SIToFP:
1469 // We might be able to do something with these but we don't right now.
1470 break;
1471 default:
1472 break;
1475 // There are MANY other foldings that we could perform here. They will
1476 // probably be added on demand, as they seem needed.
1477 return FCmpInst::BAD_FCMP_PREDICATE;
1480 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1481 const GlobalValue *GV2) {
1482 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1483 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1484 return true;
1485 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1486 Type *Ty = GVar->getValueType();
1487 // A global with opaque type might end up being zero sized.
1488 if (!Ty->isSized())
1489 return true;
1490 // A global with an empty type might lie at the address of any other
1491 // global.
1492 if (Ty->isEmptyTy())
1493 return true;
1495 return false;
1497 // Don't try to decide equality of aliases.
1498 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1499 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1500 return ICmpInst::ICMP_NE;
1501 return ICmpInst::BAD_ICMP_PREDICATE;
1504 /// This function determines if there is anything we can decide about the two
1505 /// constants provided. This doesn't need to handle simple things like integer
1506 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1507 /// If we can determine that the two constants have a particular relation to
1508 /// each other, we should return the corresponding ICmp predicate, otherwise
1509 /// return ICmpInst::BAD_ICMP_PREDICATE.
1511 /// To simplify this code we canonicalize the relation so that the first
1512 /// operand is always the most "complex" of the two. We consider simple
1513 /// constants (like ConstantInt) to be the simplest, followed by
1514 /// GlobalValues, followed by ConstantExpr's (the most complex).
1516 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1517 bool isSigned) {
1518 assert(V1->getType() == V2->getType() &&
1519 "Cannot compare different types of values!");
1520 if (V1 == V2) return ICmpInst::ICMP_EQ;
1522 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1523 !isa<BlockAddress>(V1)) {
1524 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1525 !isa<BlockAddress>(V2)) {
1526 // We distilled this down to a simple case, use the standard constant
1527 // folder.
1528 ConstantInt *R = nullptr;
1529 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1530 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1531 if (R && !R->isZero())
1532 return pred;
1533 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1534 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1535 if (R && !R->isZero())
1536 return pred;
1537 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1538 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1539 if (R && !R->isZero())
1540 return pred;
1542 // If we couldn't figure it out, bail.
1543 return ICmpInst::BAD_ICMP_PREDICATE;
1546 // If the first operand is simple, swap operands.
1547 ICmpInst::Predicate SwappedRelation =
1548 evaluateICmpRelation(V2, V1, isSigned);
1549 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1550 return ICmpInst::getSwappedPredicate(SwappedRelation);
1552 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1553 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1554 ICmpInst::Predicate SwappedRelation =
1555 evaluateICmpRelation(V2, V1, isSigned);
1556 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1557 return ICmpInst::getSwappedPredicate(SwappedRelation);
1558 return ICmpInst::BAD_ICMP_PREDICATE;
1561 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1562 // constant (which, since the types must match, means that it's a
1563 // ConstantPointerNull).
1564 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1565 return areGlobalsPotentiallyEqual(GV, GV2);
1566 } else if (isa<BlockAddress>(V2)) {
1567 return ICmpInst::ICMP_NE; // Globals never equal labels.
1568 } else {
1569 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1570 // GlobalVals can never be null unless they have external weak linkage.
1571 // We don't try to evaluate aliases here.
1572 // NOTE: We should not be doing this constant folding if null pointer
1573 // is considered valid for the function. But currently there is no way to
1574 // query it from the Constant type.
1575 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1576 !NullPointerIsDefined(nullptr /* F */,
1577 GV->getType()->getAddressSpace()))
1578 return ICmpInst::ICMP_NE;
1580 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1581 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1582 ICmpInst::Predicate SwappedRelation =
1583 evaluateICmpRelation(V2, V1, isSigned);
1584 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1585 return ICmpInst::getSwappedPredicate(SwappedRelation);
1586 return ICmpInst::BAD_ICMP_PREDICATE;
1589 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1590 // constant (which, since the types must match, means that it is a
1591 // ConstantPointerNull).
1592 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1593 // Block address in another function can't equal this one, but block
1594 // addresses in the current function might be the same if blocks are
1595 // empty.
1596 if (BA2->getFunction() != BA->getFunction())
1597 return ICmpInst::ICMP_NE;
1598 } else {
1599 // Block addresses aren't null, don't equal the address of globals.
1600 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1601 "Canonicalization guarantee!");
1602 return ICmpInst::ICMP_NE;
1604 } else {
1605 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1606 // constantexpr, a global, block address, or a simple constant.
1607 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1608 Constant *CE1Op0 = CE1->getOperand(0);
1610 switch (CE1->getOpcode()) {
1611 case Instruction::Trunc:
1612 case Instruction::FPTrunc:
1613 case Instruction::FPExt:
1614 case Instruction::FPToUI:
1615 case Instruction::FPToSI:
1616 break; // We can't evaluate floating point casts or truncations.
1618 case Instruction::UIToFP:
1619 case Instruction::SIToFP:
1620 case Instruction::BitCast:
1621 case Instruction::ZExt:
1622 case Instruction::SExt:
1623 // We can't evaluate floating point casts or truncations.
1624 if (CE1Op0->getType()->isFPOrFPVectorTy())
1625 break;
1627 // If the cast is not actually changing bits, and the second operand is a
1628 // null pointer, do the comparison with the pre-casted value.
1629 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1630 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1631 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1632 return evaluateICmpRelation(CE1Op0,
1633 Constant::getNullValue(CE1Op0->getType()),
1634 isSigned);
1636 break;
1638 case Instruction::GetElementPtr: {
1639 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1640 // Ok, since this is a getelementptr, we know that the constant has a
1641 // pointer type. Check the various cases.
1642 if (isa<ConstantPointerNull>(V2)) {
1643 // If we are comparing a GEP to a null pointer, check to see if the base
1644 // of the GEP equals the null pointer.
1645 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1646 if (GV->hasExternalWeakLinkage())
1647 // Weak linkage GVals could be zero or not. We're comparing that
1648 // to null pointer so its greater-or-equal
1649 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1650 else
1651 // If its not weak linkage, the GVal must have a non-zero address
1652 // so the result is greater-than
1653 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1654 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1655 // If we are indexing from a null pointer, check to see if we have any
1656 // non-zero indices.
1657 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1658 if (!CE1->getOperand(i)->isNullValue())
1659 // Offsetting from null, must not be equal.
1660 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1661 // Only zero indexes from null, must still be zero.
1662 return ICmpInst::ICMP_EQ;
1664 // Otherwise, we can't really say if the first operand is null or not.
1665 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1666 if (isa<ConstantPointerNull>(CE1Op0)) {
1667 if (GV2->hasExternalWeakLinkage())
1668 // Weak linkage GVals could be zero or not. We're comparing it to
1669 // a null pointer, so its less-or-equal
1670 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1671 else
1672 // If its not weak linkage, the GVal must have a non-zero address
1673 // so the result is less-than
1674 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1675 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1676 if (GV == GV2) {
1677 // If this is a getelementptr of the same global, then it must be
1678 // different. Because the types must match, the getelementptr could
1679 // only have at most one index, and because we fold getelementptr's
1680 // with a single zero index, it must be nonzero.
1681 assert(CE1->getNumOperands() == 2 &&
1682 !CE1->getOperand(1)->isNullValue() &&
1683 "Surprising getelementptr!");
1684 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1685 } else {
1686 if (CE1GEP->hasAllZeroIndices())
1687 return areGlobalsPotentiallyEqual(GV, GV2);
1688 return ICmpInst::BAD_ICMP_PREDICATE;
1691 } else {
1692 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1693 Constant *CE2Op0 = CE2->getOperand(0);
1695 // There are MANY other foldings that we could perform here. They will
1696 // probably be added on demand, as they seem needed.
1697 switch (CE2->getOpcode()) {
1698 default: break;
1699 case Instruction::GetElementPtr:
1700 // By far the most common case to handle is when the base pointers are
1701 // obviously to the same global.
1702 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1703 // Don't know relative ordering, but check for inequality.
1704 if (CE1Op0 != CE2Op0) {
1705 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1706 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1707 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1708 cast<GlobalValue>(CE2Op0));
1709 return ICmpInst::BAD_ICMP_PREDICATE;
1711 // Ok, we know that both getelementptr instructions are based on the
1712 // same global. From this, we can precisely determine the relative
1713 // ordering of the resultant pointers.
1714 unsigned i = 1;
1716 // The logic below assumes that the result of the comparison
1717 // can be determined by finding the first index that differs.
1718 // This doesn't work if there is over-indexing in any
1719 // subsequent indices, so check for that case first.
1720 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1721 !CE2->isGEPWithNoNotionalOverIndexing())
1722 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1724 // Compare all of the operands the GEP's have in common.
1725 gep_type_iterator GTI = gep_type_begin(CE1);
1726 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1727 ++i, ++GTI)
1728 switch (IdxCompare(CE1->getOperand(i),
1729 CE2->getOperand(i), GTI.getIndexedType())) {
1730 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1731 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1732 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1735 // Ok, we ran out of things they have in common. If any leftovers
1736 // are non-zero then we have a difference, otherwise we are equal.
1737 for (; i < CE1->getNumOperands(); ++i)
1738 if (!CE1->getOperand(i)->isNullValue()) {
1739 if (isa<ConstantInt>(CE1->getOperand(i)))
1740 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1741 else
1742 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1745 for (; i < CE2->getNumOperands(); ++i)
1746 if (!CE2->getOperand(i)->isNullValue()) {
1747 if (isa<ConstantInt>(CE2->getOperand(i)))
1748 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1749 else
1750 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1752 return ICmpInst::ICMP_EQ;
1756 break;
1758 default:
1759 break;
1763 return ICmpInst::BAD_ICMP_PREDICATE;
1766 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1767 Constant *C1, Constant *C2) {
1768 Type *ResultTy;
1769 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1770 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1771 VT->getNumElements());
1772 else
1773 ResultTy = Type::getInt1Ty(C1->getContext());
1775 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1776 if (pred == FCmpInst::FCMP_FALSE)
1777 return Constant::getNullValue(ResultTy);
1779 if (pred == FCmpInst::FCMP_TRUE)
1780 return Constant::getAllOnesValue(ResultTy);
1782 // Handle some degenerate cases first
1783 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1784 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1785 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1786 // For EQ and NE, we can always pick a value for the undef to make the
1787 // predicate pass or fail, so we can return undef.
1788 // Also, if both operands are undef, we can return undef for int comparison.
1789 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1790 return UndefValue::get(ResultTy);
1792 // Otherwise, for integer compare, pick the same value as the non-undef
1793 // operand, and fold it to true or false.
1794 if (isIntegerPredicate)
1795 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1797 // Choosing NaN for the undef will always make unordered comparison succeed
1798 // and ordered comparison fails.
1799 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1802 // icmp eq/ne(null,GV) -> false/true
1803 if (C1->isNullValue()) {
1804 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1805 // Don't try to evaluate aliases. External weak GV can be null.
1806 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1807 !NullPointerIsDefined(nullptr /* F */,
1808 GV->getType()->getAddressSpace())) {
1809 if (pred == ICmpInst::ICMP_EQ)
1810 return ConstantInt::getFalse(C1->getContext());
1811 else if (pred == ICmpInst::ICMP_NE)
1812 return ConstantInt::getTrue(C1->getContext());
1814 // icmp eq/ne(GV,null) -> false/true
1815 } else if (C2->isNullValue()) {
1816 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1817 // Don't try to evaluate aliases. External weak GV can be null.
1818 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1819 !NullPointerIsDefined(nullptr /* F */,
1820 GV->getType()->getAddressSpace())) {
1821 if (pred == ICmpInst::ICMP_EQ)
1822 return ConstantInt::getFalse(C1->getContext());
1823 else if (pred == ICmpInst::ICMP_NE)
1824 return ConstantInt::getTrue(C1->getContext());
1828 // If the comparison is a comparison between two i1's, simplify it.
1829 if (C1->getType()->isIntegerTy(1)) {
1830 switch(pred) {
1831 case ICmpInst::ICMP_EQ:
1832 if (isa<ConstantInt>(C2))
1833 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1834 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1835 case ICmpInst::ICMP_NE:
1836 return ConstantExpr::getXor(C1, C2);
1837 default:
1838 break;
1842 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1843 const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1844 const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1845 switch (pred) {
1846 default: llvm_unreachable("Invalid ICmp Predicate");
1847 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1848 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1849 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1850 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1851 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1852 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1853 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1854 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1855 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1856 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1858 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1859 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1860 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1861 APFloat::cmpResult R = C1V.compare(C2V);
1862 switch (pred) {
1863 default: llvm_unreachable("Invalid FCmp Predicate");
1864 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1865 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1866 case FCmpInst::FCMP_UNO:
1867 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1868 case FCmpInst::FCMP_ORD:
1869 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1870 case FCmpInst::FCMP_UEQ:
1871 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1872 R==APFloat::cmpEqual);
1873 case FCmpInst::FCMP_OEQ:
1874 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1875 case FCmpInst::FCMP_UNE:
1876 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1877 case FCmpInst::FCMP_ONE:
1878 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1879 R==APFloat::cmpGreaterThan);
1880 case FCmpInst::FCMP_ULT:
1881 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1882 R==APFloat::cmpLessThan);
1883 case FCmpInst::FCMP_OLT:
1884 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1885 case FCmpInst::FCMP_UGT:
1886 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1887 R==APFloat::cmpGreaterThan);
1888 case FCmpInst::FCMP_OGT:
1889 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1890 case FCmpInst::FCMP_ULE:
1891 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1892 case FCmpInst::FCMP_OLE:
1893 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1894 R==APFloat::cmpEqual);
1895 case FCmpInst::FCMP_UGE:
1896 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1897 case FCmpInst::FCMP_OGE:
1898 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1899 R==APFloat::cmpEqual);
1901 } else if (C1->getType()->isVectorTy()) {
1902 // If we can constant fold the comparison of each element, constant fold
1903 // the whole vector comparison.
1904 SmallVector<Constant*, 4> ResElts;
1905 Type *Ty = IntegerType::get(C1->getContext(), 32);
1906 // Compare the elements, producing an i1 result or constant expr.
1907 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1908 Constant *C1E =
1909 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1910 Constant *C2E =
1911 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1913 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1916 return ConstantVector::get(ResElts);
1919 if (C1->getType()->isFloatingPointTy() &&
1920 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1921 // infinite recursive loop
1922 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1923 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1924 switch (evaluateFCmpRelation(C1, C2)) {
1925 default: llvm_unreachable("Unknown relation!");
1926 case FCmpInst::FCMP_UNO:
1927 case FCmpInst::FCMP_ORD:
1928 case FCmpInst::FCMP_UNE:
1929 case FCmpInst::FCMP_ULT:
1930 case FCmpInst::FCMP_UGT:
1931 case FCmpInst::FCMP_ULE:
1932 case FCmpInst::FCMP_UGE:
1933 case FCmpInst::FCMP_TRUE:
1934 case FCmpInst::FCMP_FALSE:
1935 case FCmpInst::BAD_FCMP_PREDICATE:
1936 break; // Couldn't determine anything about these constants.
1937 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1938 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1939 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1940 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1941 break;
1942 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1943 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1944 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1945 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1946 break;
1947 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1948 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1949 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1950 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1951 break;
1952 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1953 // We can only partially decide this relation.
1954 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1955 Result = 0;
1956 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1957 Result = 1;
1958 break;
1959 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1960 // We can only partially decide this relation.
1961 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1962 Result = 0;
1963 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1964 Result = 1;
1965 break;
1966 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1967 // We can only partially decide this relation.
1968 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1969 Result = 0;
1970 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1971 Result = 1;
1972 break;
1973 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
1974 // We can only partially decide this relation.
1975 if (pred == FCmpInst::FCMP_ONE)
1976 Result = 0;
1977 else if (pred == FCmpInst::FCMP_UEQ)
1978 Result = 1;
1979 break;
1982 // If we evaluated the result, return it now.
1983 if (Result != -1)
1984 return ConstantInt::get(ResultTy, Result);
1986 } else {
1987 // Evaluate the relation between the two constants, per the predicate.
1988 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1989 switch (evaluateICmpRelation(C1, C2,
1990 CmpInst::isSigned((CmpInst::Predicate)pred))) {
1991 default: llvm_unreachable("Unknown relational!");
1992 case ICmpInst::BAD_ICMP_PREDICATE:
1993 break; // Couldn't determine anything about these constants.
1994 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1995 // If we know the constants are equal, we can decide the result of this
1996 // computation precisely.
1997 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1998 break;
1999 case ICmpInst::ICMP_ULT:
2000 switch (pred) {
2001 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
2002 Result = 1; break;
2003 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
2004 Result = 0; break;
2006 break;
2007 case ICmpInst::ICMP_SLT:
2008 switch (pred) {
2009 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
2010 Result = 1; break;
2011 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
2012 Result = 0; break;
2014 break;
2015 case ICmpInst::ICMP_UGT:
2016 switch (pred) {
2017 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
2018 Result = 1; break;
2019 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2020 Result = 0; break;
2022 break;
2023 case ICmpInst::ICMP_SGT:
2024 switch (pred) {
2025 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2026 Result = 1; break;
2027 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2028 Result = 0; break;
2030 break;
2031 case ICmpInst::ICMP_ULE:
2032 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2033 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2034 break;
2035 case ICmpInst::ICMP_SLE:
2036 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2037 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2038 break;
2039 case ICmpInst::ICMP_UGE:
2040 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2041 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2042 break;
2043 case ICmpInst::ICMP_SGE:
2044 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2045 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2046 break;
2047 case ICmpInst::ICMP_NE:
2048 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2049 if (pred == ICmpInst::ICMP_NE) Result = 1;
2050 break;
2053 // If we evaluated the result, return it now.
2054 if (Result != -1)
2055 return ConstantInt::get(ResultTy, Result);
2057 // If the right hand side is a bitcast, try using its inverse to simplify
2058 // it by moving it to the left hand side. We can't do this if it would turn
2059 // a vector compare into a scalar compare or visa versa, or if it would turn
2060 // the operands into FP values.
2061 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2062 Constant *CE2Op0 = CE2->getOperand(0);
2063 if (CE2->getOpcode() == Instruction::BitCast &&
2064 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
2065 !CE2Op0->getType()->isFPOrFPVectorTy()) {
2066 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2067 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2071 // If the left hand side is an extension, try eliminating it.
2072 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2073 if ((CE1->getOpcode() == Instruction::SExt &&
2074 ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2075 (CE1->getOpcode() == Instruction::ZExt &&
2076 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2077 Constant *CE1Op0 = CE1->getOperand(0);
2078 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2079 if (CE1Inverse == CE1Op0) {
2080 // Check whether we can safely truncate the right hand side.
2081 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2082 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2083 C2->getType()) == C2)
2084 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2089 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2090 (C1->isNullValue() && !C2->isNullValue())) {
2091 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2092 // other way if possible.
2093 // Also, if C1 is null and C2 isn't, flip them around.
2094 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2095 return ConstantExpr::getICmp(pred, C2, C1);
2098 return nullptr;
2101 /// Test whether the given sequence of *normalized* indices is "inbounds".
2102 template<typename IndexTy>
2103 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2104 // No indices means nothing that could be out of bounds.
2105 if (Idxs.empty()) return true;
2107 // If the first index is zero, it's in bounds.
2108 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2110 // If the first index is one and all the rest are zero, it's in bounds,
2111 // by the one-past-the-end rule.
2112 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2113 if (!CI->isOne())
2114 return false;
2115 } else {
2116 auto *CV = cast<ConstantDataVector>(Idxs[0]);
2117 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2118 if (!CI || !CI->isOne())
2119 return false;
2122 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2123 if (!cast<Constant>(Idxs[i])->isNullValue())
2124 return false;
2125 return true;
2128 /// Test whether a given ConstantInt is in-range for a SequentialType.
2129 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2130 const ConstantInt *CI) {
2131 // We cannot bounds check the index if it doesn't fit in an int64_t.
2132 if (CI->getValue().getMinSignedBits() > 64)
2133 return false;
2135 // A negative index or an index past the end of our sequential type is
2136 // considered out-of-range.
2137 int64_t IndexVal = CI->getSExtValue();
2138 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2139 return false;
2141 // Otherwise, it is in-range.
2142 return true;
2145 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2146 bool InBounds,
2147 Optional<unsigned> InRangeIndex,
2148 ArrayRef<Value *> Idxs) {
2149 if (Idxs.empty()) return C;
2151 Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2152 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2154 if (isa<UndefValue>(C))
2155 return UndefValue::get(GEPTy);
2157 Constant *Idx0 = cast<Constant>(Idxs[0]);
2158 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2159 return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2160 ? ConstantVector::getSplat(
2161 cast<VectorType>(GEPTy)->getNumElements(), C)
2162 : C;
2164 if (C->isNullValue()) {
2165 bool isNull = true;
2166 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2167 if (!isa<UndefValue>(Idxs[i]) &&
2168 !cast<Constant>(Idxs[i])->isNullValue()) {
2169 isNull = false;
2170 break;
2172 if (isNull) {
2173 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2174 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2176 assert(Ty && "Invalid indices for GEP!");
2177 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2178 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2179 if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2180 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2182 // The GEP returns a vector of pointers when one of more of
2183 // its arguments is a vector.
2184 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2185 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2186 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements());
2187 break;
2191 return Constant::getNullValue(GEPTy);
2195 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2196 // Combine Indices - If the source pointer to this getelementptr instruction
2197 // is a getelementptr instruction, combine the indices of the two
2198 // getelementptr instructions into a single instruction.
2200 if (CE->getOpcode() == Instruction::GetElementPtr) {
2201 gep_type_iterator LastI = gep_type_end(CE);
2202 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2203 I != E; ++I)
2204 LastI = I;
2206 // We cannot combine indices if doing so would take us outside of an
2207 // array or vector. Doing otherwise could trick us if we evaluated such a
2208 // GEP as part of a load.
2210 // e.g. Consider if the original GEP was:
2211 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2212 // i32 0, i32 0, i64 0)
2214 // If we then tried to offset it by '8' to get to the third element,
2215 // an i8, we should *not* get:
2216 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2217 // i32 0, i32 0, i64 8)
2219 // This GEP tries to index array element '8 which runs out-of-bounds.
2220 // Subsequent evaluation would get confused and produce erroneous results.
2222 // The following prohibits such a GEP from being formed by checking to see
2223 // if the index is in-range with respect to an array.
2224 // TODO: This code may be extended to handle vectors as well.
2225 bool PerformFold = false;
2226 if (Idx0->isNullValue())
2227 PerformFold = true;
2228 else if (LastI.isSequential())
2229 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2230 PerformFold = (!LastI.isBoundedSequential() ||
2231 isIndexInRangeOfArrayType(
2232 LastI.getSequentialNumElements(), CI)) &&
2233 !CE->getOperand(CE->getNumOperands() - 1)
2234 ->getType()
2235 ->isVectorTy();
2237 if (PerformFold) {
2238 SmallVector<Value*, 16> NewIndices;
2239 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2240 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2242 // Add the last index of the source with the first index of the new GEP.
2243 // Make sure to handle the case when they are actually different types.
2244 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2245 // Otherwise it must be an array.
2246 if (!Idx0->isNullValue()) {
2247 Type *IdxTy = Combined->getType();
2248 if (IdxTy != Idx0->getType()) {
2249 unsigned CommonExtendedWidth =
2250 std::max(IdxTy->getIntegerBitWidth(),
2251 Idx0->getType()->getIntegerBitWidth());
2252 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2254 Type *CommonTy =
2255 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2256 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2257 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2258 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2259 } else {
2260 Combined =
2261 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2265 NewIndices.push_back(Combined);
2266 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2268 // The combined GEP normally inherits its index inrange attribute from
2269 // the inner GEP, but if the inner GEP's last index was adjusted by the
2270 // outer GEP, any inbounds attribute on that index is invalidated.
2271 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex();
2272 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue())
2273 IRIndex = None;
2275 return ConstantExpr::getGetElementPtr(
2276 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2277 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(),
2278 IRIndex);
2282 // Attempt to fold casts to the same type away. For example, folding:
2284 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2285 // i64 0, i64 0)
2286 // into:
2288 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2290 // Don't fold if the cast is changing address spaces.
2291 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2292 PointerType *SrcPtrTy =
2293 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2294 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2295 if (SrcPtrTy && DstPtrTy) {
2296 ArrayType *SrcArrayTy =
2297 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2298 ArrayType *DstArrayTy =
2299 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2300 if (SrcArrayTy && DstArrayTy
2301 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2302 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2303 return ConstantExpr::getGetElementPtr(SrcArrayTy,
2304 (Constant *)CE->getOperand(0),
2305 Idxs, InBounds, InRangeIndex);
2310 // Check to see if any array indices are not within the corresponding
2311 // notional array or vector bounds. If so, try to determine if they can be
2312 // factored out into preceding dimensions.
2313 SmallVector<Constant *, 8> NewIdxs;
2314 Type *Ty = PointeeTy;
2315 Type *Prev = C->getType();
2316 bool Unknown =
2317 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2318 for (unsigned i = 1, e = Idxs.size(); i != e;
2319 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2320 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2321 // We don't know if it's in range or not.
2322 Unknown = true;
2323 continue;
2325 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2326 // Skip if the type of the previous index is not supported.
2327 continue;
2328 if (InRangeIndex && i == *InRangeIndex + 1) {
2329 // If an index is marked inrange, we cannot apply this canonicalization to
2330 // the following index, as that will cause the inrange index to point to
2331 // the wrong element.
2332 continue;
2334 if (isa<StructType>(Ty)) {
2335 // The verify makes sure that GEPs into a struct are in range.
2336 continue;
2338 auto *STy = cast<SequentialType>(Ty);
2339 if (isa<VectorType>(STy)) {
2340 // There can be awkward padding in after a non-power of two vector.
2341 Unknown = true;
2342 continue;
2344 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2345 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2346 // It's in range, skip to the next index.
2347 continue;
2348 if (CI->getSExtValue() < 0) {
2349 // It's out of range and negative, don't try to factor it.
2350 Unknown = true;
2351 continue;
2353 } else {
2354 auto *CV = cast<ConstantDataVector>(Idxs[i]);
2355 bool InRange = true;
2356 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2357 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2358 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2359 if (CI->getSExtValue() < 0) {
2360 Unknown = true;
2361 break;
2364 if (InRange || Unknown)
2365 // It's in range, skip to the next index.
2366 // It's out of range and negative, don't try to factor it.
2367 continue;
2369 if (isa<StructType>(Prev)) {
2370 // It's out of range, but the prior dimension is a struct
2371 // so we can't do anything about it.
2372 Unknown = true;
2373 continue;
2375 // It's out of range, but we can factor it into the prior
2376 // dimension.
2377 NewIdxs.resize(Idxs.size());
2378 // Determine the number of elements in our sequential type.
2379 uint64_t NumElements = STy->getArrayNumElements();
2381 // Expand the current index or the previous index to a vector from a scalar
2382 // if necessary.
2383 Constant *CurrIdx = cast<Constant>(Idxs[i]);
2384 auto *PrevIdx =
2385 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2386 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2387 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2388 bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2390 if (!IsCurrIdxVector && IsPrevIdxVector)
2391 CurrIdx = ConstantDataVector::getSplat(
2392 PrevIdx->getType()->getVectorNumElements(), CurrIdx);
2394 if (!IsPrevIdxVector && IsCurrIdxVector)
2395 PrevIdx = ConstantDataVector::getSplat(
2396 CurrIdx->getType()->getVectorNumElements(), PrevIdx);
2398 Constant *Factor =
2399 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2400 if (UseVector)
2401 Factor = ConstantDataVector::getSplat(
2402 IsPrevIdxVector ? PrevIdx->getType()->getVectorNumElements()
2403 : CurrIdx->getType()->getVectorNumElements(),
2404 Factor);
2406 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2408 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2410 unsigned CommonExtendedWidth =
2411 std::max(PrevIdx->getType()->getScalarSizeInBits(),
2412 Div->getType()->getScalarSizeInBits());
2413 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2415 // Before adding, extend both operands to i64 to avoid
2416 // overflow trouble.
2417 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2418 if (UseVector)
2419 ExtendedTy = VectorType::get(
2420 ExtendedTy, IsPrevIdxVector
2421 ? PrevIdx->getType()->getVectorNumElements()
2422 : CurrIdx->getType()->getVectorNumElements());
2424 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2425 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2427 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2428 Div = ConstantExpr::getSExt(Div, ExtendedTy);
2430 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2433 // If we did any factoring, start over with the adjusted indices.
2434 if (!NewIdxs.empty()) {
2435 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2436 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2437 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2438 InRangeIndex);
2441 // If all indices are known integers and normalized, we can do a simple
2442 // check for the "inbounds" property.
2443 if (!Unknown && !InBounds)
2444 if (auto *GV = dyn_cast<GlobalVariable>(C))
2445 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2446 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2447 /*InBounds=*/true, InRangeIndex);
2449 return nullptr;