[AArch64] Crypto requires FP.
[llvm-core.git] / lib / IR / ConstantFold.cpp
blobbba230677ebf711fba4538ce26fc7ffe07dd6561
1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements folding of constants for LLVM. This implements the
11 // (internal) ConstantFold.h interface, which is used by the
12 // ConstantExpr::get* methods to automatically fold constants when possible.
14 // The current constant folding implementation is implemented in two pieces: the
15 // pieces that don't need DataLayout, and the pieces that do. This is to avoid
16 // a dependence in IR on Target.
18 //===----------------------------------------------------------------------===//
20 #include "ConstantFold.h"
21 #include "llvm/ADT/APSInt.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GetElementPtrTypeIterator.h"
27 #include "llvm/IR/GlobalAlias.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/Instructions.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 /// @brief 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 = V.lshr(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->isAllOnesValue())
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 unsigned ShAmt = Amt->getZExtValue();
272 // Cannot analyze non-byte shifts.
273 if ((ShAmt & 7) != 0)
274 return nullptr;
275 ShAmt >>= 3;
277 // If the extract is known to be all zeros, return zero.
278 if (ByteStart >= CSize-ShAmt)
279 return Constant::getNullValue(IntegerType::get(CE->getContext(),
280 ByteSize*8));
281 // If the extract is known to be fully in the input, extract it.
282 if (ByteStart+ByteSize+ShAmt <= CSize)
283 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
285 // TODO: Handle the 'partially zero' case.
286 return nullptr;
289 case Instruction::Shl: {
290 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
291 if (!Amt)
292 return nullptr;
293 unsigned ShAmt = Amt->getZExtValue();
294 // Cannot analyze non-byte shifts.
295 if ((ShAmt & 7) != 0)
296 return nullptr;
297 ShAmt >>= 3;
299 // If the extract is known to be all zeros, return zero.
300 if (ByteStart+ByteSize <= ShAmt)
301 return Constant::getNullValue(IntegerType::get(CE->getContext(),
302 ByteSize*8));
303 // If the extract is known to be fully in the input, extract it.
304 if (ByteStart >= ShAmt)
305 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
307 // TODO: Handle the 'partially zero' case.
308 return nullptr;
311 case Instruction::ZExt: {
312 unsigned SrcBitSize =
313 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
315 // If extracting something that is completely zero, return 0.
316 if (ByteStart*8 >= SrcBitSize)
317 return Constant::getNullValue(IntegerType::get(CE->getContext(),
318 ByteSize*8));
320 // If exactly extracting the input, return it.
321 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
322 return CE->getOperand(0);
324 // If extracting something completely in the input, if if the input is a
325 // multiple of 8 bits, recurse.
326 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
327 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
329 // Otherwise, if extracting a subset of the input, which is not multiple of
330 // 8 bits, do a shift and trunc to get the bits.
331 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
332 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
333 Constant *Res = CE->getOperand(0);
334 if (ByteStart)
335 Res = ConstantExpr::getLShr(Res,
336 ConstantInt::get(Res->getType(), ByteStart*8));
337 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
338 ByteSize*8));
341 // TODO: Handle the 'partially zero' case.
342 return nullptr;
347 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known
348 /// factors factored out. If Folded is false, return null if no factoring was
349 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
350 /// top-level folder.
351 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy,
352 bool Folded) {
353 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
354 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
355 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
356 return ConstantExpr::getNUWMul(E, N);
359 if (StructType *STy = dyn_cast<StructType>(Ty))
360 if (!STy->isPacked()) {
361 unsigned NumElems = STy->getNumElements();
362 // An empty struct has size zero.
363 if (NumElems == 0)
364 return ConstantExpr::getNullValue(DestTy);
365 // Check for a struct with all members having the same size.
366 Constant *MemberSize =
367 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
368 bool AllSame = true;
369 for (unsigned i = 1; i != NumElems; ++i)
370 if (MemberSize !=
371 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
372 AllSame = false;
373 break;
375 if (AllSame) {
376 Constant *N = ConstantInt::get(DestTy, NumElems);
377 return ConstantExpr::getNUWMul(MemberSize, N);
381 // Pointer size doesn't depend on the pointee type, so canonicalize them
382 // to an arbitrary pointee.
383 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
384 if (!PTy->getElementType()->isIntegerTy(1))
385 return
386 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
387 PTy->getAddressSpace()),
388 DestTy, true);
390 // If there's no interesting folding happening, bail so that we don't create
391 // a constant that looks like it needs folding but really doesn't.
392 if (!Folded)
393 return nullptr;
395 // Base case: Get a regular sizeof expression.
396 Constant *C = ConstantExpr::getSizeOf(Ty);
397 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
398 DestTy, false),
399 C, DestTy);
400 return C;
403 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known
404 /// factors factored out. If Folded is false, return null if no factoring was
405 /// possible, to avoid endlessly bouncing an unfoldable expression back into the
406 /// top-level folder.
407 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy,
408 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,
473 Type *DestTy,
474 bool Folded) {
475 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
476 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
477 DestTy, false),
478 FieldNo, DestTy);
479 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
480 return ConstantExpr::getNUWMul(E, N);
483 if (StructType *STy = dyn_cast<StructType>(Ty))
484 if (!STy->isPacked()) {
485 unsigned NumElems = STy->getNumElements();
486 // An empty struct has no members.
487 if (NumElems == 0)
488 return nullptr;
489 // Check for a struct with all members having the same size.
490 Constant *MemberSize =
491 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
492 bool AllSame = true;
493 for (unsigned i = 1; i != NumElems; ++i)
494 if (MemberSize !=
495 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
496 AllSame = false;
497 break;
499 if (AllSame) {
500 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
501 false,
502 DestTy,
503 false),
504 FieldNo, DestTy);
505 return ConstantExpr::getNUWMul(MemberSize, N);
509 // If there's no interesting folding happening, bail so that we don't create
510 // a constant that looks like it needs folding but really doesn't.
511 if (!Folded)
512 return nullptr;
514 // Base case: Get a regular offsetof expression.
515 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
516 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
517 DestTy, false),
518 C, DestTy);
519 return C;
522 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
523 Type *DestTy) {
524 if (isa<UndefValue>(V)) {
525 // zext(undef) = 0, because the top bits will be zero.
526 // sext(undef) = 0, because the top bits will all be the same.
527 // [us]itofp(undef) = 0, because the result value is bounded.
528 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
529 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
530 return Constant::getNullValue(DestTy);
531 return UndefValue::get(DestTy);
534 if (V->isNullValue() && !DestTy->isX86_MMXTy() &&
535 opc != Instruction::AddrSpaceCast)
536 return Constant::getNullValue(DestTy);
538 // If the cast operand is a constant expression, there's a few things we can
539 // do to try to simplify it.
540 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
541 if (CE->isCast()) {
542 // Try hard to fold cast of cast because they are often eliminable.
543 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
544 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
545 } else if (CE->getOpcode() == Instruction::GetElementPtr &&
546 // Do not fold addrspacecast (gep 0, .., 0). It might make the
547 // addrspacecast uncanonicalized.
548 opc != Instruction::AddrSpaceCast &&
549 // Do not fold bitcast (gep) with inrange index, as this loses
550 // information.
551 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue()) {
552 // If all of the indexes in the GEP are null values, there is no pointer
553 // adjustment going on. We might as well cast the source pointer.
554 bool isAllNull = true;
555 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
556 if (!CE->getOperand(i)->isNullValue()) {
557 isAllNull = false;
558 break;
560 if (isAllNull)
561 // This is casting one pointer type to another, always BitCast
562 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
566 // If the cast operand is a constant vector, perform the cast by
567 // operating on each element. In the cast of bitcasts, the element
568 // count may be mismatched; don't attempt to handle that here.
569 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
570 DestTy->isVectorTy() &&
571 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) {
572 SmallVector<Constant*, 16> res;
573 VectorType *DestVecTy = cast<VectorType>(DestTy);
574 Type *DstEltTy = DestVecTy->getElementType();
575 Type *Ty = IntegerType::get(V->getContext(), 32);
576 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) {
577 Constant *C =
578 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
579 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
581 return ConstantVector::get(res);
584 // We actually have to do a cast now. Perform the cast according to the
585 // opcode specified.
586 switch (opc) {
587 default:
588 llvm_unreachable("Failed to cast constant expression");
589 case Instruction::FPTrunc:
590 case Instruction::FPExt:
591 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
592 bool ignored;
593 APFloat Val = FPC->getValueAPF();
594 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
595 DestTy->isFloatTy() ? APFloat::IEEEsingle() :
596 DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
597 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
598 DestTy->isFP128Ty() ? APFloat::IEEEquad() :
599 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
600 APFloat::Bogus(),
601 APFloat::rmNearestTiesToEven, &ignored);
602 return ConstantFP::get(V->getContext(), Val);
604 return nullptr; // Can't fold.
605 case Instruction::FPToUI:
606 case Instruction::FPToSI:
607 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
608 const APFloat &V = FPC->getValueAPF();
609 bool ignored;
610 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
611 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
612 if (APFloat::opInvalidOp ==
613 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
614 // Undefined behavior invoked - the destination type can't represent
615 // the input constant.
616 return UndefValue::get(DestTy);
618 return ConstantInt::get(FPC->getContext(), IntVal);
620 return nullptr; // Can't fold.
621 case Instruction::IntToPtr: //always treated as unsigned
622 if (V->isNullValue()) // Is it an integral null value?
623 return ConstantPointerNull::get(cast<PointerType>(DestTy));
624 return nullptr; // Other pointer types cannot be casted
625 case Instruction::PtrToInt: // always treated as unsigned
626 // Is it a null pointer value?
627 if (V->isNullValue())
628 return ConstantInt::get(DestTy, 0);
629 // If this is a sizeof-like expression, pull out multiplications by
630 // known factors to expose them to subsequent folding. If it's an
631 // alignof-like expression, factor out known factors.
632 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
633 if (CE->getOpcode() == Instruction::GetElementPtr &&
634 CE->getOperand(0)->isNullValue()) {
635 GEPOperator *GEPO = cast<GEPOperator>(CE);
636 Type *Ty = GEPO->getSourceElementType();
637 if (CE->getNumOperands() == 2) {
638 // Handle a sizeof-like expression.
639 Constant *Idx = CE->getOperand(1);
640 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
641 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
642 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
643 DestTy, false),
644 Idx, DestTy);
645 return ConstantExpr::getMul(C, Idx);
647 } else if (CE->getNumOperands() == 3 &&
648 CE->getOperand(1)->isNullValue()) {
649 // Handle an alignof-like expression.
650 if (StructType *STy = dyn_cast<StructType>(Ty))
651 if (!STy->isPacked()) {
652 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
653 if (CI->isOne() &&
654 STy->getNumElements() == 2 &&
655 STy->getElementType(0)->isIntegerTy(1)) {
656 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
659 // Handle an offsetof-like expression.
660 if (Ty->isStructTy() || Ty->isArrayTy()) {
661 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
662 DestTy, false))
663 return C;
667 // Other pointer types cannot be casted
668 return nullptr;
669 case Instruction::UIToFP:
670 case Instruction::SIToFP:
671 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
672 const APInt &api = CI->getValue();
673 APFloat apf(DestTy->getFltSemantics(),
674 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
675 if (APFloat::opOverflow &
676 apf.convertFromAPInt(api, opc==Instruction::SIToFP,
677 APFloat::rmNearestTiesToEven)) {
678 // Undefined behavior invoked - the destination type can't represent
679 // the input constant.
680 return UndefValue::get(DestTy);
682 return ConstantFP::get(V->getContext(), apf);
684 return nullptr;
685 case Instruction::ZExt:
686 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
687 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
688 return ConstantInt::get(V->getContext(),
689 CI->getValue().zext(BitWidth));
691 return nullptr;
692 case Instruction::SExt:
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().sext(BitWidth));
698 return nullptr;
699 case Instruction::Trunc: {
700 if (V->getType()->isVectorTy())
701 return nullptr;
703 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
704 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
705 return ConstantInt::get(V->getContext(),
706 CI->getValue().trunc(DestBitWidth));
709 // The input must be a constantexpr. See if we can simplify this based on
710 // the bytes we are demanding. Only do this if the source and dest are an
711 // even multiple of a byte.
712 if ((DestBitWidth & 7) == 0 &&
713 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
714 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
715 return Res;
717 return nullptr;
719 case Instruction::BitCast:
720 return FoldBitCast(V, DestTy);
721 case Instruction::AddrSpaceCast:
722 return nullptr;
726 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
727 Constant *V1, Constant *V2) {
728 // Check for i1 and vector true/false conditions.
729 if (Cond->isNullValue()) return V2;
730 if (Cond->isAllOnesValue()) return V1;
732 // If the condition is a vector constant, fold the result elementwise.
733 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
734 SmallVector<Constant*, 16> Result;
735 Type *Ty = IntegerType::get(CondV->getContext(), 32);
736 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){
737 Constant *V;
738 Constant *V1Element = ConstantExpr::getExtractElement(V1,
739 ConstantInt::get(Ty, i));
740 Constant *V2Element = ConstantExpr::getExtractElement(V2,
741 ConstantInt::get(Ty, i));
742 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i));
743 if (V1Element == V2Element) {
744 V = V1Element;
745 } else if (isa<UndefValue>(Cond)) {
746 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
747 } else {
748 if (!isa<ConstantInt>(Cond)) break;
749 V = Cond->isNullValue() ? V2Element : V1Element;
751 Result.push_back(V);
754 // If we were able to build the vector, return it.
755 if (Result.size() == V1->getType()->getVectorNumElements())
756 return ConstantVector::get(Result);
759 if (isa<UndefValue>(Cond)) {
760 if (isa<UndefValue>(V1)) return V1;
761 return V2;
763 if (isa<UndefValue>(V1)) return V2;
764 if (isa<UndefValue>(V2)) return V1;
765 if (V1 == V2) return V1;
767 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
768 if (TrueVal->getOpcode() == Instruction::Select)
769 if (TrueVal->getOperand(0) == Cond)
770 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
772 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
773 if (FalseVal->getOpcode() == Instruction::Select)
774 if (FalseVal->getOperand(0) == Cond)
775 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
778 return nullptr;
781 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
782 Constant *Idx) {
783 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
784 return UndefValue::get(Val->getType()->getVectorElementType());
785 if (Val->isNullValue()) // ee(zero, x) -> zero
786 return Constant::getNullValue(Val->getType()->getVectorElementType());
787 // ee({w,x,y,z}, undef) -> undef
788 if (isa<UndefValue>(Idx))
789 return UndefValue::get(Val->getType()->getVectorElementType());
791 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
792 // ee({w,x,y,z}, wrong_value) -> undef
793 if (CIdx->uge(Val->getType()->getVectorNumElements()))
794 return UndefValue::get(Val->getType()->getVectorElementType());
795 return Val->getAggregateElement(CIdx->getZExtValue());
797 return nullptr;
800 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
801 Constant *Elt,
802 Constant *Idx) {
803 if (isa<UndefValue>(Idx))
804 return UndefValue::get(Val->getType());
806 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
807 if (!CIdx) return nullptr;
809 unsigned NumElts = Val->getType()->getVectorNumElements();
810 if (CIdx->uge(NumElts))
811 return UndefValue::get(Val->getType());
813 SmallVector<Constant*, 16> Result;
814 Result.reserve(NumElts);
815 auto *Ty = Type::getInt32Ty(Val->getContext());
816 uint64_t IdxVal = CIdx->getZExtValue();
817 for (unsigned i = 0; i != NumElts; ++i) {
818 if (i == IdxVal) {
819 Result.push_back(Elt);
820 continue;
823 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
824 Result.push_back(C);
827 return ConstantVector::get(Result);
830 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
831 Constant *V2,
832 Constant *Mask) {
833 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
834 Type *EltTy = V1->getType()->getVectorElementType();
836 // Undefined shuffle mask -> undefined value.
837 if (isa<UndefValue>(Mask))
838 return UndefValue::get(VectorType::get(EltTy, MaskNumElts));
840 // Don't break the bitcode reader hack.
841 if (isa<ConstantExpr>(Mask)) return nullptr;
843 unsigned SrcNumElts = V1->getType()->getVectorNumElements();
845 // Loop over the shuffle mask, evaluating each element.
846 SmallVector<Constant*, 32> Result;
847 for (unsigned i = 0; i != MaskNumElts; ++i) {
848 int Elt = ShuffleVectorInst::getMaskValue(Mask, i);
849 if (Elt == -1) {
850 Result.push_back(UndefValue::get(EltTy));
851 continue;
853 Constant *InElt;
854 if (unsigned(Elt) >= SrcNumElts*2)
855 InElt = UndefValue::get(EltTy);
856 else if (unsigned(Elt) >= SrcNumElts) {
857 Type *Ty = IntegerType::get(V2->getContext(), 32);
858 InElt =
859 ConstantExpr::getExtractElement(V2,
860 ConstantInt::get(Ty, Elt - SrcNumElts));
861 } else {
862 Type *Ty = IntegerType::get(V1->getContext(), 32);
863 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
865 Result.push_back(InElt);
868 return ConstantVector::get(Result);
871 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
872 ArrayRef<unsigned> Idxs) {
873 // Base case: no indices, so return the entire value.
874 if (Idxs.empty())
875 return Agg;
877 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
878 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
880 return nullptr;
883 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
884 Constant *Val,
885 ArrayRef<unsigned> Idxs) {
886 // Base case: no indices, so replace the entire value.
887 if (Idxs.empty())
888 return Val;
890 unsigned NumElts;
891 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
892 NumElts = ST->getNumElements();
893 else
894 NumElts = cast<SequentialType>(Agg->getType())->getNumElements();
896 SmallVector<Constant*, 32> Result;
897 for (unsigned i = 0; i != NumElts; ++i) {
898 Constant *C = Agg->getAggregateElement(i);
899 if (!C) return nullptr;
901 if (Idxs[0] == i)
902 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
904 Result.push_back(C);
907 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
908 return ConstantStruct::get(ST, Result);
909 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType()))
910 return ConstantArray::get(AT, Result);
911 return ConstantVector::get(Result);
915 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
916 Constant *C1, Constant *C2) {
917 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
919 // Handle UndefValue up front.
920 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
921 switch (static_cast<Instruction::BinaryOps>(Opcode)) {
922 case Instruction::Xor:
923 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
924 // Handle undef ^ undef -> 0 special case. This is a common
925 // idiom (misuse).
926 return Constant::getNullValue(C1->getType());
927 LLVM_FALLTHROUGH;
928 case Instruction::Add:
929 case Instruction::Sub:
930 return UndefValue::get(C1->getType());
931 case Instruction::And:
932 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
933 return C1;
934 return Constant::getNullValue(C1->getType()); // undef & X -> 0
935 case Instruction::Mul: {
936 // undef * undef -> undef
937 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
938 return C1;
939 const APInt *CV;
940 // X * undef -> undef if X is odd
941 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
942 if ((*CV)[0])
943 return UndefValue::get(C1->getType());
945 // X * undef -> 0 otherwise
946 return Constant::getNullValue(C1->getType());
948 case Instruction::SDiv:
949 case Instruction::UDiv:
950 // X / undef -> undef
951 if (isa<UndefValue>(C2))
952 return C2;
953 // undef / 0 -> undef
954 // undef / 1 -> undef
955 if (match(C2, m_Zero()) || match(C2, m_One()))
956 return C1;
957 // undef / X -> 0 otherwise
958 return Constant::getNullValue(C1->getType());
959 case Instruction::URem:
960 case Instruction::SRem:
961 // X % undef -> undef
962 if (match(C2, m_Undef()))
963 return C2;
964 // undef % 0 -> undef
965 if (match(C2, m_Zero()))
966 return C1;
967 // undef % X -> 0 otherwise
968 return Constant::getNullValue(C1->getType());
969 case Instruction::Or: // X | undef -> -1
970 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
971 return C1;
972 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
973 case Instruction::LShr:
974 // X >>l undef -> undef
975 if (isa<UndefValue>(C2))
976 return C2;
977 // undef >>l 0 -> undef
978 if (match(C2, m_Zero()))
979 return C1;
980 // undef >>l X -> 0
981 return Constant::getNullValue(C1->getType());
982 case Instruction::AShr:
983 // X >>a undef -> undef
984 if (isa<UndefValue>(C2))
985 return C2;
986 // undef >>a 0 -> undef
987 if (match(C2, m_Zero()))
988 return C1;
989 // TODO: undef >>a X -> undef if the shift is exact
990 // undef >>a X -> 0
991 return Constant::getNullValue(C1->getType());
992 case Instruction::Shl:
993 // X << undef -> undef
994 if (isa<UndefValue>(C2))
995 return C2;
996 // undef << 0 -> undef
997 if (match(C2, m_Zero()))
998 return C1;
999 // undef << X -> 0
1000 return Constant::getNullValue(C1->getType());
1001 case Instruction::FAdd:
1002 case Instruction::FSub:
1003 case Instruction::FMul:
1004 case Instruction::FDiv:
1005 case Instruction::FRem:
1006 // TODO: UNDEF handling for binary float instructions.
1007 return nullptr;
1008 case Instruction::BinaryOpsEnd:
1009 llvm_unreachable("Invalid BinaryOp");
1013 // At this point neither constant should be an UndefValue.
1014 assert(!isa<UndefValue>(C1) && !isa<UndefValue>(C2) &&
1015 "Unexpected UndefValue");
1017 // Handle simplifications when the RHS is a constant int.
1018 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1019 switch (Opcode) {
1020 case Instruction::Add:
1021 if (CI2->equalsInt(0)) return C1; // X + 0 == X
1022 break;
1023 case Instruction::Sub:
1024 if (CI2->equalsInt(0)) return C1; // X - 0 == X
1025 break;
1026 case Instruction::Mul:
1027 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
1028 if (CI2->equalsInt(1))
1029 return C1; // X * 1 == X
1030 break;
1031 case Instruction::UDiv:
1032 case Instruction::SDiv:
1033 if (CI2->equalsInt(1))
1034 return C1; // X / 1 == X
1035 if (CI2->equalsInt(0))
1036 return UndefValue::get(CI2->getType()); // X / 0 == undef
1037 break;
1038 case Instruction::URem:
1039 case Instruction::SRem:
1040 if (CI2->equalsInt(1))
1041 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1042 if (CI2->equalsInt(0))
1043 return UndefValue::get(CI2->getType()); // X % 0 == undef
1044 break;
1045 case Instruction::And:
1046 if (CI2->isZero()) return C2; // X & 0 == 0
1047 if (CI2->isAllOnesValue())
1048 return C1; // X & -1 == X
1050 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1051 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1052 if (CE1->getOpcode() == Instruction::ZExt) {
1053 unsigned DstWidth = CI2->getType()->getBitWidth();
1054 unsigned SrcWidth =
1055 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1056 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1057 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1058 return C1;
1061 // If and'ing the address of a global with a constant, fold it.
1062 if (CE1->getOpcode() == Instruction::PtrToInt &&
1063 isa<GlobalValue>(CE1->getOperand(0))) {
1064 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1066 // Functions are at least 4-byte aligned.
1067 unsigned GVAlign = GV->getAlignment();
1068 if (isa<Function>(GV))
1069 GVAlign = std::max(GVAlign, 4U);
1071 if (GVAlign > 1) {
1072 unsigned DstWidth = CI2->getType()->getBitWidth();
1073 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1074 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1076 // If checking bits we know are clear, return zero.
1077 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1078 return Constant::getNullValue(CI2->getType());
1082 break;
1083 case Instruction::Or:
1084 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1085 if (CI2->isAllOnesValue())
1086 return C2; // X | -1 == -1
1087 break;
1088 case Instruction::Xor:
1089 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1091 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1092 switch (CE1->getOpcode()) {
1093 default: break;
1094 case Instruction::ICmp:
1095 case Instruction::FCmp:
1096 // cmp pred ^ true -> cmp !pred
1097 assert(CI2->equalsInt(1));
1098 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1099 pred = CmpInst::getInversePredicate(pred);
1100 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1101 CE1->getOperand(1));
1104 break;
1105 case Instruction::AShr:
1106 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1107 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1108 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1109 return ConstantExpr::getLShr(C1, C2);
1110 break;
1112 } else if (isa<ConstantInt>(C1)) {
1113 // If C1 is a ConstantInt and C2 is not, swap the operands.
1114 if (Instruction::isCommutative(Opcode))
1115 return ConstantExpr::get(Opcode, C2, C1);
1118 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1119 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1120 const APInt &C1V = CI1->getValue();
1121 const APInt &C2V = CI2->getValue();
1122 switch (Opcode) {
1123 default:
1124 break;
1125 case Instruction::Add:
1126 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1127 case Instruction::Sub:
1128 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1129 case Instruction::Mul:
1130 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1131 case Instruction::UDiv:
1132 assert(!CI2->isNullValue() && "Div by zero handled above");
1133 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1134 case Instruction::SDiv:
1135 assert(!CI2->isNullValue() && "Div by zero handled above");
1136 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1137 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1138 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1139 case Instruction::URem:
1140 assert(!CI2->isNullValue() && "Div by zero handled above");
1141 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1142 case Instruction::SRem:
1143 assert(!CI2->isNullValue() && "Div by zero handled above");
1144 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1145 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1146 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1147 case Instruction::And:
1148 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1149 case Instruction::Or:
1150 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1151 case Instruction::Xor:
1152 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1153 case Instruction::Shl:
1154 if (C2V.ult(C1V.getBitWidth()))
1155 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1156 return UndefValue::get(C1->getType()); // too big shift is undef
1157 case Instruction::LShr:
1158 if (C2V.ult(C1V.getBitWidth()))
1159 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1160 return UndefValue::get(C1->getType()); // too big shift is undef
1161 case Instruction::AShr:
1162 if (C2V.ult(C1V.getBitWidth()))
1163 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1164 return UndefValue::get(C1->getType()); // too big shift is undef
1168 switch (Opcode) {
1169 case Instruction::SDiv:
1170 case Instruction::UDiv:
1171 case Instruction::URem:
1172 case Instruction::SRem:
1173 case Instruction::LShr:
1174 case Instruction::AShr:
1175 case Instruction::Shl:
1176 if (CI1->equalsInt(0)) return C1;
1177 break;
1178 default:
1179 break;
1181 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1182 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1183 const APFloat &C1V = CFP1->getValueAPF();
1184 const APFloat &C2V = CFP2->getValueAPF();
1185 APFloat C3V = C1V; // copy for modification
1186 switch (Opcode) {
1187 default:
1188 break;
1189 case Instruction::FAdd:
1190 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1191 return ConstantFP::get(C1->getContext(), C3V);
1192 case Instruction::FSub:
1193 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1194 return ConstantFP::get(C1->getContext(), C3V);
1195 case Instruction::FMul:
1196 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1197 return ConstantFP::get(C1->getContext(), C3V);
1198 case Instruction::FDiv:
1199 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1200 return ConstantFP::get(C1->getContext(), C3V);
1201 case Instruction::FRem:
1202 (void)C3V.mod(C2V);
1203 return ConstantFP::get(C1->getContext(), C3V);
1206 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1207 // Perform elementwise folding.
1208 SmallVector<Constant*, 16> Result;
1209 Type *Ty = IntegerType::get(VTy->getContext(), 32);
1210 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1211 Constant *ExtractIdx = ConstantInt::get(Ty, i);
1212 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1213 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1215 // If any element of a divisor vector is zero, the whole op is undef.
1216 if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv ||
1217 Opcode == Instruction::SRem || Opcode == Instruction::URem) &&
1218 RHS->isNullValue())
1219 return UndefValue::get(VTy);
1221 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1224 return ConstantVector::get(Result);
1227 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1228 // There are many possible foldings we could do here. We should probably
1229 // at least fold add of a pointer with an integer into the appropriate
1230 // getelementptr. This will improve alias analysis a bit.
1232 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1233 // (a + (b + c)).
1234 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1235 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1236 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1237 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1239 } else if (isa<ConstantExpr>(C2)) {
1240 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1241 // other way if possible.
1242 if (Instruction::isCommutative(Opcode))
1243 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1246 // i1 can be simplified in many cases.
1247 if (C1->getType()->isIntegerTy(1)) {
1248 switch (Opcode) {
1249 case Instruction::Add:
1250 case Instruction::Sub:
1251 return ConstantExpr::getXor(C1, C2);
1252 case Instruction::Mul:
1253 return ConstantExpr::getAnd(C1, C2);
1254 case Instruction::Shl:
1255 case Instruction::LShr:
1256 case Instruction::AShr:
1257 // We can assume that C2 == 0. If it were one the result would be
1258 // undefined because the shift value is as large as the bitwidth.
1259 return C1;
1260 case Instruction::SDiv:
1261 case Instruction::UDiv:
1262 // We can assume that C2 == 1. If it were zero the result would be
1263 // undefined through division by zero.
1264 return C1;
1265 case Instruction::URem:
1266 case Instruction::SRem:
1267 // We can assume that C2 == 1. If it were zero the result would be
1268 // undefined through division by zero.
1269 return ConstantInt::getFalse(C1->getContext());
1270 default:
1271 break;
1275 // We don't know how to fold this.
1276 return nullptr;
1279 /// This type is zero-sized if it's an array or structure of zero-sized types.
1280 /// The only leaf zero-sized type is an empty structure.
1281 static bool isMaybeZeroSizedType(Type *Ty) {
1282 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1283 if (STy->isOpaque()) return true; // Can't say.
1285 // If all of elements have zero size, this does too.
1286 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1287 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1288 return true;
1290 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1291 return isMaybeZeroSizedType(ATy->getElementType());
1293 return false;
1296 /// Compare the two constants as though they were getelementptr indices.
1297 /// This allows coercion of the types to be the same thing.
1299 /// If the two constants are the "same" (after coercion), return 0. If the
1300 /// first is less than the second, return -1, if the second is less than the
1301 /// first, return 1. If the constants are not integral, return -2.
1303 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1304 if (C1 == C2) return 0;
1306 // Ok, we found a different index. If they are not ConstantInt, we can't do
1307 // anything with them.
1308 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1309 return -2; // don't know!
1311 // We cannot compare the indices if they don't fit in an int64_t.
1312 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1313 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1314 return -2; // don't know!
1316 // Ok, we have two differing integer indices. Sign extend them to be the same
1317 // type.
1318 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1319 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1321 if (C1Val == C2Val) return 0; // They are equal
1323 // If the type being indexed over is really just a zero sized type, there is
1324 // no pointer difference being made here.
1325 if (isMaybeZeroSizedType(ElTy))
1326 return -2; // dunno.
1328 // If they are really different, now that they are the same type, then we
1329 // found a difference!
1330 if (C1Val < C2Val)
1331 return -1;
1332 else
1333 return 1;
1336 /// This function determines if there is anything we can decide about the two
1337 /// constants provided. This doesn't need to handle simple things like
1338 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1339 /// If we can determine that the two constants have a particular relation to
1340 /// each other, we should return the corresponding FCmpInst predicate,
1341 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1342 /// ConstantFoldCompareInstruction.
1344 /// To simplify this code we canonicalize the relation so that the first
1345 /// operand is always the most "complex" of the two. We consider ConstantFP
1346 /// to be the simplest, and ConstantExprs to be the most complex.
1347 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1348 assert(V1->getType() == V2->getType() &&
1349 "Cannot compare values of different types!");
1351 // Handle degenerate case quickly
1352 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1354 if (!isa<ConstantExpr>(V1)) {
1355 if (!isa<ConstantExpr>(V2)) {
1356 // Simple case, use the standard constant folder.
1357 ConstantInt *R = nullptr;
1358 R = dyn_cast<ConstantInt>(
1359 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1360 if (R && !R->isZero())
1361 return FCmpInst::FCMP_OEQ;
1362 R = dyn_cast<ConstantInt>(
1363 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1364 if (R && !R->isZero())
1365 return FCmpInst::FCMP_OLT;
1366 R = dyn_cast<ConstantInt>(
1367 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1368 if (R && !R->isZero())
1369 return FCmpInst::FCMP_OGT;
1371 // Nothing more we can do
1372 return FCmpInst::BAD_FCMP_PREDICATE;
1375 // If the first operand is simple and second is ConstantExpr, swap operands.
1376 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1377 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1378 return FCmpInst::getSwappedPredicate(SwappedRelation);
1379 } else {
1380 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1381 // constantexpr or a simple constant.
1382 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1383 switch (CE1->getOpcode()) {
1384 case Instruction::FPTrunc:
1385 case Instruction::FPExt:
1386 case Instruction::UIToFP:
1387 case Instruction::SIToFP:
1388 // We might be able to do something with these but we don't right now.
1389 break;
1390 default:
1391 break;
1394 // There are MANY other foldings that we could perform here. They will
1395 // probably be added on demand, as they seem needed.
1396 return FCmpInst::BAD_FCMP_PREDICATE;
1399 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1400 const GlobalValue *GV2) {
1401 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1402 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage())
1403 return true;
1404 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1405 Type *Ty = GVar->getValueType();
1406 // A global with opaque type might end up being zero sized.
1407 if (!Ty->isSized())
1408 return true;
1409 // A global with an empty type might lie at the address of any other
1410 // global.
1411 if (Ty->isEmptyTy())
1412 return true;
1414 return false;
1416 // Don't try to decide equality of aliases.
1417 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1418 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1419 return ICmpInst::ICMP_NE;
1420 return ICmpInst::BAD_ICMP_PREDICATE;
1423 /// This function determines if there is anything we can decide about the two
1424 /// constants provided. This doesn't need to handle simple things like integer
1425 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1426 /// If we can determine that the two constants have a particular relation to
1427 /// each other, we should return the corresponding ICmp predicate, otherwise
1428 /// return ICmpInst::BAD_ICMP_PREDICATE.
1430 /// To simplify this code we canonicalize the relation so that the first
1431 /// operand is always the most "complex" of the two. We consider simple
1432 /// constants (like ConstantInt) to be the simplest, followed by
1433 /// GlobalValues, followed by ConstantExpr's (the most complex).
1435 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1436 bool isSigned) {
1437 assert(V1->getType() == V2->getType() &&
1438 "Cannot compare different types of values!");
1439 if (V1 == V2) return ICmpInst::ICMP_EQ;
1441 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1442 !isa<BlockAddress>(V1)) {
1443 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1444 !isa<BlockAddress>(V2)) {
1445 // We distilled this down to a simple case, use the standard constant
1446 // folder.
1447 ConstantInt *R = nullptr;
1448 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1449 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1450 if (R && !R->isZero())
1451 return pred;
1452 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1453 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1454 if (R && !R->isZero())
1455 return pred;
1456 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1457 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1458 if (R && !R->isZero())
1459 return pred;
1461 // If we couldn't figure it out, bail.
1462 return ICmpInst::BAD_ICMP_PREDICATE;
1465 // If the first operand is simple, swap operands.
1466 ICmpInst::Predicate SwappedRelation =
1467 evaluateICmpRelation(V2, V1, isSigned);
1468 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1469 return ICmpInst::getSwappedPredicate(SwappedRelation);
1471 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1472 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1473 ICmpInst::Predicate SwappedRelation =
1474 evaluateICmpRelation(V2, V1, isSigned);
1475 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1476 return ICmpInst::getSwappedPredicate(SwappedRelation);
1477 return ICmpInst::BAD_ICMP_PREDICATE;
1480 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1481 // constant (which, since the types must match, means that it's a
1482 // ConstantPointerNull).
1483 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1484 return areGlobalsPotentiallyEqual(GV, GV2);
1485 } else if (isa<BlockAddress>(V2)) {
1486 return ICmpInst::ICMP_NE; // Globals never equal labels.
1487 } else {
1488 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1489 // GlobalVals can never be null unless they have external weak linkage.
1490 // We don't try to evaluate aliases here.
1491 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1492 return ICmpInst::ICMP_NE;
1494 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1495 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1496 ICmpInst::Predicate SwappedRelation =
1497 evaluateICmpRelation(V2, V1, isSigned);
1498 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1499 return ICmpInst::getSwappedPredicate(SwappedRelation);
1500 return ICmpInst::BAD_ICMP_PREDICATE;
1503 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1504 // constant (which, since the types must match, means that it is a
1505 // ConstantPointerNull).
1506 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1507 // Block address in another function can't equal this one, but block
1508 // addresses in the current function might be the same if blocks are
1509 // empty.
1510 if (BA2->getFunction() != BA->getFunction())
1511 return ICmpInst::ICMP_NE;
1512 } else {
1513 // Block addresses aren't null, don't equal the address of globals.
1514 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1515 "Canonicalization guarantee!");
1516 return ICmpInst::ICMP_NE;
1518 } else {
1519 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1520 // constantexpr, a global, block address, or a simple constant.
1521 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1522 Constant *CE1Op0 = CE1->getOperand(0);
1524 switch (CE1->getOpcode()) {
1525 case Instruction::Trunc:
1526 case Instruction::FPTrunc:
1527 case Instruction::FPExt:
1528 case Instruction::FPToUI:
1529 case Instruction::FPToSI:
1530 break; // We can't evaluate floating point casts or truncations.
1532 case Instruction::UIToFP:
1533 case Instruction::SIToFP:
1534 case Instruction::BitCast:
1535 case Instruction::ZExt:
1536 case Instruction::SExt:
1537 // We can't evaluate floating point casts or truncations.
1538 if (CE1Op0->getType()->isFloatingPointTy())
1539 break;
1541 // If the cast is not actually changing bits, and the second operand is a
1542 // null pointer, do the comparison with the pre-casted value.
1543 if (V2->isNullValue() &&
1544 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1545 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1546 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1547 return evaluateICmpRelation(CE1Op0,
1548 Constant::getNullValue(CE1Op0->getType()),
1549 isSigned);
1551 break;
1553 case Instruction::GetElementPtr: {
1554 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1555 // Ok, since this is a getelementptr, we know that the constant has a
1556 // pointer type. Check the various cases.
1557 if (isa<ConstantPointerNull>(V2)) {
1558 // If we are comparing a GEP to a null pointer, check to see if the base
1559 // of the GEP equals the null pointer.
1560 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1561 if (GV->hasExternalWeakLinkage())
1562 // Weak linkage GVals could be zero or not. We're comparing that
1563 // to null pointer so its greater-or-equal
1564 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1565 else
1566 // If its not weak linkage, the GVal must have a non-zero address
1567 // so the result is greater-than
1568 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1569 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1570 // If we are indexing from a null pointer, check to see if we have any
1571 // non-zero indices.
1572 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1573 if (!CE1->getOperand(i)->isNullValue())
1574 // Offsetting from null, must not be equal.
1575 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1576 // Only zero indexes from null, must still be zero.
1577 return ICmpInst::ICMP_EQ;
1579 // Otherwise, we can't really say if the first operand is null or not.
1580 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1581 if (isa<ConstantPointerNull>(CE1Op0)) {
1582 if (GV2->hasExternalWeakLinkage())
1583 // Weak linkage GVals could be zero or not. We're comparing it to
1584 // a null pointer, so its less-or-equal
1585 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1586 else
1587 // If its not weak linkage, the GVal must have a non-zero address
1588 // so the result is less-than
1589 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1590 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1591 if (GV == GV2) {
1592 // If this is a getelementptr of the same global, then it must be
1593 // different. Because the types must match, the getelementptr could
1594 // only have at most one index, and because we fold getelementptr's
1595 // with a single zero index, it must be nonzero.
1596 assert(CE1->getNumOperands() == 2 &&
1597 !CE1->getOperand(1)->isNullValue() &&
1598 "Surprising getelementptr!");
1599 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1600 } else {
1601 if (CE1GEP->hasAllZeroIndices())
1602 return areGlobalsPotentiallyEqual(GV, GV2);
1603 return ICmpInst::BAD_ICMP_PREDICATE;
1606 } else {
1607 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1608 Constant *CE2Op0 = CE2->getOperand(0);
1610 // There are MANY other foldings that we could perform here. They will
1611 // probably be added on demand, as they seem needed.
1612 switch (CE2->getOpcode()) {
1613 default: break;
1614 case Instruction::GetElementPtr:
1615 // By far the most common case to handle is when the base pointers are
1616 // obviously to the same global.
1617 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1618 // Don't know relative ordering, but check for inequality.
1619 if (CE1Op0 != CE2Op0) {
1620 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1621 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1622 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1623 cast<GlobalValue>(CE2Op0));
1624 return ICmpInst::BAD_ICMP_PREDICATE;
1626 // Ok, we know that both getelementptr instructions are based on the
1627 // same global. From this, we can precisely determine the relative
1628 // ordering of the resultant pointers.
1629 unsigned i = 1;
1631 // The logic below assumes that the result of the comparison
1632 // can be determined by finding the first index that differs.
1633 // This doesn't work if there is over-indexing in any
1634 // subsequent indices, so check for that case first.
1635 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1636 !CE2->isGEPWithNoNotionalOverIndexing())
1637 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1639 // Compare all of the operands the GEP's have in common.
1640 gep_type_iterator GTI = gep_type_begin(CE1);
1641 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1642 ++i, ++GTI)
1643 switch (IdxCompare(CE1->getOperand(i),
1644 CE2->getOperand(i), GTI.getIndexedType())) {
1645 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1646 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1647 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1650 // Ok, we ran out of things they have in common. If any leftovers
1651 // are non-zero then we have a difference, otherwise we are equal.
1652 for (; i < CE1->getNumOperands(); ++i)
1653 if (!CE1->getOperand(i)->isNullValue()) {
1654 if (isa<ConstantInt>(CE1->getOperand(i)))
1655 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1656 else
1657 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1660 for (; i < CE2->getNumOperands(); ++i)
1661 if (!CE2->getOperand(i)->isNullValue()) {
1662 if (isa<ConstantInt>(CE2->getOperand(i)))
1663 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1664 else
1665 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1667 return ICmpInst::ICMP_EQ;
1672 default:
1673 break;
1677 return ICmpInst::BAD_ICMP_PREDICATE;
1680 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1681 Constant *C1, Constant *C2) {
1682 Type *ResultTy;
1683 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1684 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1685 VT->getNumElements());
1686 else
1687 ResultTy = Type::getInt1Ty(C1->getContext());
1689 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1690 if (pred == FCmpInst::FCMP_FALSE)
1691 return Constant::getNullValue(ResultTy);
1693 if (pred == FCmpInst::FCMP_TRUE)
1694 return Constant::getAllOnesValue(ResultTy);
1696 // Handle some degenerate cases first
1697 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1698 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1699 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1700 // For EQ and NE, we can always pick a value for the undef to make the
1701 // predicate pass or fail, so we can return undef.
1702 // Also, if both operands are undef, we can return undef for int comparison.
1703 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1704 return UndefValue::get(ResultTy);
1706 // Otherwise, for integer compare, pick the same value as the non-undef
1707 // operand, and fold it to true or false.
1708 if (isIntegerPredicate)
1709 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1711 // Choosing NaN for the undef will always make unordered comparison succeed
1712 // and ordered comparison fails.
1713 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1716 // icmp eq/ne(null,GV) -> false/true
1717 if (C1->isNullValue()) {
1718 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1719 // Don't try to evaluate aliases. External weak GV can be null.
1720 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1721 if (pred == ICmpInst::ICMP_EQ)
1722 return ConstantInt::getFalse(C1->getContext());
1723 else if (pred == ICmpInst::ICMP_NE)
1724 return ConstantInt::getTrue(C1->getContext());
1726 // icmp eq/ne(GV,null) -> false/true
1727 } else if (C2->isNullValue()) {
1728 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1729 // Don't try to evaluate aliases. External weak GV can be null.
1730 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1731 if (pred == ICmpInst::ICMP_EQ)
1732 return ConstantInt::getFalse(C1->getContext());
1733 else if (pred == ICmpInst::ICMP_NE)
1734 return ConstantInt::getTrue(C1->getContext());
1738 // If the comparison is a comparison between two i1's, simplify it.
1739 if (C1->getType()->isIntegerTy(1)) {
1740 switch(pred) {
1741 case ICmpInst::ICMP_EQ:
1742 if (isa<ConstantInt>(C2))
1743 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1744 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1745 case ICmpInst::ICMP_NE:
1746 return ConstantExpr::getXor(C1, C2);
1747 default:
1748 break;
1752 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1753 const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1754 const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1755 switch (pred) {
1756 default: llvm_unreachable("Invalid ICmp Predicate");
1757 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1758 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1759 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1760 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1761 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1762 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1763 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1764 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1765 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1766 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1768 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1769 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1770 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1771 APFloat::cmpResult R = C1V.compare(C2V);
1772 switch (pred) {
1773 default: llvm_unreachable("Invalid FCmp Predicate");
1774 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1775 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1776 case FCmpInst::FCMP_UNO:
1777 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1778 case FCmpInst::FCMP_ORD:
1779 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1780 case FCmpInst::FCMP_UEQ:
1781 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1782 R==APFloat::cmpEqual);
1783 case FCmpInst::FCMP_OEQ:
1784 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1785 case FCmpInst::FCMP_UNE:
1786 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1787 case FCmpInst::FCMP_ONE:
1788 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1789 R==APFloat::cmpGreaterThan);
1790 case FCmpInst::FCMP_ULT:
1791 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1792 R==APFloat::cmpLessThan);
1793 case FCmpInst::FCMP_OLT:
1794 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1795 case FCmpInst::FCMP_UGT:
1796 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1797 R==APFloat::cmpGreaterThan);
1798 case FCmpInst::FCMP_OGT:
1799 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1800 case FCmpInst::FCMP_ULE:
1801 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1802 case FCmpInst::FCMP_OLE:
1803 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1804 R==APFloat::cmpEqual);
1805 case FCmpInst::FCMP_UGE:
1806 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1807 case FCmpInst::FCMP_OGE:
1808 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1809 R==APFloat::cmpEqual);
1811 } else if (C1->getType()->isVectorTy()) {
1812 // If we can constant fold the comparison of each element, constant fold
1813 // the whole vector comparison.
1814 SmallVector<Constant*, 4> ResElts;
1815 Type *Ty = IntegerType::get(C1->getContext(), 32);
1816 // Compare the elements, producing an i1 result or constant expr.
1817 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){
1818 Constant *C1E =
1819 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i));
1820 Constant *C2E =
1821 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i));
1823 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1826 return ConstantVector::get(ResElts);
1829 if (C1->getType()->isFloatingPointTy() &&
1830 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1831 // infinite recursive loop
1832 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1833 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1834 switch (evaluateFCmpRelation(C1, C2)) {
1835 default: llvm_unreachable("Unknown relation!");
1836 case FCmpInst::FCMP_UNO:
1837 case FCmpInst::FCMP_ORD:
1838 case FCmpInst::FCMP_UEQ:
1839 case FCmpInst::FCMP_UNE:
1840 case FCmpInst::FCMP_ULT:
1841 case FCmpInst::FCMP_UGT:
1842 case FCmpInst::FCMP_ULE:
1843 case FCmpInst::FCMP_UGE:
1844 case FCmpInst::FCMP_TRUE:
1845 case FCmpInst::FCMP_FALSE:
1846 case FCmpInst::BAD_FCMP_PREDICATE:
1847 break; // Couldn't determine anything about these constants.
1848 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1849 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1850 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1851 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1852 break;
1853 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1854 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1855 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1856 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1857 break;
1858 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1859 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1860 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1861 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1862 break;
1863 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1864 // We can only partially decide this relation.
1865 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1866 Result = 0;
1867 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1868 Result = 1;
1869 break;
1870 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1871 // We can only partially decide this relation.
1872 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1873 Result = 0;
1874 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1875 Result = 1;
1876 break;
1877 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1878 // We can only partially decide this relation.
1879 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1880 Result = 0;
1881 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1882 Result = 1;
1883 break;
1886 // If we evaluated the result, return it now.
1887 if (Result != -1)
1888 return ConstantInt::get(ResultTy, Result);
1890 } else {
1891 // Evaluate the relation between the two constants, per the predicate.
1892 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1893 switch (evaluateICmpRelation(C1, C2,
1894 CmpInst::isSigned((CmpInst::Predicate)pred))) {
1895 default: llvm_unreachable("Unknown relational!");
1896 case ICmpInst::BAD_ICMP_PREDICATE:
1897 break; // Couldn't determine anything about these constants.
1898 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1899 // If we know the constants are equal, we can decide the result of this
1900 // computation precisely.
1901 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1902 break;
1903 case ICmpInst::ICMP_ULT:
1904 switch (pred) {
1905 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1906 Result = 1; break;
1907 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1908 Result = 0; break;
1910 break;
1911 case ICmpInst::ICMP_SLT:
1912 switch (pred) {
1913 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1914 Result = 1; break;
1915 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1916 Result = 0; break;
1918 break;
1919 case ICmpInst::ICMP_UGT:
1920 switch (pred) {
1921 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1922 Result = 1; break;
1923 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1924 Result = 0; break;
1926 break;
1927 case ICmpInst::ICMP_SGT:
1928 switch (pred) {
1929 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1930 Result = 1; break;
1931 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1932 Result = 0; break;
1934 break;
1935 case ICmpInst::ICMP_ULE:
1936 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1937 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1938 break;
1939 case ICmpInst::ICMP_SLE:
1940 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1941 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1942 break;
1943 case ICmpInst::ICMP_UGE:
1944 if (pred == ICmpInst::ICMP_ULT) Result = 0;
1945 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
1946 break;
1947 case ICmpInst::ICMP_SGE:
1948 if (pred == ICmpInst::ICMP_SLT) Result = 0;
1949 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
1950 break;
1951 case ICmpInst::ICMP_NE:
1952 if (pred == ICmpInst::ICMP_EQ) Result = 0;
1953 if (pred == ICmpInst::ICMP_NE) Result = 1;
1954 break;
1957 // If we evaluated the result, return it now.
1958 if (Result != -1)
1959 return ConstantInt::get(ResultTy, Result);
1961 // If the right hand side is a bitcast, try using its inverse to simplify
1962 // it by moving it to the left hand side. We can't do this if it would turn
1963 // a vector compare into a scalar compare or visa versa.
1964 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
1965 Constant *CE2Op0 = CE2->getOperand(0);
1966 if (CE2->getOpcode() == Instruction::BitCast &&
1967 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
1968 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
1969 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
1973 // If the left hand side is an extension, try eliminating it.
1974 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1975 if ((CE1->getOpcode() == Instruction::SExt &&
1976 ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
1977 (CE1->getOpcode() == Instruction::ZExt &&
1978 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
1979 Constant *CE1Op0 = CE1->getOperand(0);
1980 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
1981 if (CE1Inverse == CE1Op0) {
1982 // Check whether we can safely truncate the right hand side.
1983 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
1984 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
1985 C2->getType()) == C2)
1986 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
1991 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
1992 (C1->isNullValue() && !C2->isNullValue())) {
1993 // If C2 is a constant expr and C1 isn't, flip them around and fold the
1994 // other way if possible.
1995 // Also, if C1 is null and C2 isn't, flip them around.
1996 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1997 return ConstantExpr::getICmp(pred, C2, C1);
2000 return nullptr;
2003 /// Test whether the given sequence of *normalized* indices is "inbounds".
2004 template<typename IndexTy>
2005 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2006 // No indices means nothing that could be out of bounds.
2007 if (Idxs.empty()) return true;
2009 // If the first index is zero, it's in bounds.
2010 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2012 // If the first index is one and all the rest are zero, it's in bounds,
2013 // by the one-past-the-end rule.
2014 if (!cast<ConstantInt>(Idxs[0])->isOne())
2015 return false;
2016 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2017 if (!cast<Constant>(Idxs[i])->isNullValue())
2018 return false;
2019 return true;
2022 /// Test whether a given ConstantInt is in-range for a SequentialType.
2023 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2024 const ConstantInt *CI) {
2025 // We cannot bounds check the index if it doesn't fit in an int64_t.
2026 if (CI->getValue().getActiveBits() > 64)
2027 return false;
2029 // A negative index or an index past the end of our sequential type is
2030 // considered out-of-range.
2031 int64_t IndexVal = CI->getSExtValue();
2032 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2033 return false;
2035 // Otherwise, it is in-range.
2036 return true;
2039 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2040 bool InBounds,
2041 Optional<unsigned> InRangeIndex,
2042 ArrayRef<Value *> Idxs) {
2043 if (Idxs.empty()) return C;
2044 Constant *Idx0 = cast<Constant>(Idxs[0]);
2045 if ((Idxs.size() == 1 && Idx0->isNullValue()))
2046 return C;
2048 if (isa<UndefValue>(C)) {
2049 Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2050 C, makeArrayRef((Value * const *)Idxs.data(), Idxs.size()));
2051 return UndefValue::get(GEPTy);
2054 if (C->isNullValue()) {
2055 bool isNull = true;
2056 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2057 if (!cast<Constant>(Idxs[i])->isNullValue()) {
2058 isNull = false;
2059 break;
2061 if (isNull) {
2062 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2063 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2065 assert(Ty && "Invalid indices for GEP!");
2066 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2067 if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2068 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
2069 return Constant::getNullValue(GEPTy);
2073 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2074 // Combine Indices - If the source pointer to this getelementptr instruction
2075 // is a getelementptr instruction, combine the indices of the two
2076 // getelementptr instructions into a single instruction.
2078 if (CE->getOpcode() == Instruction::GetElementPtr) {
2079 gep_type_iterator LastI = gep_type_end(CE);
2080 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2081 I != E; ++I)
2082 LastI = I;
2084 // We cannot combine indices if doing so would take us outside of an
2085 // array or vector. Doing otherwise could trick us if we evaluated such a
2086 // GEP as part of a load.
2088 // e.g. Consider if the original GEP was:
2089 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2090 // i32 0, i32 0, i64 0)
2092 // If we then tried to offset it by '8' to get to the third element,
2093 // an i8, we should *not* get:
2094 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2095 // i32 0, i32 0, i64 8)
2097 // This GEP tries to index array element '8 which runs out-of-bounds.
2098 // Subsequent evaluation would get confused and produce erroneous results.
2100 // The following prohibits such a GEP from being formed by checking to see
2101 // if the index is in-range with respect to an array or vector.
2102 bool PerformFold = false;
2103 if (Idx0->isNullValue())
2104 PerformFold = true;
2105 else if (LastI.isSequential())
2106 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0))
2107 PerformFold =
2108 !LastI.isBoundedSequential() ||
2109 isIndexInRangeOfArrayType(LastI.getSequentialNumElements(), CI);
2111 if (PerformFold) {
2112 SmallVector<Value*, 16> NewIndices;
2113 NewIndices.reserve(Idxs.size() + CE->getNumOperands());
2114 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1);
2116 // Add the last index of the source with the first index of the new GEP.
2117 // Make sure to handle the case when they are actually different types.
2118 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2119 // Otherwise it must be an array.
2120 if (!Idx0->isNullValue()) {
2121 Type *IdxTy = Combined->getType();
2122 if (IdxTy != Idx0->getType()) {
2123 unsigned CommonExtendedWidth =
2124 std::max(IdxTy->getIntegerBitWidth(),
2125 Idx0->getType()->getIntegerBitWidth());
2126 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2128 Type *CommonTy =
2129 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth);
2130 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2131 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy);
2132 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2133 } else {
2134 Combined =
2135 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2139 NewIndices.push_back(Combined);
2140 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2142 // The combined GEP normally inherits its index inrange attribute from
2143 // the inner GEP, but if the inner GEP's last index was adjusted by the
2144 // outer GEP, any inbounds attribute on that index is invalidated.
2145 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex();
2146 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue())
2147 IRIndex = None;
2149 return ConstantExpr::getGetElementPtr(
2150 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0),
2151 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(),
2152 IRIndex);
2156 // Attempt to fold casts to the same type away. For example, folding:
2158 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2159 // i64 0, i64 0)
2160 // into:
2162 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2164 // Don't fold if the cast is changing address spaces.
2165 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2166 PointerType *SrcPtrTy =
2167 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2168 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2169 if (SrcPtrTy && DstPtrTy) {
2170 ArrayType *SrcArrayTy =
2171 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2172 ArrayType *DstArrayTy =
2173 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2174 if (SrcArrayTy && DstArrayTy
2175 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2176 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2177 return ConstantExpr::getGetElementPtr(SrcArrayTy,
2178 (Constant *)CE->getOperand(0),
2179 Idxs, InBounds, InRangeIndex);
2184 // Check to see if any array indices are not within the corresponding
2185 // notional array or vector bounds. If so, try to determine if they can be
2186 // factored out into preceding dimensions.
2187 SmallVector<Constant *, 8> NewIdxs;
2188 Type *Ty = PointeeTy;
2189 Type *Prev = C->getType();
2190 bool Unknown = !isa<ConstantInt>(Idxs[0]);
2191 for (unsigned i = 1, e = Idxs.size(); i != e;
2192 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2193 auto *CI = dyn_cast<ConstantInt>(Idxs[i]);
2194 if (!CI) {
2195 // We don't know if it's in range or not.
2196 Unknown = true;
2197 continue;
2199 if (InRangeIndex && i == *InRangeIndex + 1) {
2200 // If an index is marked inrange, we cannot apply this canonicalization to
2201 // the following index, as that will cause the inrange index to point to
2202 // the wrong element.
2203 continue;
2205 if (isa<StructType>(Ty)) {
2206 // The verify makes sure that GEPs into a struct are in range.
2207 continue;
2209 auto *STy = cast<SequentialType>(Ty);
2210 if (isa<VectorType>(STy)) {
2211 // There can be awkward padding in after a non-power of two vector.
2212 Unknown = true;
2213 continue;
2215 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2216 // It's in range, skip to the next index.
2217 continue;
2218 if (isa<StructType>(Prev)) {
2219 // It's out of range, but the prior dimension is a struct
2220 // so we can't do anything about it.
2221 Unknown = true;
2222 continue;
2224 if (CI->getSExtValue() < 0) {
2225 // It's out of range and negative, don't try to factor it.
2226 Unknown = true;
2227 continue;
2229 // It's out of range, but we can factor it into the prior
2230 // dimension.
2231 NewIdxs.resize(Idxs.size());
2232 // Determine the number of elements in our sequential type.
2233 uint64_t NumElements = STy->getArrayNumElements();
2235 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements);
2236 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2238 Constant *PrevIdx = NewIdxs[i-1] ? NewIdxs[i-1] :
2239 cast<Constant>(Idxs[i - 1]);
2240 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2242 unsigned CommonExtendedWidth =
2243 std::max(PrevIdx->getType()->getIntegerBitWidth(),
2244 Div->getType()->getIntegerBitWidth());
2245 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2247 // Before adding, extend both operands to i64 to avoid
2248 // overflow trouble.
2249 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth))
2250 PrevIdx = ConstantExpr::getSExt(
2251 PrevIdx, Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2252 if (!Div->getType()->isIntegerTy(CommonExtendedWidth))
2253 Div = ConstantExpr::getSExt(
2254 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth));
2256 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2259 // If we did any factoring, start over with the adjusted indices.
2260 if (!NewIdxs.empty()) {
2261 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2262 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2263 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2264 InRangeIndex);
2267 // If all indices are known integers and normalized, we can do a simple
2268 // check for the "inbounds" property.
2269 if (!Unknown && !InBounds)
2270 if (auto *GV = dyn_cast<GlobalVariable>(C))
2271 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2272 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2273 /*InBounds=*/true, InRangeIndex);
2275 return nullptr;