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.
12 //
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.
16 //
17 //===----------------------------------------------------------------------===//
18
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;
37
38 //===----------------------------------------------------------------------===//
39 // ConstantFold*Instruction Implementations
40 //===----------------------------------------------------------------------===//
41
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.
BitCastConstantVector(Constant * CV,VectorType * DstTy)45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
46
47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49
50 // Do not iterate on scalable vector. The num of elements is unknown at
51 // compile-time.
52 if (isa<ScalableVectorType>(DstTy))
53 return nullptr;
54
55 // If this cast changes element count then we can't handle it here:
56 // doing so requires endianness information. This should be handled by
57 // Analysis/ConstantFolding.cpp
58 unsigned NumElts = cast<FixedVectorType>(DstTy)->getNumElements();
59 if (NumElts != cast<FixedVectorType>(CV->getType())->getNumElements())
60 return nullptr;
61
62 Type *DstEltTy = DstTy->getElementType();
63 // Fast path for splatted constants.
64 if (Constant *Splat = CV->getSplatValue()) {
65 return ConstantVector::getSplat(DstTy->getElementCount(),
66 ConstantExpr::getBitCast(Splat, DstEltTy));
67 }
68
69 SmallVector<Constant*, 16> Result;
70 Type *Ty = IntegerType::get(CV->getContext(), 32);
71 for (unsigned i = 0; i != NumElts; ++i) {
72 Constant *C =
73 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
74 C = ConstantExpr::getBitCast(C, DstEltTy);
75 Result.push_back(C);
76 }
77
78 return ConstantVector::get(Result);
79 }
80
81 /// This function determines which opcode to use to fold two constant cast
82 /// expressions together. It uses CastInst::isEliminableCastPair to determine
83 /// the opcode. Consequently its just a wrapper around that function.
84 /// Determine if it is valid to fold a cast of a cast
85 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)86 foldConstantCastPair(
87 unsigned opc, ///< opcode of the second cast constant expression
88 ConstantExpr *Op, ///< the first cast constant expression
89 Type *DstTy ///< destination type of the first cast
90 ) {
91 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
92 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
93 assert(CastInst::isCast(opc) && "Invalid cast opcode");
94
95 // The types and opcodes for the two Cast constant expressions
96 Type *SrcTy = Op->getOperand(0)->getType();
97 Type *MidTy = Op->getType();
98 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
99 Instruction::CastOps secondOp = Instruction::CastOps(opc);
100
101 // Assume that pointers are never more than 64 bits wide, and only use this
102 // for the middle type. Otherwise we could end up folding away illegal
103 // bitcasts between address spaces with different sizes.
104 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
105
106 // Let CastInst::isEliminableCastPair do the heavy lifting.
107 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
108 nullptr, FakeIntPtrTy, nullptr);
109 }
110
FoldBitCast(Constant * V,Type * DestTy)111 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
112 Type *SrcTy = V->getType();
113 if (SrcTy == DestTy)
114 return V; // no-op cast
115
116 // Check to see if we are casting a pointer to an aggregate to a pointer to
117 // the first element. If so, return the appropriate GEP instruction.
118 if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
119 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
120 if (PTy->getAddressSpace() == DPTy->getAddressSpace() &&
121 !PTy->isOpaque() && !DPTy->isOpaque() &&
122 PTy->getElementType()->isSized()) {
123 SmallVector<Value*, 8> IdxList;
124 Value *Zero =
125 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
126 IdxList.push_back(Zero);
127 Type *ElTy = PTy->getElementType();
128 while (ElTy && ElTy != DPTy->getElementType()) {
129 ElTy = GetElementPtrInst::getTypeAtIndex(ElTy, (uint64_t)0);
130 IdxList.push_back(Zero);
131 }
132
133 if (ElTy == DPTy->getElementType())
134 // This GEP is inbounds because all indices are zero.
135 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
136 V, IdxList);
137 }
138
139 // Handle casts from one vector constant to another. We know that the src
140 // and dest type have the same size (otherwise its an illegal cast).
141 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
142 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
143 assert(DestPTy->getPrimitiveSizeInBits() ==
144 SrcTy->getPrimitiveSizeInBits() &&
145 "Not cast between same sized vectors!");
146 SrcTy = nullptr;
147 // First, check for null. Undef is already handled.
148 if (isa<ConstantAggregateZero>(V))
149 return Constant::getNullValue(DestTy);
150
151 // Handle ConstantVector and ConstantAggregateVector.
152 return BitCastConstantVector(V, DestPTy);
153 }
154
155 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
156 // This allows for other simplifications (although some of them
157 // can only be handled by Analysis/ConstantFolding.cpp).
158 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
159 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
160 }
161
162 // Finally, implement bitcast folding now. The code below doesn't handle
163 // bitcast right.
164 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
165 return ConstantPointerNull::get(cast<PointerType>(DestTy));
166
167 // Handle integral constant input.
168 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
169 if (DestTy->isIntegerTy())
170 // Integral -> Integral. This is a no-op because the bit widths must
171 // be the same. Consequently, we just fold to V.
172 return V;
173
174 // See note below regarding the PPC_FP128 restriction.
175 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
176 return ConstantFP::get(DestTy->getContext(),
177 APFloat(DestTy->getFltSemantics(),
178 CI->getValue()));
179
180 // Otherwise, can't fold this (vector?)
181 return nullptr;
182 }
183
184 // Handle ConstantFP input: FP -> Integral.
185 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
186 // PPC_FP128 is really the sum of two consecutive doubles, where the first
187 // double is always stored first in memory, regardless of the target
188 // endianness. The memory layout of i128, however, depends on the target
189 // endianness, and so we can't fold this without target endianness
190 // information. This should instead be handled by
191 // Analysis/ConstantFolding.cpp
192 if (FP->getType()->isPPC_FP128Ty())
193 return nullptr;
194
195 // Make sure dest type is compatible with the folded integer constant.
196 if (!DestTy->isIntegerTy())
197 return nullptr;
198
199 return ConstantInt::get(FP->getContext(),
200 FP->getValueAPF().bitcastToAPInt());
201 }
202
203 return nullptr;
204 }
205
206
207 /// V is an integer constant which only has a subset of its bytes used.
208 /// The bytes used are indicated by ByteStart (which is the first byte used,
209 /// counting from the least significant byte) and ByteSize, which is the number
210 /// of bytes used.
211 ///
212 /// This function analyzes the specified constant to see if the specified byte
213 /// range can be returned as a simplified constant. If so, the constant is
214 /// returned, otherwise null is returned.
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)215 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
216 unsigned ByteSize) {
217 assert(C->getType()->isIntegerTy() &&
218 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
219 "Non-byte sized integer input");
220 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
221 assert(ByteSize && "Must be accessing some piece");
222 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
223 assert(ByteSize != CSize && "Should not extract everything");
224
225 // Constant Integers are simple.
226 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
227 APInt V = CI->getValue();
228 if (ByteStart)
229 V.lshrInPlace(ByteStart*8);
230 V = V.trunc(ByteSize*8);
231 return ConstantInt::get(CI->getContext(), V);
232 }
233
234 // In the input is a constant expr, we might be able to recursively simplify.
235 // If not, we definitely can't do anything.
236 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
237 if (!CE) return nullptr;
238
239 switch (CE->getOpcode()) {
240 default: return nullptr;
241 case Instruction::Or: {
242 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
243 if (!RHS)
244 return nullptr;
245
246 // X | -1 -> -1.
247 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
248 if (RHSC->isMinusOne())
249 return RHSC;
250
251 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
252 if (!LHS)
253 return nullptr;
254 return ConstantExpr::getOr(LHS, RHS);
255 }
256 case Instruction::And: {
257 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
258 if (!RHS)
259 return nullptr;
260
261 // X & 0 -> 0.
262 if (RHS->isNullValue())
263 return RHS;
264
265 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
266 if (!LHS)
267 return nullptr;
268 return ConstantExpr::getAnd(LHS, RHS);
269 }
270 case Instruction::LShr: {
271 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
272 if (!Amt)
273 return nullptr;
274 APInt ShAmt = Amt->getValue();
275 // Cannot analyze non-byte shifts.
276 if ((ShAmt & 7) != 0)
277 return nullptr;
278 ShAmt.lshrInPlace(3);
279
280 // If the extract is known to be all zeros, return zero.
281 if (ShAmt.uge(CSize - ByteStart))
282 return Constant::getNullValue(
283 IntegerType::get(CE->getContext(), ByteSize * 8));
284 // If the extract is known to be fully in the input, extract it.
285 if (ShAmt.ule(CSize - (ByteStart + ByteSize)))
286 return ExtractConstantBytes(CE->getOperand(0),
287 ByteStart + ShAmt.getZExtValue(), ByteSize);
288
289 // TODO: Handle the 'partially zero' case.
290 return nullptr;
291 }
292
293 case Instruction::Shl: {
294 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
295 if (!Amt)
296 return nullptr;
297 APInt ShAmt = Amt->getValue();
298 // Cannot analyze non-byte shifts.
299 if ((ShAmt & 7) != 0)
300 return nullptr;
301 ShAmt.lshrInPlace(3);
302
303 // If the extract is known to be all zeros, return zero.
304 if (ShAmt.uge(ByteStart + ByteSize))
305 return Constant::getNullValue(
306 IntegerType::get(CE->getContext(), ByteSize * 8));
307 // If the extract is known to be fully in the input, extract it.
308 if (ShAmt.ule(ByteStart))
309 return ExtractConstantBytes(CE->getOperand(0),
310 ByteStart - ShAmt.getZExtValue(), ByteSize);
311
312 // TODO: Handle the 'partially zero' case.
313 return nullptr;
314 }
315
316 case Instruction::ZExt: {
317 unsigned SrcBitSize =
318 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
319
320 // If extracting something that is completely zero, return 0.
321 if (ByteStart*8 >= SrcBitSize)
322 return Constant::getNullValue(IntegerType::get(CE->getContext(),
323 ByteSize*8));
324
325 // If exactly extracting the input, return it.
326 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
327 return CE->getOperand(0);
328
329 // If extracting something completely in the input, if the input is a
330 // multiple of 8 bits, recurse.
331 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
332 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
333
334 // Otherwise, if extracting a subset of the input, which is not multiple of
335 // 8 bits, do a shift and trunc to get the bits.
336 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
337 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
338 Constant *Res = CE->getOperand(0);
339 if (ByteStart)
340 Res = ConstantExpr::getLShr(Res,
341 ConstantInt::get(Res->getType(), ByteStart*8));
342 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
343 ByteSize*8));
344 }
345
346 // TODO: Handle the 'partially zero' case.
347 return nullptr;
348 }
349 }
350 }
351
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)352 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
353 Type *DestTy) {
354 if (isa<PoisonValue>(V))
355 return PoisonValue::get(DestTy);
356
357 if (isa<UndefValue>(V)) {
358 // zext(undef) = 0, because the top bits will be zero.
359 // sext(undef) = 0, because the top bits will all be the same.
360 // [us]itofp(undef) = 0, because the result value is bounded.
361 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
362 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
363 return Constant::getNullValue(DestTy);
364 return UndefValue::get(DestTy);
365 }
366
367 if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() &&
368 opc != Instruction::AddrSpaceCast)
369 return Constant::getNullValue(DestTy);
370
371 // If the cast operand is a constant expression, there's a few things we can
372 // do to try to simplify it.
373 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
374 if (CE->isCast()) {
375 // Try hard to fold cast of cast because they are often eliminable.
376 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
377 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
378 } else if (CE->getOpcode() == Instruction::GetElementPtr &&
379 // Do not fold addrspacecast (gep 0, .., 0). It might make the
380 // addrspacecast uncanonicalized.
381 opc != Instruction::AddrSpaceCast &&
382 // Do not fold bitcast (gep) with inrange index, as this loses
383 // information.
384 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() &&
385 // Do not fold if the gep type is a vector, as bitcasting
386 // operand 0 of a vector gep will result in a bitcast between
387 // different sizes.
388 !CE->getType()->isVectorTy()) {
389 // If all of the indexes in the GEP are null values, there is no pointer
390 // adjustment going on. We might as well cast the source pointer.
391 bool isAllNull = true;
392 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
393 if (!CE->getOperand(i)->isNullValue()) {
394 isAllNull = false;
395 break;
396 }
397 if (isAllNull)
398 // This is casting one pointer type to another, always BitCast
399 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
400 }
401 }
402
403 // If the cast operand is a constant vector, perform the cast by
404 // operating on each element. In the cast of bitcasts, the element
405 // count may be mismatched; don't attempt to handle that here.
406 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
407 DestTy->isVectorTy() &&
408 cast<FixedVectorType>(DestTy)->getNumElements() ==
409 cast<FixedVectorType>(V->getType())->getNumElements()) {
410 VectorType *DestVecTy = cast<VectorType>(DestTy);
411 Type *DstEltTy = DestVecTy->getElementType();
412 // Fast path for splatted constants.
413 if (Constant *Splat = V->getSplatValue()) {
414 return ConstantVector::getSplat(
415 cast<VectorType>(DestTy)->getElementCount(),
416 ConstantExpr::getCast(opc, Splat, DstEltTy));
417 }
418 SmallVector<Constant *, 16> res;
419 Type *Ty = IntegerType::get(V->getContext(), 32);
420 for (unsigned i = 0,
421 e = cast<FixedVectorType>(V->getType())->getNumElements();
422 i != e; ++i) {
423 Constant *C =
424 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
425 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
426 }
427 return ConstantVector::get(res);
428 }
429
430 // We actually have to do a cast now. Perform the cast according to the
431 // opcode specified.
432 switch (opc) {
433 default:
434 llvm_unreachable("Failed to cast constant expression");
435 case Instruction::FPTrunc:
436 case Instruction::FPExt:
437 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
438 bool ignored;
439 APFloat Val = FPC->getValueAPF();
440 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
441 DestTy->isFloatTy() ? APFloat::IEEEsingle() :
442 DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
443 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
444 DestTy->isFP128Ty() ? APFloat::IEEEquad() :
445 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
446 APFloat::Bogus(),
447 APFloat::rmNearestTiesToEven, &ignored);
448 return ConstantFP::get(V->getContext(), Val);
449 }
450 return nullptr; // Can't fold.
451 case Instruction::FPToUI:
452 case Instruction::FPToSI:
453 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
454 const APFloat &V = FPC->getValueAPF();
455 bool ignored;
456 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
457 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
458 if (APFloat::opInvalidOp ==
459 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
460 // Undefined behavior invoked - the destination type can't represent
461 // the input constant.
462 return PoisonValue::get(DestTy);
463 }
464 return ConstantInt::get(FPC->getContext(), IntVal);
465 }
466 return nullptr; // Can't fold.
467 case Instruction::IntToPtr: //always treated as unsigned
468 if (V->isNullValue()) // Is it an integral null value?
469 return ConstantPointerNull::get(cast<PointerType>(DestTy));
470 return nullptr; // Other pointer types cannot be casted
471 case Instruction::PtrToInt: // always treated as unsigned
472 // Is it a null pointer value?
473 if (V->isNullValue())
474 return ConstantInt::get(DestTy, 0);
475 // Other pointer types cannot be casted
476 return nullptr;
477 case Instruction::UIToFP:
478 case Instruction::SIToFP:
479 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
480 const APInt &api = CI->getValue();
481 APFloat apf(DestTy->getFltSemantics(),
482 APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
483 apf.convertFromAPInt(api, opc==Instruction::SIToFP,
484 APFloat::rmNearestTiesToEven);
485 return ConstantFP::get(V->getContext(), apf);
486 }
487 return nullptr;
488 case Instruction::ZExt:
489 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
490 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
491 return ConstantInt::get(V->getContext(),
492 CI->getValue().zext(BitWidth));
493 }
494 return nullptr;
495 case Instruction::SExt:
496 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
497 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
498 return ConstantInt::get(V->getContext(),
499 CI->getValue().sext(BitWidth));
500 }
501 return nullptr;
502 case Instruction::Trunc: {
503 if (V->getType()->isVectorTy())
504 return nullptr;
505
506 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
507 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
508 return ConstantInt::get(V->getContext(),
509 CI->getValue().trunc(DestBitWidth));
510 }
511
512 // The input must be a constantexpr. See if we can simplify this based on
513 // the bytes we are demanding. Only do this if the source and dest are an
514 // even multiple of a byte.
515 if ((DestBitWidth & 7) == 0 &&
516 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
517 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
518 return Res;
519
520 return nullptr;
521 }
522 case Instruction::BitCast:
523 return FoldBitCast(V, DestTy);
524 case Instruction::AddrSpaceCast:
525 return nullptr;
526 }
527 }
528
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)529 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
530 Constant *V1, Constant *V2) {
531 // Check for i1 and vector true/false conditions.
532 if (Cond->isNullValue()) return V2;
533 if (Cond->isAllOnesValue()) return V1;
534
535 // If the condition is a vector constant, fold the result elementwise.
536 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
537 auto *V1VTy = CondV->getType();
538 SmallVector<Constant*, 16> Result;
539 Type *Ty = IntegerType::get(CondV->getContext(), 32);
540 for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) {
541 Constant *V;
542 Constant *V1Element = ConstantExpr::getExtractElement(V1,
543 ConstantInt::get(Ty, i));
544 Constant *V2Element = ConstantExpr::getExtractElement(V2,
545 ConstantInt::get(Ty, i));
546 auto *Cond = cast<Constant>(CondV->getOperand(i));
547 if (isa<PoisonValue>(Cond)) {
548 V = PoisonValue::get(V1Element->getType());
549 } else if (V1Element == V2Element) {
550 V = V1Element;
551 } else if (isa<UndefValue>(Cond)) {
552 V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
553 } else {
554 if (!isa<ConstantInt>(Cond)) break;
555 V = Cond->isNullValue() ? V2Element : V1Element;
556 }
557 Result.push_back(V);
558 }
559
560 // If we were able to build the vector, return it.
561 if (Result.size() == V1VTy->getNumElements())
562 return ConstantVector::get(Result);
563 }
564
565 if (isa<PoisonValue>(Cond))
566 return PoisonValue::get(V1->getType());
567
568 if (isa<UndefValue>(Cond)) {
569 if (isa<UndefValue>(V1)) return V1;
570 return V2;
571 }
572
573 if (V1 == V2) return V1;
574
575 if (isa<PoisonValue>(V1))
576 return V2;
577 if (isa<PoisonValue>(V2))
578 return V1;
579
580 // If the true or false value is undef, we can fold to the other value as
581 // long as the other value isn't poison.
582 auto NotPoison = [](Constant *C) {
583 if (isa<PoisonValue>(C))
584 return false;
585
586 // TODO: We can analyze ConstExpr by opcode to determine if there is any
587 // possibility of poison.
588 if (isa<ConstantExpr>(C))
589 return false;
590
591 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(C) ||
592 isa<ConstantPointerNull>(C) || isa<Function>(C))
593 return true;
594
595 if (C->getType()->isVectorTy())
596 return !C->containsPoisonElement() && !C->containsConstantExpression();
597
598 // TODO: Recursively analyze aggregates or other constants.
599 return false;
600 };
601 if (isa<UndefValue>(V1) && NotPoison(V2)) return V2;
602 if (isa<UndefValue>(V2) && NotPoison(V1)) return V1;
603
604 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
605 if (TrueVal->getOpcode() == Instruction::Select)
606 if (TrueVal->getOperand(0) == Cond)
607 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
608 }
609 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
610 if (FalseVal->getOpcode() == Instruction::Select)
611 if (FalseVal->getOperand(0) == Cond)
612 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
613 }
614
615 return nullptr;
616 }
617
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)618 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
619 Constant *Idx) {
620 auto *ValVTy = cast<VectorType>(Val->getType());
621
622 // extractelt poison, C -> poison
623 // extractelt C, undef -> poison
624 if (isa<PoisonValue>(Val) || isa<UndefValue>(Idx))
625 return PoisonValue::get(ValVTy->getElementType());
626
627 // extractelt undef, C -> undef
628 if (isa<UndefValue>(Val))
629 return UndefValue::get(ValVTy->getElementType());
630
631 auto *CIdx = dyn_cast<ConstantInt>(Idx);
632 if (!CIdx)
633 return nullptr;
634
635 if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val->getType())) {
636 // ee({w,x,y,z}, wrong_value) -> poison
637 if (CIdx->uge(ValFVTy->getNumElements()))
638 return PoisonValue::get(ValFVTy->getElementType());
639 }
640
641 // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...)
642 if (auto *CE = dyn_cast<ConstantExpr>(Val)) {
643 if (auto *GEP = dyn_cast<GEPOperator>(CE)) {
644 SmallVector<Constant *, 8> Ops;
645 Ops.reserve(CE->getNumOperands());
646 for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) {
647 Constant *Op = CE->getOperand(i);
648 if (Op->getType()->isVectorTy()) {
649 Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx);
650 if (!ScalarOp)
651 return nullptr;
652 Ops.push_back(ScalarOp);
653 } else
654 Ops.push_back(Op);
655 }
656 return CE->getWithOperands(Ops, ValVTy->getElementType(), false,
657 GEP->getSourceElementType());
658 } else if (CE->getOpcode() == Instruction::InsertElement) {
659 if (const auto *IEIdx = dyn_cast<ConstantInt>(CE->getOperand(2))) {
660 if (APSInt::isSameValue(APSInt(IEIdx->getValue()),
661 APSInt(CIdx->getValue()))) {
662 return CE->getOperand(1);
663 } else {
664 return ConstantExpr::getExtractElement(CE->getOperand(0), CIdx);
665 }
666 }
667 }
668 }
669
670 // Lane < Splat minimum vector width => extractelt Splat(x), Lane -> x
671 if (CIdx->getValue().ult(ValVTy->getElementCount().getKnownMinValue())) {
672 if (Constant *SplatVal = Val->getSplatValue())
673 return SplatVal;
674 }
675
676 return Val->getAggregateElement(CIdx);
677 }
678
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)679 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
680 Constant *Elt,
681 Constant *Idx) {
682 if (isa<UndefValue>(Idx))
683 return PoisonValue::get(Val->getType());
684
685 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
686 if (!CIdx) return nullptr;
687
688 // Do not iterate on scalable vector. The num of elements is unknown at
689 // compile-time.
690 if (isa<ScalableVectorType>(Val->getType()))
691 return nullptr;
692
693 auto *ValTy = cast<FixedVectorType>(Val->getType());
694
695 unsigned NumElts = ValTy->getNumElements();
696 if (CIdx->uge(NumElts))
697 return PoisonValue::get(Val->getType());
698
699 SmallVector<Constant*, 16> Result;
700 Result.reserve(NumElts);
701 auto *Ty = Type::getInt32Ty(Val->getContext());
702 uint64_t IdxVal = CIdx->getZExtValue();
703 for (unsigned i = 0; i != NumElts; ++i) {
704 if (i == IdxVal) {
705 Result.push_back(Elt);
706 continue;
707 }
708
709 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
710 Result.push_back(C);
711 }
712
713 return ConstantVector::get(Result);
714 }
715
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,ArrayRef<int> Mask)716 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2,
717 ArrayRef<int> Mask) {
718 auto *V1VTy = cast<VectorType>(V1->getType());
719 unsigned MaskNumElts = Mask.size();
720 auto MaskEltCount =
721 ElementCount::get(MaskNumElts, isa<ScalableVectorType>(V1VTy));
722 Type *EltTy = V1VTy->getElementType();
723
724 // Undefined shuffle mask -> undefined value.
725 if (all_of(Mask, [](int Elt) { return Elt == UndefMaskElem; })) {
726 return UndefValue::get(FixedVectorType::get(EltTy, MaskNumElts));
727 }
728
729 // If the mask is all zeros this is a splat, no need to go through all
730 // elements.
731 if (all_of(Mask, [](int Elt) { return Elt == 0; }) &&
732 !MaskEltCount.isScalable()) {
733 Type *Ty = IntegerType::get(V1->getContext(), 32);
734 Constant *Elt =
735 ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, 0));
736 return ConstantVector::getSplat(MaskEltCount, Elt);
737 }
738 // Do not iterate on scalable vector. The num of elements is unknown at
739 // compile-time.
740 if (isa<ScalableVectorType>(V1VTy))
741 return nullptr;
742
743 unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue();
744
745 // Loop over the shuffle mask, evaluating each element.
746 SmallVector<Constant*, 32> Result;
747 for (unsigned i = 0; i != MaskNumElts; ++i) {
748 int Elt = Mask[i];
749 if (Elt == -1) {
750 Result.push_back(UndefValue::get(EltTy));
751 continue;
752 }
753 Constant *InElt;
754 if (unsigned(Elt) >= SrcNumElts*2)
755 InElt = UndefValue::get(EltTy);
756 else if (unsigned(Elt) >= SrcNumElts) {
757 Type *Ty = IntegerType::get(V2->getContext(), 32);
758 InElt =
759 ConstantExpr::getExtractElement(V2,
760 ConstantInt::get(Ty, Elt - SrcNumElts));
761 } else {
762 Type *Ty = IntegerType::get(V1->getContext(), 32);
763 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
764 }
765 Result.push_back(InElt);
766 }
767
768 return ConstantVector::get(Result);
769 }
770
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)771 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
772 ArrayRef<unsigned> Idxs) {
773 // Base case: no indices, so return the entire value.
774 if (Idxs.empty())
775 return Agg;
776
777 if (Constant *C = Agg->getAggregateElement(Idxs[0]))
778 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
779
780 return nullptr;
781 }
782
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)783 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
784 Constant *Val,
785 ArrayRef<unsigned> Idxs) {
786 // Base case: no indices, so replace the entire value.
787 if (Idxs.empty())
788 return Val;
789
790 unsigned NumElts;
791 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
792 NumElts = ST->getNumElements();
793 else
794 NumElts = cast<ArrayType>(Agg->getType())->getNumElements();
795
796 SmallVector<Constant*, 32> Result;
797 for (unsigned i = 0; i != NumElts; ++i) {
798 Constant *C = Agg->getAggregateElement(i);
799 if (!C) return nullptr;
800
801 if (Idxs[0] == i)
802 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
803
804 Result.push_back(C);
805 }
806
807 if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
808 return ConstantStruct::get(ST, Result);
809 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Result);
810 }
811
ConstantFoldUnaryInstruction(unsigned Opcode,Constant * C)812 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) {
813 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected");
814
815 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
816 // vectors are always evaluated per element.
817 bool IsScalableVector = isa<ScalableVectorType>(C->getType());
818 bool HasScalarUndefOrScalableVectorUndef =
819 (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C);
820
821 if (HasScalarUndefOrScalableVectorUndef) {
822 switch (static_cast<Instruction::UnaryOps>(Opcode)) {
823 case Instruction::FNeg:
824 return C; // -undef -> undef
825 case Instruction::UnaryOpsEnd:
826 llvm_unreachable("Invalid UnaryOp");
827 }
828 }
829
830 // Constant should not be UndefValue, unless these are vector constants.
831 assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue");
832 // We only have FP UnaryOps right now.
833 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp");
834
835 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
836 const APFloat &CV = CFP->getValueAPF();
837 switch (Opcode) {
838 default:
839 break;
840 case Instruction::FNeg:
841 return ConstantFP::get(C->getContext(), neg(CV));
842 }
843 } else if (auto *VTy = dyn_cast<FixedVectorType>(C->getType())) {
844
845 Type *Ty = IntegerType::get(VTy->getContext(), 32);
846 // Fast path for splatted constants.
847 if (Constant *Splat = C->getSplatValue()) {
848 Constant *Elt = ConstantExpr::get(Opcode, Splat);
849 return ConstantVector::getSplat(VTy->getElementCount(), Elt);
850 }
851
852 // Fold each element and create a vector constant from those constants.
853 SmallVector<Constant *, 16> Result;
854 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
855 Constant *ExtractIdx = ConstantInt::get(Ty, i);
856 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx);
857
858 Result.push_back(ConstantExpr::get(Opcode, Elt));
859 }
860
861 return ConstantVector::get(Result);
862 }
863
864 // We don't know how to fold this.
865 return nullptr;
866 }
867
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)868 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
869 Constant *C2) {
870 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
871
872 // Simplify BinOps with their identity values first. They are no-ops and we
873 // can always return the other value, including undef or poison values.
874 // FIXME: remove unnecessary duplicated identity patterns below.
875 // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops,
876 // like X << 0 = X.
877 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType());
878 if (Identity) {
879 if (C1 == Identity)
880 return C2;
881 if (C2 == Identity)
882 return C1;
883 }
884
885 // Binary operations propagate poison.
886 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2))
887 return PoisonValue::get(C1->getType());
888
889 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
890 // vectors are always evaluated per element.
891 bool IsScalableVector = isa<ScalableVectorType>(C1->getType());
892 bool HasScalarUndefOrScalableVectorUndef =
893 (!C1->getType()->isVectorTy() || IsScalableVector) &&
894 (isa<UndefValue>(C1) || isa<UndefValue>(C2));
895 if (HasScalarUndefOrScalableVectorUndef) {
896 switch (static_cast<Instruction::BinaryOps>(Opcode)) {
897 case Instruction::Xor:
898 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
899 // Handle undef ^ undef -> 0 special case. This is a common
900 // idiom (misuse).
901 return Constant::getNullValue(C1->getType());
902 LLVM_FALLTHROUGH;
903 case Instruction::Add:
904 case Instruction::Sub:
905 return UndefValue::get(C1->getType());
906 case Instruction::And:
907 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
908 return C1;
909 return Constant::getNullValue(C1->getType()); // undef & X -> 0
910 case Instruction::Mul: {
911 // undef * undef -> undef
912 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
913 return C1;
914 const APInt *CV;
915 // X * undef -> undef if X is odd
916 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
917 if ((*CV)[0])
918 return UndefValue::get(C1->getType());
919
920 // X * undef -> 0 otherwise
921 return Constant::getNullValue(C1->getType());
922 }
923 case Instruction::SDiv:
924 case Instruction::UDiv:
925 // X / undef -> poison
926 // X / 0 -> poison
927 if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
928 return PoisonValue::get(C2->getType());
929 // undef / 1 -> undef
930 if (match(C2, m_One()))
931 return C1;
932 // undef / X -> 0 otherwise
933 return Constant::getNullValue(C1->getType());
934 case Instruction::URem:
935 case Instruction::SRem:
936 // X % undef -> poison
937 // X % 0 -> poison
938 if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
939 return PoisonValue::get(C2->getType());
940 // undef % X -> 0 otherwise
941 return Constant::getNullValue(C1->getType());
942 case Instruction::Or: // X | undef -> -1
943 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
944 return C1;
945 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
946 case Instruction::LShr:
947 // X >>l undef -> poison
948 if (isa<UndefValue>(C2))
949 return PoisonValue::get(C2->getType());
950 // undef >>l 0 -> undef
951 if (match(C2, m_Zero()))
952 return C1;
953 // undef >>l X -> 0
954 return Constant::getNullValue(C1->getType());
955 case Instruction::AShr:
956 // X >>a undef -> poison
957 if (isa<UndefValue>(C2))
958 return PoisonValue::get(C2->getType());
959 // undef >>a 0 -> undef
960 if (match(C2, m_Zero()))
961 return C1;
962 // TODO: undef >>a X -> poison if the shift is exact
963 // undef >>a X -> 0
964 return Constant::getNullValue(C1->getType());
965 case Instruction::Shl:
966 // X << undef -> undef
967 if (isa<UndefValue>(C2))
968 return PoisonValue::get(C2->getType());
969 // undef << 0 -> undef
970 if (match(C2, m_Zero()))
971 return C1;
972 // undef << X -> 0
973 return Constant::getNullValue(C1->getType());
974 case Instruction::FSub:
975 // -0.0 - undef --> undef (consistent with "fneg undef")
976 if (match(C1, m_NegZeroFP()) && isa<UndefValue>(C2))
977 return C2;
978 LLVM_FALLTHROUGH;
979 case Instruction::FAdd:
980 case Instruction::FMul:
981 case Instruction::FDiv:
982 case Instruction::FRem:
983 // [any flop] undef, undef -> undef
984 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
985 return C1;
986 // [any flop] C, undef -> NaN
987 // [any flop] undef, C -> NaN
988 // We could potentially specialize NaN/Inf constants vs. 'normal'
989 // constants (possibly differently depending on opcode and operand). This
990 // would allow returning undef sometimes. But it is always safe to fold to
991 // NaN because we can choose the undef operand as NaN, and any FP opcode
992 // with a NaN operand will propagate NaN.
993 return ConstantFP::getNaN(C1->getType());
994 case Instruction::BinaryOpsEnd:
995 llvm_unreachable("Invalid BinaryOp");
996 }
997 }
998
999 // Neither constant should be UndefValue, unless these are vector constants.
1000 assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue");
1001
1002 // Handle simplifications when the RHS is a constant int.
1003 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1004 switch (Opcode) {
1005 case Instruction::Add:
1006 if (CI2->isZero()) return C1; // X + 0 == X
1007 break;
1008 case Instruction::Sub:
1009 if (CI2->isZero()) return C1; // X - 0 == X
1010 break;
1011 case Instruction::Mul:
1012 if (CI2->isZero()) return C2; // X * 0 == 0
1013 if (CI2->isOne())
1014 return C1; // X * 1 == X
1015 break;
1016 case Instruction::UDiv:
1017 case Instruction::SDiv:
1018 if (CI2->isOne())
1019 return C1; // X / 1 == X
1020 if (CI2->isZero())
1021 return PoisonValue::get(CI2->getType()); // X / 0 == poison
1022 break;
1023 case Instruction::URem:
1024 case Instruction::SRem:
1025 if (CI2->isOne())
1026 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1027 if (CI2->isZero())
1028 return PoisonValue::get(CI2->getType()); // X % 0 == poison
1029 break;
1030 case Instruction::And:
1031 if (CI2->isZero()) return C2; // X & 0 == 0
1032 if (CI2->isMinusOne())
1033 return C1; // X & -1 == X
1034
1035 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1036 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1037 if (CE1->getOpcode() == Instruction::ZExt) {
1038 unsigned DstWidth = CI2->getType()->getBitWidth();
1039 unsigned SrcWidth =
1040 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1041 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1042 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1043 return C1;
1044 }
1045
1046 // If and'ing the address of a global with a constant, fold it.
1047 if (CE1->getOpcode() == Instruction::PtrToInt &&
1048 isa<GlobalValue>(CE1->getOperand(0))) {
1049 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1050
1051 MaybeAlign GVAlign;
1052
1053 if (Module *TheModule = GV->getParent()) {
1054 const DataLayout &DL = TheModule->getDataLayout();
1055 GVAlign = GV->getPointerAlignment(DL);
1056
1057 // If the function alignment is not specified then assume that it
1058 // is 4.
1059 // This is dangerous; on x86, the alignment of the pointer
1060 // corresponds to the alignment of the function, but might be less
1061 // than 4 if it isn't explicitly specified.
1062 // However, a fix for this behaviour was reverted because it
1063 // increased code size (see https://reviews.llvm.org/D55115)
1064 // FIXME: This code should be deleted once existing targets have
1065 // appropriate defaults
1066 if (isa<Function>(GV) && !DL.getFunctionPtrAlign())
1067 GVAlign = Align(4);
1068 } else if (isa<Function>(GV)) {
1069 // Without a datalayout we have to assume the worst case: that the
1070 // function pointer isn't aligned at all.
1071 GVAlign = llvm::None;
1072 } else if (isa<GlobalVariable>(GV)) {
1073 GVAlign = cast<GlobalVariable>(GV)->getAlign();
1074 }
1075
1076 if (GVAlign && *GVAlign > 1) {
1077 unsigned DstWidth = CI2->getType()->getBitWidth();
1078 unsigned SrcWidth = std::min(DstWidth, Log2(*GVAlign));
1079 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1080
1081 // If checking bits we know are clear, return zero.
1082 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1083 return Constant::getNullValue(CI2->getType());
1084 }
1085 }
1086 }
1087 break;
1088 case Instruction::Or:
1089 if (CI2->isZero()) return C1; // X | 0 == X
1090 if (CI2->isMinusOne())
1091 return C2; // X | -1 == -1
1092 break;
1093 case Instruction::Xor:
1094 if (CI2->isZero()) return C1; // X ^ 0 == X
1095
1096 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1097 switch (CE1->getOpcode()) {
1098 default: break;
1099 case Instruction::ICmp:
1100 case Instruction::FCmp:
1101 // cmp pred ^ true -> cmp !pred
1102 assert(CI2->isOne());
1103 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1104 pred = CmpInst::getInversePredicate(pred);
1105 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1106 CE1->getOperand(1));
1107 }
1108 }
1109 break;
1110 case Instruction::AShr:
1111 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1112 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1113 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1114 return ConstantExpr::getLShr(C1, C2);
1115 break;
1116 }
1117 } else if (isa<ConstantInt>(C1)) {
1118 // If C1 is a ConstantInt and C2 is not, swap the operands.
1119 if (Instruction::isCommutative(Opcode))
1120 return ConstantExpr::get(Opcode, C2, C1);
1121 }
1122
1123 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1124 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1125 const APInt &C1V = CI1->getValue();
1126 const APInt &C2V = CI2->getValue();
1127 switch (Opcode) {
1128 default:
1129 break;
1130 case Instruction::Add:
1131 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1132 case Instruction::Sub:
1133 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1134 case Instruction::Mul:
1135 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1136 case Instruction::UDiv:
1137 assert(!CI2->isZero() && "Div by zero handled above");
1138 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1139 case Instruction::SDiv:
1140 assert(!CI2->isZero() && "Div by zero handled above");
1141 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1142 return PoisonValue::get(CI1->getType()); // MIN_INT / -1 -> poison
1143 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1144 case Instruction::URem:
1145 assert(!CI2->isZero() && "Div by zero handled above");
1146 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1147 case Instruction::SRem:
1148 assert(!CI2->isZero() && "Div by zero handled above");
1149 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1150 return PoisonValue::get(CI1->getType()); // MIN_INT % -1 -> poison
1151 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1152 case Instruction::And:
1153 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1154 case Instruction::Or:
1155 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1156 case Instruction::Xor:
1157 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1158 case Instruction::Shl:
1159 if (C2V.ult(C1V.getBitWidth()))
1160 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1161 return PoisonValue::get(C1->getType()); // too big shift is poison
1162 case Instruction::LShr:
1163 if (C2V.ult(C1V.getBitWidth()))
1164 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1165 return PoisonValue::get(C1->getType()); // too big shift is poison
1166 case Instruction::AShr:
1167 if (C2V.ult(C1V.getBitWidth()))
1168 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1169 return PoisonValue::get(C1->getType()); // too big shift is poison
1170 }
1171 }
1172
1173 switch (Opcode) {
1174 case Instruction::SDiv:
1175 case Instruction::UDiv:
1176 case Instruction::URem:
1177 case Instruction::SRem:
1178 case Instruction::LShr:
1179 case Instruction::AShr:
1180 case Instruction::Shl:
1181 if (CI1->isZero()) return C1;
1182 break;
1183 default:
1184 break;
1185 }
1186 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1187 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1188 const APFloat &C1V = CFP1->getValueAPF();
1189 const APFloat &C2V = CFP2->getValueAPF();
1190 APFloat C3V = C1V; // copy for modification
1191 switch (Opcode) {
1192 default:
1193 break;
1194 case Instruction::FAdd:
1195 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1196 return ConstantFP::get(C1->getContext(), C3V);
1197 case Instruction::FSub:
1198 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1199 return ConstantFP::get(C1->getContext(), C3V);
1200 case Instruction::FMul:
1201 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1202 return ConstantFP::get(C1->getContext(), C3V);
1203 case Instruction::FDiv:
1204 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1205 return ConstantFP::get(C1->getContext(), C3V);
1206 case Instruction::FRem:
1207 (void)C3V.mod(C2V);
1208 return ConstantFP::get(C1->getContext(), C3V);
1209 }
1210 }
1211 } else if (auto *VTy = dyn_cast<VectorType>(C1->getType())) {
1212 // Fast path for splatted constants.
1213 if (Constant *C2Splat = C2->getSplatValue()) {
1214 if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue())
1215 return PoisonValue::get(VTy);
1216 if (Constant *C1Splat = C1->getSplatValue()) {
1217 return ConstantVector::getSplat(
1218 VTy->getElementCount(),
1219 ConstantExpr::get(Opcode, C1Splat, C2Splat));
1220 }
1221 }
1222
1223 if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
1224 // Fold each element and create a vector constant from those constants.
1225 SmallVector<Constant*, 16> Result;
1226 Type *Ty = IntegerType::get(FVTy->getContext(), 32);
1227 for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) {
1228 Constant *ExtractIdx = ConstantInt::get(Ty, i);
1229 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1230 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1231
1232 // If any element of a divisor vector is zero, the whole op is poison.
1233 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1234 return PoisonValue::get(VTy);
1235
1236 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1237 }
1238
1239 return ConstantVector::get(Result);
1240 }
1241 }
1242
1243 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1244 // There are many possible foldings we could do here. We should probably
1245 // at least fold add of a pointer with an integer into the appropriate
1246 // getelementptr. This will improve alias analysis a bit.
1247
1248 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1249 // (a + (b + c)).
1250 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1251 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1252 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1253 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1254 }
1255 } else if (isa<ConstantExpr>(C2)) {
1256 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1257 // other way if possible.
1258 if (Instruction::isCommutative(Opcode))
1259 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1260 }
1261
1262 // i1 can be simplified in many cases.
1263 if (C1->getType()->isIntegerTy(1)) {
1264 switch (Opcode) {
1265 case Instruction::Add:
1266 case Instruction::Sub:
1267 return ConstantExpr::getXor(C1, C2);
1268 case Instruction::Mul:
1269 return ConstantExpr::getAnd(C1, C2);
1270 case Instruction::Shl:
1271 case Instruction::LShr:
1272 case Instruction::AShr:
1273 // We can assume that C2 == 0. If it were one the result would be
1274 // undefined because the shift value is as large as the bitwidth.
1275 return C1;
1276 case Instruction::SDiv:
1277 case Instruction::UDiv:
1278 // We can assume that C2 == 1. If it were zero the result would be
1279 // undefined through division by zero.
1280 return C1;
1281 case Instruction::URem:
1282 case Instruction::SRem:
1283 // We can assume that C2 == 1. If it were zero the result would be
1284 // undefined through division by zero.
1285 return ConstantInt::getFalse(C1->getContext());
1286 default:
1287 break;
1288 }
1289 }
1290
1291 // We don't know how to fold this.
1292 return nullptr;
1293 }
1294
1295 /// This type is zero-sized if it's an array or structure of zero-sized types.
1296 /// The only leaf zero-sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1297 static bool isMaybeZeroSizedType(Type *Ty) {
1298 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1299 if (STy->isOpaque()) return true; // Can't say.
1300
1301 // If all of elements have zero size, this does too.
1302 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1303 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1304 return true;
1305
1306 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1307 return isMaybeZeroSizedType(ATy->getElementType());
1308 }
1309 return false;
1310 }
1311
1312 /// Compare the two constants as though they were getelementptr indices.
1313 /// This allows coercion of the types to be the same thing.
1314 ///
1315 /// If the two constants are the "same" (after coercion), return 0. If the
1316 /// first is less than the second, return -1, if the second is less than the
1317 /// first, return 1. If the constants are not integral, return -2.
1318 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1319 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1320 if (C1 == C2) return 0;
1321
1322 // Ok, we found a different index. If they are not ConstantInt, we can't do
1323 // anything with them.
1324 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1325 return -2; // don't know!
1326
1327 // We cannot compare the indices if they don't fit in an int64_t.
1328 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1329 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1330 return -2; // don't know!
1331
1332 // Ok, we have two differing integer indices. Sign extend them to be the same
1333 // type.
1334 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1335 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1336
1337 if (C1Val == C2Val) return 0; // They are equal
1338
1339 // If the type being indexed over is really just a zero sized type, there is
1340 // no pointer difference being made here.
1341 if (isMaybeZeroSizedType(ElTy))
1342 return -2; // dunno.
1343
1344 // If they are really different, now that they are the same type, then we
1345 // found a difference!
1346 if (C1Val < C2Val)
1347 return -1;
1348 else
1349 return 1;
1350 }
1351
1352 /// This function determines if there is anything we can decide about the two
1353 /// constants provided. This doesn't need to handle simple things like
1354 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1355 /// If we can determine that the two constants have a particular relation to
1356 /// each other, we should return the corresponding FCmpInst predicate,
1357 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1358 /// ConstantFoldCompareInstruction.
1359 ///
1360 /// To simplify this code we canonicalize the relation so that the first
1361 /// operand is always the most "complex" of the two. We consider ConstantFP
1362 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1363 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1364 assert(V1->getType() == V2->getType() &&
1365 "Cannot compare values of different types!");
1366
1367 // We do not know if a constant expression will evaluate to a number or NaN.
1368 // Therefore, we can only say that the relation is unordered or equal.
1369 if (V1 == V2) return FCmpInst::FCMP_UEQ;
1370
1371 if (!isa<ConstantExpr>(V1)) {
1372 if (!isa<ConstantExpr>(V2)) {
1373 // Simple case, use the standard constant folder.
1374 ConstantInt *R = nullptr;
1375 R = dyn_cast<ConstantInt>(
1376 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1377 if (R && !R->isZero())
1378 return FCmpInst::FCMP_OEQ;
1379 R = dyn_cast<ConstantInt>(
1380 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1381 if (R && !R->isZero())
1382 return FCmpInst::FCMP_OLT;
1383 R = dyn_cast<ConstantInt>(
1384 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1385 if (R && !R->isZero())
1386 return FCmpInst::FCMP_OGT;
1387
1388 // Nothing more we can do
1389 return FCmpInst::BAD_FCMP_PREDICATE;
1390 }
1391
1392 // If the first operand is simple and second is ConstantExpr, swap operands.
1393 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1394 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1395 return FCmpInst::getSwappedPredicate(SwappedRelation);
1396 } else {
1397 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1398 // constantexpr or a simple constant.
1399 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1400 switch (CE1->getOpcode()) {
1401 case Instruction::FPTrunc:
1402 case Instruction::FPExt:
1403 case Instruction::UIToFP:
1404 case Instruction::SIToFP:
1405 // We might be able to do something with these but we don't right now.
1406 break;
1407 default:
1408 break;
1409 }
1410 }
1411 // There are MANY other foldings that we could perform here. They will
1412 // probably be added on demand, as they seem needed.
1413 return FCmpInst::BAD_FCMP_PREDICATE;
1414 }
1415
areGlobalsPotentiallyEqual(const GlobalValue * GV1,const GlobalValue * GV2)1416 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1417 const GlobalValue *GV2) {
1418 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1419 if (GV->isInterposable() || GV->hasGlobalUnnamedAddr())
1420 return true;
1421 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1422 Type *Ty = GVar->getValueType();
1423 // A global with opaque type might end up being zero sized.
1424 if (!Ty->isSized())
1425 return true;
1426 // A global with an empty type might lie at the address of any other
1427 // global.
1428 if (Ty->isEmptyTy())
1429 return true;
1430 }
1431 return false;
1432 };
1433 // Don't try to decide equality of aliases.
1434 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1435 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1436 return ICmpInst::ICMP_NE;
1437 return ICmpInst::BAD_ICMP_PREDICATE;
1438 }
1439
1440 /// This function determines if there is anything we can decide about the two
1441 /// constants provided. This doesn't need to handle simple things like integer
1442 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1443 /// If we can determine that the two constants have a particular relation to
1444 /// each other, we should return the corresponding ICmp predicate, otherwise
1445 /// return ICmpInst::BAD_ICMP_PREDICATE.
1446 ///
1447 /// To simplify this code we canonicalize the relation so that the first
1448 /// operand is always the most "complex" of the two. We consider simple
1449 /// constants (like ConstantInt) to be the simplest, followed by
1450 /// GlobalValues, followed by ConstantExpr's (the most complex).
1451 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1452 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1453 bool isSigned) {
1454 assert(V1->getType() == V2->getType() &&
1455 "Cannot compare different types of values!");
1456 if (V1 == V2) return ICmpInst::ICMP_EQ;
1457
1458 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1459 !isa<BlockAddress>(V1)) {
1460 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1461 !isa<BlockAddress>(V2)) {
1462 // We distilled this down to a simple case, use the standard constant
1463 // folder.
1464 ConstantInt *R = nullptr;
1465 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1466 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1467 if (R && !R->isZero())
1468 return pred;
1469 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1470 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1471 if (R && !R->isZero())
1472 return pred;
1473 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1474 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1475 if (R && !R->isZero())
1476 return pred;
1477
1478 // If we couldn't figure it out, bail.
1479 return ICmpInst::BAD_ICMP_PREDICATE;
1480 }
1481
1482 // If the first operand is simple, swap operands.
1483 ICmpInst::Predicate SwappedRelation =
1484 evaluateICmpRelation(V2, V1, isSigned);
1485 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1486 return ICmpInst::getSwappedPredicate(SwappedRelation);
1487
1488 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1489 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1490 ICmpInst::Predicate SwappedRelation =
1491 evaluateICmpRelation(V2, V1, isSigned);
1492 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1493 return ICmpInst::getSwappedPredicate(SwappedRelation);
1494 return ICmpInst::BAD_ICMP_PREDICATE;
1495 }
1496
1497 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1498 // constant (which, since the types must match, means that it's a
1499 // ConstantPointerNull).
1500 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1501 return areGlobalsPotentiallyEqual(GV, GV2);
1502 } else if (isa<BlockAddress>(V2)) {
1503 return ICmpInst::ICMP_NE; // Globals never equal labels.
1504 } else {
1505 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1506 // GlobalVals can never be null unless they have external weak linkage.
1507 // We don't try to evaluate aliases here.
1508 // NOTE: We should not be doing this constant folding if null pointer
1509 // is considered valid for the function. But currently there is no way to
1510 // query it from the Constant type.
1511 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1512 !NullPointerIsDefined(nullptr /* F */,
1513 GV->getType()->getAddressSpace()))
1514 return ICmpInst::ICMP_UGT;
1515 }
1516 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1517 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1518 ICmpInst::Predicate SwappedRelation =
1519 evaluateICmpRelation(V2, V1, isSigned);
1520 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1521 return ICmpInst::getSwappedPredicate(SwappedRelation);
1522 return ICmpInst::BAD_ICMP_PREDICATE;
1523 }
1524
1525 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1526 // constant (which, since the types must match, means that it is a
1527 // ConstantPointerNull).
1528 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1529 // Block address in another function can't equal this one, but block
1530 // addresses in the current function might be the same if blocks are
1531 // empty.
1532 if (BA2->getFunction() != BA->getFunction())
1533 return ICmpInst::ICMP_NE;
1534 } else {
1535 // Block addresses aren't null, don't equal the address of globals.
1536 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1537 "Canonicalization guarantee!");
1538 return ICmpInst::ICMP_NE;
1539 }
1540 } else {
1541 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1542 // constantexpr, a global, block address, or a simple constant.
1543 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1544 Constant *CE1Op0 = CE1->getOperand(0);
1545
1546 switch (CE1->getOpcode()) {
1547 case Instruction::Trunc:
1548 case Instruction::FPTrunc:
1549 case Instruction::FPExt:
1550 case Instruction::FPToUI:
1551 case Instruction::FPToSI:
1552 break; // We can't evaluate floating point casts or truncations.
1553
1554 case Instruction::BitCast:
1555 // If this is a global value cast, check to see if the RHS is also a
1556 // GlobalValue.
1557 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0))
1558 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2))
1559 return areGlobalsPotentiallyEqual(GV, GV2);
1560 LLVM_FALLTHROUGH;
1561 case Instruction::UIToFP:
1562 case Instruction::SIToFP:
1563 case Instruction::ZExt:
1564 case Instruction::SExt:
1565 // We can't evaluate floating point casts or truncations.
1566 if (CE1Op0->getType()->isFPOrFPVectorTy())
1567 break;
1568
1569 // If the cast is not actually changing bits, and the second operand is a
1570 // null pointer, do the comparison with the pre-casted value.
1571 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1572 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1573 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1574 return evaluateICmpRelation(CE1Op0,
1575 Constant::getNullValue(CE1Op0->getType()),
1576 isSigned);
1577 }
1578 break;
1579
1580 case Instruction::GetElementPtr: {
1581 GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1582 // Ok, since this is a getelementptr, we know that the constant has a
1583 // pointer type. Check the various cases.
1584 if (isa<ConstantPointerNull>(V2)) {
1585 // If we are comparing a GEP to a null pointer, check to see if the base
1586 // of the GEP equals the null pointer.
1587 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1588 // If its not weak linkage, the GVal must have a non-zero address
1589 // so the result is greater-than
1590 if (!GV->hasExternalWeakLinkage())
1591 return ICmpInst::ICMP_UGT;
1592 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1593 // If we are indexing from a null pointer, check to see if we have any
1594 // non-zero indices.
1595 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1596 if (!CE1->getOperand(i)->isNullValue())
1597 // Offsetting from null, must not be equal.
1598 return ICmpInst::ICMP_UGT;
1599 // Only zero indexes from null, must still be zero.
1600 return ICmpInst::ICMP_EQ;
1601 }
1602 // Otherwise, we can't really say if the first operand is null or not.
1603 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1604 if (isa<ConstantPointerNull>(CE1Op0)) {
1605 // If its not weak linkage, the GVal must have a non-zero address
1606 // so the result is less-than
1607 if (!GV2->hasExternalWeakLinkage())
1608 return ICmpInst::ICMP_ULT;
1609 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1610 if (GV == GV2) {
1611 // If this is a getelementptr of the same global, then it must be
1612 // different. Because the types must match, the getelementptr could
1613 // only have at most one index, and because we fold getelementptr's
1614 // with a single zero index, it must be nonzero.
1615 assert(CE1->getNumOperands() == 2 &&
1616 !CE1->getOperand(1)->isNullValue() &&
1617 "Surprising getelementptr!");
1618 return ICmpInst::ICMP_UGT;
1619 } else {
1620 if (CE1GEP->hasAllZeroIndices())
1621 return areGlobalsPotentiallyEqual(GV, GV2);
1622 return ICmpInst::BAD_ICMP_PREDICATE;
1623 }
1624 }
1625 } else {
1626 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1627 Constant *CE2Op0 = CE2->getOperand(0);
1628
1629 // There are MANY other foldings that we could perform here. They will
1630 // probably be added on demand, as they seem needed.
1631 switch (CE2->getOpcode()) {
1632 default: break;
1633 case Instruction::GetElementPtr:
1634 // By far the most common case to handle is when the base pointers are
1635 // obviously to the same global.
1636 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1637 // Don't know relative ordering, but check for inequality.
1638 if (CE1Op0 != CE2Op0) {
1639 GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1640 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1641 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1642 cast<GlobalValue>(CE2Op0));
1643 return ICmpInst::BAD_ICMP_PREDICATE;
1644 }
1645 // Ok, we know that both getelementptr instructions are based on the
1646 // same global. From this, we can precisely determine the relative
1647 // ordering of the resultant pointers.
1648 unsigned i = 1;
1649
1650 // The logic below assumes that the result of the comparison
1651 // can be determined by finding the first index that differs.
1652 // This doesn't work if there is over-indexing in any
1653 // subsequent indices, so check for that case first.
1654 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1655 !CE2->isGEPWithNoNotionalOverIndexing())
1656 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1657
1658 // Compare all of the operands the GEP's have in common.
1659 gep_type_iterator GTI = gep_type_begin(CE1);
1660 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1661 ++i, ++GTI)
1662 switch (IdxCompare(CE1->getOperand(i),
1663 CE2->getOperand(i), GTI.getIndexedType())) {
1664 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1665 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1666 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1667 }
1668
1669 // Ok, we ran out of things they have in common. If any leftovers
1670 // are non-zero then we have a difference, otherwise we are equal.
1671 for (; i < CE1->getNumOperands(); ++i)
1672 if (!CE1->getOperand(i)->isNullValue()) {
1673 if (isa<ConstantInt>(CE1->getOperand(i)))
1674 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1675 else
1676 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1677 }
1678
1679 for (; i < CE2->getNumOperands(); ++i)
1680 if (!CE2->getOperand(i)->isNullValue()) {
1681 if (isa<ConstantInt>(CE2->getOperand(i)))
1682 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1683 else
1684 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1685 }
1686 return ICmpInst::ICMP_EQ;
1687 }
1688 }
1689 }
1690 break;
1691 }
1692 default:
1693 break;
1694 }
1695 }
1696
1697 return ICmpInst::BAD_ICMP_PREDICATE;
1698 }
1699
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1700 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1701 Constant *C1, Constant *C2) {
1702 Type *ResultTy;
1703 if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1704 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1705 VT->getElementCount());
1706 else
1707 ResultTy = Type::getInt1Ty(C1->getContext());
1708
1709 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1710 if (pred == FCmpInst::FCMP_FALSE)
1711 return Constant::getNullValue(ResultTy);
1712
1713 if (pred == FCmpInst::FCMP_TRUE)
1714 return Constant::getAllOnesValue(ResultTy);
1715
1716 // Handle some degenerate cases first
1717 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2))
1718 return PoisonValue::get(ResultTy);
1719
1720 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1721 CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1722 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1723 // For EQ and NE, we can always pick a value for the undef to make the
1724 // predicate pass or fail, so we can return undef.
1725 // Also, if both operands are undef, we can return undef for int comparison.
1726 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1727 return UndefValue::get(ResultTy);
1728
1729 // Otherwise, for integer compare, pick the same value as the non-undef
1730 // operand, and fold it to true or false.
1731 if (isIntegerPredicate)
1732 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1733
1734 // Choosing NaN for the undef will always make unordered comparison succeed
1735 // and ordered comparison fails.
1736 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1737 }
1738
1739 // icmp eq/ne(null,GV) -> false/true
1740 if (C1->isNullValue()) {
1741 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1742 // Don't try to evaluate aliases. External weak GV can be null.
1743 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1744 !NullPointerIsDefined(nullptr /* F */,
1745 GV->getType()->getAddressSpace())) {
1746 if (pred == ICmpInst::ICMP_EQ)
1747 return ConstantInt::getFalse(C1->getContext());
1748 else if (pred == ICmpInst::ICMP_NE)
1749 return ConstantInt::getTrue(C1->getContext());
1750 }
1751 // icmp eq/ne(GV,null) -> false/true
1752 } else if (C2->isNullValue()) {
1753 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) {
1754 // Don't try to evaluate aliases. External weak GV can be null.
1755 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1756 !NullPointerIsDefined(nullptr /* F */,
1757 GV->getType()->getAddressSpace())) {
1758 if (pred == ICmpInst::ICMP_EQ)
1759 return ConstantInt::getFalse(C1->getContext());
1760 else if (pred == ICmpInst::ICMP_NE)
1761 return ConstantInt::getTrue(C1->getContext());
1762 }
1763 }
1764
1765 // The caller is expected to commute the operands if the constant expression
1766 // is C2.
1767 // C1 >= 0 --> true
1768 if (pred == ICmpInst::ICMP_UGE)
1769 return Constant::getAllOnesValue(ResultTy);
1770 // C1 < 0 --> false
1771 if (pred == ICmpInst::ICMP_ULT)
1772 return Constant::getNullValue(ResultTy);
1773 }
1774
1775 // If the comparison is a comparison between two i1's, simplify it.
1776 if (C1->getType()->isIntegerTy(1)) {
1777 switch(pred) {
1778 case ICmpInst::ICMP_EQ:
1779 if (isa<ConstantInt>(C2))
1780 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1781 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1782 case ICmpInst::ICMP_NE:
1783 return ConstantExpr::getXor(C1, C2);
1784 default:
1785 break;
1786 }
1787 }
1788
1789 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1790 const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1791 const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1792 switch (pred) {
1793 default: llvm_unreachable("Invalid ICmp Predicate");
1794 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1795 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1796 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1797 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1798 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1799 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1800 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1801 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1802 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1803 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1804 }
1805 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1806 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1807 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1808 APFloat::cmpResult R = C1V.compare(C2V);
1809 switch (pred) {
1810 default: llvm_unreachable("Invalid FCmp Predicate");
1811 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1812 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1813 case FCmpInst::FCMP_UNO:
1814 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1815 case FCmpInst::FCMP_ORD:
1816 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1817 case FCmpInst::FCMP_UEQ:
1818 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1819 R==APFloat::cmpEqual);
1820 case FCmpInst::FCMP_OEQ:
1821 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1822 case FCmpInst::FCMP_UNE:
1823 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1824 case FCmpInst::FCMP_ONE:
1825 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1826 R==APFloat::cmpGreaterThan);
1827 case FCmpInst::FCMP_ULT:
1828 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1829 R==APFloat::cmpLessThan);
1830 case FCmpInst::FCMP_OLT:
1831 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1832 case FCmpInst::FCMP_UGT:
1833 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1834 R==APFloat::cmpGreaterThan);
1835 case FCmpInst::FCMP_OGT:
1836 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1837 case FCmpInst::FCMP_ULE:
1838 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1839 case FCmpInst::FCMP_OLE:
1840 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1841 R==APFloat::cmpEqual);
1842 case FCmpInst::FCMP_UGE:
1843 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1844 case FCmpInst::FCMP_OGE:
1845 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1846 R==APFloat::cmpEqual);
1847 }
1848 } else if (auto *C1VTy = dyn_cast<VectorType>(C1->getType())) {
1849
1850 // Fast path for splatted constants.
1851 if (Constant *C1Splat = C1->getSplatValue())
1852 if (Constant *C2Splat = C2->getSplatValue())
1853 return ConstantVector::getSplat(
1854 C1VTy->getElementCount(),
1855 ConstantExpr::getCompare(pred, C1Splat, C2Splat));
1856
1857 // Do not iterate on scalable vector. The number of elements is unknown at
1858 // compile-time.
1859 if (isa<ScalableVectorType>(C1VTy))
1860 return nullptr;
1861
1862 // If we can constant fold the comparison of each element, constant fold
1863 // the whole vector comparison.
1864 SmallVector<Constant*, 4> ResElts;
1865 Type *Ty = IntegerType::get(C1->getContext(), 32);
1866 // Compare the elements, producing an i1 result or constant expr.
1867 for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue();
1868 I != E; ++I) {
1869 Constant *C1E =
1870 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, I));
1871 Constant *C2E =
1872 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, I));
1873
1874 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1875 }
1876
1877 return ConstantVector::get(ResElts);
1878 }
1879
1880 if (C1->getType()->isFloatingPointTy() &&
1881 // Only call evaluateFCmpRelation if we have a constant expr to avoid
1882 // infinite recursive loop
1883 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1884 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1885 switch (evaluateFCmpRelation(C1, C2)) {
1886 default: llvm_unreachable("Unknown relation!");
1887 case FCmpInst::FCMP_UNO:
1888 case FCmpInst::FCMP_ORD:
1889 case FCmpInst::FCMP_UNE:
1890 case FCmpInst::FCMP_ULT:
1891 case FCmpInst::FCMP_UGT:
1892 case FCmpInst::FCMP_ULE:
1893 case FCmpInst::FCMP_UGE:
1894 case FCmpInst::FCMP_TRUE:
1895 case FCmpInst::FCMP_FALSE:
1896 case FCmpInst::BAD_FCMP_PREDICATE:
1897 break; // Couldn't determine anything about these constants.
1898 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1899 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1900 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1901 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1902 break;
1903 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1904 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1905 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1906 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1907 break;
1908 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1909 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1910 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1911 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1912 break;
1913 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1914 // We can only partially decide this relation.
1915 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1916 Result = 0;
1917 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1918 Result = 1;
1919 break;
1920 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1921 // We can only partially decide this relation.
1922 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1923 Result = 0;
1924 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1925 Result = 1;
1926 break;
1927 case FCmpInst::FCMP_ONE: // We know that C1 != C2
1928 // We can only partially decide this relation.
1929 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1930 Result = 0;
1931 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1932 Result = 1;
1933 break;
1934 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
1935 // We can only partially decide this relation.
1936 if (pred == FCmpInst::FCMP_ONE)
1937 Result = 0;
1938 else if (pred == FCmpInst::FCMP_UEQ)
1939 Result = 1;
1940 break;
1941 }
1942
1943 // If we evaluated the result, return it now.
1944 if (Result != -1)
1945 return ConstantInt::get(ResultTy, Result);
1946
1947 } else {
1948 // Evaluate the relation between the two constants, per the predicate.
1949 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1950 switch (evaluateICmpRelation(C1, C2,
1951 CmpInst::isSigned((CmpInst::Predicate)pred))) {
1952 default: llvm_unreachable("Unknown relational!");
1953 case ICmpInst::BAD_ICMP_PREDICATE:
1954 break; // Couldn't determine anything about these constants.
1955 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1956 // If we know the constants are equal, we can decide the result of this
1957 // computation precisely.
1958 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1959 break;
1960 case ICmpInst::ICMP_ULT:
1961 switch (pred) {
1962 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1963 Result = 1; break;
1964 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1965 Result = 0; break;
1966 }
1967 break;
1968 case ICmpInst::ICMP_SLT:
1969 switch (pred) {
1970 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1971 Result = 1; break;
1972 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1973 Result = 0; break;
1974 }
1975 break;
1976 case ICmpInst::ICMP_UGT:
1977 switch (pred) {
1978 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1979 Result = 1; break;
1980 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1981 Result = 0; break;
1982 }
1983 break;
1984 case ICmpInst::ICMP_SGT:
1985 switch (pred) {
1986 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1987 Result = 1; break;
1988 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1989 Result = 0; break;
1990 }
1991 break;
1992 case ICmpInst::ICMP_ULE:
1993 if (pred == ICmpInst::ICMP_UGT) Result = 0;
1994 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1995 break;
1996 case ICmpInst::ICMP_SLE:
1997 if (pred == ICmpInst::ICMP_SGT) Result = 0;
1998 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
1999 break;
2000 case ICmpInst::ICMP_UGE:
2001 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2002 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2003 break;
2004 case ICmpInst::ICMP_SGE:
2005 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2006 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2007 break;
2008 case ICmpInst::ICMP_NE:
2009 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2010 if (pred == ICmpInst::ICMP_NE) Result = 1;
2011 break;
2012 }
2013
2014 // If we evaluated the result, return it now.
2015 if (Result != -1)
2016 return ConstantInt::get(ResultTy, Result);
2017
2018 // If the right hand side is a bitcast, try using its inverse to simplify
2019 // it by moving it to the left hand side. We can't do this if it would turn
2020 // a vector compare into a scalar compare or visa versa, or if it would turn
2021 // the operands into FP values.
2022 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2023 Constant *CE2Op0 = CE2->getOperand(0);
2024 if (CE2->getOpcode() == Instruction::BitCast &&
2025 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
2026 !CE2Op0->getType()->isFPOrFPVectorTy()) {
2027 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2028 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2029 }
2030 }
2031
2032 // If the left hand side is an extension, try eliminating it.
2033 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2034 if ((CE1->getOpcode() == Instruction::SExt &&
2035 ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2036 (CE1->getOpcode() == Instruction::ZExt &&
2037 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2038 Constant *CE1Op0 = CE1->getOperand(0);
2039 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2040 if (CE1Inverse == CE1Op0) {
2041 // Check whether we can safely truncate the right hand side.
2042 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2043 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2044 C2->getType()) == C2)
2045 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2046 }
2047 }
2048 }
2049
2050 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2051 (C1->isNullValue() && !C2->isNullValue())) {
2052 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2053 // other way if possible.
2054 // Also, if C1 is null and C2 isn't, flip them around.
2055 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2056 return ConstantExpr::getICmp(pred, C2, C1);
2057 }
2058 }
2059 return nullptr;
2060 }
2061
2062 /// Test whether the given sequence of *normalized* indices is "inbounds".
2063 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)2064 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2065 // No indices means nothing that could be out of bounds.
2066 if (Idxs.empty()) return true;
2067
2068 // If the first index is zero, it's in bounds.
2069 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2070
2071 // If the first index is one and all the rest are zero, it's in bounds,
2072 // by the one-past-the-end rule.
2073 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2074 if (!CI->isOne())
2075 return false;
2076 } else {
2077 auto *CV = cast<ConstantDataVector>(Idxs[0]);
2078 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2079 if (!CI || !CI->isOne())
2080 return false;
2081 }
2082
2083 for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2084 if (!cast<Constant>(Idxs[i])->isNullValue())
2085 return false;
2086 return true;
2087 }
2088
2089 /// Test whether a given ConstantInt is in-range for a SequentialType.
isIndexInRangeOfArrayType(uint64_t NumElements,const ConstantInt * CI)2090 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2091 const ConstantInt *CI) {
2092 // We cannot bounds check the index if it doesn't fit in an int64_t.
2093 if (CI->getValue().getMinSignedBits() > 64)
2094 return false;
2095
2096 // A negative index or an index past the end of our sequential type is
2097 // considered out-of-range.
2098 int64_t IndexVal = CI->getSExtValue();
2099 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2100 return false;
2101
2102 // Otherwise, it is in-range.
2103 return true;
2104 }
2105
2106 // Combine Indices - If the source pointer to this getelementptr instruction
2107 // is a getelementptr instruction, combine the indices of the two
2108 // getelementptr instructions into a single instruction.
foldGEPOfGEP(GEPOperator * GEP,Type * PointeeTy,bool InBounds,ArrayRef<Value * > Idxs)2109 static Constant *foldGEPOfGEP(GEPOperator *GEP, Type *PointeeTy, bool InBounds,
2110 ArrayRef<Value *> Idxs) {
2111 if (PointeeTy != GEP->getResultElementType())
2112 return nullptr;
2113
2114 Constant *Idx0 = cast<Constant>(Idxs[0]);
2115 if (Idx0->isNullValue()) {
2116 // Handle the simple case of a zero index.
2117 SmallVector<Value*, 16> NewIndices;
2118 NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
2119 NewIndices.append(GEP->idx_begin(), GEP->idx_end());
2120 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2121 return ConstantExpr::getGetElementPtr(
2122 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
2123 NewIndices, InBounds && GEP->isInBounds(), GEP->getInRangeIndex());
2124 }
2125
2126 gep_type_iterator LastI = gep_type_end(GEP);
2127 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
2128 I != E; ++I)
2129 LastI = I;
2130
2131 // We cannot combine indices if doing so would take us outside of an
2132 // array or vector. Doing otherwise could trick us if we evaluated such a
2133 // GEP as part of a load.
2134 //
2135 // e.g. Consider if the original GEP was:
2136 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2137 // i32 0, i32 0, i64 0)
2138 //
2139 // If we then tried to offset it by '8' to get to the third element,
2140 // an i8, we should *not* get:
2141 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2142 // i32 0, i32 0, i64 8)
2143 //
2144 // This GEP tries to index array element '8 which runs out-of-bounds.
2145 // Subsequent evaluation would get confused and produce erroneous results.
2146 //
2147 // The following prohibits such a GEP from being formed by checking to see
2148 // if the index is in-range with respect to an array.
2149 if (!LastI.isSequential())
2150 return nullptr;
2151 ConstantInt *CI = dyn_cast<ConstantInt>(Idx0);
2152 if (!CI)
2153 return nullptr;
2154 if (LastI.isBoundedSequential() &&
2155 !isIndexInRangeOfArrayType(LastI.getSequentialNumElements(), CI))
2156 return nullptr;
2157
2158 // TODO: This code may be extended to handle vectors as well.
2159 auto *LastIdx = cast<Constant>(GEP->getOperand(GEP->getNumOperands()-1));
2160 Type *LastIdxTy = LastIdx->getType();
2161 if (LastIdxTy->isVectorTy())
2162 return nullptr;
2163
2164 SmallVector<Value*, 16> NewIndices;
2165 NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
2166 NewIndices.append(GEP->idx_begin(), GEP->idx_end() - 1);
2167
2168 // Add the last index of the source with the first index of the new GEP.
2169 // Make sure to handle the case when they are actually different types.
2170 if (LastIdxTy != Idx0->getType()) {
2171 unsigned CommonExtendedWidth =
2172 std::max(LastIdxTy->getIntegerBitWidth(),
2173 Idx0->getType()->getIntegerBitWidth());
2174 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2175
2176 Type *CommonTy =
2177 Type::getIntNTy(LastIdxTy->getContext(), CommonExtendedWidth);
2178 Idx0 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2179 LastIdx = ConstantExpr::getSExtOrBitCast(LastIdx, CommonTy);
2180 }
2181
2182 NewIndices.push_back(ConstantExpr::get(Instruction::Add, Idx0, LastIdx));
2183 NewIndices.append(Idxs.begin() + 1, Idxs.end());
2184
2185 // The combined GEP normally inherits its index inrange attribute from
2186 // the inner GEP, but if the inner GEP's last index was adjusted by the
2187 // outer GEP, any inbounds attribute on that index is invalidated.
2188 Optional<unsigned> IRIndex = GEP->getInRangeIndex();
2189 if (IRIndex && *IRIndex == GEP->getNumIndices() - 1)
2190 IRIndex = None;
2191
2192 return ConstantExpr::getGetElementPtr(
2193 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
2194 NewIndices, InBounds && GEP->isInBounds(), IRIndex);
2195 }
2196
ConstantFoldGetElementPtr(Type * PointeeTy,Constant * C,bool InBounds,Optional<unsigned> InRangeIndex,ArrayRef<Value * > Idxs)2197 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2198 bool InBounds,
2199 Optional<unsigned> InRangeIndex,
2200 ArrayRef<Value *> Idxs) {
2201 if (Idxs.empty()) return C;
2202
2203 Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2204 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2205
2206 if (isa<PoisonValue>(C))
2207 return PoisonValue::get(GEPTy);
2208
2209 if (isa<UndefValue>(C))
2210 // If inbounds, we can choose an out-of-bounds pointer as a base pointer.
2211 return InBounds ? PoisonValue::get(GEPTy) : UndefValue::get(GEPTy);
2212
2213 Constant *Idx0 = cast<Constant>(Idxs[0]);
2214 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2215 return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2216 ? ConstantVector::getSplat(
2217 cast<VectorType>(GEPTy)->getElementCount(), C)
2218 : C;
2219
2220 if (C->isNullValue()) {
2221 bool isNull = true;
2222 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2223 if (!isa<UndefValue>(Idxs[i]) &&
2224 !cast<Constant>(Idxs[i])->isNullValue()) {
2225 isNull = false;
2226 break;
2227 }
2228 if (isNull) {
2229 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2230 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2231
2232 assert(Ty && "Invalid indices for GEP!");
2233 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2234 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2235 if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2236 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2237
2238 // The GEP returns a vector of pointers when one of more of
2239 // its arguments is a vector.
2240 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2241 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2242 assert((!isa<VectorType>(GEPTy) || isa<ScalableVectorType>(GEPTy) ==
2243 isa<ScalableVectorType>(VT)) &&
2244 "Mismatched GEPTy vector types");
2245 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2246 break;
2247 }
2248 }
2249
2250 return Constant::getNullValue(GEPTy);
2251 }
2252 }
2253
2254 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2255 if (auto *GEP = dyn_cast<GEPOperator>(CE))
2256 if (Constant *C = foldGEPOfGEP(GEP, PointeeTy, InBounds, Idxs))
2257 return C;
2258
2259 // Attempt to fold casts to the same type away. For example, folding:
2260 //
2261 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2262 // i64 0, i64 0)
2263 // into:
2264 //
2265 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2266 //
2267 // Don't fold if the cast is changing address spaces.
2268 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2269 PointerType *SrcPtrTy =
2270 dyn_cast<PointerType>(CE->getOperand(0)->getType());
2271 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2272 if (SrcPtrTy && DstPtrTy) {
2273 ArrayType *SrcArrayTy =
2274 dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2275 ArrayType *DstArrayTy =
2276 dyn_cast<ArrayType>(DstPtrTy->getElementType());
2277 if (SrcArrayTy && DstArrayTy
2278 && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2279 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2280 return ConstantExpr::getGetElementPtr(SrcArrayTy,
2281 (Constant *)CE->getOperand(0),
2282 Idxs, InBounds, InRangeIndex);
2283 }
2284 }
2285 }
2286
2287 // Check to see if any array indices are not within the corresponding
2288 // notional array or vector bounds. If so, try to determine if they can be
2289 // factored out into preceding dimensions.
2290 SmallVector<Constant *, 8> NewIdxs;
2291 Type *Ty = PointeeTy;
2292 Type *Prev = C->getType();
2293 auto GEPIter = gep_type_begin(PointeeTy, Idxs);
2294 bool Unknown =
2295 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2296 for (unsigned i = 1, e = Idxs.size(); i != e;
2297 Prev = Ty, Ty = (++GEPIter).getIndexedType(), ++i) {
2298 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2299 // We don't know if it's in range or not.
2300 Unknown = true;
2301 continue;
2302 }
2303 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2304 // Skip if the type of the previous index is not supported.
2305 continue;
2306 if (InRangeIndex && i == *InRangeIndex + 1) {
2307 // If an index is marked inrange, we cannot apply this canonicalization to
2308 // the following index, as that will cause the inrange index to point to
2309 // the wrong element.
2310 continue;
2311 }
2312 if (isa<StructType>(Ty)) {
2313 // The verify makes sure that GEPs into a struct are in range.
2314 continue;
2315 }
2316 if (isa<VectorType>(Ty)) {
2317 // There can be awkward padding in after a non-power of two vector.
2318 Unknown = true;
2319 continue;
2320 }
2321 auto *STy = cast<ArrayType>(Ty);
2322 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2323 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2324 // It's in range, skip to the next index.
2325 continue;
2326 if (CI->getSExtValue() < 0) {
2327 // It's out of range and negative, don't try to factor it.
2328 Unknown = true;
2329 continue;
2330 }
2331 } else {
2332 auto *CV = cast<ConstantDataVector>(Idxs[i]);
2333 bool InRange = true;
2334 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2335 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2336 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2337 if (CI->getSExtValue() < 0) {
2338 Unknown = true;
2339 break;
2340 }
2341 }
2342 if (InRange || Unknown)
2343 // It's in range, skip to the next index.
2344 // It's out of range and negative, don't try to factor it.
2345 continue;
2346 }
2347 if (isa<StructType>(Prev)) {
2348 // It's out of range, but the prior dimension is a struct
2349 // so we can't do anything about it.
2350 Unknown = true;
2351 continue;
2352 }
2353 // It's out of range, but we can factor it into the prior
2354 // dimension.
2355 NewIdxs.resize(Idxs.size());
2356 // Determine the number of elements in our sequential type.
2357 uint64_t NumElements = STy->getArrayNumElements();
2358
2359 // Expand the current index or the previous index to a vector from a scalar
2360 // if necessary.
2361 Constant *CurrIdx = cast<Constant>(Idxs[i]);
2362 auto *PrevIdx =
2363 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2364 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2365 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2366 bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2367
2368 if (!IsCurrIdxVector && IsPrevIdxVector)
2369 CurrIdx = ConstantDataVector::getSplat(
2370 cast<FixedVectorType>(PrevIdx->getType())->getNumElements(), CurrIdx);
2371
2372 if (!IsPrevIdxVector && IsCurrIdxVector)
2373 PrevIdx = ConstantDataVector::getSplat(
2374 cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), PrevIdx);
2375
2376 Constant *Factor =
2377 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2378 if (UseVector)
2379 Factor = ConstantDataVector::getSplat(
2380 IsPrevIdxVector
2381 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2382 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements(),
2383 Factor);
2384
2385 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2386
2387 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2388
2389 unsigned CommonExtendedWidth =
2390 std::max(PrevIdx->getType()->getScalarSizeInBits(),
2391 Div->getType()->getScalarSizeInBits());
2392 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2393
2394 // Before adding, extend both operands to i64 to avoid
2395 // overflow trouble.
2396 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2397 if (UseVector)
2398 ExtendedTy = FixedVectorType::get(
2399 ExtendedTy,
2400 IsPrevIdxVector
2401 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2402 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements());
2403
2404 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2405 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2406
2407 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2408 Div = ConstantExpr::getSExt(Div, ExtendedTy);
2409
2410 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2411 }
2412
2413 // If we did any factoring, start over with the adjusted indices.
2414 if (!NewIdxs.empty()) {
2415 for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2416 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2417 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2418 InRangeIndex);
2419 }
2420
2421 // If all indices are known integers and normalized, we can do a simple
2422 // check for the "inbounds" property.
2423 if (!Unknown && !InBounds)
2424 if (auto *GV = dyn_cast<GlobalVariable>(C))
2425 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2426 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2427 /*InBounds=*/true, InRangeIndex);
2428
2429 return nullptr;
2430 }
2431