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 using namespace llvm; 34 using namespace llvm::PatternMatch; 35 36 //===----------------------------------------------------------------------===// 37 // ConstantFold*Instruction Implementations 38 //===----------------------------------------------------------------------===// 39 40 /// Convert the specified vector Constant node to the specified vector type. 41 /// At this point, we know that the elements of the input vector constant are 42 /// all simple integer or FP values. 43 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 44 45 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 46 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 47 48 // Do not iterate on scalable vector. The num of elements is unknown at 49 // compile-time. 50 if (isa<ScalableVectorType>(DstTy)) 51 return nullptr; 52 53 // If this cast changes element count then we can't handle it here: 54 // doing so requires endianness information. This should be handled by 55 // Analysis/ConstantFolding.cpp 56 unsigned NumElts = cast<FixedVectorType>(DstTy)->getNumElements(); 57 if (NumElts != cast<FixedVectorType>(CV->getType())->getNumElements()) 58 return nullptr; 59 60 Type *DstEltTy = DstTy->getElementType(); 61 // Fast path for splatted constants. 62 if (Constant *Splat = CV->getSplatValue()) { 63 return ConstantVector::getSplat(DstTy->getElementCount(), 64 ConstantExpr::getBitCast(Splat, DstEltTy)); 65 } 66 67 SmallVector<Constant*, 16> Result; 68 Type *Ty = IntegerType::get(CV->getContext(), 32); 69 for (unsigned i = 0; i != NumElts; ++i) { 70 Constant *C = 71 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 72 C = ConstantExpr::getBitCast(C, DstEltTy); 73 Result.push_back(C); 74 } 75 76 return ConstantVector::get(Result); 77 } 78 79 /// This function determines which opcode to use to fold two constant cast 80 /// expressions together. It uses CastInst::isEliminableCastPair to determine 81 /// the opcode. Consequently its just a wrapper around that function. 82 /// Determine if it is valid to fold a cast of a cast 83 static unsigned 84 foldConstantCastPair( 85 unsigned opc, ///< opcode of the second cast constant expression 86 ConstantExpr *Op, ///< the first cast constant expression 87 Type *DstTy ///< destination type of the first cast 88 ) { 89 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 90 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 91 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 92 93 // The types and opcodes for the two Cast constant expressions 94 Type *SrcTy = Op->getOperand(0)->getType(); 95 Type *MidTy = Op->getType(); 96 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 97 Instruction::CastOps secondOp = Instruction::CastOps(opc); 98 99 // Assume that pointers are never more than 64 bits wide, and only use this 100 // for the middle type. Otherwise we could end up folding away illegal 101 // bitcasts between address spaces with different sizes. 102 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 103 104 // Let CastInst::isEliminableCastPair do the heavy lifting. 105 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 106 nullptr, FakeIntPtrTy, nullptr); 107 } 108 109 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 110 Type *SrcTy = V->getType(); 111 if (SrcTy == DestTy) 112 return V; // no-op cast 113 114 // Check to see if we are casting a pointer to an aggregate to a pointer to 115 // the first element. If so, return the appropriate GEP instruction. 116 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 117 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 118 if (PTy->getAddressSpace() == DPTy->getAddressSpace() && 119 !PTy->isOpaque() && !DPTy->isOpaque() && 120 PTy->getNonOpaquePointerElementType()->isSized()) { 121 SmallVector<Value*, 8> IdxList; 122 Value *Zero = 123 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 124 IdxList.push_back(Zero); 125 Type *ElTy = PTy->getNonOpaquePointerElementType(); 126 while (ElTy && ElTy != DPTy->getNonOpaquePointerElementType()) { 127 ElTy = GetElementPtrInst::getTypeAtIndex(ElTy, (uint64_t)0); 128 IdxList.push_back(Zero); 129 } 130 131 if (ElTy == DPTy->getNonOpaquePointerElementType()) 132 // This GEP is inbounds because all indices are zero. 133 return ConstantExpr::getInBoundsGetElementPtr( 134 PTy->getNonOpaquePointerElementType(), V, IdxList); 135 } 136 137 // Handle casts from one vector constant to another. We know that the src 138 // and dest type have the same size (otherwise its an illegal cast). 139 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 140 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 141 assert(DestPTy->getPrimitiveSizeInBits() == 142 SrcTy->getPrimitiveSizeInBits() && 143 "Not cast between same sized vectors!"); 144 SrcTy = nullptr; 145 // First, check for null. Undef is already handled. 146 if (isa<ConstantAggregateZero>(V)) 147 return Constant::getNullValue(DestTy); 148 149 // Handle ConstantVector and ConstantAggregateVector. 150 return BitCastConstantVector(V, DestPTy); 151 } 152 153 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 154 // This allows for other simplifications (although some of them 155 // can only be handled by Analysis/ConstantFolding.cpp). 156 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 157 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 158 } 159 160 // Finally, implement bitcast folding now. The code below doesn't handle 161 // bitcast right. 162 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 163 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 164 165 // Handle integral constant input. 166 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 167 if (DestTy->isIntegerTy()) 168 // Integral -> Integral. This is a no-op because the bit widths must 169 // be the same. Consequently, we just fold to V. 170 return V; 171 172 // See note below regarding the PPC_FP128 restriction. 173 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 174 return ConstantFP::get(DestTy->getContext(), 175 APFloat(DestTy->getFltSemantics(), 176 CI->getValue())); 177 178 // Otherwise, can't fold this (vector?) 179 return nullptr; 180 } 181 182 // Handle ConstantFP input: FP -> Integral. 183 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 184 // PPC_FP128 is really the sum of two consecutive doubles, where the first 185 // double is always stored first in memory, regardless of the target 186 // endianness. The memory layout of i128, however, depends on the target 187 // endianness, and so we can't fold this without target endianness 188 // information. This should instead be handled by 189 // Analysis/ConstantFolding.cpp 190 if (FP->getType()->isPPC_FP128Ty()) 191 return nullptr; 192 193 // Make sure dest type is compatible with the folded integer constant. 194 if (!DestTy->isIntegerTy()) 195 return nullptr; 196 197 return ConstantInt::get(FP->getContext(), 198 FP->getValueAPF().bitcastToAPInt()); 199 } 200 201 return nullptr; 202 } 203 204 205 /// V is an integer constant which only has a subset of its bytes used. 206 /// The bytes used are indicated by ByteStart (which is the first byte used, 207 /// counting from the least significant byte) and ByteSize, which is the number 208 /// of bytes used. 209 /// 210 /// This function analyzes the specified constant to see if the specified byte 211 /// range can be returned as a simplified constant. If so, the constant is 212 /// returned, otherwise null is returned. 213 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 214 unsigned ByteSize) { 215 assert(C->getType()->isIntegerTy() && 216 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 217 "Non-byte sized integer input"); 218 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 219 assert(ByteSize && "Must be accessing some piece"); 220 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 221 assert(ByteSize != CSize && "Should not extract everything"); 222 223 // Constant Integers are simple. 224 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 225 APInt V = CI->getValue(); 226 if (ByteStart) 227 V.lshrInPlace(ByteStart*8); 228 V = V.trunc(ByteSize*8); 229 return ConstantInt::get(CI->getContext(), V); 230 } 231 232 // In the input is a constant expr, we might be able to recursively simplify. 233 // If not, we definitely can't do anything. 234 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 235 if (!CE) return nullptr; 236 237 switch (CE->getOpcode()) { 238 default: return nullptr; 239 case Instruction::Or: { 240 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 241 if (!RHS) 242 return nullptr; 243 244 // X | -1 -> -1. 245 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 246 if (RHSC->isMinusOne()) 247 return RHSC; 248 249 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 250 if (!LHS) 251 return nullptr; 252 return ConstantExpr::getOr(LHS, RHS); 253 } 254 case Instruction::And: { 255 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 256 if (!RHS) 257 return nullptr; 258 259 // X & 0 -> 0. 260 if (RHS->isNullValue()) 261 return RHS; 262 263 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 264 if (!LHS) 265 return nullptr; 266 return ConstantExpr::getAnd(LHS, RHS); 267 } 268 case Instruction::LShr: { 269 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 270 if (!Amt) 271 return nullptr; 272 APInt ShAmt = Amt->getValue(); 273 // Cannot analyze non-byte shifts. 274 if ((ShAmt & 7) != 0) 275 return nullptr; 276 ShAmt.lshrInPlace(3); 277 278 // If the extract is known to be all zeros, return zero. 279 if (ShAmt.uge(CSize - ByteStart)) 280 return Constant::getNullValue( 281 IntegerType::get(CE->getContext(), ByteSize * 8)); 282 // If the extract is known to be fully in the input, extract it. 283 if (ShAmt.ule(CSize - (ByteStart + ByteSize))) 284 return ExtractConstantBytes(CE->getOperand(0), 285 ByteStart + ShAmt.getZExtValue(), ByteSize); 286 287 // TODO: Handle the 'partially zero' case. 288 return nullptr; 289 } 290 291 case Instruction::Shl: { 292 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 293 if (!Amt) 294 return nullptr; 295 APInt ShAmt = Amt->getValue(); 296 // Cannot analyze non-byte shifts. 297 if ((ShAmt & 7) != 0) 298 return nullptr; 299 ShAmt.lshrInPlace(3); 300 301 // If the extract is known to be all zeros, return zero. 302 if (ShAmt.uge(ByteStart + ByteSize)) 303 return Constant::getNullValue( 304 IntegerType::get(CE->getContext(), ByteSize * 8)); 305 // If the extract is known to be fully in the input, extract it. 306 if (ShAmt.ule(ByteStart)) 307 return ExtractConstantBytes(CE->getOperand(0), 308 ByteStart - ShAmt.getZExtValue(), ByteSize); 309 310 // TODO: Handle the 'partially zero' case. 311 return nullptr; 312 } 313 314 case Instruction::ZExt: { 315 unsigned SrcBitSize = 316 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 317 318 // If extracting something that is completely zero, return 0. 319 if (ByteStart*8 >= SrcBitSize) 320 return Constant::getNullValue(IntegerType::get(CE->getContext(), 321 ByteSize*8)); 322 323 // If exactly extracting the input, return it. 324 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 325 return CE->getOperand(0); 326 327 // If extracting something completely in the input, if the input is a 328 // multiple of 8 bits, recurse. 329 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 330 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 331 332 // Otherwise, if extracting a subset of the input, which is not multiple of 333 // 8 bits, do a shift and trunc to get the bits. 334 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 335 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 336 Constant *Res = CE->getOperand(0); 337 if (ByteStart) 338 Res = ConstantExpr::getLShr(Res, 339 ConstantInt::get(Res->getType(), ByteStart*8)); 340 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 341 ByteSize*8)); 342 } 343 344 // TODO: Handle the 'partially zero' case. 345 return nullptr; 346 } 347 } 348 } 349 350 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 351 Type *DestTy) { 352 if (isa<PoisonValue>(V)) 353 return PoisonValue::get(DestTy); 354 355 if (isa<UndefValue>(V)) { 356 // zext(undef) = 0, because the top bits will be zero. 357 // sext(undef) = 0, because the top bits will all be the same. 358 // [us]itofp(undef) = 0, because the result value is bounded. 359 if (opc == Instruction::ZExt || opc == Instruction::SExt || 360 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 361 return Constant::getNullValue(DestTy); 362 return UndefValue::get(DestTy); 363 } 364 365 if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() && 366 opc != Instruction::AddrSpaceCast) 367 return Constant::getNullValue(DestTy); 368 369 // If the cast operand is a constant expression, there's a few things we can 370 // do to try to simplify it. 371 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 372 if (CE->isCast()) { 373 // Try hard to fold cast of cast because they are often eliminable. 374 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 375 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 376 } else if (CE->getOpcode() == Instruction::GetElementPtr && 377 // Do not fold addrspacecast (gep 0, .., 0). It might make the 378 // addrspacecast uncanonicalized. 379 opc != Instruction::AddrSpaceCast && 380 // Do not fold bitcast (gep) with inrange index, as this loses 381 // information. 382 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() && 383 // Do not fold if the gep type is a vector, as bitcasting 384 // operand 0 of a vector gep will result in a bitcast between 385 // different sizes. 386 !CE->getType()->isVectorTy()) { 387 // If all of the indexes in the GEP are null values, there is no pointer 388 // adjustment going on. We might as well cast the source pointer. 389 bool isAllNull = true; 390 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 391 if (!CE->getOperand(i)->isNullValue()) { 392 isAllNull = false; 393 break; 394 } 395 if (isAllNull) 396 // This is casting one pointer type to another, always BitCast 397 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 398 } 399 } 400 401 // If the cast operand is a constant vector, perform the cast by 402 // operating on each element. In the cast of bitcasts, the element 403 // count may be mismatched; don't attempt to handle that here. 404 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 405 DestTy->isVectorTy() && 406 cast<FixedVectorType>(DestTy)->getNumElements() == 407 cast<FixedVectorType>(V->getType())->getNumElements()) { 408 VectorType *DestVecTy = cast<VectorType>(DestTy); 409 Type *DstEltTy = DestVecTy->getElementType(); 410 // Fast path for splatted constants. 411 if (Constant *Splat = V->getSplatValue()) { 412 return ConstantVector::getSplat( 413 cast<VectorType>(DestTy)->getElementCount(), 414 ConstantExpr::getCast(opc, Splat, DstEltTy)); 415 } 416 SmallVector<Constant *, 16> res; 417 Type *Ty = IntegerType::get(V->getContext(), 32); 418 for (unsigned i = 0, 419 e = cast<FixedVectorType>(V->getType())->getNumElements(); 420 i != e; ++i) { 421 Constant *C = 422 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 423 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 424 } 425 return ConstantVector::get(res); 426 } 427 428 // We actually have to do a cast now. Perform the cast according to the 429 // opcode specified. 430 switch (opc) { 431 default: 432 llvm_unreachable("Failed to cast constant expression"); 433 case Instruction::FPTrunc: 434 case Instruction::FPExt: 435 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 436 bool ignored; 437 APFloat Val = FPC->getValueAPF(); 438 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() : 439 DestTy->isFloatTy() ? APFloat::IEEEsingle() : 440 DestTy->isDoubleTy() ? APFloat::IEEEdouble() : 441 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() : 442 DestTy->isFP128Ty() ? APFloat::IEEEquad() : 443 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() : 444 APFloat::Bogus(), 445 APFloat::rmNearestTiesToEven, &ignored); 446 return ConstantFP::get(V->getContext(), Val); 447 } 448 return nullptr; // Can't fold. 449 case Instruction::FPToUI: 450 case Instruction::FPToSI: 451 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 452 const APFloat &V = FPC->getValueAPF(); 453 bool ignored; 454 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 455 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI); 456 if (APFloat::opInvalidOp == 457 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) { 458 // Undefined behavior invoked - the destination type can't represent 459 // the input constant. 460 return PoisonValue::get(DestTy); 461 } 462 return ConstantInt::get(FPC->getContext(), IntVal); 463 } 464 return nullptr; // Can't fold. 465 case Instruction::IntToPtr: //always treated as unsigned 466 if (V->isNullValue()) // Is it an integral null value? 467 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 468 return nullptr; // Other pointer types cannot be casted 469 case Instruction::PtrToInt: // always treated as unsigned 470 // Is it a null pointer value? 471 if (V->isNullValue()) 472 return ConstantInt::get(DestTy, 0); 473 // Other pointer types cannot be casted 474 return nullptr; 475 case Instruction::UIToFP: 476 case Instruction::SIToFP: 477 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 478 const APInt &api = CI->getValue(); 479 APFloat apf(DestTy->getFltSemantics(), 480 APInt::getZero(DestTy->getPrimitiveSizeInBits())); 481 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 482 APFloat::rmNearestTiesToEven); 483 return ConstantFP::get(V->getContext(), apf); 484 } 485 return nullptr; 486 case Instruction::ZExt: 487 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 488 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 489 return ConstantInt::get(V->getContext(), 490 CI->getValue().zext(BitWidth)); 491 } 492 return nullptr; 493 case Instruction::SExt: 494 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 495 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 496 return ConstantInt::get(V->getContext(), 497 CI->getValue().sext(BitWidth)); 498 } 499 return nullptr; 500 case Instruction::Trunc: { 501 if (V->getType()->isVectorTy()) 502 return nullptr; 503 504 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 505 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 506 return ConstantInt::get(V->getContext(), 507 CI->getValue().trunc(DestBitWidth)); 508 } 509 510 // The input must be a constantexpr. See if we can simplify this based on 511 // the bytes we are demanding. Only do this if the source and dest are an 512 // even multiple of a byte. 513 if ((DestBitWidth & 7) == 0 && 514 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 515 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 516 return Res; 517 518 return nullptr; 519 } 520 case Instruction::BitCast: 521 return FoldBitCast(V, DestTy); 522 case Instruction::AddrSpaceCast: 523 return nullptr; 524 } 525 } 526 527 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 528 Constant *V1, Constant *V2) { 529 // Check for i1 and vector true/false conditions. 530 if (Cond->isNullValue()) return V2; 531 if (Cond->isAllOnesValue()) return V1; 532 533 // If the condition is a vector constant, fold the result elementwise. 534 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 535 auto *V1VTy = CondV->getType(); 536 SmallVector<Constant*, 16> Result; 537 Type *Ty = IntegerType::get(CondV->getContext(), 32); 538 for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) { 539 Constant *V; 540 Constant *V1Element = ConstantExpr::getExtractElement(V1, 541 ConstantInt::get(Ty, i)); 542 Constant *V2Element = ConstantExpr::getExtractElement(V2, 543 ConstantInt::get(Ty, i)); 544 auto *Cond = cast<Constant>(CondV->getOperand(i)); 545 if (isa<PoisonValue>(Cond)) { 546 V = PoisonValue::get(V1Element->getType()); 547 } else if (V1Element == V2Element) { 548 V = V1Element; 549 } else if (isa<UndefValue>(Cond)) { 550 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 551 } else { 552 if (!isa<ConstantInt>(Cond)) break; 553 V = Cond->isNullValue() ? V2Element : V1Element; 554 } 555 Result.push_back(V); 556 } 557 558 // If we were able to build the vector, return it. 559 if (Result.size() == V1VTy->getNumElements()) 560 return ConstantVector::get(Result); 561 } 562 563 if (isa<PoisonValue>(Cond)) 564 return PoisonValue::get(V1->getType()); 565 566 if (isa<UndefValue>(Cond)) { 567 if (isa<UndefValue>(V1)) return V1; 568 return V2; 569 } 570 571 if (V1 == V2) return V1; 572 573 if (isa<PoisonValue>(V1)) 574 return V2; 575 if (isa<PoisonValue>(V2)) 576 return V1; 577 578 // If the true or false value is undef, we can fold to the other value as 579 // long as the other value isn't poison. 580 auto NotPoison = [](Constant *C) { 581 if (isa<PoisonValue>(C)) 582 return false; 583 584 // TODO: We can analyze ConstExpr by opcode to determine if there is any 585 // possibility of poison. 586 if (isa<ConstantExpr>(C)) 587 return false; 588 589 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(C) || 590 isa<ConstantPointerNull>(C) || isa<Function>(C)) 591 return true; 592 593 if (C->getType()->isVectorTy()) 594 return !C->containsPoisonElement() && !C->containsConstantExpression(); 595 596 // TODO: Recursively analyze aggregates or other constants. 597 return false; 598 }; 599 if (isa<UndefValue>(V1) && NotPoison(V2)) return V2; 600 if (isa<UndefValue>(V2) && NotPoison(V1)) return V1; 601 602 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 603 if (TrueVal->getOpcode() == Instruction::Select) 604 if (TrueVal->getOperand(0) == Cond) 605 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 606 } 607 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 608 if (FalseVal->getOpcode() == Instruction::Select) 609 if (FalseVal->getOperand(0) == Cond) 610 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 611 } 612 613 return nullptr; 614 } 615 616 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 617 Constant *Idx) { 618 auto *ValVTy = cast<VectorType>(Val->getType()); 619 620 // extractelt poison, C -> poison 621 // extractelt C, undef -> poison 622 if (isa<PoisonValue>(Val) || isa<UndefValue>(Idx)) 623 return PoisonValue::get(ValVTy->getElementType()); 624 625 // extractelt undef, C -> undef 626 if (isa<UndefValue>(Val)) 627 return UndefValue::get(ValVTy->getElementType()); 628 629 auto *CIdx = dyn_cast<ConstantInt>(Idx); 630 if (!CIdx) 631 return nullptr; 632 633 if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val->getType())) { 634 // ee({w,x,y,z}, wrong_value) -> poison 635 if (CIdx->uge(ValFVTy->getNumElements())) 636 return PoisonValue::get(ValFVTy->getElementType()); 637 } 638 639 // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...) 640 if (auto *CE = dyn_cast<ConstantExpr>(Val)) { 641 if (auto *GEP = dyn_cast<GEPOperator>(CE)) { 642 SmallVector<Constant *, 8> Ops; 643 Ops.reserve(CE->getNumOperands()); 644 for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) { 645 Constant *Op = CE->getOperand(i); 646 if (Op->getType()->isVectorTy()) { 647 Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx); 648 if (!ScalarOp) 649 return nullptr; 650 Ops.push_back(ScalarOp); 651 } else 652 Ops.push_back(Op); 653 } 654 return CE->getWithOperands(Ops, ValVTy->getElementType(), false, 655 GEP->getSourceElementType()); 656 } else if (CE->getOpcode() == Instruction::InsertElement) { 657 if (const auto *IEIdx = dyn_cast<ConstantInt>(CE->getOperand(2))) { 658 if (APSInt::isSameValue(APSInt(IEIdx->getValue()), 659 APSInt(CIdx->getValue()))) { 660 return CE->getOperand(1); 661 } else { 662 return ConstantExpr::getExtractElement(CE->getOperand(0), CIdx); 663 } 664 } 665 } 666 } 667 668 if (Constant *C = Val->getAggregateElement(CIdx)) 669 return C; 670 671 // Lane < Splat minimum vector width => extractelt Splat(x), Lane -> x 672 if (CIdx->getValue().ult(ValVTy->getElementCount().getKnownMinValue())) { 673 if (Constant *SplatVal = Val->getSplatValue()) 674 return SplatVal; 675 } 676 677 return nullptr; 678 } 679 680 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 681 Constant *Elt, 682 Constant *Idx) { 683 if (isa<UndefValue>(Idx)) 684 return PoisonValue::get(Val->getType()); 685 686 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 687 if (!CIdx) return nullptr; 688 689 // Do not iterate on scalable vector. The num of elements is unknown at 690 // compile-time. 691 if (isa<ScalableVectorType>(Val->getType())) 692 return nullptr; 693 694 auto *ValTy = cast<FixedVectorType>(Val->getType()); 695 696 unsigned NumElts = ValTy->getNumElements(); 697 if (CIdx->uge(NumElts)) 698 return PoisonValue::get(Val->getType()); 699 700 SmallVector<Constant*, 16> Result; 701 Result.reserve(NumElts); 702 auto *Ty = Type::getInt32Ty(Val->getContext()); 703 uint64_t IdxVal = CIdx->getZExtValue(); 704 for (unsigned i = 0; i != NumElts; ++i) { 705 if (i == IdxVal) { 706 Result.push_back(Elt); 707 continue; 708 } 709 710 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 711 Result.push_back(C); 712 } 713 714 return ConstantVector::get(Result); 715 } 716 717 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2, 718 ArrayRef<int> Mask) { 719 auto *V1VTy = cast<VectorType>(V1->getType()); 720 unsigned MaskNumElts = Mask.size(); 721 auto MaskEltCount = 722 ElementCount::get(MaskNumElts, isa<ScalableVectorType>(V1VTy)); 723 Type *EltTy = V1VTy->getElementType(); 724 725 // Undefined shuffle mask -> undefined value. 726 if (all_of(Mask, [](int Elt) { return Elt == UndefMaskElem; })) { 727 return UndefValue::get(FixedVectorType::get(EltTy, MaskNumElts)); 728 } 729 730 // If the mask is all zeros this is a splat, no need to go through all 731 // elements. 732 if (all_of(Mask, [](int Elt) { return Elt == 0; })) { 733 Type *Ty = IntegerType::get(V1->getContext(), 32); 734 Constant *Elt = 735 ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, 0)); 736 737 if (Elt->isNullValue()) { 738 auto *VTy = VectorType::get(EltTy, MaskEltCount); 739 return ConstantAggregateZero::get(VTy); 740 } else if (!MaskEltCount.isScalable()) 741 return ConstantVector::getSplat(MaskEltCount, Elt); 742 } 743 // Do not iterate on scalable vector. The num of elements is unknown at 744 // compile-time. 745 if (isa<ScalableVectorType>(V1VTy)) 746 return nullptr; 747 748 unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue(); 749 750 // Loop over the shuffle mask, evaluating each element. 751 SmallVector<Constant*, 32> Result; 752 for (unsigned i = 0; i != MaskNumElts; ++i) { 753 int Elt = Mask[i]; 754 if (Elt == -1) { 755 Result.push_back(UndefValue::get(EltTy)); 756 continue; 757 } 758 Constant *InElt; 759 if (unsigned(Elt) >= SrcNumElts*2) 760 InElt = UndefValue::get(EltTy); 761 else if (unsigned(Elt) >= SrcNumElts) { 762 Type *Ty = IntegerType::get(V2->getContext(), 32); 763 InElt = 764 ConstantExpr::getExtractElement(V2, 765 ConstantInt::get(Ty, Elt - SrcNumElts)); 766 } else { 767 Type *Ty = IntegerType::get(V1->getContext(), 32); 768 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 769 } 770 Result.push_back(InElt); 771 } 772 773 return ConstantVector::get(Result); 774 } 775 776 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 777 ArrayRef<unsigned> Idxs) { 778 // Base case: no indices, so return the entire value. 779 if (Idxs.empty()) 780 return Agg; 781 782 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 783 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 784 785 return nullptr; 786 } 787 788 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 789 Constant *Val, 790 ArrayRef<unsigned> Idxs) { 791 // Base case: no indices, so replace the entire value. 792 if (Idxs.empty()) 793 return Val; 794 795 unsigned NumElts; 796 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 797 NumElts = ST->getNumElements(); 798 else 799 NumElts = cast<ArrayType>(Agg->getType())->getNumElements(); 800 801 SmallVector<Constant*, 32> Result; 802 for (unsigned i = 0; i != NumElts; ++i) { 803 Constant *C = Agg->getAggregateElement(i); 804 if (!C) return nullptr; 805 806 if (Idxs[0] == i) 807 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 808 809 Result.push_back(C); 810 } 811 812 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 813 return ConstantStruct::get(ST, Result); 814 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Result); 815 } 816 817 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) { 818 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected"); 819 820 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 821 // vectors are always evaluated per element. 822 bool IsScalableVector = isa<ScalableVectorType>(C->getType()); 823 bool HasScalarUndefOrScalableVectorUndef = 824 (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C); 825 826 if (HasScalarUndefOrScalableVectorUndef) { 827 switch (static_cast<Instruction::UnaryOps>(Opcode)) { 828 case Instruction::FNeg: 829 return C; // -undef -> undef 830 case Instruction::UnaryOpsEnd: 831 llvm_unreachable("Invalid UnaryOp"); 832 } 833 } 834 835 // Constant should not be UndefValue, unless these are vector constants. 836 assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue"); 837 // We only have FP UnaryOps right now. 838 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp"); 839 840 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 841 const APFloat &CV = CFP->getValueAPF(); 842 switch (Opcode) { 843 default: 844 break; 845 case Instruction::FNeg: 846 return ConstantFP::get(C->getContext(), neg(CV)); 847 } 848 } else if (auto *VTy = dyn_cast<FixedVectorType>(C->getType())) { 849 850 Type *Ty = IntegerType::get(VTy->getContext(), 32); 851 // Fast path for splatted constants. 852 if (Constant *Splat = C->getSplatValue()) { 853 Constant *Elt = ConstantExpr::get(Opcode, Splat); 854 return ConstantVector::getSplat(VTy->getElementCount(), Elt); 855 } 856 857 // Fold each element and create a vector constant from those constants. 858 SmallVector<Constant *, 16> Result; 859 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 860 Constant *ExtractIdx = ConstantInt::get(Ty, i); 861 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx); 862 863 Result.push_back(ConstantExpr::get(Opcode, Elt)); 864 } 865 866 return ConstantVector::get(Result); 867 } 868 869 // We don't know how to fold this. 870 return nullptr; 871 } 872 873 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, 874 Constant *C2) { 875 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 876 877 // Simplify BinOps with their identity values first. They are no-ops and we 878 // can always return the other value, including undef or poison values. 879 // FIXME: remove unnecessary duplicated identity patterns below. 880 // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops, 881 // like X << 0 = X. 882 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType()); 883 if (Identity) { 884 if (C1 == Identity) 885 return C2; 886 if (C2 == Identity) 887 return C1; 888 } 889 890 // Binary operations propagate poison. 891 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2)) 892 return PoisonValue::get(C1->getType()); 893 894 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 895 // vectors are always evaluated per element. 896 bool IsScalableVector = isa<ScalableVectorType>(C1->getType()); 897 bool HasScalarUndefOrScalableVectorUndef = 898 (!C1->getType()->isVectorTy() || IsScalableVector) && 899 (isa<UndefValue>(C1) || isa<UndefValue>(C2)); 900 if (HasScalarUndefOrScalableVectorUndef) { 901 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 902 case Instruction::Xor: 903 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 904 // Handle undef ^ undef -> 0 special case. This is a common 905 // idiom (misuse). 906 return Constant::getNullValue(C1->getType()); 907 LLVM_FALLTHROUGH; 908 case Instruction::Add: 909 case Instruction::Sub: 910 return UndefValue::get(C1->getType()); 911 case Instruction::And: 912 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 913 return C1; 914 return Constant::getNullValue(C1->getType()); // undef & X -> 0 915 case Instruction::Mul: { 916 // undef * undef -> undef 917 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 918 return C1; 919 const APInt *CV; 920 // X * undef -> undef if X is odd 921 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 922 if ((*CV)[0]) 923 return UndefValue::get(C1->getType()); 924 925 // X * undef -> 0 otherwise 926 return Constant::getNullValue(C1->getType()); 927 } 928 case Instruction::SDiv: 929 case Instruction::UDiv: 930 // X / undef -> poison 931 // X / 0 -> poison 932 if (match(C2, m_CombineOr(m_Undef(), m_Zero()))) 933 return PoisonValue::get(C2->getType()); 934 // undef / 1 -> undef 935 if (match(C2, m_One())) 936 return C1; 937 // undef / X -> 0 otherwise 938 return Constant::getNullValue(C1->getType()); 939 case Instruction::URem: 940 case Instruction::SRem: 941 // X % undef -> poison 942 // X % 0 -> poison 943 if (match(C2, m_CombineOr(m_Undef(), m_Zero()))) 944 return PoisonValue::get(C2->getType()); 945 // undef % X -> 0 otherwise 946 return Constant::getNullValue(C1->getType()); 947 case Instruction::Or: // X | undef -> -1 948 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 949 return C1; 950 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 951 case Instruction::LShr: 952 // X >>l undef -> poison 953 if (isa<UndefValue>(C2)) 954 return PoisonValue::get(C2->getType()); 955 // undef >>l 0 -> undef 956 if (match(C2, m_Zero())) 957 return C1; 958 // undef >>l X -> 0 959 return Constant::getNullValue(C1->getType()); 960 case Instruction::AShr: 961 // X >>a undef -> poison 962 if (isa<UndefValue>(C2)) 963 return PoisonValue::get(C2->getType()); 964 // undef >>a 0 -> undef 965 if (match(C2, m_Zero())) 966 return C1; 967 // TODO: undef >>a X -> poison if the shift is exact 968 // undef >>a X -> 0 969 return Constant::getNullValue(C1->getType()); 970 case Instruction::Shl: 971 // X << undef -> undef 972 if (isa<UndefValue>(C2)) 973 return PoisonValue::get(C2->getType()); 974 // undef << 0 -> undef 975 if (match(C2, m_Zero())) 976 return C1; 977 // undef << X -> 0 978 return Constant::getNullValue(C1->getType()); 979 case Instruction::FSub: 980 // -0.0 - undef --> undef (consistent with "fneg undef") 981 if (match(C1, m_NegZeroFP()) && isa<UndefValue>(C2)) 982 return C2; 983 LLVM_FALLTHROUGH; 984 case Instruction::FAdd: 985 case Instruction::FMul: 986 case Instruction::FDiv: 987 case Instruction::FRem: 988 // [any flop] undef, undef -> undef 989 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 990 return C1; 991 // [any flop] C, undef -> NaN 992 // [any flop] undef, C -> NaN 993 // We could potentially specialize NaN/Inf constants vs. 'normal' 994 // constants (possibly differently depending on opcode and operand). This 995 // would allow returning undef sometimes. But it is always safe to fold to 996 // NaN because we can choose the undef operand as NaN, and any FP opcode 997 // with a NaN operand will propagate NaN. 998 return ConstantFP::getNaN(C1->getType()); 999 case Instruction::BinaryOpsEnd: 1000 llvm_unreachable("Invalid BinaryOp"); 1001 } 1002 } 1003 1004 // Neither constant should be UndefValue, unless these are vector constants. 1005 assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue"); 1006 1007 // Handle simplifications when the RHS is a constant int. 1008 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1009 switch (Opcode) { 1010 case Instruction::Add: 1011 if (CI2->isZero()) return C1; // X + 0 == X 1012 break; 1013 case Instruction::Sub: 1014 if (CI2->isZero()) return C1; // X - 0 == X 1015 break; 1016 case Instruction::Mul: 1017 if (CI2->isZero()) return C2; // X * 0 == 0 1018 if (CI2->isOne()) 1019 return C1; // X * 1 == X 1020 break; 1021 case Instruction::UDiv: 1022 case Instruction::SDiv: 1023 if (CI2->isOne()) 1024 return C1; // X / 1 == X 1025 if (CI2->isZero()) 1026 return PoisonValue::get(CI2->getType()); // X / 0 == poison 1027 break; 1028 case Instruction::URem: 1029 case Instruction::SRem: 1030 if (CI2->isOne()) 1031 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1032 if (CI2->isZero()) 1033 return PoisonValue::get(CI2->getType()); // X % 0 == poison 1034 break; 1035 case Instruction::And: 1036 if (CI2->isZero()) return C2; // X & 0 == 0 1037 if (CI2->isMinusOne()) 1038 return C1; // X & -1 == X 1039 1040 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1041 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1042 if (CE1->getOpcode() == Instruction::ZExt) { 1043 unsigned DstWidth = CI2->getType()->getBitWidth(); 1044 unsigned SrcWidth = 1045 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1046 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1047 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1048 return C1; 1049 } 1050 1051 // If and'ing the address of a global with a constant, fold it. 1052 if (CE1->getOpcode() == Instruction::PtrToInt && 1053 isa<GlobalValue>(CE1->getOperand(0))) { 1054 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1055 1056 MaybeAlign GVAlign; 1057 1058 if (Module *TheModule = GV->getParent()) { 1059 const DataLayout &DL = TheModule->getDataLayout(); 1060 GVAlign = GV->getPointerAlignment(DL); 1061 1062 // If the function alignment is not specified then assume that it 1063 // is 4. 1064 // This is dangerous; on x86, the alignment of the pointer 1065 // corresponds to the alignment of the function, but might be less 1066 // than 4 if it isn't explicitly specified. 1067 // However, a fix for this behaviour was reverted because it 1068 // increased code size (see https://reviews.llvm.org/D55115) 1069 // FIXME: This code should be deleted once existing targets have 1070 // appropriate defaults 1071 if (isa<Function>(GV) && !DL.getFunctionPtrAlign()) 1072 GVAlign = Align(4); 1073 } else if (isa<Function>(GV)) { 1074 // Without a datalayout we have to assume the worst case: that the 1075 // function pointer isn't aligned at all. 1076 GVAlign = llvm::None; 1077 } else if (isa<GlobalVariable>(GV)) { 1078 GVAlign = cast<GlobalVariable>(GV)->getAlign(); 1079 } 1080 1081 if (GVAlign && *GVAlign > 1) { 1082 unsigned DstWidth = CI2->getType()->getBitWidth(); 1083 unsigned SrcWidth = std::min(DstWidth, Log2(*GVAlign)); 1084 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1085 1086 // If checking bits we know are clear, return zero. 1087 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1088 return Constant::getNullValue(CI2->getType()); 1089 } 1090 } 1091 } 1092 break; 1093 case Instruction::Or: 1094 if (CI2->isZero()) return C1; // X | 0 == X 1095 if (CI2->isMinusOne()) 1096 return C2; // X | -1 == -1 1097 break; 1098 case Instruction::Xor: 1099 if (CI2->isZero()) return C1; // X ^ 0 == X 1100 1101 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1102 switch (CE1->getOpcode()) { 1103 default: break; 1104 case Instruction::ICmp: 1105 case Instruction::FCmp: 1106 // cmp pred ^ true -> cmp !pred 1107 assert(CI2->isOne()); 1108 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1109 pred = CmpInst::getInversePredicate(pred); 1110 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1111 CE1->getOperand(1)); 1112 } 1113 } 1114 break; 1115 case Instruction::AShr: 1116 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1117 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1118 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1119 return ConstantExpr::getLShr(C1, C2); 1120 break; 1121 } 1122 } else if (isa<ConstantInt>(C1)) { 1123 // If C1 is a ConstantInt and C2 is not, swap the operands. 1124 if (Instruction::isCommutative(Opcode)) 1125 return ConstantExpr::get(Opcode, C2, C1); 1126 } 1127 1128 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1129 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1130 const APInt &C1V = CI1->getValue(); 1131 const APInt &C2V = CI2->getValue(); 1132 switch (Opcode) { 1133 default: 1134 break; 1135 case Instruction::Add: 1136 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1137 case Instruction::Sub: 1138 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1139 case Instruction::Mul: 1140 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1141 case Instruction::UDiv: 1142 assert(!CI2->isZero() && "Div by zero handled above"); 1143 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1144 case Instruction::SDiv: 1145 assert(!CI2->isZero() && "Div by zero handled above"); 1146 if (C2V.isAllOnes() && C1V.isMinSignedValue()) 1147 return PoisonValue::get(CI1->getType()); // MIN_INT / -1 -> poison 1148 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1149 case Instruction::URem: 1150 assert(!CI2->isZero() && "Div by zero handled above"); 1151 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1152 case Instruction::SRem: 1153 assert(!CI2->isZero() && "Div by zero handled above"); 1154 if (C2V.isAllOnes() && C1V.isMinSignedValue()) 1155 return PoisonValue::get(CI1->getType()); // MIN_INT % -1 -> poison 1156 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1157 case Instruction::And: 1158 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1159 case Instruction::Or: 1160 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1161 case Instruction::Xor: 1162 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1163 case Instruction::Shl: 1164 if (C2V.ult(C1V.getBitWidth())) 1165 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1166 return PoisonValue::get(C1->getType()); // too big shift is poison 1167 case Instruction::LShr: 1168 if (C2V.ult(C1V.getBitWidth())) 1169 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1170 return PoisonValue::get(C1->getType()); // too big shift is poison 1171 case Instruction::AShr: 1172 if (C2V.ult(C1V.getBitWidth())) 1173 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1174 return PoisonValue::get(C1->getType()); // too big shift is poison 1175 } 1176 } 1177 1178 switch (Opcode) { 1179 case Instruction::SDiv: 1180 case Instruction::UDiv: 1181 case Instruction::URem: 1182 case Instruction::SRem: 1183 case Instruction::LShr: 1184 case Instruction::AShr: 1185 case Instruction::Shl: 1186 if (CI1->isZero()) return C1; 1187 break; 1188 default: 1189 break; 1190 } 1191 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1192 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1193 const APFloat &C1V = CFP1->getValueAPF(); 1194 const APFloat &C2V = CFP2->getValueAPF(); 1195 APFloat C3V = C1V; // copy for modification 1196 switch (Opcode) { 1197 default: 1198 break; 1199 case Instruction::FAdd: 1200 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1201 return ConstantFP::get(C1->getContext(), C3V); 1202 case Instruction::FSub: 1203 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1204 return ConstantFP::get(C1->getContext(), C3V); 1205 case Instruction::FMul: 1206 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1207 return ConstantFP::get(C1->getContext(), C3V); 1208 case Instruction::FDiv: 1209 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1210 return ConstantFP::get(C1->getContext(), C3V); 1211 case Instruction::FRem: 1212 (void)C3V.mod(C2V); 1213 return ConstantFP::get(C1->getContext(), C3V); 1214 } 1215 } 1216 } else if (auto *VTy = dyn_cast<VectorType>(C1->getType())) { 1217 // Fast path for splatted constants. 1218 if (Constant *C2Splat = C2->getSplatValue()) { 1219 if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue()) 1220 return PoisonValue::get(VTy); 1221 if (Constant *C1Splat = C1->getSplatValue()) { 1222 return ConstantVector::getSplat( 1223 VTy->getElementCount(), 1224 ConstantExpr::get(Opcode, C1Splat, C2Splat)); 1225 } 1226 } 1227 1228 if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) { 1229 // Fold each element and create a vector constant from those constants. 1230 SmallVector<Constant*, 16> Result; 1231 Type *Ty = IntegerType::get(FVTy->getContext(), 32); 1232 for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) { 1233 Constant *ExtractIdx = ConstantInt::get(Ty, i); 1234 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx); 1235 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx); 1236 1237 // If any element of a divisor vector is zero, the whole op is poison. 1238 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue()) 1239 return PoisonValue::get(VTy); 1240 1241 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1242 } 1243 1244 return ConstantVector::get(Result); 1245 } 1246 } 1247 1248 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1249 // There are many possible foldings we could do here. We should probably 1250 // at least fold add of a pointer with an integer into the appropriate 1251 // getelementptr. This will improve alias analysis a bit. 1252 1253 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1254 // (a + (b + c)). 1255 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1256 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1257 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1258 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1259 } 1260 } else if (isa<ConstantExpr>(C2)) { 1261 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1262 // other way if possible. 1263 if (Instruction::isCommutative(Opcode)) 1264 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1265 } 1266 1267 // i1 can be simplified in many cases. 1268 if (C1->getType()->isIntegerTy(1)) { 1269 switch (Opcode) { 1270 case Instruction::Add: 1271 case Instruction::Sub: 1272 return ConstantExpr::getXor(C1, C2); 1273 case Instruction::Mul: 1274 return ConstantExpr::getAnd(C1, C2); 1275 case Instruction::Shl: 1276 case Instruction::LShr: 1277 case Instruction::AShr: 1278 // We can assume that C2 == 0. If it were one the result would be 1279 // undefined because the shift value is as large as the bitwidth. 1280 return C1; 1281 case Instruction::SDiv: 1282 case Instruction::UDiv: 1283 // We can assume that C2 == 1. If it were zero the result would be 1284 // undefined through division by zero. 1285 return C1; 1286 case Instruction::URem: 1287 case Instruction::SRem: 1288 // We can assume that C2 == 1. If it were zero the result would be 1289 // undefined through division by zero. 1290 return ConstantInt::getFalse(C1->getContext()); 1291 default: 1292 break; 1293 } 1294 } 1295 1296 // We don't know how to fold this. 1297 return nullptr; 1298 } 1299 1300 /// This function determines if there is anything we can decide about the two 1301 /// constants provided. This doesn't need to handle simple things like 1302 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1303 /// If we can determine that the two constants have a particular relation to 1304 /// each other, we should return the corresponding FCmpInst predicate, 1305 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1306 /// ConstantFoldCompareInstruction. 1307 /// 1308 /// To simplify this code we canonicalize the relation so that the first 1309 /// operand is always the most "complex" of the two. We consider ConstantFP 1310 /// to be the simplest, and ConstantExprs to be the most complex. 1311 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1312 assert(V1->getType() == V2->getType() && 1313 "Cannot compare values of different types!"); 1314 1315 // We do not know if a constant expression will evaluate to a number or NaN. 1316 // Therefore, we can only say that the relation is unordered or equal. 1317 if (V1 == V2) return FCmpInst::FCMP_UEQ; 1318 1319 if (!isa<ConstantExpr>(V1)) { 1320 if (!isa<ConstantExpr>(V2)) { 1321 // Simple case, use the standard constant folder. 1322 ConstantInt *R = nullptr; 1323 R = dyn_cast<ConstantInt>( 1324 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1325 if (R && !R->isZero()) 1326 return FCmpInst::FCMP_OEQ; 1327 R = dyn_cast<ConstantInt>( 1328 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1329 if (R && !R->isZero()) 1330 return FCmpInst::FCMP_OLT; 1331 R = dyn_cast<ConstantInt>( 1332 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1333 if (R && !R->isZero()) 1334 return FCmpInst::FCMP_OGT; 1335 1336 // Nothing more we can do 1337 return FCmpInst::BAD_FCMP_PREDICATE; 1338 } 1339 1340 // If the first operand is simple and second is ConstantExpr, swap operands. 1341 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1342 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1343 return FCmpInst::getSwappedPredicate(SwappedRelation); 1344 } else { 1345 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1346 // constantexpr or a simple constant. 1347 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1348 switch (CE1->getOpcode()) { 1349 case Instruction::FPTrunc: 1350 case Instruction::FPExt: 1351 case Instruction::UIToFP: 1352 case Instruction::SIToFP: 1353 // We might be able to do something with these but we don't right now. 1354 break; 1355 default: 1356 break; 1357 } 1358 } 1359 // There are MANY other foldings that we could perform here. They will 1360 // probably be added on demand, as they seem needed. 1361 return FCmpInst::BAD_FCMP_PREDICATE; 1362 } 1363 1364 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1365 const GlobalValue *GV2) { 1366 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1367 if (GV->isInterposable() || GV->hasGlobalUnnamedAddr()) 1368 return true; 1369 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1370 Type *Ty = GVar->getValueType(); 1371 // A global with opaque type might end up being zero sized. 1372 if (!Ty->isSized()) 1373 return true; 1374 // A global with an empty type might lie at the address of any other 1375 // global. 1376 if (Ty->isEmptyTy()) 1377 return true; 1378 } 1379 return false; 1380 }; 1381 // Don't try to decide equality of aliases. 1382 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1383 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1384 return ICmpInst::ICMP_NE; 1385 return ICmpInst::BAD_ICMP_PREDICATE; 1386 } 1387 1388 /// This function determines if there is anything we can decide about the two 1389 /// constants provided. This doesn't need to handle simple things like integer 1390 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1391 /// If we can determine that the two constants have a particular relation to 1392 /// each other, we should return the corresponding ICmp predicate, otherwise 1393 /// return ICmpInst::BAD_ICMP_PREDICATE. 1394 /// 1395 /// To simplify this code we canonicalize the relation so that the first 1396 /// operand is always the most "complex" of the two. We consider simple 1397 /// constants (like ConstantInt) to be the simplest, followed by 1398 /// GlobalValues, followed by ConstantExpr's (the most complex). 1399 /// 1400 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1401 bool isSigned) { 1402 assert(V1->getType() == V2->getType() && 1403 "Cannot compare different types of values!"); 1404 if (V1 == V2) return ICmpInst::ICMP_EQ; 1405 1406 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1407 !isa<BlockAddress>(V1)) { 1408 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1409 !isa<BlockAddress>(V2)) { 1410 // We distilled this down to a simple case, use the standard constant 1411 // folder. 1412 ConstantInt *R = nullptr; 1413 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1414 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1415 if (R && !R->isZero()) 1416 return pred; 1417 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1418 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1419 if (R && !R->isZero()) 1420 return pred; 1421 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1422 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1423 if (R && !R->isZero()) 1424 return pred; 1425 1426 // If we couldn't figure it out, bail. 1427 return ICmpInst::BAD_ICMP_PREDICATE; 1428 } 1429 1430 // If the first operand is simple, swap operands. 1431 ICmpInst::Predicate SwappedRelation = 1432 evaluateICmpRelation(V2, V1, isSigned); 1433 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1434 return ICmpInst::getSwappedPredicate(SwappedRelation); 1435 1436 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1437 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1438 ICmpInst::Predicate SwappedRelation = 1439 evaluateICmpRelation(V2, V1, isSigned); 1440 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1441 return ICmpInst::getSwappedPredicate(SwappedRelation); 1442 return ICmpInst::BAD_ICMP_PREDICATE; 1443 } 1444 1445 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1446 // constant (which, since the types must match, means that it's a 1447 // ConstantPointerNull). 1448 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1449 return areGlobalsPotentiallyEqual(GV, GV2); 1450 } else if (isa<BlockAddress>(V2)) { 1451 return ICmpInst::ICMP_NE; // Globals never equal labels. 1452 } else { 1453 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1454 // GlobalVals can never be null unless they have external weak linkage. 1455 // We don't try to evaluate aliases here. 1456 // NOTE: We should not be doing this constant folding if null pointer 1457 // is considered valid for the function. But currently there is no way to 1458 // query it from the Constant type. 1459 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) && 1460 !NullPointerIsDefined(nullptr /* F */, 1461 GV->getType()->getAddressSpace())) 1462 return ICmpInst::ICMP_UGT; 1463 } 1464 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1465 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1466 ICmpInst::Predicate SwappedRelation = 1467 evaluateICmpRelation(V2, V1, isSigned); 1468 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1469 return ICmpInst::getSwappedPredicate(SwappedRelation); 1470 return ICmpInst::BAD_ICMP_PREDICATE; 1471 } 1472 1473 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1474 // constant (which, since the types must match, means that it is a 1475 // ConstantPointerNull). 1476 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1477 // Block address in another function can't equal this one, but block 1478 // addresses in the current function might be the same if blocks are 1479 // empty. 1480 if (BA2->getFunction() != BA->getFunction()) 1481 return ICmpInst::ICMP_NE; 1482 } else { 1483 // Block addresses aren't null, don't equal the address of globals. 1484 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1485 "Canonicalization guarantee!"); 1486 return ICmpInst::ICMP_NE; 1487 } 1488 } else { 1489 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1490 // constantexpr, a global, block address, or a simple constant. 1491 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1492 Constant *CE1Op0 = CE1->getOperand(0); 1493 1494 switch (CE1->getOpcode()) { 1495 case Instruction::Trunc: 1496 case Instruction::FPTrunc: 1497 case Instruction::FPExt: 1498 case Instruction::FPToUI: 1499 case Instruction::FPToSI: 1500 break; // We can't evaluate floating point casts or truncations. 1501 1502 case Instruction::BitCast: 1503 // If this is a global value cast, check to see if the RHS is also a 1504 // GlobalValue. 1505 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) 1506 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) 1507 return areGlobalsPotentiallyEqual(GV, GV2); 1508 LLVM_FALLTHROUGH; 1509 case Instruction::UIToFP: 1510 case Instruction::SIToFP: 1511 case Instruction::ZExt: 1512 case Instruction::SExt: 1513 // We can't evaluate floating point casts or truncations. 1514 if (CE1Op0->getType()->isFPOrFPVectorTy()) 1515 break; 1516 1517 // If the cast is not actually changing bits, and the second operand is a 1518 // null pointer, do the comparison with the pre-casted value. 1519 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) { 1520 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1521 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1522 return evaluateICmpRelation(CE1Op0, 1523 Constant::getNullValue(CE1Op0->getType()), 1524 isSigned); 1525 } 1526 break; 1527 1528 case Instruction::GetElementPtr: { 1529 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1530 // Ok, since this is a getelementptr, we know that the constant has a 1531 // pointer type. Check the various cases. 1532 if (isa<ConstantPointerNull>(V2)) { 1533 // If we are comparing a GEP to a null pointer, check to see if the base 1534 // of the GEP equals the null pointer. 1535 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1536 // If its not weak linkage, the GVal must have a non-zero address 1537 // so the result is greater-than 1538 if (!GV->hasExternalWeakLinkage() && CE1GEP->isInBounds()) 1539 return ICmpInst::ICMP_UGT; 1540 } 1541 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1542 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1543 if (GV != GV2) { 1544 if (CE1GEP->hasAllZeroIndices()) 1545 return areGlobalsPotentiallyEqual(GV, GV2); 1546 return ICmpInst::BAD_ICMP_PREDICATE; 1547 } 1548 } 1549 } else if (const auto *CE2GEP = dyn_cast<GEPOperator>(V2)) { 1550 // By far the most common case to handle is when the base pointers are 1551 // obviously to the same global. 1552 const Constant *CE2Op0 = cast<Constant>(CE2GEP->getPointerOperand()); 1553 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1554 // Don't know relative ordering, but check for inequality. 1555 if (CE1Op0 != CE2Op0) { 1556 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1557 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1558 cast<GlobalValue>(CE2Op0)); 1559 return ICmpInst::BAD_ICMP_PREDICATE; 1560 } 1561 } 1562 } 1563 break; 1564 } 1565 default: 1566 break; 1567 } 1568 } 1569 1570 return ICmpInst::BAD_ICMP_PREDICATE; 1571 } 1572 1573 Constant *llvm::ConstantFoldCompareInstruction(CmpInst::Predicate Predicate, 1574 Constant *C1, Constant *C2) { 1575 Type *ResultTy; 1576 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1577 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1578 VT->getElementCount()); 1579 else 1580 ResultTy = Type::getInt1Ty(C1->getContext()); 1581 1582 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1583 if (Predicate == FCmpInst::FCMP_FALSE) 1584 return Constant::getNullValue(ResultTy); 1585 1586 if (Predicate == FCmpInst::FCMP_TRUE) 1587 return Constant::getAllOnesValue(ResultTy); 1588 1589 // Handle some degenerate cases first 1590 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2)) 1591 return PoisonValue::get(ResultTy); 1592 1593 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1594 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1595 // For EQ and NE, we can always pick a value for the undef to make the 1596 // predicate pass or fail, so we can return undef. 1597 // Also, if both operands are undef, we can return undef for int comparison. 1598 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1599 return UndefValue::get(ResultTy); 1600 1601 // Otherwise, for integer compare, pick the same value as the non-undef 1602 // operand, and fold it to true or false. 1603 if (isIntegerPredicate) 1604 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1605 1606 // Choosing NaN for the undef will always make unordered comparison succeed 1607 // and ordered comparison fails. 1608 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1609 } 1610 1611 // icmp eq/ne(null,GV) -> false/true 1612 if (C1->isNullValue()) { 1613 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1614 // Don't try to evaluate aliases. External weak GV can be null. 1615 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1616 !NullPointerIsDefined(nullptr /* F */, 1617 GV->getType()->getAddressSpace())) { 1618 if (Predicate == ICmpInst::ICMP_EQ) 1619 return ConstantInt::getFalse(C1->getContext()); 1620 else if (Predicate == ICmpInst::ICMP_NE) 1621 return ConstantInt::getTrue(C1->getContext()); 1622 } 1623 // icmp eq/ne(GV,null) -> false/true 1624 } else if (C2->isNullValue()) { 1625 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) { 1626 // Don't try to evaluate aliases. External weak GV can be null. 1627 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1628 !NullPointerIsDefined(nullptr /* F */, 1629 GV->getType()->getAddressSpace())) { 1630 if (Predicate == ICmpInst::ICMP_EQ) 1631 return ConstantInt::getFalse(C1->getContext()); 1632 else if (Predicate == ICmpInst::ICMP_NE) 1633 return ConstantInt::getTrue(C1->getContext()); 1634 } 1635 } 1636 1637 // The caller is expected to commute the operands if the constant expression 1638 // is C2. 1639 // C1 >= 0 --> true 1640 if (Predicate == ICmpInst::ICMP_UGE) 1641 return Constant::getAllOnesValue(ResultTy); 1642 // C1 < 0 --> false 1643 if (Predicate == ICmpInst::ICMP_ULT) 1644 return Constant::getNullValue(ResultTy); 1645 } 1646 1647 // If the comparison is a comparison between two i1's, simplify it. 1648 if (C1->getType()->isIntegerTy(1)) { 1649 switch (Predicate) { 1650 case ICmpInst::ICMP_EQ: 1651 if (isa<ConstantInt>(C2)) 1652 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1653 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1654 case ICmpInst::ICMP_NE: 1655 return ConstantExpr::getXor(C1, C2); 1656 default: 1657 break; 1658 } 1659 } 1660 1661 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1662 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1663 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1664 return ConstantInt::get(ResultTy, ICmpInst::compare(V1, V2, Predicate)); 1665 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1666 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1667 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1668 return ConstantInt::get(ResultTy, FCmpInst::compare(C1V, C2V, Predicate)); 1669 } else if (auto *C1VTy = dyn_cast<VectorType>(C1->getType())) { 1670 1671 // Fast path for splatted constants. 1672 if (Constant *C1Splat = C1->getSplatValue()) 1673 if (Constant *C2Splat = C2->getSplatValue()) 1674 return ConstantVector::getSplat( 1675 C1VTy->getElementCount(), 1676 ConstantExpr::getCompare(Predicate, C1Splat, C2Splat)); 1677 1678 // Do not iterate on scalable vector. The number of elements is unknown at 1679 // compile-time. 1680 if (isa<ScalableVectorType>(C1VTy)) 1681 return nullptr; 1682 1683 // If we can constant fold the comparison of each element, constant fold 1684 // the whole vector comparison. 1685 SmallVector<Constant*, 4> ResElts; 1686 Type *Ty = IntegerType::get(C1->getContext(), 32); 1687 // Compare the elements, producing an i1 result or constant expr. 1688 for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue(); 1689 I != E; ++I) { 1690 Constant *C1E = 1691 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, I)); 1692 Constant *C2E = 1693 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, I)); 1694 1695 ResElts.push_back(ConstantExpr::getCompare(Predicate, C1E, C2E)); 1696 } 1697 1698 return ConstantVector::get(ResElts); 1699 } 1700 1701 if (C1->getType()->isFloatingPointTy() && 1702 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1703 // infinite recursive loop 1704 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1705 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1706 switch (evaluateFCmpRelation(C1, C2)) { 1707 default: llvm_unreachable("Unknown relation!"); 1708 case FCmpInst::FCMP_UNO: 1709 case FCmpInst::FCMP_ORD: 1710 case FCmpInst::FCMP_UNE: 1711 case FCmpInst::FCMP_ULT: 1712 case FCmpInst::FCMP_UGT: 1713 case FCmpInst::FCMP_ULE: 1714 case FCmpInst::FCMP_UGE: 1715 case FCmpInst::FCMP_TRUE: 1716 case FCmpInst::FCMP_FALSE: 1717 case FCmpInst::BAD_FCMP_PREDICATE: 1718 break; // Couldn't determine anything about these constants. 1719 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1720 Result = 1721 (Predicate == FCmpInst::FCMP_UEQ || Predicate == FCmpInst::FCMP_OEQ || 1722 Predicate == FCmpInst::FCMP_ULE || Predicate == FCmpInst::FCMP_OLE || 1723 Predicate == FCmpInst::FCMP_UGE || Predicate == FCmpInst::FCMP_OGE); 1724 break; 1725 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1726 Result = 1727 (Predicate == FCmpInst::FCMP_UNE || Predicate == FCmpInst::FCMP_ONE || 1728 Predicate == FCmpInst::FCMP_ULT || Predicate == FCmpInst::FCMP_OLT || 1729 Predicate == FCmpInst::FCMP_ULE || Predicate == FCmpInst::FCMP_OLE); 1730 break; 1731 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1732 Result = 1733 (Predicate == FCmpInst::FCMP_UNE || Predicate == FCmpInst::FCMP_ONE || 1734 Predicate == FCmpInst::FCMP_UGT || Predicate == FCmpInst::FCMP_OGT || 1735 Predicate == FCmpInst::FCMP_UGE || Predicate == FCmpInst::FCMP_OGE); 1736 break; 1737 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1738 // We can only partially decide this relation. 1739 if (Predicate == FCmpInst::FCMP_UGT || Predicate == FCmpInst::FCMP_OGT) 1740 Result = 0; 1741 else if (Predicate == FCmpInst::FCMP_ULT || 1742 Predicate == FCmpInst::FCMP_OLT) 1743 Result = 1; 1744 break; 1745 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1746 // We can only partially decide this relation. 1747 if (Predicate == FCmpInst::FCMP_ULT || Predicate == FCmpInst::FCMP_OLT) 1748 Result = 0; 1749 else if (Predicate == FCmpInst::FCMP_UGT || 1750 Predicate == FCmpInst::FCMP_OGT) 1751 Result = 1; 1752 break; 1753 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1754 // We can only partially decide this relation. 1755 if (Predicate == FCmpInst::FCMP_OEQ || Predicate == FCmpInst::FCMP_UEQ) 1756 Result = 0; 1757 else if (Predicate == FCmpInst::FCMP_ONE || 1758 Predicate == FCmpInst::FCMP_UNE) 1759 Result = 1; 1760 break; 1761 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2). 1762 // We can only partially decide this relation. 1763 if (Predicate == FCmpInst::FCMP_ONE) 1764 Result = 0; 1765 else if (Predicate == FCmpInst::FCMP_UEQ) 1766 Result = 1; 1767 break; 1768 } 1769 1770 // If we evaluated the result, return it now. 1771 if (Result != -1) 1772 return ConstantInt::get(ResultTy, Result); 1773 1774 } else { 1775 // Evaluate the relation between the two constants, per the predicate. 1776 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1777 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(Predicate))) { 1778 default: llvm_unreachable("Unknown relational!"); 1779 case ICmpInst::BAD_ICMP_PREDICATE: 1780 break; // Couldn't determine anything about these constants. 1781 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1782 // If we know the constants are equal, we can decide the result of this 1783 // computation precisely. 1784 Result = ICmpInst::isTrueWhenEqual(Predicate); 1785 break; 1786 case ICmpInst::ICMP_ULT: 1787 switch (Predicate) { 1788 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1789 Result = 1; break; 1790 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1791 Result = 0; break; 1792 default: 1793 break; 1794 } 1795 break; 1796 case ICmpInst::ICMP_SLT: 1797 switch (Predicate) { 1798 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1799 Result = 1; break; 1800 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1801 Result = 0; break; 1802 default: 1803 break; 1804 } 1805 break; 1806 case ICmpInst::ICMP_UGT: 1807 switch (Predicate) { 1808 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1809 Result = 1; break; 1810 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1811 Result = 0; break; 1812 default: 1813 break; 1814 } 1815 break; 1816 case ICmpInst::ICMP_SGT: 1817 switch (Predicate) { 1818 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1819 Result = 1; break; 1820 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1821 Result = 0; break; 1822 default: 1823 break; 1824 } 1825 break; 1826 case ICmpInst::ICMP_ULE: 1827 if (Predicate == ICmpInst::ICMP_UGT) 1828 Result = 0; 1829 if (Predicate == ICmpInst::ICMP_ULT || Predicate == ICmpInst::ICMP_ULE) 1830 Result = 1; 1831 break; 1832 case ICmpInst::ICMP_SLE: 1833 if (Predicate == ICmpInst::ICMP_SGT) 1834 Result = 0; 1835 if (Predicate == ICmpInst::ICMP_SLT || Predicate == ICmpInst::ICMP_SLE) 1836 Result = 1; 1837 break; 1838 case ICmpInst::ICMP_UGE: 1839 if (Predicate == ICmpInst::ICMP_ULT) 1840 Result = 0; 1841 if (Predicate == ICmpInst::ICMP_UGT || Predicate == ICmpInst::ICMP_UGE) 1842 Result = 1; 1843 break; 1844 case ICmpInst::ICMP_SGE: 1845 if (Predicate == ICmpInst::ICMP_SLT) 1846 Result = 0; 1847 if (Predicate == ICmpInst::ICMP_SGT || Predicate == ICmpInst::ICMP_SGE) 1848 Result = 1; 1849 break; 1850 case ICmpInst::ICMP_NE: 1851 if (Predicate == ICmpInst::ICMP_EQ) 1852 Result = 0; 1853 if (Predicate == ICmpInst::ICMP_NE) 1854 Result = 1; 1855 break; 1856 } 1857 1858 // If we evaluated the result, return it now. 1859 if (Result != -1) 1860 return ConstantInt::get(ResultTy, Result); 1861 1862 // If the right hand side is a bitcast, try using its inverse to simplify 1863 // it by moving it to the left hand side. We can't do this if it would turn 1864 // a vector compare into a scalar compare or visa versa, or if it would turn 1865 // the operands into FP values. 1866 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1867 Constant *CE2Op0 = CE2->getOperand(0); 1868 if (CE2->getOpcode() == Instruction::BitCast && 1869 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() && 1870 !CE2Op0->getType()->isFPOrFPVectorTy()) { 1871 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1872 return ConstantExpr::getICmp(Predicate, Inverse, CE2Op0); 1873 } 1874 } 1875 1876 // If the left hand side is an extension, try eliminating it. 1877 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1878 if ((CE1->getOpcode() == Instruction::SExt && 1879 ICmpInst::isSigned(Predicate)) || 1880 (CE1->getOpcode() == Instruction::ZExt && 1881 !ICmpInst::isSigned(Predicate))) { 1882 Constant *CE1Op0 = CE1->getOperand(0); 1883 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1884 if (CE1Inverse == CE1Op0) { 1885 // Check whether we can safely truncate the right hand side. 1886 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1887 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1888 C2->getType()) == C2) 1889 return ConstantExpr::getICmp(Predicate, CE1Inverse, C2Inverse); 1890 } 1891 } 1892 } 1893 1894 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1895 (C1->isNullValue() && !C2->isNullValue())) { 1896 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1897 // other way if possible. 1898 // Also, if C1 is null and C2 isn't, flip them around. 1899 Predicate = ICmpInst::getSwappedPredicate(Predicate); 1900 return ConstantExpr::getICmp(Predicate, C2, C1); 1901 } 1902 } 1903 return nullptr; 1904 } 1905 1906 /// Test whether the given sequence of *normalized* indices is "inbounds". 1907 template<typename IndexTy> 1908 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1909 // No indices means nothing that could be out of bounds. 1910 if (Idxs.empty()) return true; 1911 1912 // If the first index is zero, it's in bounds. 1913 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1914 1915 // If the first index is one and all the rest are zero, it's in bounds, 1916 // by the one-past-the-end rule. 1917 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) { 1918 if (!CI->isOne()) 1919 return false; 1920 } else { 1921 auto *CV = cast<ConstantDataVector>(Idxs[0]); 1922 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue()); 1923 if (!CI || !CI->isOne()) 1924 return false; 1925 } 1926 1927 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1928 if (!cast<Constant>(Idxs[i])->isNullValue()) 1929 return false; 1930 return true; 1931 } 1932 1933 /// Test whether a given ConstantInt is in-range for a SequentialType. 1934 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 1935 const ConstantInt *CI) { 1936 // We cannot bounds check the index if it doesn't fit in an int64_t. 1937 if (CI->getValue().getMinSignedBits() > 64) 1938 return false; 1939 1940 // A negative index or an index past the end of our sequential type is 1941 // considered out-of-range. 1942 int64_t IndexVal = CI->getSExtValue(); 1943 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 1944 return false; 1945 1946 // Otherwise, it is in-range. 1947 return true; 1948 } 1949 1950 // Combine Indices - If the source pointer to this getelementptr instruction 1951 // is a getelementptr instruction, combine the indices of the two 1952 // getelementptr instructions into a single instruction. 1953 static Constant *foldGEPOfGEP(GEPOperator *GEP, Type *PointeeTy, bool InBounds, 1954 ArrayRef<Value *> Idxs) { 1955 if (PointeeTy != GEP->getResultElementType()) 1956 return nullptr; 1957 1958 Constant *Idx0 = cast<Constant>(Idxs[0]); 1959 if (Idx0->isNullValue()) { 1960 // Handle the simple case of a zero index. 1961 SmallVector<Value*, 16> NewIndices; 1962 NewIndices.reserve(Idxs.size() + GEP->getNumIndices()); 1963 NewIndices.append(GEP->idx_begin(), GEP->idx_end()); 1964 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 1965 return ConstantExpr::getGetElementPtr( 1966 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()), 1967 NewIndices, InBounds && GEP->isInBounds(), GEP->getInRangeIndex()); 1968 } 1969 1970 gep_type_iterator LastI = gep_type_end(GEP); 1971 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP); 1972 I != E; ++I) 1973 LastI = I; 1974 1975 // We can't combine GEPs if the last index is a struct type. 1976 if (!LastI.isSequential()) 1977 return nullptr; 1978 // We could perform the transform with non-constant index, but prefer leaving 1979 // it as GEP of GEP rather than GEP of add for now. 1980 ConstantInt *CI = dyn_cast<ConstantInt>(Idx0); 1981 if (!CI) 1982 return nullptr; 1983 1984 // TODO: This code may be extended to handle vectors as well. 1985 auto *LastIdx = cast<Constant>(GEP->getOperand(GEP->getNumOperands()-1)); 1986 Type *LastIdxTy = LastIdx->getType(); 1987 if (LastIdxTy->isVectorTy()) 1988 return nullptr; 1989 1990 SmallVector<Value*, 16> NewIndices; 1991 NewIndices.reserve(Idxs.size() + GEP->getNumIndices()); 1992 NewIndices.append(GEP->idx_begin(), GEP->idx_end() - 1); 1993 1994 // Add the last index of the source with the first index of the new GEP. 1995 // Make sure to handle the case when they are actually different types. 1996 if (LastIdxTy != Idx0->getType()) { 1997 unsigned CommonExtendedWidth = 1998 std::max(LastIdxTy->getIntegerBitWidth(), 1999 Idx0->getType()->getIntegerBitWidth()); 2000 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2001 2002 Type *CommonTy = 2003 Type::getIntNTy(LastIdxTy->getContext(), CommonExtendedWidth); 2004 Idx0 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2005 LastIdx = ConstantExpr::getSExtOrBitCast(LastIdx, CommonTy); 2006 } 2007 2008 NewIndices.push_back(ConstantExpr::get(Instruction::Add, Idx0, LastIdx)); 2009 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2010 2011 // The combined GEP normally inherits its index inrange attribute from 2012 // the inner GEP, but if the inner GEP's last index was adjusted by the 2013 // outer GEP, any inbounds attribute on that index is invalidated. 2014 Optional<unsigned> IRIndex = GEP->getInRangeIndex(); 2015 if (IRIndex && *IRIndex == GEP->getNumIndices() - 1) 2016 IRIndex = None; 2017 2018 return ConstantExpr::getGetElementPtr( 2019 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()), 2020 NewIndices, InBounds && GEP->isInBounds(), IRIndex); 2021 } 2022 2023 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 2024 bool InBounds, 2025 Optional<unsigned> InRangeIndex, 2026 ArrayRef<Value *> Idxs) { 2027 if (Idxs.empty()) return C; 2028 2029 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2030 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size())); 2031 2032 if (isa<PoisonValue>(C)) 2033 return PoisonValue::get(GEPTy); 2034 2035 if (isa<UndefValue>(C)) 2036 // If inbounds, we can choose an out-of-bounds pointer as a base pointer. 2037 return InBounds ? PoisonValue::get(GEPTy) : UndefValue::get(GEPTy); 2038 2039 Constant *Idx0 = cast<Constant>(Idxs[0]); 2040 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0))) 2041 return GEPTy->isVectorTy() && !C->getType()->isVectorTy() 2042 ? ConstantVector::getSplat( 2043 cast<VectorType>(GEPTy)->getElementCount(), C) 2044 : C; 2045 2046 if (C->isNullValue()) { 2047 bool isNull = true; 2048 for (Value *Idx : Idxs) 2049 if (!isa<UndefValue>(Idx) && !cast<Constant>(Idx)->isNullValue()) { 2050 isNull = false; 2051 break; 2052 } 2053 if (isNull) { 2054 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2055 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2056 2057 assert(Ty && "Invalid indices for GEP!"); 2058 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2059 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2060 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2061 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount()); 2062 2063 // The GEP returns a vector of pointers when one of more of 2064 // its arguments is a vector. 2065 for (Value *Idx : Idxs) { 2066 if (auto *VT = dyn_cast<VectorType>(Idx->getType())) { 2067 assert((!isa<VectorType>(GEPTy) || isa<ScalableVectorType>(GEPTy) == 2068 isa<ScalableVectorType>(VT)) && 2069 "Mismatched GEPTy vector types"); 2070 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount()); 2071 break; 2072 } 2073 } 2074 2075 return Constant::getNullValue(GEPTy); 2076 } 2077 } 2078 2079 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2080 if (auto *GEP = dyn_cast<GEPOperator>(CE)) 2081 if (Constant *C = foldGEPOfGEP(GEP, PointeeTy, InBounds, Idxs)) 2082 return C; 2083 2084 // Attempt to fold casts to the same type away. For example, folding: 2085 // 2086 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2087 // i64 0, i64 0) 2088 // into: 2089 // 2090 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2091 // 2092 // Don't fold if the cast is changing address spaces. 2093 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2094 PointerType *SrcPtrTy = 2095 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2096 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2097 if (SrcPtrTy && DstPtrTy && !SrcPtrTy->isOpaque() && 2098 !DstPtrTy->isOpaque()) { 2099 ArrayType *SrcArrayTy = 2100 dyn_cast<ArrayType>(SrcPtrTy->getNonOpaquePointerElementType()); 2101 ArrayType *DstArrayTy = 2102 dyn_cast<ArrayType>(DstPtrTy->getNonOpaquePointerElementType()); 2103 if (SrcArrayTy && DstArrayTy 2104 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2105 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2106 return ConstantExpr::getGetElementPtr(SrcArrayTy, 2107 (Constant *)CE->getOperand(0), 2108 Idxs, InBounds, InRangeIndex); 2109 } 2110 } 2111 } 2112 2113 // Check to see if any array indices are not within the corresponding 2114 // notional array or vector bounds. If so, try to determine if they can be 2115 // factored out into preceding dimensions. 2116 SmallVector<Constant *, 8> NewIdxs; 2117 Type *Ty = PointeeTy; 2118 Type *Prev = C->getType(); 2119 auto GEPIter = gep_type_begin(PointeeTy, Idxs); 2120 bool Unknown = 2121 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]); 2122 for (unsigned i = 1, e = Idxs.size(); i != e; 2123 Prev = Ty, Ty = (++GEPIter).getIndexedType(), ++i) { 2124 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) { 2125 // We don't know if it's in range or not. 2126 Unknown = true; 2127 continue; 2128 } 2129 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1])) 2130 // Skip if the type of the previous index is not supported. 2131 continue; 2132 if (InRangeIndex && i == *InRangeIndex + 1) { 2133 // If an index is marked inrange, we cannot apply this canonicalization to 2134 // the following index, as that will cause the inrange index to point to 2135 // the wrong element. 2136 continue; 2137 } 2138 if (isa<StructType>(Ty)) { 2139 // The verify makes sure that GEPs into a struct are in range. 2140 continue; 2141 } 2142 if (isa<VectorType>(Ty)) { 2143 // There can be awkward padding in after a non-power of two vector. 2144 Unknown = true; 2145 continue; 2146 } 2147 auto *STy = cast<ArrayType>(Ty); 2148 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2149 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2150 // It's in range, skip to the next index. 2151 continue; 2152 if (CI->isNegative()) { 2153 // It's out of range and negative, don't try to factor it. 2154 Unknown = true; 2155 continue; 2156 } 2157 } else { 2158 auto *CV = cast<ConstantDataVector>(Idxs[i]); 2159 bool InRange = true; 2160 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) { 2161 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I)); 2162 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI); 2163 if (CI->isNegative()) { 2164 Unknown = true; 2165 break; 2166 } 2167 } 2168 if (InRange || Unknown) 2169 // It's in range, skip to the next index. 2170 // It's out of range and negative, don't try to factor it. 2171 continue; 2172 } 2173 if (isa<StructType>(Prev)) { 2174 // It's out of range, but the prior dimension is a struct 2175 // so we can't do anything about it. 2176 Unknown = true; 2177 continue; 2178 } 2179 // It's out of range, but we can factor it into the prior 2180 // dimension. 2181 NewIdxs.resize(Idxs.size()); 2182 // Determine the number of elements in our sequential type. 2183 uint64_t NumElements = STy->getArrayNumElements(); 2184 2185 // Expand the current index or the previous index to a vector from a scalar 2186 // if necessary. 2187 Constant *CurrIdx = cast<Constant>(Idxs[i]); 2188 auto *PrevIdx = 2189 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]); 2190 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy(); 2191 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy(); 2192 bool UseVector = IsCurrIdxVector || IsPrevIdxVector; 2193 2194 if (!IsCurrIdxVector && IsPrevIdxVector) 2195 CurrIdx = ConstantDataVector::getSplat( 2196 cast<FixedVectorType>(PrevIdx->getType())->getNumElements(), CurrIdx); 2197 2198 if (!IsPrevIdxVector && IsCurrIdxVector) 2199 PrevIdx = ConstantDataVector::getSplat( 2200 cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), PrevIdx); 2201 2202 Constant *Factor = 2203 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements); 2204 if (UseVector) 2205 Factor = ConstantDataVector::getSplat( 2206 IsPrevIdxVector 2207 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements() 2208 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), 2209 Factor); 2210 2211 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor); 2212 2213 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor); 2214 2215 unsigned CommonExtendedWidth = 2216 std::max(PrevIdx->getType()->getScalarSizeInBits(), 2217 Div->getType()->getScalarSizeInBits()); 2218 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2219 2220 // Before adding, extend both operands to i64 to avoid 2221 // overflow trouble. 2222 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth); 2223 if (UseVector) 2224 ExtendedTy = FixedVectorType::get( 2225 ExtendedTy, 2226 IsPrevIdxVector 2227 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements() 2228 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements()); 2229 2230 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2231 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy); 2232 2233 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2234 Div = ConstantExpr::getSExt(Div, ExtendedTy); 2235 2236 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2237 } 2238 2239 // If we did any factoring, start over with the adjusted indices. 2240 if (!NewIdxs.empty()) { 2241 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2242 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2243 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2244 InRangeIndex); 2245 } 2246 2247 // If all indices are known integers and normalized, we can do a simple 2248 // check for the "inbounds" property. 2249 if (!Unknown && !InBounds) 2250 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2251 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2252 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2253 /*InBounds=*/true, InRangeIndex); 2254 2255 return nullptr; 2256 } 2257