ConstantFold.cpp revision 280031
1//===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements folding of constants for LLVM. This implements the 11// (internal) ConstantFold.h interface, which is used by the 12// ConstantExpr::get* methods to automatically fold constants when possible. 13// 14// The current constant folding implementation is implemented in two pieces: the 15// pieces that don't need DataLayout, and the pieces that do. This is to avoid 16// a dependence in IR on Target. 17// 18//===----------------------------------------------------------------------===// 19 20#include "ConstantFold.h" 21#include "llvm/ADT/SmallVector.h" 22#include "llvm/IR/Constants.h" 23#include "llvm/IR/DerivedTypes.h" 24#include "llvm/IR/Function.h" 25#include "llvm/IR/GetElementPtrTypeIterator.h" 26#include "llvm/IR/GlobalAlias.h" 27#include "llvm/IR/GlobalVariable.h" 28#include "llvm/IR/Instructions.h" 29#include "llvm/IR/Operator.h" 30#include "llvm/IR/PatternMatch.h" 31#include "llvm/Support/Compiler.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/ManagedStatic.h" 34#include "llvm/Support/MathExtras.h" 35#include <limits> 36using namespace llvm; 37using namespace llvm::PatternMatch; 38 39//===----------------------------------------------------------------------===// 40// ConstantFold*Instruction Implementations 41//===----------------------------------------------------------------------===// 42 43/// BitCastConstantVector - Convert the specified vector Constant node to the 44/// specified vector type. At this point, we know that the elements of the 45/// input vector constant are all simple integer or FP values. 46static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 47 48 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 49 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 50 51 // If this cast changes element count then we can't handle it here: 52 // doing so requires endianness information. This should be handled by 53 // Analysis/ConstantFolding.cpp 54 unsigned NumElts = DstTy->getNumElements(); 55 if (NumElts != CV->getType()->getVectorNumElements()) 56 return nullptr; 57 58 Type *DstEltTy = DstTy->getElementType(); 59 60 SmallVector<Constant*, 16> Result; 61 Type *Ty = IntegerType::get(CV->getContext(), 32); 62 for (unsigned i = 0; i != NumElts; ++i) { 63 Constant *C = 64 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 65 C = ConstantExpr::getBitCast(C, DstEltTy); 66 Result.push_back(C); 67 } 68 69 return ConstantVector::get(Result); 70} 71 72/// This function determines which opcode to use to fold two constant cast 73/// expressions together. It uses CastInst::isEliminableCastPair to determine 74/// the opcode. Consequently its just a wrapper around that function. 75/// @brief Determine if it is valid to fold a cast of a cast 76static unsigned 77foldConstantCastPair( 78 unsigned opc, ///< opcode of the second cast constant expression 79 ConstantExpr *Op, ///< the first cast constant expression 80 Type *DstTy ///< destination type of the first cast 81) { 82 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 83 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 84 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 85 86 // The the types and opcodes for the two Cast constant expressions 87 Type *SrcTy = Op->getOperand(0)->getType(); 88 Type *MidTy = Op->getType(); 89 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 90 Instruction::CastOps secondOp = Instruction::CastOps(opc); 91 92 // Assume that pointers are never more than 64 bits wide, and only use this 93 // for the middle type. Otherwise we could end up folding away illegal 94 // bitcasts between address spaces with different sizes. 95 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 96 97 // Let CastInst::isEliminableCastPair do the heavy lifting. 98 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 99 nullptr, FakeIntPtrTy, nullptr); 100} 101 102static Constant *FoldBitCast(Constant *V, Type *DestTy) { 103 Type *SrcTy = V->getType(); 104 if (SrcTy == DestTy) 105 return V; // no-op cast 106 107 // Check to see if we are casting a pointer to an aggregate to a pointer to 108 // the first element. If so, return the appropriate GEP instruction. 109 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 110 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 111 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 112 && DPTy->getElementType()->isSized()) { 113 SmallVector<Value*, 8> IdxList; 114 Value *Zero = 115 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 116 IdxList.push_back(Zero); 117 Type *ElTy = PTy->getElementType(); 118 while (ElTy != DPTy->getElementType()) { 119 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 120 if (STy->getNumElements() == 0) break; 121 ElTy = STy->getElementType(0); 122 IdxList.push_back(Zero); 123 } else if (SequentialType *STy = 124 dyn_cast<SequentialType>(ElTy)) { 125 if (ElTy->isPointerTy()) break; // Can't index into pointers! 126 ElTy = STy->getElementType(); 127 IdxList.push_back(Zero); 128 } else { 129 break; 130 } 131 } 132 133 if (ElTy == DPTy->getElementType()) 134 // This GEP is inbounds because all indices are zero. 135 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); 136 } 137 138 // Handle casts from one vector constant to another. We know that the src 139 // and dest type have the same size (otherwise its an illegal cast). 140 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 141 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 142 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 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 if (DestTy->isFloatingPointTy()) 173 return ConstantFP::get(DestTy->getContext(), 174 APFloat(DestTy->getFltSemantics(), 175 CI->getValue())); 176 177 // Otherwise, can't fold this (vector?) 178 return nullptr; 179 } 180 181 // Handle ConstantFP input: FP -> Integral. 182 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 183 return ConstantInt::get(FP->getContext(), 184 FP->getValueAPF().bitcastToAPInt()); 185 186 return nullptr; 187} 188 189 190/// ExtractConstantBytes - V is an integer constant which only has a subset of 191/// its bytes used. The bytes used are indicated by ByteStart (which is the 192/// first byte used, counting from the least significant byte) and ByteSize, 193/// which is the number of bytes used. 194/// 195/// This function analyzes the specified constant to see if the specified byte 196/// range can be returned as a simplified constant. If so, the constant is 197/// returned, otherwise null is returned. 198/// 199static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 200 unsigned ByteSize) { 201 assert(C->getType()->isIntegerTy() && 202 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 203 "Non-byte sized integer input"); 204 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 205 assert(ByteSize && "Must be accessing some piece"); 206 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 207 assert(ByteSize != CSize && "Should not extract everything"); 208 209 // Constant Integers are simple. 210 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 211 APInt V = CI->getValue(); 212 if (ByteStart) 213 V = V.lshr(ByteStart*8); 214 V = V.trunc(ByteSize*8); 215 return ConstantInt::get(CI->getContext(), V); 216 } 217 218 // In the input is a constant expr, we might be able to recursively simplify. 219 // If not, we definitely can't do anything. 220 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 221 if (!CE) return nullptr; 222 223 switch (CE->getOpcode()) { 224 default: return nullptr; 225 case Instruction::Or: { 226 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 227 if (!RHS) 228 return nullptr; 229 230 // X | -1 -> -1. 231 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 232 if (RHSC->isAllOnesValue()) 233 return RHSC; 234 235 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 236 if (!LHS) 237 return nullptr; 238 return ConstantExpr::getOr(LHS, RHS); 239 } 240 case Instruction::And: { 241 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 242 if (!RHS) 243 return nullptr; 244 245 // X & 0 -> 0. 246 if (RHS->isNullValue()) 247 return RHS; 248 249 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 250 if (!LHS) 251 return nullptr; 252 return ConstantExpr::getAnd(LHS, RHS); 253 } 254 case Instruction::LShr: { 255 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 256 if (!Amt) 257 return nullptr; 258 unsigned ShAmt = Amt->getZExtValue(); 259 // Cannot analyze non-byte shifts. 260 if ((ShAmt & 7) != 0) 261 return nullptr; 262 ShAmt >>= 3; 263 264 // If the extract is known to be all zeros, return zero. 265 if (ByteStart >= CSize-ShAmt) 266 return Constant::getNullValue(IntegerType::get(CE->getContext(), 267 ByteSize*8)); 268 // If the extract is known to be fully in the input, extract it. 269 if (ByteStart+ByteSize+ShAmt <= CSize) 270 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 271 272 // TODO: Handle the 'partially zero' case. 273 return nullptr; 274 } 275 276 case Instruction::Shl: { 277 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 278 if (!Amt) 279 return nullptr; 280 unsigned ShAmt = Amt->getZExtValue(); 281 // Cannot analyze non-byte shifts. 282 if ((ShAmt & 7) != 0) 283 return nullptr; 284 ShAmt >>= 3; 285 286 // If the extract is known to be all zeros, return zero. 287 if (ByteStart+ByteSize <= ShAmt) 288 return Constant::getNullValue(IntegerType::get(CE->getContext(), 289 ByteSize*8)); 290 // If the extract is known to be fully in the input, extract it. 291 if (ByteStart >= ShAmt) 292 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 293 294 // TODO: Handle the 'partially zero' case. 295 return nullptr; 296 } 297 298 case Instruction::ZExt: { 299 unsigned SrcBitSize = 300 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 301 302 // If extracting something that is completely zero, return 0. 303 if (ByteStart*8 >= SrcBitSize) 304 return Constant::getNullValue(IntegerType::get(CE->getContext(), 305 ByteSize*8)); 306 307 // If exactly extracting the input, return it. 308 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 309 return CE->getOperand(0); 310 311 // If extracting something completely in the input, if if the input is a 312 // multiple of 8 bits, recurse. 313 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 314 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 315 316 // Otherwise, if extracting a subset of the input, which is not multiple of 317 // 8 bits, do a shift and trunc to get the bits. 318 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 319 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 320 Constant *Res = CE->getOperand(0); 321 if (ByteStart) 322 Res = ConstantExpr::getLShr(Res, 323 ConstantInt::get(Res->getType(), ByteStart*8)); 324 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 325 ByteSize*8)); 326 } 327 328 // TODO: Handle the 'partially zero' case. 329 return nullptr; 330 } 331 } 332} 333 334/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 335/// on Ty, with any known factors factored out. If Folded is false, 336/// return null if no factoring was possible, to avoid endlessly 337/// bouncing an unfoldable expression back into the top-level folder. 338/// 339static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 340 bool Folded) { 341 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 342 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 343 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 344 return ConstantExpr::getNUWMul(E, N); 345 } 346 347 if (StructType *STy = dyn_cast<StructType>(Ty)) 348 if (!STy->isPacked()) { 349 unsigned NumElems = STy->getNumElements(); 350 // An empty struct has size zero. 351 if (NumElems == 0) 352 return ConstantExpr::getNullValue(DestTy); 353 // Check for a struct with all members having the same size. 354 Constant *MemberSize = 355 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 356 bool AllSame = true; 357 for (unsigned i = 1; i != NumElems; ++i) 358 if (MemberSize != 359 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 360 AllSame = false; 361 break; 362 } 363 if (AllSame) { 364 Constant *N = ConstantInt::get(DestTy, NumElems); 365 return ConstantExpr::getNUWMul(MemberSize, N); 366 } 367 } 368 369 // Pointer size doesn't depend on the pointee type, so canonicalize them 370 // to an arbitrary pointee. 371 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 372 if (!PTy->getElementType()->isIntegerTy(1)) 373 return 374 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 375 PTy->getAddressSpace()), 376 DestTy, true); 377 378 // If there's no interesting folding happening, bail so that we don't create 379 // a constant that looks like it needs folding but really doesn't. 380 if (!Folded) 381 return nullptr; 382 383 // Base case: Get a regular sizeof expression. 384 Constant *C = ConstantExpr::getSizeOf(Ty); 385 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 386 DestTy, false), 387 C, DestTy); 388 return C; 389} 390 391/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 392/// on Ty, with any known factors factored out. If Folded is false, 393/// return null if no factoring was possible, to avoid endlessly 394/// bouncing an unfoldable expression back into the top-level folder. 395/// 396static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 397 bool Folded) { 398 // The alignment of an array is equal to the alignment of the 399 // array element. Note that this is not always true for vectors. 400 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 401 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 402 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 403 DestTy, 404 false), 405 C, DestTy); 406 return C; 407 } 408 409 if (StructType *STy = dyn_cast<StructType>(Ty)) { 410 // Packed structs always have an alignment of 1. 411 if (STy->isPacked()) 412 return ConstantInt::get(DestTy, 1); 413 414 // Otherwise, struct alignment is the maximum alignment of any member. 415 // Without target data, we can't compare much, but we can check to see 416 // if all the members have the same alignment. 417 unsigned NumElems = STy->getNumElements(); 418 // An empty struct has minimal alignment. 419 if (NumElems == 0) 420 return ConstantInt::get(DestTy, 1); 421 // Check for a struct with all members having the same alignment. 422 Constant *MemberAlign = 423 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 424 bool AllSame = true; 425 for (unsigned i = 1; i != NumElems; ++i) 426 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 427 AllSame = false; 428 break; 429 } 430 if (AllSame) 431 return MemberAlign; 432 } 433 434 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 435 // to an arbitrary pointee. 436 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 437 if (!PTy->getElementType()->isIntegerTy(1)) 438 return 439 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 440 1), 441 PTy->getAddressSpace()), 442 DestTy, true); 443 444 // If there's no interesting folding happening, bail so that we don't create 445 // a constant that looks like it needs folding but really doesn't. 446 if (!Folded) 447 return nullptr; 448 449 // Base case: Get a regular alignof expression. 450 Constant *C = ConstantExpr::getAlignOf(Ty); 451 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 452 DestTy, false), 453 C, DestTy); 454 return C; 455} 456 457/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 458/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 459/// return null if no factoring was possible, to avoid endlessly 460/// bouncing an unfoldable expression back into the top-level folder. 461/// 462static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 463 Type *DestTy, 464 bool Folded) { 465 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 466 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 467 DestTy, false), 468 FieldNo, DestTy); 469 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 470 return ConstantExpr::getNUWMul(E, N); 471 } 472 473 if (StructType *STy = dyn_cast<StructType>(Ty)) 474 if (!STy->isPacked()) { 475 unsigned NumElems = STy->getNumElements(); 476 // An empty struct has no members. 477 if (NumElems == 0) 478 return nullptr; 479 // Check for a struct with all members having the same size. 480 Constant *MemberSize = 481 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 482 bool AllSame = true; 483 for (unsigned i = 1; i != NumElems; ++i) 484 if (MemberSize != 485 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 486 AllSame = false; 487 break; 488 } 489 if (AllSame) { 490 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 491 false, 492 DestTy, 493 false), 494 FieldNo, DestTy); 495 return ConstantExpr::getNUWMul(MemberSize, N); 496 } 497 } 498 499 // If there's no interesting folding happening, bail so that we don't create 500 // a constant that looks like it needs folding but really doesn't. 501 if (!Folded) 502 return nullptr; 503 504 // Base case: Get a regular offsetof expression. 505 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 506 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 507 DestTy, false), 508 C, DestTy); 509 return C; 510} 511 512Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 513 Type *DestTy) { 514 if (isa<UndefValue>(V)) { 515 // zext(undef) = 0, because the top bits will be zero. 516 // sext(undef) = 0, because the top bits will all be the same. 517 // [us]itofp(undef) = 0, because the result value is bounded. 518 if (opc == Instruction::ZExt || opc == Instruction::SExt || 519 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 520 return Constant::getNullValue(DestTy); 521 return UndefValue::get(DestTy); 522 } 523 524 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 525 return Constant::getNullValue(DestTy); 526 527 // If the cast operand is a constant expression, there's a few things we can 528 // do to try to simplify it. 529 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 530 if (CE->isCast()) { 531 // Try hard to fold cast of cast because they are often eliminable. 532 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 533 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 534 } else if (CE->getOpcode() == Instruction::GetElementPtr && 535 // Do not fold addrspacecast (gep 0, .., 0). It might make the 536 // addrspacecast uncanonicalized. 537 opc != Instruction::AddrSpaceCast) { 538 // If all of the indexes in the GEP are null values, there is no pointer 539 // adjustment going on. We might as well cast the source pointer. 540 bool isAllNull = true; 541 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 542 if (!CE->getOperand(i)->isNullValue()) { 543 isAllNull = false; 544 break; 545 } 546 if (isAllNull) 547 // This is casting one pointer type to another, always BitCast 548 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 549 } 550 } 551 552 // If the cast operand is a constant vector, perform the cast by 553 // operating on each element. In the cast of bitcasts, the element 554 // count may be mismatched; don't attempt to handle that here. 555 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 556 DestTy->isVectorTy() && 557 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 558 SmallVector<Constant*, 16> res; 559 VectorType *DestVecTy = cast<VectorType>(DestTy); 560 Type *DstEltTy = DestVecTy->getElementType(); 561 Type *Ty = IntegerType::get(V->getContext(), 32); 562 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 563 Constant *C = 564 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 565 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 566 } 567 return ConstantVector::get(res); 568 } 569 570 // We actually have to do a cast now. Perform the cast according to the 571 // opcode specified. 572 switch (opc) { 573 default: 574 llvm_unreachable("Failed to cast constant expression"); 575 case Instruction::FPTrunc: 576 case Instruction::FPExt: 577 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 578 bool ignored; 579 APFloat Val = FPC->getValueAPF(); 580 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 581 DestTy->isFloatTy() ? APFloat::IEEEsingle : 582 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 583 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 584 DestTy->isFP128Ty() ? APFloat::IEEEquad : 585 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 586 APFloat::Bogus, 587 APFloat::rmNearestTiesToEven, &ignored); 588 return ConstantFP::get(V->getContext(), Val); 589 } 590 return nullptr; // Can't fold. 591 case Instruction::FPToUI: 592 case Instruction::FPToSI: 593 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 594 const APFloat &V = FPC->getValueAPF(); 595 bool ignored; 596 uint64_t x[2]; 597 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 598 if (APFloat::opInvalidOp == 599 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 600 APFloat::rmTowardZero, &ignored)) { 601 // Undefined behavior invoked - the destination type can't represent 602 // the input constant. 603 return UndefValue::get(DestTy); 604 } 605 APInt Val(DestBitWidth, x); 606 return ConstantInt::get(FPC->getContext(), Val); 607 } 608 return nullptr; // Can't fold. 609 case Instruction::IntToPtr: //always treated as unsigned 610 if (V->isNullValue()) // Is it an integral null value? 611 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 612 return nullptr; // Other pointer types cannot be casted 613 case Instruction::PtrToInt: // always treated as unsigned 614 // Is it a null pointer value? 615 if (V->isNullValue()) 616 return ConstantInt::get(DestTy, 0); 617 // If this is a sizeof-like expression, pull out multiplications by 618 // known factors to expose them to subsequent folding. If it's an 619 // alignof-like expression, factor out known factors. 620 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 621 if (CE->getOpcode() == Instruction::GetElementPtr && 622 CE->getOperand(0)->isNullValue()) { 623 Type *Ty = 624 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 625 if (CE->getNumOperands() == 2) { 626 // Handle a sizeof-like expression. 627 Constant *Idx = CE->getOperand(1); 628 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 629 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 630 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 631 DestTy, false), 632 Idx, DestTy); 633 return ConstantExpr::getMul(C, Idx); 634 } 635 } else if (CE->getNumOperands() == 3 && 636 CE->getOperand(1)->isNullValue()) { 637 // Handle an alignof-like expression. 638 if (StructType *STy = dyn_cast<StructType>(Ty)) 639 if (!STy->isPacked()) { 640 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 641 if (CI->isOne() && 642 STy->getNumElements() == 2 && 643 STy->getElementType(0)->isIntegerTy(1)) { 644 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 645 } 646 } 647 // Handle an offsetof-like expression. 648 if (Ty->isStructTy() || Ty->isArrayTy()) { 649 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 650 DestTy, false)) 651 return C; 652 } 653 } 654 } 655 // Other pointer types cannot be casted 656 return nullptr; 657 case Instruction::UIToFP: 658 case Instruction::SIToFP: 659 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 660 APInt api = CI->getValue(); 661 APFloat apf(DestTy->getFltSemantics(), 662 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 663 if (APFloat::opOverflow & 664 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 665 APFloat::rmNearestTiesToEven)) { 666 // Undefined behavior invoked - the destination type can't represent 667 // the input constant. 668 return UndefValue::get(DestTy); 669 } 670 return ConstantFP::get(V->getContext(), apf); 671 } 672 return nullptr; 673 case Instruction::ZExt: 674 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 675 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 676 return ConstantInt::get(V->getContext(), 677 CI->getValue().zext(BitWidth)); 678 } 679 return nullptr; 680 case Instruction::SExt: 681 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 682 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 683 return ConstantInt::get(V->getContext(), 684 CI->getValue().sext(BitWidth)); 685 } 686 return nullptr; 687 case Instruction::Trunc: { 688 if (V->getType()->isVectorTy()) 689 return nullptr; 690 691 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 692 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 693 return ConstantInt::get(V->getContext(), 694 CI->getValue().trunc(DestBitWidth)); 695 } 696 697 // The input must be a constantexpr. See if we can simplify this based on 698 // the bytes we are demanding. Only do this if the source and dest are an 699 // even multiple of a byte. 700 if ((DestBitWidth & 7) == 0 && 701 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 702 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 703 return Res; 704 705 return nullptr; 706 } 707 case Instruction::BitCast: 708 return FoldBitCast(V, DestTy); 709 case Instruction::AddrSpaceCast: 710 return nullptr; 711 } 712} 713 714Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 715 Constant *V1, Constant *V2) { 716 // Check for i1 and vector true/false conditions. 717 if (Cond->isNullValue()) return V2; 718 if (Cond->isAllOnesValue()) return V1; 719 720 // If the condition is a vector constant, fold the result elementwise. 721 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 722 SmallVector<Constant*, 16> Result; 723 Type *Ty = IntegerType::get(CondV->getContext(), 32); 724 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 725 Constant *V; 726 Constant *V1Element = ConstantExpr::getExtractElement(V1, 727 ConstantInt::get(Ty, i)); 728 Constant *V2Element = ConstantExpr::getExtractElement(V2, 729 ConstantInt::get(Ty, i)); 730 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 731 if (V1Element == V2Element) { 732 V = V1Element; 733 } else if (isa<UndefValue>(Cond)) { 734 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 735 } else { 736 if (!isa<ConstantInt>(Cond)) break; 737 V = Cond->isNullValue() ? V2Element : V1Element; 738 } 739 Result.push_back(V); 740 } 741 742 // If we were able to build the vector, return it. 743 if (Result.size() == V1->getType()->getVectorNumElements()) 744 return ConstantVector::get(Result); 745 } 746 747 if (isa<UndefValue>(Cond)) { 748 if (isa<UndefValue>(V1)) return V1; 749 return V2; 750 } 751 if (isa<UndefValue>(V1)) return V2; 752 if (isa<UndefValue>(V2)) return V1; 753 if (V1 == V2) return V1; 754 755 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 756 if (TrueVal->getOpcode() == Instruction::Select) 757 if (TrueVal->getOperand(0) == Cond) 758 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 759 } 760 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 761 if (FalseVal->getOpcode() == Instruction::Select) 762 if (FalseVal->getOperand(0) == Cond) 763 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 764 } 765 766 return nullptr; 767} 768 769Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 770 Constant *Idx) { 771 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 772 return UndefValue::get(Val->getType()->getVectorElementType()); 773 if (Val->isNullValue()) // ee(zero, x) -> zero 774 return Constant::getNullValue(Val->getType()->getVectorElementType()); 775 // ee({w,x,y,z}, undef) -> undef 776 if (isa<UndefValue>(Idx)) 777 return UndefValue::get(Val->getType()->getVectorElementType()); 778 779 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 780 uint64_t Index = CIdx->getZExtValue(); 781 // ee({w,x,y,z}, wrong_value) -> undef 782 if (Index >= Val->getType()->getVectorNumElements()) 783 return UndefValue::get(Val->getType()->getVectorElementType()); 784 return Val->getAggregateElement(Index); 785 } 786 return nullptr; 787} 788 789Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 790 Constant *Elt, 791 Constant *Idx) { 792 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 793 if (!CIdx) return nullptr; 794 const APInt &IdxVal = CIdx->getValue(); 795 796 SmallVector<Constant*, 16> Result; 797 Type *Ty = IntegerType::get(Val->getContext(), 32); 798 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ 799 if (i == IdxVal) { 800 Result.push_back(Elt); 801 continue; 802 } 803 804 Constant *C = 805 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 806 Result.push_back(C); 807 } 808 809 return ConstantVector::get(Result); 810} 811 812Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 813 Constant *V2, 814 Constant *Mask) { 815 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 816 Type *EltTy = V1->getType()->getVectorElementType(); 817 818 // Undefined shuffle mask -> undefined value. 819 if (isa<UndefValue>(Mask)) 820 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 821 822 // Don't break the bitcode reader hack. 823 if (isa<ConstantExpr>(Mask)) return nullptr; 824 825 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 826 827 // Loop over the shuffle mask, evaluating each element. 828 SmallVector<Constant*, 32> Result; 829 for (unsigned i = 0; i != MaskNumElts; ++i) { 830 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 831 if (Elt == -1) { 832 Result.push_back(UndefValue::get(EltTy)); 833 continue; 834 } 835 Constant *InElt; 836 if (unsigned(Elt) >= SrcNumElts*2) 837 InElt = UndefValue::get(EltTy); 838 else if (unsigned(Elt) >= SrcNumElts) { 839 Type *Ty = IntegerType::get(V2->getContext(), 32); 840 InElt = 841 ConstantExpr::getExtractElement(V2, 842 ConstantInt::get(Ty, Elt - SrcNumElts)); 843 } else { 844 Type *Ty = IntegerType::get(V1->getContext(), 32); 845 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 846 } 847 Result.push_back(InElt); 848 } 849 850 return ConstantVector::get(Result); 851} 852 853Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 854 ArrayRef<unsigned> Idxs) { 855 // Base case: no indices, so return the entire value. 856 if (Idxs.empty()) 857 return Agg; 858 859 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 860 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 861 862 return nullptr; 863} 864 865Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 866 Constant *Val, 867 ArrayRef<unsigned> Idxs) { 868 // Base case: no indices, so replace the entire value. 869 if (Idxs.empty()) 870 return Val; 871 872 unsigned NumElts; 873 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 874 NumElts = ST->getNumElements(); 875 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 876 NumElts = AT->getNumElements(); 877 else 878 NumElts = Agg->getType()->getVectorNumElements(); 879 880 SmallVector<Constant*, 32> Result; 881 for (unsigned i = 0; i != NumElts; ++i) { 882 Constant *C = Agg->getAggregateElement(i); 883 if (!C) return nullptr; 884 885 if (Idxs[0] == i) 886 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 887 888 Result.push_back(C); 889 } 890 891 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 892 return ConstantStruct::get(ST, Result); 893 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 894 return ConstantArray::get(AT, Result); 895 return ConstantVector::get(Result); 896} 897 898 899Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 900 Constant *C1, Constant *C2) { 901 // Handle UndefValue up front. 902 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 903 switch (Opcode) { 904 case Instruction::Xor: 905 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 906 // Handle undef ^ undef -> 0 special case. This is a common 907 // idiom (misuse). 908 return Constant::getNullValue(C1->getType()); 909 // Fallthrough 910 case Instruction::Add: 911 case Instruction::Sub: 912 return UndefValue::get(C1->getType()); 913 case Instruction::And: 914 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 915 return C1; 916 return Constant::getNullValue(C1->getType()); // undef & X -> 0 917 case Instruction::Mul: { 918 // undef * undef -> undef 919 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 920 return C1; 921 const APInt *CV; 922 // X * undef -> undef if X is odd 923 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 924 if ((*CV)[0]) 925 return UndefValue::get(C1->getType()); 926 927 // X * undef -> 0 otherwise 928 return Constant::getNullValue(C1->getType()); 929 } 930 case Instruction::SDiv: 931 case Instruction::UDiv: 932 // X / undef -> undef 933 if (match(C1, m_Zero())) 934 return C2; 935 // undef / 0 -> undef 936 // undef / 1 -> undef 937 if (match(C2, m_Zero()) || match(C2, m_One())) 938 return C1; 939 // undef / X -> 0 otherwise 940 return Constant::getNullValue(C1->getType()); 941 case Instruction::URem: 942 case Instruction::SRem: 943 // X % undef -> undef 944 if (match(C2, m_Undef())) 945 return C2; 946 // undef % 0 -> undef 947 if (match(C2, m_Zero())) 948 return C1; 949 // undef % X -> 0 otherwise 950 return Constant::getNullValue(C1->getType()); 951 case Instruction::Or: // X | undef -> -1 952 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 953 return C1; 954 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 955 case Instruction::LShr: 956 // X >>l undef -> undef 957 if (isa<UndefValue>(C2)) 958 return C2; 959 // undef >>l 0 -> undef 960 if (match(C2, m_Zero())) 961 return C1; 962 // undef >>l X -> 0 963 return Constant::getNullValue(C1->getType()); 964 case Instruction::AShr: 965 // X >>a undef -> undef 966 if (isa<UndefValue>(C2)) 967 return C2; 968 // undef >>a 0 -> undef 969 if (match(C2, m_Zero())) 970 return C1; 971 // TODO: undef >>a X -> undef if the shift is exact 972 // undef >>a X -> 0 973 return Constant::getNullValue(C1->getType()); 974 case Instruction::Shl: 975 // X << undef -> undef 976 if (isa<UndefValue>(C2)) 977 return C2; 978 // undef << 0 -> undef 979 if (match(C2, m_Zero())) 980 return C1; 981 // undef << X -> 0 982 return Constant::getNullValue(C1->getType()); 983 } 984 } 985 986 // Handle simplifications when the RHS is a constant int. 987 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 988 switch (Opcode) { 989 case Instruction::Add: 990 if (CI2->equalsInt(0)) return C1; // X + 0 == X 991 break; 992 case Instruction::Sub: 993 if (CI2->equalsInt(0)) return C1; // X - 0 == X 994 break; 995 case Instruction::Mul: 996 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 997 if (CI2->equalsInt(1)) 998 return C1; // X * 1 == X 999 break; 1000 case Instruction::UDiv: 1001 case Instruction::SDiv: 1002 if (CI2->equalsInt(1)) 1003 return C1; // X / 1 == X 1004 if (CI2->equalsInt(0)) 1005 return UndefValue::get(CI2->getType()); // X / 0 == undef 1006 break; 1007 case Instruction::URem: 1008 case Instruction::SRem: 1009 if (CI2->equalsInt(1)) 1010 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1011 if (CI2->equalsInt(0)) 1012 return UndefValue::get(CI2->getType()); // X % 0 == undef 1013 break; 1014 case Instruction::And: 1015 if (CI2->isZero()) return C2; // X & 0 == 0 1016 if (CI2->isAllOnesValue()) 1017 return C1; // X & -1 == X 1018 1019 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1020 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1021 if (CE1->getOpcode() == Instruction::ZExt) { 1022 unsigned DstWidth = CI2->getType()->getBitWidth(); 1023 unsigned SrcWidth = 1024 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1025 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1026 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1027 return C1; 1028 } 1029 1030 // If and'ing the address of a global with a constant, fold it. 1031 if (CE1->getOpcode() == Instruction::PtrToInt && 1032 isa<GlobalValue>(CE1->getOperand(0))) { 1033 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1034 1035 // Functions are at least 4-byte aligned. 1036 unsigned GVAlign = GV->getAlignment(); 1037 if (isa<Function>(GV)) 1038 GVAlign = std::max(GVAlign, 4U); 1039 1040 if (GVAlign > 1) { 1041 unsigned DstWidth = CI2->getType()->getBitWidth(); 1042 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1043 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1044 1045 // If checking bits we know are clear, return zero. 1046 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1047 return Constant::getNullValue(CI2->getType()); 1048 } 1049 } 1050 } 1051 break; 1052 case Instruction::Or: 1053 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1054 if (CI2->isAllOnesValue()) 1055 return C2; // X | -1 == -1 1056 break; 1057 case Instruction::Xor: 1058 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1059 1060 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1061 switch (CE1->getOpcode()) { 1062 default: break; 1063 case Instruction::ICmp: 1064 case Instruction::FCmp: 1065 // cmp pred ^ true -> cmp !pred 1066 assert(CI2->equalsInt(1)); 1067 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1068 pred = CmpInst::getInversePredicate(pred); 1069 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1070 CE1->getOperand(1)); 1071 } 1072 } 1073 break; 1074 case Instruction::AShr: 1075 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1076 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1077 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1078 return ConstantExpr::getLShr(C1, C2); 1079 break; 1080 } 1081 } else if (isa<ConstantInt>(C1)) { 1082 // If C1 is a ConstantInt and C2 is not, swap the operands. 1083 if (Instruction::isCommutative(Opcode)) 1084 return ConstantExpr::get(Opcode, C2, C1); 1085 } 1086 1087 // At this point we know neither constant is an UndefValue. 1088 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1089 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1090 const APInt &C1V = CI1->getValue(); 1091 const APInt &C2V = CI2->getValue(); 1092 switch (Opcode) { 1093 default: 1094 break; 1095 case Instruction::Add: 1096 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1097 case Instruction::Sub: 1098 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1099 case Instruction::Mul: 1100 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1101 case Instruction::UDiv: 1102 assert(!CI2->isNullValue() && "Div by zero handled above"); 1103 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1104 case Instruction::SDiv: 1105 assert(!CI2->isNullValue() && "Div by zero handled above"); 1106 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1107 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1108 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1109 case Instruction::URem: 1110 assert(!CI2->isNullValue() && "Div by zero handled above"); 1111 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1112 case Instruction::SRem: 1113 assert(!CI2->isNullValue() && "Div by zero handled above"); 1114 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1115 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1116 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1117 case Instruction::And: 1118 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1119 case Instruction::Or: 1120 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1121 case Instruction::Xor: 1122 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1123 case Instruction::Shl: { 1124 uint32_t shiftAmt = C2V.getZExtValue(); 1125 if (shiftAmt < C1V.getBitWidth()) 1126 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1127 else 1128 return UndefValue::get(C1->getType()); // too big shift is undef 1129 } 1130 case Instruction::LShr: { 1131 uint32_t shiftAmt = C2V.getZExtValue(); 1132 if (shiftAmt < C1V.getBitWidth()) 1133 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1134 else 1135 return UndefValue::get(C1->getType()); // too big shift is undef 1136 } 1137 case Instruction::AShr: { 1138 uint32_t shiftAmt = C2V.getZExtValue(); 1139 if (shiftAmt < C1V.getBitWidth()) 1140 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1141 else 1142 return UndefValue::get(C1->getType()); // too big shift is undef 1143 } 1144 } 1145 } 1146 1147 switch (Opcode) { 1148 case Instruction::SDiv: 1149 case Instruction::UDiv: 1150 case Instruction::URem: 1151 case Instruction::SRem: 1152 case Instruction::LShr: 1153 case Instruction::AShr: 1154 case Instruction::Shl: 1155 if (CI1->equalsInt(0)) return C1; 1156 break; 1157 default: 1158 break; 1159 } 1160 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1161 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1162 APFloat C1V = CFP1->getValueAPF(); 1163 APFloat C2V = CFP2->getValueAPF(); 1164 APFloat C3V = C1V; // copy for modification 1165 switch (Opcode) { 1166 default: 1167 break; 1168 case Instruction::FAdd: 1169 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1170 return ConstantFP::get(C1->getContext(), C3V); 1171 case Instruction::FSub: 1172 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1173 return ConstantFP::get(C1->getContext(), C3V); 1174 case Instruction::FMul: 1175 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1176 return ConstantFP::get(C1->getContext(), C3V); 1177 case Instruction::FDiv: 1178 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1179 return ConstantFP::get(C1->getContext(), C3V); 1180 case Instruction::FRem: 1181 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1182 return ConstantFP::get(C1->getContext(), C3V); 1183 } 1184 } 1185 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1186 // Perform elementwise folding. 1187 SmallVector<Constant*, 16> Result; 1188 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1189 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1190 Constant *LHS = 1191 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1192 Constant *RHS = 1193 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1194 1195 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1196 } 1197 1198 return ConstantVector::get(Result); 1199 } 1200 1201 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1202 // There are many possible foldings we could do here. We should probably 1203 // at least fold add of a pointer with an integer into the appropriate 1204 // getelementptr. This will improve alias analysis a bit. 1205 1206 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1207 // (a + (b + c)). 1208 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1209 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1210 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1211 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1212 } 1213 } else if (isa<ConstantExpr>(C2)) { 1214 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1215 // other way if possible. 1216 if (Instruction::isCommutative(Opcode)) 1217 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1218 } 1219 1220 // i1 can be simplified in many cases. 1221 if (C1->getType()->isIntegerTy(1)) { 1222 switch (Opcode) { 1223 case Instruction::Add: 1224 case Instruction::Sub: 1225 return ConstantExpr::getXor(C1, C2); 1226 case Instruction::Mul: 1227 return ConstantExpr::getAnd(C1, C2); 1228 case Instruction::Shl: 1229 case Instruction::LShr: 1230 case Instruction::AShr: 1231 // We can assume that C2 == 0. If it were one the result would be 1232 // undefined because the shift value is as large as the bitwidth. 1233 return C1; 1234 case Instruction::SDiv: 1235 case Instruction::UDiv: 1236 // We can assume that C2 == 1. If it were zero the result would be 1237 // undefined through division by zero. 1238 return C1; 1239 case Instruction::URem: 1240 case Instruction::SRem: 1241 // We can assume that C2 == 1. If it were zero the result would be 1242 // undefined through division by zero. 1243 return ConstantInt::getFalse(C1->getContext()); 1244 default: 1245 break; 1246 } 1247 } 1248 1249 // We don't know how to fold this. 1250 return nullptr; 1251} 1252 1253/// isZeroSizedType - This type is zero sized if its an array or structure of 1254/// zero sized types. The only leaf zero sized type is an empty structure. 1255static bool isMaybeZeroSizedType(Type *Ty) { 1256 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1257 if (STy->isOpaque()) return true; // Can't say. 1258 1259 // If all of elements have zero size, this does too. 1260 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1261 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1262 return true; 1263 1264 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1265 return isMaybeZeroSizedType(ATy->getElementType()); 1266 } 1267 return false; 1268} 1269 1270/// IdxCompare - Compare the two constants as though they were getelementptr 1271/// indices. This allows coersion of the types to be the same thing. 1272/// 1273/// If the two constants are the "same" (after coersion), return 0. If the 1274/// first is less than the second, return -1, if the second is less than the 1275/// first, return 1. If the constants are not integral, return -2. 1276/// 1277static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1278 if (C1 == C2) return 0; 1279 1280 // Ok, we found a different index. If they are not ConstantInt, we can't do 1281 // anything with them. 1282 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1283 return -2; // don't know! 1284 1285 // Ok, we have two differing integer indices. Sign extend them to be the same 1286 // type. Long is always big enough, so we use it. 1287 if (!C1->getType()->isIntegerTy(64)) 1288 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1289 1290 if (!C2->getType()->isIntegerTy(64)) 1291 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1292 1293 if (C1 == C2) return 0; // They are equal 1294 1295 // If the type being indexed over is really just a zero sized type, there is 1296 // no pointer difference being made here. 1297 if (isMaybeZeroSizedType(ElTy)) 1298 return -2; // dunno. 1299 1300 // If they are really different, now that they are the same type, then we 1301 // found a difference! 1302 if (cast<ConstantInt>(C1)->getSExtValue() < 1303 cast<ConstantInt>(C2)->getSExtValue()) 1304 return -1; 1305 else 1306 return 1; 1307} 1308 1309/// evaluateFCmpRelation - This function determines if there is anything we can 1310/// decide about the two constants provided. This doesn't need to handle simple 1311/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1312/// If we can determine that the two constants have a particular relation to 1313/// each other, we should return the corresponding FCmpInst predicate, 1314/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1315/// ConstantFoldCompareInstruction. 1316/// 1317/// To simplify this code we canonicalize the relation so that the first 1318/// operand is always the most "complex" of the two. We consider ConstantFP 1319/// to be the simplest, and ConstantExprs to be the most complex. 1320static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1321 assert(V1->getType() == V2->getType() && 1322 "Cannot compare values of different types!"); 1323 1324 // Handle degenerate case quickly 1325 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1326 1327 if (!isa<ConstantExpr>(V1)) { 1328 if (!isa<ConstantExpr>(V2)) { 1329 // We distilled thisUse the standard constant folder for a few cases 1330 ConstantInt *R = nullptr; 1331 R = dyn_cast<ConstantInt>( 1332 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1333 if (R && !R->isZero()) 1334 return FCmpInst::FCMP_OEQ; 1335 R = dyn_cast<ConstantInt>( 1336 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1337 if (R && !R->isZero()) 1338 return FCmpInst::FCMP_OLT; 1339 R = dyn_cast<ConstantInt>( 1340 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1341 if (R && !R->isZero()) 1342 return FCmpInst::FCMP_OGT; 1343 1344 // Nothing more we can do 1345 return FCmpInst::BAD_FCMP_PREDICATE; 1346 } 1347 1348 // If the first operand is simple and second is ConstantExpr, swap operands. 1349 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1350 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1351 return FCmpInst::getSwappedPredicate(SwappedRelation); 1352 } else { 1353 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1354 // constantexpr or a simple constant. 1355 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1356 switch (CE1->getOpcode()) { 1357 case Instruction::FPTrunc: 1358 case Instruction::FPExt: 1359 case Instruction::UIToFP: 1360 case Instruction::SIToFP: 1361 // We might be able to do something with these but we don't right now. 1362 break; 1363 default: 1364 break; 1365 } 1366 } 1367 // There are MANY other foldings that we could perform here. They will 1368 // probably be added on demand, as they seem needed. 1369 return FCmpInst::BAD_FCMP_PREDICATE; 1370} 1371 1372static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1373 const GlobalValue *GV2) { 1374 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1375 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1376 return true; 1377 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1378 Type *Ty = GVar->getType()->getPointerElementType(); 1379 // A global with opaque type might end up being zero sized. 1380 if (!Ty->isSized()) 1381 return true; 1382 // A global with an empty type might lie at the address of any other 1383 // global. 1384 if (Ty->isEmptyTy()) 1385 return true; 1386 } 1387 return false; 1388 }; 1389 // Don't try to decide equality of aliases. 1390 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1391 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1392 return ICmpInst::ICMP_NE; 1393 return ICmpInst::BAD_ICMP_PREDICATE; 1394} 1395 1396/// evaluateICmpRelation - This function determines if there is anything we can 1397/// decide about the two constants provided. This doesn't need to handle simple 1398/// things like integer comparisons, but should instead handle ConstantExprs 1399/// and GlobalValues. If we can determine that the two constants have a 1400/// particular relation to each other, we should return the corresponding ICmp 1401/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1402/// 1403/// To simplify this code we canonicalize the relation so that the first 1404/// operand is always the most "complex" of the two. We consider simple 1405/// constants (like ConstantInt) to be the simplest, followed by 1406/// GlobalValues, followed by ConstantExpr's (the most complex). 1407/// 1408static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1409 bool isSigned) { 1410 assert(V1->getType() == V2->getType() && 1411 "Cannot compare different types of values!"); 1412 if (V1 == V2) return ICmpInst::ICMP_EQ; 1413 1414 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1415 !isa<BlockAddress>(V1)) { 1416 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1417 !isa<BlockAddress>(V2)) { 1418 // We distilled this down to a simple case, use the standard constant 1419 // folder. 1420 ConstantInt *R = nullptr; 1421 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1422 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1423 if (R && !R->isZero()) 1424 return pred; 1425 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1426 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1427 if (R && !R->isZero()) 1428 return pred; 1429 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1430 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1431 if (R && !R->isZero()) 1432 return pred; 1433 1434 // If we couldn't figure it out, bail. 1435 return ICmpInst::BAD_ICMP_PREDICATE; 1436 } 1437 1438 // If the first operand is simple, swap operands. 1439 ICmpInst::Predicate SwappedRelation = 1440 evaluateICmpRelation(V2, V1, isSigned); 1441 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1442 return ICmpInst::getSwappedPredicate(SwappedRelation); 1443 1444 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1445 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1446 ICmpInst::Predicate SwappedRelation = 1447 evaluateICmpRelation(V2, V1, isSigned); 1448 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1449 return ICmpInst::getSwappedPredicate(SwappedRelation); 1450 return ICmpInst::BAD_ICMP_PREDICATE; 1451 } 1452 1453 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1454 // constant (which, since the types must match, means that it's a 1455 // ConstantPointerNull). 1456 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1457 return areGlobalsPotentiallyEqual(GV, GV2); 1458 } else if (isa<BlockAddress>(V2)) { 1459 return ICmpInst::ICMP_NE; // Globals never equal labels. 1460 } else { 1461 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1462 // GlobalVals can never be null unless they have external weak linkage. 1463 // We don't try to evaluate aliases here. 1464 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1465 return ICmpInst::ICMP_NE; 1466 } 1467 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1468 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1469 ICmpInst::Predicate SwappedRelation = 1470 evaluateICmpRelation(V2, V1, isSigned); 1471 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1472 return ICmpInst::getSwappedPredicate(SwappedRelation); 1473 return ICmpInst::BAD_ICMP_PREDICATE; 1474 } 1475 1476 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1477 // constant (which, since the types must match, means that it is a 1478 // ConstantPointerNull). 1479 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1480 // Block address in another function can't equal this one, but block 1481 // addresses in the current function might be the same if blocks are 1482 // empty. 1483 if (BA2->getFunction() != BA->getFunction()) 1484 return ICmpInst::ICMP_NE; 1485 } else { 1486 // Block addresses aren't null, don't equal the address of globals. 1487 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1488 "Canonicalization guarantee!"); 1489 return ICmpInst::ICMP_NE; 1490 } 1491 } else { 1492 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1493 // constantexpr, a global, block address, or a simple constant. 1494 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1495 Constant *CE1Op0 = CE1->getOperand(0); 1496 1497 switch (CE1->getOpcode()) { 1498 case Instruction::Trunc: 1499 case Instruction::FPTrunc: 1500 case Instruction::FPExt: 1501 case Instruction::FPToUI: 1502 case Instruction::FPToSI: 1503 break; // We can't evaluate floating point casts or truncations. 1504 1505 case Instruction::UIToFP: 1506 case Instruction::SIToFP: 1507 case Instruction::BitCast: 1508 case Instruction::ZExt: 1509 case Instruction::SExt: 1510 // If the cast is not actually changing bits, and the second operand is a 1511 // null pointer, do the comparison with the pre-casted value. 1512 if (V2->isNullValue() && 1513 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1514 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1515 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1516 return evaluateICmpRelation(CE1Op0, 1517 Constant::getNullValue(CE1Op0->getType()), 1518 isSigned); 1519 } 1520 break; 1521 1522 case Instruction::GetElementPtr: { 1523 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1524 // Ok, since this is a getelementptr, we know that the constant has a 1525 // pointer type. Check the various cases. 1526 if (isa<ConstantPointerNull>(V2)) { 1527 // If we are comparing a GEP to a null pointer, check to see if the base 1528 // of the GEP equals the null pointer. 1529 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1530 if (GV->hasExternalWeakLinkage()) 1531 // Weak linkage GVals could be zero or not. We're comparing that 1532 // to null pointer so its greater-or-equal 1533 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1534 else 1535 // If its not weak linkage, the GVal must have a non-zero address 1536 // so the result is greater-than 1537 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1538 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1539 // If we are indexing from a null pointer, check to see if we have any 1540 // non-zero indices. 1541 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1542 if (!CE1->getOperand(i)->isNullValue()) 1543 // Offsetting from null, must not be equal. 1544 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1545 // Only zero indexes from null, must still be zero. 1546 return ICmpInst::ICMP_EQ; 1547 } 1548 // Otherwise, we can't really say if the first operand is null or not. 1549 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1550 if (isa<ConstantPointerNull>(CE1Op0)) { 1551 if (GV2->hasExternalWeakLinkage()) 1552 // Weak linkage GVals could be zero or not. We're comparing it to 1553 // a null pointer, so its less-or-equal 1554 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1555 else 1556 // If its not weak linkage, the GVal must have a non-zero address 1557 // so the result is less-than 1558 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1559 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1560 if (GV == GV2) { 1561 // If this is a getelementptr of the same global, then it must be 1562 // different. Because the types must match, the getelementptr could 1563 // only have at most one index, and because we fold getelementptr's 1564 // with a single zero index, it must be nonzero. 1565 assert(CE1->getNumOperands() == 2 && 1566 !CE1->getOperand(1)->isNullValue() && 1567 "Surprising getelementptr!"); 1568 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1569 } else { 1570 if (CE1GEP->hasAllZeroIndices()) 1571 return areGlobalsPotentiallyEqual(GV, GV2); 1572 return ICmpInst::BAD_ICMP_PREDICATE; 1573 } 1574 } 1575 } else { 1576 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1577 Constant *CE2Op0 = CE2->getOperand(0); 1578 1579 // There are MANY other foldings that we could perform here. They will 1580 // probably be added on demand, as they seem needed. 1581 switch (CE2->getOpcode()) { 1582 default: break; 1583 case Instruction::GetElementPtr: 1584 // By far the most common case to handle is when the base pointers are 1585 // obviously to the same global. 1586 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1587 // Don't know relative ordering, but check for inequality. 1588 if (CE1Op0 != CE2Op0) { 1589 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1590 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1591 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1592 cast<GlobalValue>(CE2Op0)); 1593 return ICmpInst::BAD_ICMP_PREDICATE; 1594 } 1595 // Ok, we know that both getelementptr instructions are based on the 1596 // same global. From this, we can precisely determine the relative 1597 // ordering of the resultant pointers. 1598 unsigned i = 1; 1599 1600 // The logic below assumes that the result of the comparison 1601 // can be determined by finding the first index that differs. 1602 // This doesn't work if there is over-indexing in any 1603 // subsequent indices, so check for that case first. 1604 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1605 !CE2->isGEPWithNoNotionalOverIndexing()) 1606 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1607 1608 // Compare all of the operands the GEP's have in common. 1609 gep_type_iterator GTI = gep_type_begin(CE1); 1610 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1611 ++i, ++GTI) 1612 switch (IdxCompare(CE1->getOperand(i), 1613 CE2->getOperand(i), GTI.getIndexedType())) { 1614 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1615 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1616 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1617 } 1618 1619 // Ok, we ran out of things they have in common. If any leftovers 1620 // are non-zero then we have a difference, otherwise we are equal. 1621 for (; i < CE1->getNumOperands(); ++i) 1622 if (!CE1->getOperand(i)->isNullValue()) { 1623 if (isa<ConstantInt>(CE1->getOperand(i))) 1624 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1625 else 1626 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1627 } 1628 1629 for (; i < CE2->getNumOperands(); ++i) 1630 if (!CE2->getOperand(i)->isNullValue()) { 1631 if (isa<ConstantInt>(CE2->getOperand(i))) 1632 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1633 else 1634 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1635 } 1636 return ICmpInst::ICMP_EQ; 1637 } 1638 } 1639 } 1640 } 1641 default: 1642 break; 1643 } 1644 } 1645 1646 return ICmpInst::BAD_ICMP_PREDICATE; 1647} 1648 1649Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1650 Constant *C1, Constant *C2) { 1651 Type *ResultTy; 1652 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1653 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1654 VT->getNumElements()); 1655 else 1656 ResultTy = Type::getInt1Ty(C1->getContext()); 1657 1658 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1659 if (pred == FCmpInst::FCMP_FALSE) 1660 return Constant::getNullValue(ResultTy); 1661 1662 if (pred == FCmpInst::FCMP_TRUE) 1663 return Constant::getAllOnesValue(ResultTy); 1664 1665 // Handle some degenerate cases first 1666 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1667 // For EQ and NE, we can always pick a value for the undef to make the 1668 // predicate pass or fail, so we can return undef. 1669 // Also, if both operands are undef, we can return undef. 1670 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || 1671 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1672 return UndefValue::get(ResultTy); 1673 // Otherwise, pick the same value as the non-undef operand, and fold 1674 // it to true or false. 1675 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1676 } 1677 1678 // icmp eq/ne(null,GV) -> false/true 1679 if (C1->isNullValue()) { 1680 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1681 // Don't try to evaluate aliases. External weak GV can be null. 1682 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1683 if (pred == ICmpInst::ICMP_EQ) 1684 return ConstantInt::getFalse(C1->getContext()); 1685 else if (pred == ICmpInst::ICMP_NE) 1686 return ConstantInt::getTrue(C1->getContext()); 1687 } 1688 // icmp eq/ne(GV,null) -> false/true 1689 } else if (C2->isNullValue()) { 1690 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1691 // Don't try to evaluate aliases. External weak GV can be null. 1692 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1693 if (pred == ICmpInst::ICMP_EQ) 1694 return ConstantInt::getFalse(C1->getContext()); 1695 else if (pred == ICmpInst::ICMP_NE) 1696 return ConstantInt::getTrue(C1->getContext()); 1697 } 1698 } 1699 1700 // If the comparison is a comparison between two i1's, simplify it. 1701 if (C1->getType()->isIntegerTy(1)) { 1702 switch(pred) { 1703 case ICmpInst::ICMP_EQ: 1704 if (isa<ConstantInt>(C2)) 1705 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1706 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1707 case ICmpInst::ICMP_NE: 1708 return ConstantExpr::getXor(C1, C2); 1709 default: 1710 break; 1711 } 1712 } 1713 1714 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1715 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1716 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1717 switch (pred) { 1718 default: llvm_unreachable("Invalid ICmp Predicate"); 1719 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1720 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1721 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1722 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1723 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1724 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1725 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1726 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1727 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1728 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1729 } 1730 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1731 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1732 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1733 APFloat::cmpResult R = C1V.compare(C2V); 1734 switch (pred) { 1735 default: llvm_unreachable("Invalid FCmp Predicate"); 1736 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1737 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1738 case FCmpInst::FCMP_UNO: 1739 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1740 case FCmpInst::FCMP_ORD: 1741 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1742 case FCmpInst::FCMP_UEQ: 1743 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1744 R==APFloat::cmpEqual); 1745 case FCmpInst::FCMP_OEQ: 1746 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1747 case FCmpInst::FCMP_UNE: 1748 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1749 case FCmpInst::FCMP_ONE: 1750 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1751 R==APFloat::cmpGreaterThan); 1752 case FCmpInst::FCMP_ULT: 1753 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1754 R==APFloat::cmpLessThan); 1755 case FCmpInst::FCMP_OLT: 1756 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1757 case FCmpInst::FCMP_UGT: 1758 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1759 R==APFloat::cmpGreaterThan); 1760 case FCmpInst::FCMP_OGT: 1761 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1762 case FCmpInst::FCMP_ULE: 1763 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1764 case FCmpInst::FCMP_OLE: 1765 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1766 R==APFloat::cmpEqual); 1767 case FCmpInst::FCMP_UGE: 1768 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1769 case FCmpInst::FCMP_OGE: 1770 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1771 R==APFloat::cmpEqual); 1772 } 1773 } else if (C1->getType()->isVectorTy()) { 1774 // If we can constant fold the comparison of each element, constant fold 1775 // the whole vector comparison. 1776 SmallVector<Constant*, 4> ResElts; 1777 Type *Ty = IntegerType::get(C1->getContext(), 32); 1778 // Compare the elements, producing an i1 result or constant expr. 1779 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1780 Constant *C1E = 1781 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1782 Constant *C2E = 1783 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1784 1785 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1786 } 1787 1788 return ConstantVector::get(ResElts); 1789 } 1790 1791 if (C1->getType()->isFloatingPointTy()) { 1792 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1793 switch (evaluateFCmpRelation(C1, C2)) { 1794 default: llvm_unreachable("Unknown relation!"); 1795 case FCmpInst::FCMP_UNO: 1796 case FCmpInst::FCMP_ORD: 1797 case FCmpInst::FCMP_UEQ: 1798 case FCmpInst::FCMP_UNE: 1799 case FCmpInst::FCMP_ULT: 1800 case FCmpInst::FCMP_UGT: 1801 case FCmpInst::FCMP_ULE: 1802 case FCmpInst::FCMP_UGE: 1803 case FCmpInst::FCMP_TRUE: 1804 case FCmpInst::FCMP_FALSE: 1805 case FCmpInst::BAD_FCMP_PREDICATE: 1806 break; // Couldn't determine anything about these constants. 1807 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1808 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1809 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1810 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1811 break; 1812 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1813 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1814 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1815 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1816 break; 1817 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1818 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1819 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1820 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1821 break; 1822 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1823 // We can only partially decide this relation. 1824 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1825 Result = 0; 1826 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1827 Result = 1; 1828 break; 1829 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1830 // We can only partially decide this relation. 1831 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1832 Result = 0; 1833 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1834 Result = 1; 1835 break; 1836 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1837 // We can only partially decide this relation. 1838 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1839 Result = 0; 1840 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1841 Result = 1; 1842 break; 1843 } 1844 1845 // If we evaluated the result, return it now. 1846 if (Result != -1) 1847 return ConstantInt::get(ResultTy, Result); 1848 1849 } else { 1850 // Evaluate the relation between the two constants, per the predicate. 1851 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1852 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 1853 default: llvm_unreachable("Unknown relational!"); 1854 case ICmpInst::BAD_ICMP_PREDICATE: 1855 break; // Couldn't determine anything about these constants. 1856 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1857 // If we know the constants are equal, we can decide the result of this 1858 // computation precisely. 1859 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1860 break; 1861 case ICmpInst::ICMP_ULT: 1862 switch (pred) { 1863 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1864 Result = 1; break; 1865 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1866 Result = 0; break; 1867 } 1868 break; 1869 case ICmpInst::ICMP_SLT: 1870 switch (pred) { 1871 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1872 Result = 1; break; 1873 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1874 Result = 0; break; 1875 } 1876 break; 1877 case ICmpInst::ICMP_UGT: 1878 switch (pred) { 1879 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1880 Result = 1; break; 1881 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1882 Result = 0; break; 1883 } 1884 break; 1885 case ICmpInst::ICMP_SGT: 1886 switch (pred) { 1887 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1888 Result = 1; break; 1889 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1890 Result = 0; break; 1891 } 1892 break; 1893 case ICmpInst::ICMP_ULE: 1894 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1895 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1896 break; 1897 case ICmpInst::ICMP_SLE: 1898 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1899 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1900 break; 1901 case ICmpInst::ICMP_UGE: 1902 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1903 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1904 break; 1905 case ICmpInst::ICMP_SGE: 1906 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1907 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1908 break; 1909 case ICmpInst::ICMP_NE: 1910 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1911 if (pred == ICmpInst::ICMP_NE) Result = 1; 1912 break; 1913 } 1914 1915 // If we evaluated the result, return it now. 1916 if (Result != -1) 1917 return ConstantInt::get(ResultTy, Result); 1918 1919 // If the right hand side is a bitcast, try using its inverse to simplify 1920 // it by moving it to the left hand side. We can't do this if it would turn 1921 // a vector compare into a scalar compare or visa versa. 1922 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1923 Constant *CE2Op0 = CE2->getOperand(0); 1924 if (CE2->getOpcode() == Instruction::BitCast && 1925 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1926 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1927 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1928 } 1929 } 1930 1931 // If the left hand side is an extension, try eliminating it. 1932 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1933 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 1934 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 1935 Constant *CE1Op0 = CE1->getOperand(0); 1936 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1937 if (CE1Inverse == CE1Op0) { 1938 // Check whether we can safely truncate the right hand side. 1939 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1940 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1941 C2->getType()) == C2) 1942 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1943 } 1944 } 1945 } 1946 1947 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1948 (C1->isNullValue() && !C2->isNullValue())) { 1949 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1950 // other way if possible. 1951 // Also, if C1 is null and C2 isn't, flip them around. 1952 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1953 return ConstantExpr::getICmp(pred, C2, C1); 1954 } 1955 } 1956 return nullptr; 1957} 1958 1959/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1960/// is "inbounds". 1961template<typename IndexTy> 1962static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1963 // No indices means nothing that could be out of bounds. 1964 if (Idxs.empty()) return true; 1965 1966 // If the first index is zero, it's in bounds. 1967 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1968 1969 // If the first index is one and all the rest are zero, it's in bounds, 1970 // by the one-past-the-end rule. 1971 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1972 return false; 1973 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1974 if (!cast<Constant>(Idxs[i])->isNullValue()) 1975 return false; 1976 return true; 1977} 1978 1979/// \brief Test whether a given ConstantInt is in-range for a SequentialType. 1980static bool isIndexInRangeOfSequentialType(const SequentialType *STy, 1981 const ConstantInt *CI) { 1982 if (const PointerType *PTy = dyn_cast<PointerType>(STy)) 1983 // Only handle pointers to sized types, not pointers to functions. 1984 return PTy->getElementType()->isSized(); 1985 1986 uint64_t NumElements = 0; 1987 // Determine the number of elements in our sequential type. 1988 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 1989 NumElements = ATy->getNumElements(); 1990 else if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 1991 NumElements = VTy->getNumElements(); 1992 1993 assert((isa<ArrayType>(STy) || NumElements > 0) && 1994 "didn't expect non-array type to have zero elements!"); 1995 1996 // We cannot bounds check the index if it doesn't fit in an int64_t. 1997 if (CI->getValue().getActiveBits() > 64) 1998 return false; 1999 2000 // A negative index or an index past the end of our sequential type is 2001 // considered out-of-range. 2002 int64_t IndexVal = CI->getSExtValue(); 2003 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2004 return false; 2005 2006 // Otherwise, it is in-range. 2007 return true; 2008} 2009 2010template<typename IndexTy> 2011static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 2012 bool inBounds, 2013 ArrayRef<IndexTy> Idxs) { 2014 if (Idxs.empty()) return C; 2015 Constant *Idx0 = cast<Constant>(Idxs[0]); 2016 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2017 return C; 2018 2019 if (isa<UndefValue>(C)) { 2020 PointerType *Ptr = cast<PointerType>(C->getType()); 2021 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 2022 assert(Ty && "Invalid indices for GEP!"); 2023 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2024 } 2025 2026 if (C->isNullValue()) { 2027 bool isNull = true; 2028 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2029 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2030 isNull = false; 2031 break; 2032 } 2033 if (isNull) { 2034 PointerType *Ptr = cast<PointerType>(C->getType()); 2035 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 2036 assert(Ty && "Invalid indices for GEP!"); 2037 return ConstantPointerNull::get(PointerType::get(Ty, 2038 Ptr->getAddressSpace())); 2039 } 2040 } 2041 2042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2043 // Combine Indices - If the source pointer to this getelementptr instruction 2044 // is a getelementptr instruction, combine the indices of the two 2045 // getelementptr instructions into a single instruction. 2046 // 2047 if (CE->getOpcode() == Instruction::GetElementPtr) { 2048 Type *LastTy = nullptr; 2049 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2050 I != E; ++I) 2051 LastTy = *I; 2052 2053 // We cannot combine indices if doing so would take us outside of an 2054 // array or vector. Doing otherwise could trick us if we evaluated such a 2055 // GEP as part of a load. 2056 // 2057 // e.g. Consider if the original GEP was: 2058 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2059 // i32 0, i32 0, i64 0) 2060 // 2061 // If we then tried to offset it by '8' to get to the third element, 2062 // an i8, we should *not* get: 2063 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2064 // i32 0, i32 0, i64 8) 2065 // 2066 // This GEP tries to index array element '8 which runs out-of-bounds. 2067 // Subsequent evaluation would get confused and produce erroneous results. 2068 // 2069 // The following prohibits such a GEP from being formed by checking to see 2070 // if the index is in-range with respect to an array or vector. 2071 bool PerformFold = false; 2072 if (Idx0->isNullValue()) 2073 PerformFold = true; 2074 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 2075 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2076 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 2077 2078 if (PerformFold) { 2079 SmallVector<Value*, 16> NewIndices; 2080 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2081 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2082 NewIndices.push_back(CE->getOperand(i)); 2083 2084 // Add the last index of the source with the first index of the new GEP. 2085 // Make sure to handle the case when they are actually different types. 2086 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2087 // Otherwise it must be an array. 2088 if (!Idx0->isNullValue()) { 2089 Type *IdxTy = Combined->getType(); 2090 if (IdxTy != Idx0->getType()) { 2091 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2092 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2093 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2094 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2095 } else { 2096 Combined = 2097 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2098 } 2099 } 2100 2101 NewIndices.push_back(Combined); 2102 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2103 return 2104 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, 2105 inBounds && 2106 cast<GEPOperator>(CE)->isInBounds()); 2107 } 2108 } 2109 2110 // Attempt to fold casts to the same type away. For example, folding: 2111 // 2112 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2113 // i64 0, i64 0) 2114 // into: 2115 // 2116 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2117 // 2118 // Don't fold if the cast is changing address spaces. 2119 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2120 PointerType *SrcPtrTy = 2121 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2122 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2123 if (SrcPtrTy && DstPtrTy) { 2124 ArrayType *SrcArrayTy = 2125 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2126 ArrayType *DstArrayTy = 2127 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2128 if (SrcArrayTy && DstArrayTy 2129 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2130 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2131 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), 2132 Idxs, inBounds); 2133 } 2134 } 2135 } 2136 2137 // Check to see if any array indices are not within the corresponding 2138 // notional array or vector bounds. If so, try to determine if they can be 2139 // factored out into preceding dimensions. 2140 bool Unknown = false; 2141 SmallVector<Constant *, 8> NewIdxs; 2142 Type *Ty = C->getType(); 2143 Type *Prev = nullptr; 2144 for (unsigned i = 0, e = Idxs.size(); i != e; 2145 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2146 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2147 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 2148 if (CI->getSExtValue() > 0 && 2149 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 2150 if (isa<SequentialType>(Prev)) { 2151 // It's out of range, but we can factor it into the prior 2152 // dimension. 2153 NewIdxs.resize(Idxs.size()); 2154 uint64_t NumElements = 0; 2155 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2156 NumElements = ATy->getNumElements(); 2157 else 2158 NumElements = cast<VectorType>(Ty)->getNumElements(); 2159 2160 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2161 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2162 2163 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2164 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2165 2166 // Before adding, extend both operands to i64 to avoid 2167 // overflow trouble. 2168 if (!PrevIdx->getType()->isIntegerTy(64)) 2169 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2170 Type::getInt64Ty(Div->getContext())); 2171 if (!Div->getType()->isIntegerTy(64)) 2172 Div = ConstantExpr::getSExt(Div, 2173 Type::getInt64Ty(Div->getContext())); 2174 2175 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2176 } else { 2177 // It's out of range, but the prior dimension is a struct 2178 // so we can't do anything about it. 2179 Unknown = true; 2180 } 2181 } 2182 } else { 2183 // We don't know if it's in range or not. 2184 Unknown = true; 2185 } 2186 } 2187 2188 // If we did any factoring, start over with the adjusted indices. 2189 if (!NewIdxs.empty()) { 2190 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2191 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2192 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); 2193 } 2194 2195 // If all indices are known integers and normalized, we can do a simple 2196 // check for the "inbounds" property. 2197 if (!Unknown && !inBounds) 2198 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2199 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2200 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); 2201 2202 return nullptr; 2203} 2204 2205Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2206 bool inBounds, 2207 ArrayRef<Constant *> Idxs) { 2208 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2209} 2210 2211Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2212 bool inBounds, 2213 ArrayRef<Value *> Idxs) { 2214 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2215} 2216