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 TargetData, and the pieces that do. This is to avoid 16 // a dependence in VMCore on Target. 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "ConstantFold.h" 21 #include "llvm/Constants.h" 22 #include "llvm/Instructions.h" 23 #include "llvm/DerivedTypes.h" 24 #include "llvm/Function.h" 25 #include "llvm/GlobalAlias.h" 26 #include "llvm/GlobalVariable.h" 27 #include "llvm/Operator.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/Support/Compiler.h" 30 #include "llvm/Support/ErrorHandling.h" 31 #include "llvm/Support/GetElementPtrTypeIterator.h" 32 #include "llvm/Support/ManagedStatic.h" 33 #include "llvm/Support/MathExtras.h" 34 #include <limits> 35 using namespace llvm; 36 37 //===----------------------------------------------------------------------===// 38 // ConstantFold*Instruction Implementations 39 //===----------------------------------------------------------------------===// 40 41 /// BitCastConstantVector - Convert the specified ConstantVector node to the 42 /// specified vector type. At this point, we know that the elements of the 43 /// input vector constant are all simple integer or FP values. 44 static Constant *BitCastConstantVector(ConstantVector *CV, 45 VectorType *DstTy) { 46 47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 48 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 49 50 // If this cast changes element count then we can't handle it here: 51 // doing so requires endianness information. This should be handled by 52 // Analysis/ConstantFolding.cpp 53 unsigned NumElts = DstTy->getNumElements(); 54 if (NumElts != CV->getNumOperands()) 55 return 0; 56 57 // Check to verify that all elements of the input are simple. 58 for (unsigned i = 0; i != NumElts; ++i) { 59 if (!isa<ConstantInt>(CV->getOperand(i)) && 60 !isa<ConstantFP>(CV->getOperand(i))) 61 return 0; 62 } 63 64 // Bitcast each element now. 65 std::vector<Constant*> Result; 66 Type *DstEltTy = DstTy->getElementType(); 67 for (unsigned i = 0; i != NumElts; ++i) 68 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 69 DstEltTy)); 70 return ConstantVector::get(Result); 71 } 72 73 /// This function determines which opcode to use to fold two constant cast 74 /// expressions together. It uses CastInst::isEliminableCastPair to determine 75 /// the opcode. Consequently its just a wrapper around that function. 76 /// @brief Determine if it is valid to fold a cast of a cast 77 static unsigned 78 foldConstantCastPair( 79 unsigned opc, ///< opcode of the second cast constant expression 80 ConstantExpr *Op, ///< the first cast constant expression 81 Type *DstTy ///< desintation type of the first cast 82 ) { 83 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 84 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 85 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 86 87 // The the types and opcodes for the two Cast constant expressions 88 Type *SrcTy = Op->getOperand(0)->getType(); 89 Type *MidTy = Op->getType(); 90 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 91 Instruction::CastOps secondOp = Instruction::CastOps(opc); 92 93 // Let CastInst::isEliminableCastPair do the heavy lifting. 94 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 95 Type::getInt64Ty(DstTy->getContext())); 96 } 97 98 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 99 Type *SrcTy = V->getType(); 100 if (SrcTy == DestTy) 101 return V; // no-op cast 102 103 // Check to see if we are casting a pointer to an aggregate to a pointer to 104 // the first element. If so, return the appropriate GEP instruction. 105 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 106 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 107 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 108 SmallVector<Value*, 8> IdxList; 109 Value *Zero = 110 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 111 IdxList.push_back(Zero); 112 Type *ElTy = PTy->getElementType(); 113 while (ElTy != DPTy->getElementType()) { 114 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 115 if (STy->getNumElements() == 0) break; 116 ElTy = STy->getElementType(0); 117 IdxList.push_back(Zero); 118 } else if (SequentialType *STy = 119 dyn_cast<SequentialType>(ElTy)) { 120 if (ElTy->isPointerTy()) break; // Can't index into pointers! 121 ElTy = STy->getElementType(); 122 IdxList.push_back(Zero); 123 } else { 124 break; 125 } 126 } 127 128 if (ElTy == DPTy->getElementType()) 129 // This GEP is inbounds because all indices are zero. 130 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], 131 IdxList.size()); 132 } 133 134 // Handle casts from one vector constant to another. We know that the src 135 // and dest type have the same size (otherwise its an illegal cast). 136 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 137 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 138 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 139 "Not cast between same sized vectors!"); 140 SrcTy = NULL; 141 // First, check for null. Undef is already handled. 142 if (isa<ConstantAggregateZero>(V)) 143 return Constant::getNullValue(DestTy); 144 145 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 146 return BitCastConstantVector(CV, DestPTy); 147 } 148 149 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 150 // This allows for other simplifications (although some of them 151 // can only be handled by Analysis/ConstantFolding.cpp). 152 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 153 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 154 } 155 156 // Finally, implement bitcast folding now. The code below doesn't handle 157 // bitcast right. 158 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 159 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 160 161 // Handle integral constant input. 162 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 163 if (DestTy->isIntegerTy()) 164 // Integral -> Integral. This is a no-op because the bit widths must 165 // be the same. Consequently, we just fold to V. 166 return V; 167 168 if (DestTy->isFloatingPointTy()) 169 return ConstantFP::get(DestTy->getContext(), 170 APFloat(CI->getValue(), 171 !DestTy->isPPC_FP128Ty())); 172 173 // Otherwise, can't fold this (vector?) 174 return 0; 175 } 176 177 // Handle ConstantFP input: FP -> Integral. 178 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 179 return ConstantInt::get(FP->getContext(), 180 FP->getValueAPF().bitcastToAPInt()); 181 182 return 0; 183 } 184 185 186 /// ExtractConstantBytes - V is an integer constant which only has a subset of 187 /// its bytes used. The bytes used are indicated by ByteStart (which is the 188 /// first byte used, counting from the least significant byte) and ByteSize, 189 /// which is the number of bytes used. 190 /// 191 /// This function analyzes the specified constant to see if the specified byte 192 /// range can be returned as a simplified constant. If so, the constant is 193 /// returned, otherwise null is returned. 194 /// 195 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 196 unsigned ByteSize) { 197 assert(C->getType()->isIntegerTy() && 198 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 199 "Non-byte sized integer input"); 200 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 201 assert(ByteSize && "Must be accessing some piece"); 202 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 203 assert(ByteSize != CSize && "Should not extract everything"); 204 205 // Constant Integers are simple. 206 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 207 APInt V = CI->getValue(); 208 if (ByteStart) 209 V = V.lshr(ByteStart*8); 210 V = V.trunc(ByteSize*8); 211 return ConstantInt::get(CI->getContext(), V); 212 } 213 214 // In the input is a constant expr, we might be able to recursively simplify. 215 // If not, we definitely can't do anything. 216 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 217 if (CE == 0) return 0; 218 219 switch (CE->getOpcode()) { 220 default: return 0; 221 case Instruction::Or: { 222 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 223 if (RHS == 0) 224 return 0; 225 226 // X | -1 -> -1. 227 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 228 if (RHSC->isAllOnesValue()) 229 return RHSC; 230 231 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 232 if (LHS == 0) 233 return 0; 234 return ConstantExpr::getOr(LHS, RHS); 235 } 236 case Instruction::And: { 237 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 238 if (RHS == 0) 239 return 0; 240 241 // X & 0 -> 0. 242 if (RHS->isNullValue()) 243 return RHS; 244 245 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 246 if (LHS == 0) 247 return 0; 248 return ConstantExpr::getAnd(LHS, RHS); 249 } 250 case Instruction::LShr: { 251 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 252 if (Amt == 0) 253 return 0; 254 unsigned ShAmt = Amt->getZExtValue(); 255 // Cannot analyze non-byte shifts. 256 if ((ShAmt & 7) != 0) 257 return 0; 258 ShAmt >>= 3; 259 260 // If the extract is known to be all zeros, return zero. 261 if (ByteStart >= CSize-ShAmt) 262 return Constant::getNullValue(IntegerType::get(CE->getContext(), 263 ByteSize*8)); 264 // If the extract is known to be fully in the input, extract it. 265 if (ByteStart+ByteSize+ShAmt <= CSize) 266 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 267 268 // TODO: Handle the 'partially zero' case. 269 return 0; 270 } 271 272 case Instruction::Shl: { 273 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 274 if (Amt == 0) 275 return 0; 276 unsigned ShAmt = Amt->getZExtValue(); 277 // Cannot analyze non-byte shifts. 278 if ((ShAmt & 7) != 0) 279 return 0; 280 ShAmt >>= 3; 281 282 // If the extract is known to be all zeros, return zero. 283 if (ByteStart+ByteSize <= ShAmt) 284 return Constant::getNullValue(IntegerType::get(CE->getContext(), 285 ByteSize*8)); 286 // If the extract is known to be fully in the input, extract it. 287 if (ByteStart >= ShAmt) 288 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 289 290 // TODO: Handle the 'partially zero' case. 291 return 0; 292 } 293 294 case Instruction::ZExt: { 295 unsigned SrcBitSize = 296 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 297 298 // If extracting something that is completely zero, return 0. 299 if (ByteStart*8 >= SrcBitSize) 300 return Constant::getNullValue(IntegerType::get(CE->getContext(), 301 ByteSize*8)); 302 303 // If exactly extracting the input, return it. 304 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 305 return CE->getOperand(0); 306 307 // If extracting something completely in the input, if if the input is a 308 // multiple of 8 bits, recurse. 309 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 310 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 311 312 // Otherwise, if extracting a subset of the input, which is not multiple of 313 // 8 bits, do a shift and trunc to get the bits. 314 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 315 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 316 Constant *Res = CE->getOperand(0); 317 if (ByteStart) 318 Res = ConstantExpr::getLShr(Res, 319 ConstantInt::get(Res->getType(), ByteStart*8)); 320 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 321 ByteSize*8)); 322 } 323 324 // TODO: Handle the 'partially zero' case. 325 return 0; 326 } 327 } 328 } 329 330 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 331 /// on Ty, with any known factors factored out. If Folded is false, 332 /// return null if no factoring was possible, to avoid endlessly 333 /// bouncing an unfoldable expression back into the top-level folder. 334 /// 335 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 336 bool Folded) { 337 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 338 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 339 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 340 return ConstantExpr::getNUWMul(E, N); 341 } 342 343 if (StructType *STy = dyn_cast<StructType>(Ty)) 344 if (!STy->isPacked()) { 345 unsigned NumElems = STy->getNumElements(); 346 // An empty struct has size zero. 347 if (NumElems == 0) 348 return ConstantExpr::getNullValue(DestTy); 349 // Check for a struct with all members having the same size. 350 Constant *MemberSize = 351 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 352 bool AllSame = true; 353 for (unsigned i = 1; i != NumElems; ++i) 354 if (MemberSize != 355 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 356 AllSame = false; 357 break; 358 } 359 if (AllSame) { 360 Constant *N = ConstantInt::get(DestTy, NumElems); 361 return ConstantExpr::getNUWMul(MemberSize, N); 362 } 363 } 364 365 // Pointer size doesn't depend on the pointee type, so canonicalize them 366 // to an arbitrary pointee. 367 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 368 if (!PTy->getElementType()->isIntegerTy(1)) 369 return 370 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 371 PTy->getAddressSpace()), 372 DestTy, true); 373 374 // If there's no interesting folding happening, bail so that we don't create 375 // a constant that looks like it needs folding but really doesn't. 376 if (!Folded) 377 return 0; 378 379 // Base case: Get a regular sizeof expression. 380 Constant *C = ConstantExpr::getSizeOf(Ty); 381 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 382 DestTy, false), 383 C, DestTy); 384 return C; 385 } 386 387 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 388 /// on Ty, with any known factors factored out. If Folded is false, 389 /// return null if no factoring was possible, to avoid endlessly 390 /// bouncing an unfoldable expression back into the top-level folder. 391 /// 392 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 393 bool Folded) { 394 // The alignment of an array is equal to the alignment of the 395 // array element. Note that this is not always true for vectors. 396 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 397 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 398 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 399 DestTy, 400 false), 401 C, DestTy); 402 return C; 403 } 404 405 if (StructType *STy = dyn_cast<StructType>(Ty)) { 406 // Packed structs always have an alignment of 1. 407 if (STy->isPacked()) 408 return ConstantInt::get(DestTy, 1); 409 410 // Otherwise, struct alignment is the maximum alignment of any member. 411 // Without target data, we can't compare much, but we can check to see 412 // if all the members have the same alignment. 413 unsigned NumElems = STy->getNumElements(); 414 // An empty struct has minimal alignment. 415 if (NumElems == 0) 416 return ConstantInt::get(DestTy, 1); 417 // Check for a struct with all members having the same alignment. 418 Constant *MemberAlign = 419 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 420 bool AllSame = true; 421 for (unsigned i = 1; i != NumElems; ++i) 422 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 423 AllSame = false; 424 break; 425 } 426 if (AllSame) 427 return MemberAlign; 428 } 429 430 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 431 // to an arbitrary pointee. 432 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 433 if (!PTy->getElementType()->isIntegerTy(1)) 434 return 435 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 436 1), 437 PTy->getAddressSpace()), 438 DestTy, true); 439 440 // If there's no interesting folding happening, bail so that we don't create 441 // a constant that looks like it needs folding but really doesn't. 442 if (!Folded) 443 return 0; 444 445 // Base case: Get a regular alignof expression. 446 Constant *C = ConstantExpr::getAlignOf(Ty); 447 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 448 DestTy, false), 449 C, DestTy); 450 return C; 451 } 452 453 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 454 /// on Ty and FieldNo, with any known factors factored out. If Folded is false, 455 /// return null if no factoring was possible, to avoid endlessly 456 /// bouncing an unfoldable expression back into the top-level folder. 457 /// 458 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 459 Type *DestTy, 460 bool Folded) { 461 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 462 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 463 DestTy, false), 464 FieldNo, DestTy); 465 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 466 return ConstantExpr::getNUWMul(E, N); 467 } 468 469 if (StructType *STy = dyn_cast<StructType>(Ty)) 470 if (!STy->isPacked()) { 471 unsigned NumElems = STy->getNumElements(); 472 // An empty struct has no members. 473 if (NumElems == 0) 474 return 0; 475 // Check for a struct with all members having the same size. 476 Constant *MemberSize = 477 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 478 bool AllSame = true; 479 for (unsigned i = 1; i != NumElems; ++i) 480 if (MemberSize != 481 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 482 AllSame = false; 483 break; 484 } 485 if (AllSame) { 486 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 487 false, 488 DestTy, 489 false), 490 FieldNo, DestTy); 491 return ConstantExpr::getNUWMul(MemberSize, N); 492 } 493 } 494 495 // If there's no interesting folding happening, bail so that we don't create 496 // a constant that looks like it needs folding but really doesn't. 497 if (!Folded) 498 return 0; 499 500 // Base case: Get a regular offsetof expression. 501 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 502 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 503 DestTy, false), 504 C, DestTy); 505 return C; 506 } 507 508 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 509 Type *DestTy) { 510 if (isa<UndefValue>(V)) { 511 // zext(undef) = 0, because the top bits will be zero. 512 // sext(undef) = 0, because the top bits will all be the same. 513 // [us]itofp(undef) = 0, because the result value is bounded. 514 if (opc == Instruction::ZExt || opc == Instruction::SExt || 515 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 516 return Constant::getNullValue(DestTy); 517 return UndefValue::get(DestTy); 518 } 519 520 // No compile-time operations on this type yet. 521 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) 522 return 0; 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 // If all of the indexes in the GEP are null values, there is no pointer 536 // adjustment going on. We might as well cast the source pointer. 537 bool isAllNull = true; 538 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 539 if (!CE->getOperand(i)->isNullValue()) { 540 isAllNull = false; 541 break; 542 } 543 if (isAllNull) 544 // This is casting one pointer type to another, always BitCast 545 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 546 } 547 } 548 549 // If the cast operand is a constant vector, perform the cast by 550 // operating on each element. In the cast of bitcasts, the element 551 // count may be mismatched; don't attempt to handle that here. 552 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 553 if (DestTy->isVectorTy() && 554 cast<VectorType>(DestTy)->getNumElements() == 555 CV->getType()->getNumElements()) { 556 std::vector<Constant*> res; 557 VectorType *DestVecTy = cast<VectorType>(DestTy); 558 Type *DstEltTy = DestVecTy->getElementType(); 559 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 560 res.push_back(ConstantExpr::getCast(opc, 561 CV->getOperand(i), DstEltTy)); 562 return ConstantVector::get(res); 563 } 564 565 // We actually have to do a cast now. Perform the cast according to the 566 // opcode specified. 567 switch (opc) { 568 default: 569 llvm_unreachable("Failed to cast constant expression"); 570 case Instruction::FPTrunc: 571 case Instruction::FPExt: 572 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 573 bool ignored; 574 APFloat Val = FPC->getValueAPF(); 575 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle : 576 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 577 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 578 DestTy->isFP128Ty() ? APFloat::IEEEquad : 579 APFloat::Bogus, 580 APFloat::rmNearestTiesToEven, &ignored); 581 return ConstantFP::get(V->getContext(), Val); 582 } 583 return 0; // Can't fold. 584 case Instruction::FPToUI: 585 case Instruction::FPToSI: 586 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 587 const APFloat &V = FPC->getValueAPF(); 588 bool ignored; 589 uint64_t x[2]; 590 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 591 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 592 APFloat::rmTowardZero, &ignored); 593 APInt Val(DestBitWidth, x); 594 return ConstantInt::get(FPC->getContext(), Val); 595 } 596 return 0; // Can't fold. 597 case Instruction::IntToPtr: //always treated as unsigned 598 if (V->isNullValue()) // Is it an integral null value? 599 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 600 return 0; // Other pointer types cannot be casted 601 case Instruction::PtrToInt: // always treated as unsigned 602 // Is it a null pointer value? 603 if (V->isNullValue()) 604 return ConstantInt::get(DestTy, 0); 605 // If this is a sizeof-like expression, pull out multiplications by 606 // known factors to expose them to subsequent folding. If it's an 607 // alignof-like expression, factor out known factors. 608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 609 if (CE->getOpcode() == Instruction::GetElementPtr && 610 CE->getOperand(0)->isNullValue()) { 611 Type *Ty = 612 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 613 if (CE->getNumOperands() == 2) { 614 // Handle a sizeof-like expression. 615 Constant *Idx = CE->getOperand(1); 616 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 617 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 618 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 619 DestTy, false), 620 Idx, DestTy); 621 return ConstantExpr::getMul(C, Idx); 622 } 623 } else if (CE->getNumOperands() == 3 && 624 CE->getOperand(1)->isNullValue()) { 625 // Handle an alignof-like expression. 626 if (StructType *STy = dyn_cast<StructType>(Ty)) 627 if (!STy->isPacked()) { 628 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 629 if (CI->isOne() && 630 STy->getNumElements() == 2 && 631 STy->getElementType(0)->isIntegerTy(1)) { 632 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 633 } 634 } 635 // Handle an offsetof-like expression. 636 if (Ty->isStructTy() || Ty->isArrayTy()) { 637 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 638 DestTy, false)) 639 return C; 640 } 641 } 642 } 643 // Other pointer types cannot be casted 644 return 0; 645 case Instruction::UIToFP: 646 case Instruction::SIToFP: 647 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 648 APInt api = CI->getValue(); 649 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); 650 (void)apf.convertFromAPInt(api, 651 opc==Instruction::SIToFP, 652 APFloat::rmNearestTiesToEven); 653 return ConstantFP::get(V->getContext(), apf); 654 } 655 return 0; 656 case Instruction::ZExt: 657 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 658 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 659 return ConstantInt::get(V->getContext(), 660 CI->getValue().zext(BitWidth)); 661 } 662 return 0; 663 case Instruction::SExt: 664 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 665 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 666 return ConstantInt::get(V->getContext(), 667 CI->getValue().sext(BitWidth)); 668 } 669 return 0; 670 case Instruction::Trunc: { 671 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 672 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 673 return ConstantInt::get(V->getContext(), 674 CI->getValue().trunc(DestBitWidth)); 675 } 676 677 // The input must be a constantexpr. See if we can simplify this based on 678 // the bytes we are demanding. Only do this if the source and dest are an 679 // even multiple of a byte. 680 if ((DestBitWidth & 7) == 0 && 681 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 682 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 683 return Res; 684 685 return 0; 686 } 687 case Instruction::BitCast: 688 return FoldBitCast(V, DestTy); 689 } 690 } 691 692 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 693 Constant *V1, Constant *V2) { 694 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 695 return CB->getZExtValue() ? V1 : V2; 696 697 // Check for zero aggregate and ConstantVector of zeros 698 if (Cond->isNullValue()) return V2; 699 700 if (ConstantVector* CondV = dyn_cast<ConstantVector>(Cond)) { 701 702 if (CondV->isAllOnesValue()) return V1; 703 704 VectorType *VTy = cast<VectorType>(V1->getType()); 705 ConstantVector *CP1 = dyn_cast<ConstantVector>(V1); 706 ConstantVector *CP2 = dyn_cast<ConstantVector>(V2); 707 708 if ((CP1 || isa<ConstantAggregateZero>(V1)) && 709 (CP2 || isa<ConstantAggregateZero>(V2))) { 710 711 // Find the element type of the returned vector 712 Type *EltTy = VTy->getElementType(); 713 unsigned NumElem = VTy->getNumElements(); 714 std::vector<Constant*> Res(NumElem); 715 716 bool Valid = true; 717 for (unsigned i = 0; i < NumElem; ++i) { 718 ConstantInt* c = dyn_cast<ConstantInt>(CondV->getOperand(i)); 719 if (!c) { 720 Valid = false; 721 break; 722 } 723 Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 724 Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 725 Res[i] = c->getZExtValue() ? C1 : C2; 726 } 727 // If we were able to build the vector, return it 728 if (Valid) return ConstantVector::get(Res); 729 } 730 } 731 732 733 if (isa<UndefValue>(Cond)) { 734 if (isa<UndefValue>(V1)) return V1; 735 return V2; 736 } 737 if (isa<UndefValue>(V1)) return V2; 738 if (isa<UndefValue>(V2)) return V1; 739 if (V1 == V2) return V1; 740 741 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 742 if (TrueVal->getOpcode() == Instruction::Select) 743 if (TrueVal->getOperand(0) == Cond) 744 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 745 } 746 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 747 if (FalseVal->getOpcode() == Instruction::Select) 748 if (FalseVal->getOperand(0) == Cond) 749 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 750 } 751 752 return 0; 753 } 754 755 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 756 Constant *Idx) { 757 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 758 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 759 if (Val->isNullValue()) // ee(zero, x) -> zero 760 return Constant::getNullValue( 761 cast<VectorType>(Val->getType())->getElementType()); 762 763 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 764 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 765 return CVal->getOperand(CIdx->getZExtValue()); 766 } else if (isa<UndefValue>(Idx)) { 767 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 768 return CVal->getOperand(0); 769 } 770 } 771 return 0; 772 } 773 774 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 775 Constant *Elt, 776 Constant *Idx) { 777 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 778 if (!CIdx) return 0; 779 APInt idxVal = CIdx->getValue(); 780 if (isa<UndefValue>(Val)) { 781 // Insertion of scalar constant into vector undef 782 // Optimize away insertion of undef 783 if (isa<UndefValue>(Elt)) 784 return Val; 785 // Otherwise break the aggregate undef into multiple undefs and do 786 // the insertion 787 unsigned numOps = 788 cast<VectorType>(Val->getType())->getNumElements(); 789 std::vector<Constant*> Ops; 790 Ops.reserve(numOps); 791 for (unsigned i = 0; i < numOps; ++i) { 792 Constant *Op = 793 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 794 Ops.push_back(Op); 795 } 796 return ConstantVector::get(Ops); 797 } 798 if (isa<ConstantAggregateZero>(Val)) { 799 // Insertion of scalar constant into vector aggregate zero 800 // Optimize away insertion of zero 801 if (Elt->isNullValue()) 802 return Val; 803 // Otherwise break the aggregate zero into multiple zeros and do 804 // the insertion 805 unsigned numOps = 806 cast<VectorType>(Val->getType())->getNumElements(); 807 std::vector<Constant*> Ops; 808 Ops.reserve(numOps); 809 for (unsigned i = 0; i < numOps; ++i) { 810 Constant *Op = 811 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 812 Ops.push_back(Op); 813 } 814 return ConstantVector::get(Ops); 815 } 816 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 817 // Insertion of scalar constant into vector constant 818 std::vector<Constant*> Ops; 819 Ops.reserve(CVal->getNumOperands()); 820 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 821 Constant *Op = 822 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 823 Ops.push_back(Op); 824 } 825 return ConstantVector::get(Ops); 826 } 827 828 return 0; 829 } 830 831 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 832 /// return the specified element value. Otherwise return null. 833 static Constant *GetVectorElement(Constant *C, unsigned EltNo) { 834 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) 835 return CV->getOperand(EltNo); 836 837 Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 838 if (isa<ConstantAggregateZero>(C)) 839 return Constant::getNullValue(EltTy); 840 if (isa<UndefValue>(C)) 841 return UndefValue::get(EltTy); 842 return 0; 843 } 844 845 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 846 Constant *V2, 847 Constant *Mask) { 848 // Undefined shuffle mask -> undefined value. 849 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 850 851 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 852 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 853 Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 854 855 // Loop over the shuffle mask, evaluating each element. 856 SmallVector<Constant*, 32> Result; 857 for (unsigned i = 0; i != MaskNumElts; ++i) { 858 Constant *InElt = GetVectorElement(Mask, i); 859 if (InElt == 0) return 0; 860 861 if (isa<UndefValue>(InElt)) 862 InElt = UndefValue::get(EltTy); 863 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 864 unsigned Elt = CI->getZExtValue(); 865 if (Elt >= SrcNumElts*2) 866 InElt = UndefValue::get(EltTy); 867 else if (Elt >= SrcNumElts) 868 InElt = GetVectorElement(V2, Elt - SrcNumElts); 869 else 870 InElt = GetVectorElement(V1, Elt); 871 if (InElt == 0) return 0; 872 } else { 873 // Unknown value. 874 return 0; 875 } 876 Result.push_back(InElt); 877 } 878 879 return ConstantVector::get(Result); 880 } 881 882 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 883 ArrayRef<unsigned> Idxs) { 884 // Base case: no indices, so return the entire value. 885 if (Idxs.empty()) 886 return Agg; 887 888 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 889 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 890 Idxs)); 891 892 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 893 return 894 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 895 Idxs)); 896 897 // Otherwise recurse. 898 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) 899 return ConstantFoldExtractValueInstruction(CS->getOperand(Idxs[0]), 900 Idxs.slice(1)); 901 902 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) 903 return ConstantFoldExtractValueInstruction(CA->getOperand(Idxs[0]), 904 Idxs.slice(1)); 905 ConstantVector *CV = cast<ConstantVector>(Agg); 906 return ConstantFoldExtractValueInstruction(CV->getOperand(Idxs[0]), 907 Idxs.slice(1)); 908 } 909 910 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 911 Constant *Val, 912 ArrayRef<unsigned> Idxs) { 913 // Base case: no indices, so replace the entire value. 914 if (Idxs.empty()) 915 return Val; 916 917 if (isa<UndefValue>(Agg)) { 918 // Insertion of constant into aggregate undef 919 // Optimize away insertion of undef. 920 if (isa<UndefValue>(Val)) 921 return Agg; 922 923 // Otherwise break the aggregate undef into multiple undefs and do 924 // the insertion. 925 CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 926 unsigned numOps; 927 if (ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 928 numOps = AR->getNumElements(); 929 else 930 numOps = cast<StructType>(AggTy)->getNumElements(); 931 932 std::vector<Constant*> Ops(numOps); 933 for (unsigned i = 0; i < numOps; ++i) { 934 Type *MemberTy = AggTy->getTypeAtIndex(i); 935 Constant *Op = 936 (Idxs[0] == i) ? 937 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), 938 Val, Idxs.slice(1)) : 939 UndefValue::get(MemberTy); 940 Ops[i] = Op; 941 } 942 943 if (StructType* ST = dyn_cast<StructType>(AggTy)) 944 return ConstantStruct::get(ST, Ops); 945 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 946 } 947 948 if (isa<ConstantAggregateZero>(Agg)) { 949 // Insertion of constant into aggregate zero 950 // Optimize away insertion of zero. 951 if (Val->isNullValue()) 952 return Agg; 953 954 // Otherwise break the aggregate zero into multiple zeros and do 955 // the insertion. 956 CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 957 unsigned numOps; 958 if (ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 959 numOps = AR->getNumElements(); 960 else 961 numOps = cast<StructType>(AggTy)->getNumElements(); 962 963 std::vector<Constant*> Ops(numOps); 964 for (unsigned i = 0; i < numOps; ++i) { 965 Type *MemberTy = AggTy->getTypeAtIndex(i); 966 Constant *Op = 967 (Idxs[0] == i) ? 968 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), 969 Val, Idxs.slice(1)) : 970 Constant::getNullValue(MemberTy); 971 Ops[i] = Op; 972 } 973 974 if (StructType *ST = dyn_cast<StructType>(AggTy)) 975 return ConstantStruct::get(ST, Ops); 976 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 977 } 978 979 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 980 // Insertion of constant into aggregate constant. 981 std::vector<Constant*> Ops(Agg->getNumOperands()); 982 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 983 Constant *Op = cast<Constant>(Agg->getOperand(i)); 984 if (Idxs[0] == i) 985 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs.slice(1)); 986 Ops[i] = Op; 987 } 988 989 if (StructType* ST = dyn_cast<StructType>(Agg->getType())) 990 return ConstantStruct::get(ST, Ops); 991 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 992 } 993 994 return 0; 995 } 996 997 998 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 999 Constant *C1, Constant *C2) { 1000 // No compile-time operations on this type yet. 1001 if (C1->getType()->isPPC_FP128Ty()) 1002 return 0; 1003 1004 // Handle UndefValue up front. 1005 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1006 switch (Opcode) { 1007 case Instruction::Xor: 1008 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1009 // Handle undef ^ undef -> 0 special case. This is a common 1010 // idiom (misuse). 1011 return Constant::getNullValue(C1->getType()); 1012 // Fallthrough 1013 case Instruction::Add: 1014 case Instruction::Sub: 1015 return UndefValue::get(C1->getType()); 1016 case Instruction::And: 1017 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 1018 return C1; 1019 return Constant::getNullValue(C1->getType()); // undef & X -> 0 1020 case Instruction::Mul: { 1021 ConstantInt *CI; 1022 // X * undef -> undef if X is odd or undef 1023 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) || 1024 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) || 1025 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1026 return UndefValue::get(C1->getType()); 1027 1028 // X * undef -> 0 otherwise 1029 return Constant::getNullValue(C1->getType()); 1030 } 1031 case Instruction::UDiv: 1032 case Instruction::SDiv: 1033 // undef / 1 -> undef 1034 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) 1035 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) 1036 if (CI2->isOne()) 1037 return C1; 1038 // FALL THROUGH 1039 case Instruction::URem: 1040 case Instruction::SRem: 1041 if (!isa<UndefValue>(C2)) // undef / X -> 0 1042 return Constant::getNullValue(C1->getType()); 1043 return C2; // X / undef -> undef 1044 case Instruction::Or: // X | undef -> -1 1045 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 1046 return C1; 1047 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 1048 case Instruction::LShr: 1049 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 1050 return C1; // undef lshr undef -> undef 1051 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 1052 // undef lshr X -> 0 1053 case Instruction::AShr: 1054 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 1055 return Constant::getAllOnesValue(C1->getType()); 1056 else if (isa<UndefValue>(C1)) 1057 return C1; // undef ashr undef -> undef 1058 else 1059 return C1; // X ashr undef --> X 1060 case Instruction::Shl: 1061 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 1062 return C1; // undef shl undef -> undef 1063 // undef << X -> 0 or X << undef -> 0 1064 return Constant::getNullValue(C1->getType()); 1065 } 1066 } 1067 1068 // Handle simplifications when the RHS is a constant int. 1069 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1070 switch (Opcode) { 1071 case Instruction::Add: 1072 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1073 break; 1074 case Instruction::Sub: 1075 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1076 break; 1077 case Instruction::Mul: 1078 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1079 if (CI2->equalsInt(1)) 1080 return C1; // X * 1 == X 1081 break; 1082 case Instruction::UDiv: 1083 case Instruction::SDiv: 1084 if (CI2->equalsInt(1)) 1085 return C1; // X / 1 == X 1086 if (CI2->equalsInt(0)) 1087 return UndefValue::get(CI2->getType()); // X / 0 == undef 1088 break; 1089 case Instruction::URem: 1090 case Instruction::SRem: 1091 if (CI2->equalsInt(1)) 1092 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1093 if (CI2->equalsInt(0)) 1094 return UndefValue::get(CI2->getType()); // X % 0 == undef 1095 break; 1096 case Instruction::And: 1097 if (CI2->isZero()) return C2; // X & 0 == 0 1098 if (CI2->isAllOnesValue()) 1099 return C1; // X & -1 == X 1100 1101 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1102 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1103 if (CE1->getOpcode() == Instruction::ZExt) { 1104 unsigned DstWidth = CI2->getType()->getBitWidth(); 1105 unsigned SrcWidth = 1106 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1107 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1108 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1109 return C1; 1110 } 1111 1112 // If and'ing the address of a global with a constant, fold it. 1113 if (CE1->getOpcode() == Instruction::PtrToInt && 1114 isa<GlobalValue>(CE1->getOperand(0))) { 1115 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1116 1117 // Functions are at least 4-byte aligned. 1118 unsigned GVAlign = GV->getAlignment(); 1119 if (isa<Function>(GV)) 1120 GVAlign = std::max(GVAlign, 4U); 1121 1122 if (GVAlign > 1) { 1123 unsigned DstWidth = CI2->getType()->getBitWidth(); 1124 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1125 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1126 1127 // If checking bits we know are clear, return zero. 1128 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1129 return Constant::getNullValue(CI2->getType()); 1130 } 1131 } 1132 } 1133 break; 1134 case Instruction::Or: 1135 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1136 if (CI2->isAllOnesValue()) 1137 return C2; // X | -1 == -1 1138 break; 1139 case Instruction::Xor: 1140 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1141 1142 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1143 switch (CE1->getOpcode()) { 1144 default: break; 1145 case Instruction::ICmp: 1146 case Instruction::FCmp: 1147 // cmp pred ^ true -> cmp !pred 1148 assert(CI2->equalsInt(1)); 1149 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1150 pred = CmpInst::getInversePredicate(pred); 1151 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1152 CE1->getOperand(1)); 1153 } 1154 } 1155 break; 1156 case Instruction::AShr: 1157 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1158 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1159 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1160 return ConstantExpr::getLShr(C1, C2); 1161 break; 1162 } 1163 } else if (isa<ConstantInt>(C1)) { 1164 // If C1 is a ConstantInt and C2 is not, swap the operands. 1165 if (Instruction::isCommutative(Opcode)) 1166 return ConstantExpr::get(Opcode, C2, C1); 1167 } 1168 1169 // At this point we know neither constant is an UndefValue. 1170 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1171 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1172 using namespace APIntOps; 1173 const APInt &C1V = CI1->getValue(); 1174 const APInt &C2V = CI2->getValue(); 1175 switch (Opcode) { 1176 default: 1177 break; 1178 case Instruction::Add: 1179 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1180 case Instruction::Sub: 1181 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1182 case Instruction::Mul: 1183 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1184 case Instruction::UDiv: 1185 assert(!CI2->isNullValue() && "Div by zero handled above"); 1186 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1187 case Instruction::SDiv: 1188 assert(!CI2->isNullValue() && "Div by zero handled above"); 1189 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1190 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1191 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1192 case Instruction::URem: 1193 assert(!CI2->isNullValue() && "Div by zero handled above"); 1194 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1195 case Instruction::SRem: 1196 assert(!CI2->isNullValue() && "Div by zero handled above"); 1197 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1198 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1199 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1200 case Instruction::And: 1201 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1202 case Instruction::Or: 1203 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1204 case Instruction::Xor: 1205 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1206 case Instruction::Shl: { 1207 uint32_t shiftAmt = C2V.getZExtValue(); 1208 if (shiftAmt < C1V.getBitWidth()) 1209 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1210 else 1211 return UndefValue::get(C1->getType()); // too big shift is undef 1212 } 1213 case Instruction::LShr: { 1214 uint32_t shiftAmt = C2V.getZExtValue(); 1215 if (shiftAmt < C1V.getBitWidth()) 1216 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1217 else 1218 return UndefValue::get(C1->getType()); // too big shift is undef 1219 } 1220 case Instruction::AShr: { 1221 uint32_t shiftAmt = C2V.getZExtValue(); 1222 if (shiftAmt < C1V.getBitWidth()) 1223 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1224 else 1225 return UndefValue::get(C1->getType()); // too big shift is undef 1226 } 1227 } 1228 } 1229 1230 switch (Opcode) { 1231 case Instruction::SDiv: 1232 case Instruction::UDiv: 1233 case Instruction::URem: 1234 case Instruction::SRem: 1235 case Instruction::LShr: 1236 case Instruction::AShr: 1237 case Instruction::Shl: 1238 if (CI1->equalsInt(0)) return C1; 1239 break; 1240 default: 1241 break; 1242 } 1243 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1244 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1245 APFloat C1V = CFP1->getValueAPF(); 1246 APFloat C2V = CFP2->getValueAPF(); 1247 APFloat C3V = C1V; // copy for modification 1248 switch (Opcode) { 1249 default: 1250 break; 1251 case Instruction::FAdd: 1252 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1253 return ConstantFP::get(C1->getContext(), C3V); 1254 case Instruction::FSub: 1255 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1256 return ConstantFP::get(C1->getContext(), C3V); 1257 case Instruction::FMul: 1258 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1259 return ConstantFP::get(C1->getContext(), C3V); 1260 case Instruction::FDiv: 1261 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1262 return ConstantFP::get(C1->getContext(), C3V); 1263 case Instruction::FRem: 1264 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1265 return ConstantFP::get(C1->getContext(), C3V); 1266 } 1267 } 1268 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1269 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 1270 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 1271 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 1272 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 1273 std::vector<Constant*> Res; 1274 Type* EltTy = VTy->getElementType(); 1275 Constant *C1 = 0; 1276 Constant *C2 = 0; 1277 switch (Opcode) { 1278 default: 1279 break; 1280 case Instruction::Add: 1281 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1282 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1283 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1284 Res.push_back(ConstantExpr::getAdd(C1, C2)); 1285 } 1286 return ConstantVector::get(Res); 1287 case Instruction::FAdd: 1288 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1289 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1290 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1291 Res.push_back(ConstantExpr::getFAdd(C1, C2)); 1292 } 1293 return ConstantVector::get(Res); 1294 case Instruction::Sub: 1295 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1296 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1297 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1298 Res.push_back(ConstantExpr::getSub(C1, C2)); 1299 } 1300 return ConstantVector::get(Res); 1301 case Instruction::FSub: 1302 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1303 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1304 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1305 Res.push_back(ConstantExpr::getFSub(C1, C2)); 1306 } 1307 return ConstantVector::get(Res); 1308 case Instruction::Mul: 1309 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1310 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1311 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1312 Res.push_back(ConstantExpr::getMul(C1, C2)); 1313 } 1314 return ConstantVector::get(Res); 1315 case Instruction::FMul: 1316 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1317 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1318 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1319 Res.push_back(ConstantExpr::getFMul(C1, C2)); 1320 } 1321 return ConstantVector::get(Res); 1322 case Instruction::UDiv: 1323 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1324 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1325 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1326 Res.push_back(ConstantExpr::getUDiv(C1, C2)); 1327 } 1328 return ConstantVector::get(Res); 1329 case Instruction::SDiv: 1330 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1331 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1332 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1333 Res.push_back(ConstantExpr::getSDiv(C1, C2)); 1334 } 1335 return ConstantVector::get(Res); 1336 case Instruction::FDiv: 1337 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1338 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1339 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1340 Res.push_back(ConstantExpr::getFDiv(C1, C2)); 1341 } 1342 return ConstantVector::get(Res); 1343 case Instruction::URem: 1344 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1345 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1346 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1347 Res.push_back(ConstantExpr::getURem(C1, C2)); 1348 } 1349 return ConstantVector::get(Res); 1350 case Instruction::SRem: 1351 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1352 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1353 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1354 Res.push_back(ConstantExpr::getSRem(C1, C2)); 1355 } 1356 return ConstantVector::get(Res); 1357 case Instruction::FRem: 1358 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1359 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1360 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1361 Res.push_back(ConstantExpr::getFRem(C1, C2)); 1362 } 1363 return ConstantVector::get(Res); 1364 case Instruction::And: 1365 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1366 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1367 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1368 Res.push_back(ConstantExpr::getAnd(C1, C2)); 1369 } 1370 return ConstantVector::get(Res); 1371 case Instruction::Or: 1372 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1373 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1374 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1375 Res.push_back(ConstantExpr::getOr(C1, C2)); 1376 } 1377 return ConstantVector::get(Res); 1378 case Instruction::Xor: 1379 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1380 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1381 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1382 Res.push_back(ConstantExpr::getXor(C1, C2)); 1383 } 1384 return ConstantVector::get(Res); 1385 case Instruction::LShr: 1386 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1387 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1388 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1389 Res.push_back(ConstantExpr::getLShr(C1, C2)); 1390 } 1391 return ConstantVector::get(Res); 1392 case Instruction::AShr: 1393 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1394 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1395 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1396 Res.push_back(ConstantExpr::getAShr(C1, C2)); 1397 } 1398 return ConstantVector::get(Res); 1399 case Instruction::Shl: 1400 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1401 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1402 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1403 Res.push_back(ConstantExpr::getShl(C1, C2)); 1404 } 1405 return ConstantVector::get(Res); 1406 } 1407 } 1408 } 1409 1410 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1411 // There are many possible foldings we could do here. We should probably 1412 // at least fold add of a pointer with an integer into the appropriate 1413 // getelementptr. This will improve alias analysis a bit. 1414 1415 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1416 // (a + (b + c)). 1417 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1418 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1419 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1420 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1421 } 1422 } else if (isa<ConstantExpr>(C2)) { 1423 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1424 // other way if possible. 1425 if (Instruction::isCommutative(Opcode)) 1426 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1427 } 1428 1429 // i1 can be simplified in many cases. 1430 if (C1->getType()->isIntegerTy(1)) { 1431 switch (Opcode) { 1432 case Instruction::Add: 1433 case Instruction::Sub: 1434 return ConstantExpr::getXor(C1, C2); 1435 case Instruction::Mul: 1436 return ConstantExpr::getAnd(C1, C2); 1437 case Instruction::Shl: 1438 case Instruction::LShr: 1439 case Instruction::AShr: 1440 // We can assume that C2 == 0. If it were one the result would be 1441 // undefined because the shift value is as large as the bitwidth. 1442 return C1; 1443 case Instruction::SDiv: 1444 case Instruction::UDiv: 1445 // We can assume that C2 == 1. If it were zero the result would be 1446 // undefined through division by zero. 1447 return C1; 1448 case Instruction::URem: 1449 case Instruction::SRem: 1450 // We can assume that C2 == 1. If it were zero the result would be 1451 // undefined through division by zero. 1452 return ConstantInt::getFalse(C1->getContext()); 1453 default: 1454 break; 1455 } 1456 } 1457 1458 // We don't know how to fold this. 1459 return 0; 1460 } 1461 1462 /// isZeroSizedType - This type is zero sized if its an array or structure of 1463 /// zero sized types. The only leaf zero sized type is an empty structure. 1464 static bool isMaybeZeroSizedType(Type *Ty) { 1465 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1466 if (STy->isOpaque()) return true; // Can't say. 1467 1468 // If all of elements have zero size, this does too. 1469 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1470 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1471 return true; 1472 1473 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1474 return isMaybeZeroSizedType(ATy->getElementType()); 1475 } 1476 return false; 1477 } 1478 1479 /// IdxCompare - Compare the two constants as though they were getelementptr 1480 /// indices. This allows coersion of the types to be the same thing. 1481 /// 1482 /// If the two constants are the "same" (after coersion), return 0. If the 1483 /// first is less than the second, return -1, if the second is less than the 1484 /// first, return 1. If the constants are not integral, return -2. 1485 /// 1486 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1487 if (C1 == C2) return 0; 1488 1489 // Ok, we found a different index. If they are not ConstantInt, we can't do 1490 // anything with them. 1491 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1492 return -2; // don't know! 1493 1494 // Ok, we have two differing integer indices. Sign extend them to be the same 1495 // type. Long is always big enough, so we use it. 1496 if (!C1->getType()->isIntegerTy(64)) 1497 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1498 1499 if (!C2->getType()->isIntegerTy(64)) 1500 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1501 1502 if (C1 == C2) return 0; // They are equal 1503 1504 // If the type being indexed over is really just a zero sized type, there is 1505 // no pointer difference being made here. 1506 if (isMaybeZeroSizedType(ElTy)) 1507 return -2; // dunno. 1508 1509 // If they are really different, now that they are the same type, then we 1510 // found a difference! 1511 if (cast<ConstantInt>(C1)->getSExtValue() < 1512 cast<ConstantInt>(C2)->getSExtValue()) 1513 return -1; 1514 else 1515 return 1; 1516 } 1517 1518 /// evaluateFCmpRelation - This function determines if there is anything we can 1519 /// decide about the two constants provided. This doesn't need to handle simple 1520 /// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1521 /// If we can determine that the two constants have a particular relation to 1522 /// each other, we should return the corresponding FCmpInst predicate, 1523 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1524 /// ConstantFoldCompareInstruction. 1525 /// 1526 /// To simplify this code we canonicalize the relation so that the first 1527 /// operand is always the most "complex" of the two. We consider ConstantFP 1528 /// to be the simplest, and ConstantExprs to be the most complex. 1529 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1530 assert(V1->getType() == V2->getType() && 1531 "Cannot compare values of different types!"); 1532 1533 // No compile-time operations on this type yet. 1534 if (V1->getType()->isPPC_FP128Ty()) 1535 return FCmpInst::BAD_FCMP_PREDICATE; 1536 1537 // Handle degenerate case quickly 1538 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1539 1540 if (!isa<ConstantExpr>(V1)) { 1541 if (!isa<ConstantExpr>(V2)) { 1542 // We distilled thisUse the standard constant folder for a few cases 1543 ConstantInt *R = 0; 1544 R = dyn_cast<ConstantInt>( 1545 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1546 if (R && !R->isZero()) 1547 return FCmpInst::FCMP_OEQ; 1548 R = dyn_cast<ConstantInt>( 1549 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1550 if (R && !R->isZero()) 1551 return FCmpInst::FCMP_OLT; 1552 R = dyn_cast<ConstantInt>( 1553 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1554 if (R && !R->isZero()) 1555 return FCmpInst::FCMP_OGT; 1556 1557 // Nothing more we can do 1558 return FCmpInst::BAD_FCMP_PREDICATE; 1559 } 1560 1561 // If the first operand is simple and second is ConstantExpr, swap operands. 1562 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1563 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1564 return FCmpInst::getSwappedPredicate(SwappedRelation); 1565 } else { 1566 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1567 // constantexpr or a simple constant. 1568 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1569 switch (CE1->getOpcode()) { 1570 case Instruction::FPTrunc: 1571 case Instruction::FPExt: 1572 case Instruction::UIToFP: 1573 case Instruction::SIToFP: 1574 // We might be able to do something with these but we don't right now. 1575 break; 1576 default: 1577 break; 1578 } 1579 } 1580 // There are MANY other foldings that we could perform here. They will 1581 // probably be added on demand, as they seem needed. 1582 return FCmpInst::BAD_FCMP_PREDICATE; 1583 } 1584 1585 /// evaluateICmpRelation - This function determines if there is anything we can 1586 /// decide about the two constants provided. This doesn't need to handle simple 1587 /// things like integer comparisons, but should instead handle ConstantExprs 1588 /// and GlobalValues. If we can determine that the two constants have a 1589 /// particular relation to each other, we should return the corresponding ICmp 1590 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1591 /// 1592 /// To simplify this code we canonicalize the relation so that the first 1593 /// operand is always the most "complex" of the two. We consider simple 1594 /// constants (like ConstantInt) to be the simplest, followed by 1595 /// GlobalValues, followed by ConstantExpr's (the most complex). 1596 /// 1597 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1598 bool isSigned) { 1599 assert(V1->getType() == V2->getType() && 1600 "Cannot compare different types of values!"); 1601 if (V1 == V2) return ICmpInst::ICMP_EQ; 1602 1603 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1604 !isa<BlockAddress>(V1)) { 1605 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1606 !isa<BlockAddress>(V2)) { 1607 // We distilled this down to a simple case, use the standard constant 1608 // folder. 1609 ConstantInt *R = 0; 1610 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1611 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1612 if (R && !R->isZero()) 1613 return pred; 1614 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1615 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1616 if (R && !R->isZero()) 1617 return pred; 1618 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1619 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1620 if (R && !R->isZero()) 1621 return pred; 1622 1623 // If we couldn't figure it out, bail. 1624 return ICmpInst::BAD_ICMP_PREDICATE; 1625 } 1626 1627 // If the first operand is simple, swap operands. 1628 ICmpInst::Predicate SwappedRelation = 1629 evaluateICmpRelation(V2, V1, isSigned); 1630 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1631 return ICmpInst::getSwappedPredicate(SwappedRelation); 1632 1633 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1634 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1635 ICmpInst::Predicate SwappedRelation = 1636 evaluateICmpRelation(V2, V1, isSigned); 1637 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1638 return ICmpInst::getSwappedPredicate(SwappedRelation); 1639 return ICmpInst::BAD_ICMP_PREDICATE; 1640 } 1641 1642 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1643 // constant (which, since the types must match, means that it's a 1644 // ConstantPointerNull). 1645 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1646 // Don't try to decide equality of aliases. 1647 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1648 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1649 return ICmpInst::ICMP_NE; 1650 } else if (isa<BlockAddress>(V2)) { 1651 return ICmpInst::ICMP_NE; // Globals never equal labels. 1652 } else { 1653 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1654 // GlobalVals can never be null unless they have external weak linkage. 1655 // We don't try to evaluate aliases here. 1656 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1657 return ICmpInst::ICMP_NE; 1658 } 1659 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1660 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1661 ICmpInst::Predicate SwappedRelation = 1662 evaluateICmpRelation(V2, V1, isSigned); 1663 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1664 return ICmpInst::getSwappedPredicate(SwappedRelation); 1665 return ICmpInst::BAD_ICMP_PREDICATE; 1666 } 1667 1668 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1669 // constant (which, since the types must match, means that it is a 1670 // ConstantPointerNull). 1671 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1672 // Block address in another function can't equal this one, but block 1673 // addresses in the current function might be the same if blocks are 1674 // empty. 1675 if (BA2->getFunction() != BA->getFunction()) 1676 return ICmpInst::ICMP_NE; 1677 } else { 1678 // Block addresses aren't null, don't equal the address of globals. 1679 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1680 "Canonicalization guarantee!"); 1681 return ICmpInst::ICMP_NE; 1682 } 1683 } else { 1684 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1685 // constantexpr, a global, block address, or a simple constant. 1686 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1687 Constant *CE1Op0 = CE1->getOperand(0); 1688 1689 switch (CE1->getOpcode()) { 1690 case Instruction::Trunc: 1691 case Instruction::FPTrunc: 1692 case Instruction::FPExt: 1693 case Instruction::FPToUI: 1694 case Instruction::FPToSI: 1695 break; // We can't evaluate floating point casts or truncations. 1696 1697 case Instruction::UIToFP: 1698 case Instruction::SIToFP: 1699 case Instruction::BitCast: 1700 case Instruction::ZExt: 1701 case Instruction::SExt: 1702 // If the cast is not actually changing bits, and the second operand is a 1703 // null pointer, do the comparison with the pre-casted value. 1704 if (V2->isNullValue() && 1705 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1706 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1707 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1708 return evaluateICmpRelation(CE1Op0, 1709 Constant::getNullValue(CE1Op0->getType()), 1710 isSigned); 1711 } 1712 break; 1713 1714 case Instruction::GetElementPtr: 1715 // Ok, since this is a getelementptr, we know that the constant has a 1716 // pointer type. Check the various cases. 1717 if (isa<ConstantPointerNull>(V2)) { 1718 // If we are comparing a GEP to a null pointer, check to see if the base 1719 // of the GEP equals the null pointer. 1720 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1721 if (GV->hasExternalWeakLinkage()) 1722 // Weak linkage GVals could be zero or not. We're comparing that 1723 // to null pointer so its greater-or-equal 1724 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1725 else 1726 // If its not weak linkage, the GVal must have a non-zero address 1727 // so the result is greater-than 1728 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1729 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1730 // If we are indexing from a null pointer, check to see if we have any 1731 // non-zero indices. 1732 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1733 if (!CE1->getOperand(i)->isNullValue()) 1734 // Offsetting from null, must not be equal. 1735 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1736 // Only zero indexes from null, must still be zero. 1737 return ICmpInst::ICMP_EQ; 1738 } 1739 // Otherwise, we can't really say if the first operand is null or not. 1740 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1741 if (isa<ConstantPointerNull>(CE1Op0)) { 1742 if (GV2->hasExternalWeakLinkage()) 1743 // Weak linkage GVals could be zero or not. We're comparing it to 1744 // a null pointer, so its less-or-equal 1745 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1746 else 1747 // If its not weak linkage, the GVal must have a non-zero address 1748 // so the result is less-than 1749 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1750 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1751 if (GV == GV2) { 1752 // If this is a getelementptr of the same global, then it must be 1753 // different. Because the types must match, the getelementptr could 1754 // only have at most one index, and because we fold getelementptr's 1755 // with a single zero index, it must be nonzero. 1756 assert(CE1->getNumOperands() == 2 && 1757 !CE1->getOperand(1)->isNullValue() && 1758 "Surprising getelementptr!"); 1759 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1760 } else { 1761 // If they are different globals, we don't know what the value is, 1762 // but they can't be equal. 1763 return ICmpInst::ICMP_NE; 1764 } 1765 } 1766 } else { 1767 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1768 Constant *CE2Op0 = CE2->getOperand(0); 1769 1770 // There are MANY other foldings that we could perform here. They will 1771 // probably be added on demand, as they seem needed. 1772 switch (CE2->getOpcode()) { 1773 default: break; 1774 case Instruction::GetElementPtr: 1775 // By far the most common case to handle is when the base pointers are 1776 // obviously to the same or different globals. 1777 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1778 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1779 return ICmpInst::ICMP_NE; 1780 // Ok, we know that both getelementptr instructions are based on the 1781 // same global. From this, we can precisely determine the relative 1782 // ordering of the resultant pointers. 1783 unsigned i = 1; 1784 1785 // The logic below assumes that the result of the comparison 1786 // can be determined by finding the first index that differs. 1787 // This doesn't work if there is over-indexing in any 1788 // subsequent indices, so check for that case first. 1789 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1790 !CE2->isGEPWithNoNotionalOverIndexing()) 1791 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1792 1793 // Compare all of the operands the GEP's have in common. 1794 gep_type_iterator GTI = gep_type_begin(CE1); 1795 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1796 ++i, ++GTI) 1797 switch (IdxCompare(CE1->getOperand(i), 1798 CE2->getOperand(i), GTI.getIndexedType())) { 1799 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1800 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1801 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1802 } 1803 1804 // Ok, we ran out of things they have in common. If any leftovers 1805 // are non-zero then we have a difference, otherwise we are equal. 1806 for (; i < CE1->getNumOperands(); ++i) 1807 if (!CE1->getOperand(i)->isNullValue()) { 1808 if (isa<ConstantInt>(CE1->getOperand(i))) 1809 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1810 else 1811 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1812 } 1813 1814 for (; i < CE2->getNumOperands(); ++i) 1815 if (!CE2->getOperand(i)->isNullValue()) { 1816 if (isa<ConstantInt>(CE2->getOperand(i))) 1817 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1818 else 1819 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1820 } 1821 return ICmpInst::ICMP_EQ; 1822 } 1823 } 1824 } 1825 default: 1826 break; 1827 } 1828 } 1829 1830 return ICmpInst::BAD_ICMP_PREDICATE; 1831 } 1832 1833 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1834 Constant *C1, Constant *C2) { 1835 Type *ResultTy; 1836 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1837 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1838 VT->getNumElements()); 1839 else 1840 ResultTy = Type::getInt1Ty(C1->getContext()); 1841 1842 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1843 if (pred == FCmpInst::FCMP_FALSE) 1844 return Constant::getNullValue(ResultTy); 1845 1846 if (pred == FCmpInst::FCMP_TRUE) 1847 return Constant::getAllOnesValue(ResultTy); 1848 1849 // Handle some degenerate cases first 1850 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1851 // For EQ and NE, we can always pick a value for the undef to make the 1852 // predicate pass or fail, so we can return undef. 1853 // Also, if both operands are undef, we can return undef. 1854 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || 1855 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1856 return UndefValue::get(ResultTy); 1857 // Otherwise, pick the same value as the non-undef operand, and fold 1858 // it to true or false. 1859 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1860 } 1861 1862 // No compile-time operations on this type yet. 1863 if (C1->getType()->isPPC_FP128Ty()) 1864 return 0; 1865 1866 // icmp eq/ne(null,GV) -> false/true 1867 if (C1->isNullValue()) { 1868 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1869 // Don't try to evaluate aliases. External weak GV can be null. 1870 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1871 if (pred == ICmpInst::ICMP_EQ) 1872 return ConstantInt::getFalse(C1->getContext()); 1873 else if (pred == ICmpInst::ICMP_NE) 1874 return ConstantInt::getTrue(C1->getContext()); 1875 } 1876 // icmp eq/ne(GV,null) -> false/true 1877 } else if (C2->isNullValue()) { 1878 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1879 // Don't try to evaluate aliases. External weak GV can be null. 1880 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1881 if (pred == ICmpInst::ICMP_EQ) 1882 return ConstantInt::getFalse(C1->getContext()); 1883 else if (pred == ICmpInst::ICMP_NE) 1884 return ConstantInt::getTrue(C1->getContext()); 1885 } 1886 } 1887 1888 // If the comparison is a comparison between two i1's, simplify it. 1889 if (C1->getType()->isIntegerTy(1)) { 1890 switch(pred) { 1891 case ICmpInst::ICMP_EQ: 1892 if (isa<ConstantInt>(C2)) 1893 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1894 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1895 case ICmpInst::ICMP_NE: 1896 return ConstantExpr::getXor(C1, C2); 1897 default: 1898 break; 1899 } 1900 } 1901 1902 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1903 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1904 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1905 switch (pred) { 1906 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1907 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1908 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1909 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1910 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1911 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1912 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1913 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1914 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1915 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1916 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1917 } 1918 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1919 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1920 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1921 APFloat::cmpResult R = C1V.compare(C2V); 1922 switch (pred) { 1923 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1924 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1925 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1926 case FCmpInst::FCMP_UNO: 1927 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1928 case FCmpInst::FCMP_ORD: 1929 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1930 case FCmpInst::FCMP_UEQ: 1931 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1932 R==APFloat::cmpEqual); 1933 case FCmpInst::FCMP_OEQ: 1934 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1935 case FCmpInst::FCMP_UNE: 1936 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1937 case FCmpInst::FCMP_ONE: 1938 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1939 R==APFloat::cmpGreaterThan); 1940 case FCmpInst::FCMP_ULT: 1941 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1942 R==APFloat::cmpLessThan); 1943 case FCmpInst::FCMP_OLT: 1944 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1945 case FCmpInst::FCMP_UGT: 1946 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1947 R==APFloat::cmpGreaterThan); 1948 case FCmpInst::FCMP_OGT: 1949 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1950 case FCmpInst::FCMP_ULE: 1951 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1952 case FCmpInst::FCMP_OLE: 1953 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1954 R==APFloat::cmpEqual); 1955 case FCmpInst::FCMP_UGE: 1956 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1957 case FCmpInst::FCMP_OGE: 1958 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1959 R==APFloat::cmpEqual); 1960 } 1961 } else if (C1->getType()->isVectorTy()) { 1962 SmallVector<Constant*, 16> C1Elts, C2Elts; 1963 C1->getVectorElements(C1Elts); 1964 C2->getVectorElements(C2Elts); 1965 if (C1Elts.empty() || C2Elts.empty()) 1966 return 0; 1967 1968 // If we can constant fold the comparison of each element, constant fold 1969 // the whole vector comparison. 1970 SmallVector<Constant*, 4> ResElts; 1971 // Compare the elements, producing an i1 result or constant expr. 1972 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) 1973 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1974 1975 return ConstantVector::get(ResElts); 1976 } 1977 1978 if (C1->getType()->isFloatingPointTy()) { 1979 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1980 switch (evaluateFCmpRelation(C1, C2)) { 1981 default: llvm_unreachable("Unknown relation!"); 1982 case FCmpInst::FCMP_UNO: 1983 case FCmpInst::FCMP_ORD: 1984 case FCmpInst::FCMP_UEQ: 1985 case FCmpInst::FCMP_UNE: 1986 case FCmpInst::FCMP_ULT: 1987 case FCmpInst::FCMP_UGT: 1988 case FCmpInst::FCMP_ULE: 1989 case FCmpInst::FCMP_UGE: 1990 case FCmpInst::FCMP_TRUE: 1991 case FCmpInst::FCMP_FALSE: 1992 case FCmpInst::BAD_FCMP_PREDICATE: 1993 break; // Couldn't determine anything about these constants. 1994 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1995 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1996 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1997 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1998 break; 1999 case FCmpInst::FCMP_OLT: // We know that C1 < C2 2000 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2001 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 2002 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 2003 break; 2004 case FCmpInst::FCMP_OGT: // We know that C1 > C2 2005 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2006 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 2007 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 2008 break; 2009 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 2010 // We can only partially decide this relation. 2011 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2012 Result = 0; 2013 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2014 Result = 1; 2015 break; 2016 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 2017 // We can only partially decide this relation. 2018 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2019 Result = 0; 2020 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2021 Result = 1; 2022 break; 2023 case FCmpInst::FCMP_ONE: // We know that C1 != C2 2024 // We can only partially decide this relation. 2025 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 2026 Result = 0; 2027 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 2028 Result = 1; 2029 break; 2030 } 2031 2032 // If we evaluated the result, return it now. 2033 if (Result != -1) 2034 return ConstantInt::get(ResultTy, Result); 2035 2036 } else { 2037 // Evaluate the relation between the two constants, per the predicate. 2038 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 2039 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 2040 default: llvm_unreachable("Unknown relational!"); 2041 case ICmpInst::BAD_ICMP_PREDICATE: 2042 break; // Couldn't determine anything about these constants. 2043 case ICmpInst::ICMP_EQ: // We know the constants are equal! 2044 // If we know the constants are equal, we can decide the result of this 2045 // computation precisely. 2046 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 2047 break; 2048 case ICmpInst::ICMP_ULT: 2049 switch (pred) { 2050 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 2051 Result = 1; break; 2052 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 2053 Result = 0; break; 2054 } 2055 break; 2056 case ICmpInst::ICMP_SLT: 2057 switch (pred) { 2058 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 2059 Result = 1; break; 2060 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 2061 Result = 0; break; 2062 } 2063 break; 2064 case ICmpInst::ICMP_UGT: 2065 switch (pred) { 2066 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 2067 Result = 1; break; 2068 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 2069 Result = 0; break; 2070 } 2071 break; 2072 case ICmpInst::ICMP_SGT: 2073 switch (pred) { 2074 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 2075 Result = 1; break; 2076 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 2077 Result = 0; break; 2078 } 2079 break; 2080 case ICmpInst::ICMP_ULE: 2081 if (pred == ICmpInst::ICMP_UGT) Result = 0; 2082 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 2083 break; 2084 case ICmpInst::ICMP_SLE: 2085 if (pred == ICmpInst::ICMP_SGT) Result = 0; 2086 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 2087 break; 2088 case ICmpInst::ICMP_UGE: 2089 if (pred == ICmpInst::ICMP_ULT) Result = 0; 2090 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 2091 break; 2092 case ICmpInst::ICMP_SGE: 2093 if (pred == ICmpInst::ICMP_SLT) Result = 0; 2094 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 2095 break; 2096 case ICmpInst::ICMP_NE: 2097 if (pred == ICmpInst::ICMP_EQ) Result = 0; 2098 if (pred == ICmpInst::ICMP_NE) Result = 1; 2099 break; 2100 } 2101 2102 // If we evaluated the result, return it now. 2103 if (Result != -1) 2104 return ConstantInt::get(ResultTy, Result); 2105 2106 // If the right hand side is a bitcast, try using its inverse to simplify 2107 // it by moving it to the left hand side. We can't do this if it would turn 2108 // a vector compare into a scalar compare or visa versa. 2109 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 2110 Constant *CE2Op0 = CE2->getOperand(0); 2111 if (CE2->getOpcode() == Instruction::BitCast && 2112 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 2113 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 2114 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 2115 } 2116 } 2117 2118 // If the left hand side is an extension, try eliminating it. 2119 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 2120 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 2121 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 2122 Constant *CE1Op0 = CE1->getOperand(0); 2123 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2124 if (CE1Inverse == CE1Op0) { 2125 // Check whether we can safely truncate the right hand side. 2126 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2127 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { 2128 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2129 } 2130 } 2131 } 2132 } 2133 2134 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2135 (C1->isNullValue() && !C2->isNullValue())) { 2136 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2137 // other way if possible. 2138 // Also, if C1 is null and C2 isn't, flip them around. 2139 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2140 return ConstantExpr::getICmp(pred, C2, C1); 2141 } 2142 } 2143 return 0; 2144 } 2145 2146 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices 2147 /// is "inbounds". 2148 template<typename IndexTy> 2149 static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) { 2150 // No indices means nothing that could be out of bounds. 2151 if (NumIdx == 0) return true; 2152 2153 // If the first index is zero, it's in bounds. 2154 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2155 2156 // If the first index is one and all the rest are zero, it's in bounds, 2157 // by the one-past-the-end rule. 2158 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2159 return false; 2160 for (unsigned i = 1, e = NumIdx; i != e; ++i) 2161 if (!cast<Constant>(Idxs[i])->isNullValue()) 2162 return false; 2163 return true; 2164 } 2165 2166 template<typename IndexTy> 2167 static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 2168 bool inBounds, 2169 ArrayRef<IndexTy> Idxs) { 2170 if (Idxs.empty()) return C; 2171 Constant *Idx0 = cast<Constant>(Idxs[0]); 2172 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2173 return C; 2174 2175 if (isa<UndefValue>(C)) { 2176 PointerType *Ptr = cast<PointerType>(C->getType()); 2177 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs.begin(), Idxs.end()); 2178 assert(Ty != 0 && "Invalid indices for GEP!"); 2179 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2180 } 2181 2182 if (C->isNullValue()) { 2183 bool isNull = true; 2184 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2185 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2186 isNull = false; 2187 break; 2188 } 2189 if (isNull) { 2190 PointerType *Ptr = cast<PointerType>(C->getType()); 2191 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs.begin(), 2192 Idxs.end()); 2193 assert(Ty != 0 && "Invalid indices for GEP!"); 2194 return ConstantPointerNull::get(PointerType::get(Ty, 2195 Ptr->getAddressSpace())); 2196 } 2197 } 2198 2199 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2200 // Combine Indices - If the source pointer to this getelementptr instruction 2201 // is a getelementptr instruction, combine the indices of the two 2202 // getelementptr instructions into a single instruction. 2203 // 2204 if (CE->getOpcode() == Instruction::GetElementPtr) { 2205 Type *LastTy = 0; 2206 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2207 I != E; ++I) 2208 LastTy = *I; 2209 2210 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) { 2211 SmallVector<Value*, 16> NewIndices; 2212 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2213 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2214 NewIndices.push_back(CE->getOperand(i)); 2215 2216 // Add the last index of the source with the first index of the new GEP. 2217 // Make sure to handle the case when they are actually different types. 2218 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2219 // Otherwise it must be an array. 2220 if (!Idx0->isNullValue()) { 2221 Type *IdxTy = Combined->getType(); 2222 if (IdxTy != Idx0->getType()) { 2223 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2224 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2225 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2226 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2227 } else { 2228 Combined = 2229 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2230 } 2231 } 2232 2233 NewIndices.push_back(Combined); 2234 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2235 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 2236 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 2237 &NewIndices[0], 2238 NewIndices.size()) : 2239 ConstantExpr::getGetElementPtr(CE->getOperand(0), 2240 &NewIndices[0], 2241 NewIndices.size()); 2242 } 2243 } 2244 2245 // Implement folding of: 2246 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2247 // i64 0, i64 0) 2248 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2249 // 2250 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2251 if (PointerType *SPT = 2252 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 2253 if (ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 2254 if (ArrayType *CAT = 2255 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 2256 if (CAT->getElementType() == SAT->getElementType()) 2257 return inBounds ? 2258 ConstantExpr::getInBoundsGetElementPtr( 2259 (Constant*)CE->getOperand(0), Idxs.data(), Idxs.size()) : 2260 ConstantExpr::getGetElementPtr( 2261 (Constant*)CE->getOperand(0), Idxs.data(), Idxs.size()); 2262 } 2263 } 2264 2265 // Check to see if any array indices are not within the corresponding 2266 // notional array bounds. If so, try to determine if they can be factored 2267 // out into preceding dimensions. 2268 bool Unknown = false; 2269 SmallVector<Constant *, 8> NewIdxs; 2270 Type *Ty = C->getType(); 2271 Type *Prev = 0; 2272 for (unsigned i = 0, e = Idxs.size(); i != e; 2273 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2274 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2275 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2276 if (ATy->getNumElements() <= INT64_MAX && 2277 ATy->getNumElements() != 0 && 2278 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 2279 if (isa<SequentialType>(Prev)) { 2280 // It's out of range, but we can factor it into the prior 2281 // dimension. 2282 NewIdxs.resize(Idxs.size()); 2283 ConstantInt *Factor = ConstantInt::get(CI->getType(), 2284 ATy->getNumElements()); 2285 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2286 2287 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2288 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2289 2290 // Before adding, extend both operands to i64 to avoid 2291 // overflow trouble. 2292 if (!PrevIdx->getType()->isIntegerTy(64)) 2293 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2294 Type::getInt64Ty(Div->getContext())); 2295 if (!Div->getType()->isIntegerTy(64)) 2296 Div = ConstantExpr::getSExt(Div, 2297 Type::getInt64Ty(Div->getContext())); 2298 2299 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2300 } else { 2301 // It's out of range, but the prior dimension is a struct 2302 // so we can't do anything about it. 2303 Unknown = true; 2304 } 2305 } 2306 } else { 2307 // We don't know if it's in range or not. 2308 Unknown = true; 2309 } 2310 } 2311 2312 // If we did any factoring, start over with the adjusted indices. 2313 if (!NewIdxs.empty()) { 2314 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2315 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2316 return inBounds ? 2317 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), 2318 NewIdxs.size()) : 2319 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); 2320 } 2321 2322 // If all indices are known integers and normalized, we can do a simple 2323 // check for the "inbounds" property. 2324 if (!Unknown && !inBounds && 2325 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs.data(), Idxs.size())) 2326 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs.data(), Idxs.size()); 2327 2328 return 0; 2329 } 2330 2331 Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2332 bool inBounds, 2333 ArrayRef<Constant *> Idxs) { 2334 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2335 } 2336 2337 Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2338 bool inBounds, 2339 ArrayRef<Value *> Idxs) { 2340 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2341 } 2342