ConstantFold.cpp revision fdfeb6976f07ad10d809b922ed7376ba2a3539be
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/ADT/SmallVector.h" 28#include "llvm/Support/Compiler.h" 29#include "llvm/Support/ErrorHandling.h" 30#include "llvm/Support/GetElementPtrTypeIterator.h" 31#include "llvm/Support/ManagedStatic.h" 32#include "llvm/Support/MathExtras.h" 33#include <limits> 34using namespace llvm; 35 36//===----------------------------------------------------------------------===// 37// ConstantFold*Instruction Implementations 38//===----------------------------------------------------------------------===// 39 40/// BitCastConstantVector - Convert the specified ConstantVector node to the 41/// specified vector type. At this point, we know that the elements of the 42/// input vector constant are all simple integer or FP values. 43static Constant *BitCastConstantVector(ConstantVector *CV, 44 const VectorType *DstTy) { 45 // If this cast changes element count then we can't handle it here: 46 // doing so requires endianness information. This should be handled by 47 // Analysis/ConstantFolding.cpp 48 unsigned NumElts = DstTy->getNumElements(); 49 if (NumElts != CV->getNumOperands()) 50 return 0; 51 52 // Check to verify that all elements of the input are simple. 53 for (unsigned i = 0; i != NumElts; ++i) { 54 if (!isa<ConstantInt>(CV->getOperand(i)) && 55 !isa<ConstantFP>(CV->getOperand(i))) 56 return 0; 57 } 58 59 // Bitcast each element now. 60 std::vector<Constant*> Result; 61 const Type *DstEltTy = DstTy->getElementType(); 62 for (unsigned i = 0; i != NumElts; ++i) 63 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 64 DstEltTy)); 65 return ConstantVector::get(Result); 66} 67 68/// This function determines which opcode to use to fold two constant cast 69/// expressions together. It uses CastInst::isEliminableCastPair to determine 70/// the opcode. Consequently its just a wrapper around that function. 71/// @brief Determine if it is valid to fold a cast of a cast 72static unsigned 73foldConstantCastPair( 74 unsigned opc, ///< opcode of the second cast constant expression 75 ConstantExpr *Op, ///< the first cast constant expression 76 const Type *DstTy ///< desintation type of the first cast 77) { 78 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 79 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 80 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 81 82 // The the types and opcodes for the two Cast constant expressions 83 const Type *SrcTy = Op->getOperand(0)->getType(); 84 const Type *MidTy = Op->getType(); 85 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 86 Instruction::CastOps secondOp = Instruction::CastOps(opc); 87 88 // Let CastInst::isEliminableCastPair do the heavy lifting. 89 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 90 Type::getInt64Ty(DstTy->getContext())); 91} 92 93static Constant *FoldBitCast(Constant *V, const Type *DestTy) { 94 const Type *SrcTy = V->getType(); 95 if (SrcTy == DestTy) 96 return V; // no-op cast 97 98 // Check to see if we are casting a pointer to an aggregate to a pointer to 99 // the first element. If so, return the appropriate GEP instruction. 100 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) 101 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 102 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 103 SmallVector<Value*, 8> IdxList; 104 Value *Zero = 105 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 106 IdxList.push_back(Zero); 107 const Type *ElTy = PTy->getElementType(); 108 while (ElTy != DPTy->getElementType()) { 109 if (const StructType *STy = dyn_cast<StructType>(ElTy)) { 110 if (STy->getNumElements() == 0) break; 111 ElTy = STy->getElementType(0); 112 IdxList.push_back(Zero); 113 } else if (const SequentialType *STy = 114 dyn_cast<SequentialType>(ElTy)) { 115 if (isa<PointerType>(ElTy)) break; // Can't index into pointers! 116 ElTy = STy->getElementType(); 117 IdxList.push_back(Zero); 118 } else { 119 break; 120 } 121 } 122 123 if (ElTy == DPTy->getElementType()) 124 // This GEP is inbounds because all indices are zero. 125 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], 126 IdxList.size()); 127 } 128 129 // Handle casts from one vector constant to another. We know that the src 130 // and dest type have the same size (otherwise its an illegal cast). 131 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 132 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 133 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 134 "Not cast between same sized vectors!"); 135 SrcTy = NULL; 136 // First, check for null. Undef is already handled. 137 if (isa<ConstantAggregateZero>(V)) 138 return Constant::getNullValue(DestTy); 139 140 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 141 return BitCastConstantVector(CV, DestPTy); 142 } 143 144 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 145 // This allows for other simplifications (although some of them 146 // can only be handled by Analysis/ConstantFolding.cpp). 147 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 148 return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy); 149 } 150 151 // Finally, implement bitcast folding now. The code below doesn't handle 152 // bitcast right. 153 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 154 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 155 156 // Handle integral constant input. 157 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 158 if (DestTy->isInteger()) 159 // Integral -> Integral. This is a no-op because the bit widths must 160 // be the same. Consequently, we just fold to V. 161 return V; 162 163 if (DestTy->isFloatingPoint()) 164 return ConstantFP::get(DestTy->getContext(), 165 APFloat(CI->getValue(), 166 !DestTy->isPPC_FP128Ty())); 167 168 // Otherwise, can't fold this (vector?) 169 return 0; 170 } 171 172 // Handle ConstantFP input: FP -> Integral. 173 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 174 return ConstantInt::get(FP->getContext(), 175 FP->getValueAPF().bitcastToAPInt()); 176 177 return 0; 178} 179 180 181/// ExtractConstantBytes - V is an integer constant which only has a subset of 182/// its bytes used. The bytes used are indicated by ByteStart (which is the 183/// first byte used, counting from the least significant byte) and ByteSize, 184/// which is the number of bytes used. 185/// 186/// This function analyzes the specified constant to see if the specified byte 187/// range can be returned as a simplified constant. If so, the constant is 188/// returned, otherwise null is returned. 189/// 190static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 191 unsigned ByteSize) { 192 assert(isa<IntegerType>(C->getType()) && 193 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 194 "Non-byte sized integer input"); 195 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 196 assert(ByteSize && "Must be accessing some piece"); 197 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 198 assert(ByteSize != CSize && "Should not extract everything"); 199 200 // Constant Integers are simple. 201 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 202 APInt V = CI->getValue(); 203 if (ByteStart) 204 V = V.lshr(ByteStart*8); 205 V.trunc(ByteSize*8); 206 return ConstantInt::get(CI->getContext(), V); 207 } 208 209 // In the input is a constant expr, we might be able to recursively simplify. 210 // If not, we definitely can't do anything. 211 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 212 if (CE == 0) return 0; 213 214 switch (CE->getOpcode()) { 215 default: return 0; 216 case Instruction::Or: { 217 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 218 if (RHS == 0) 219 return 0; 220 221 // X | -1 -> -1. 222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 223 if (RHSC->isAllOnesValue()) 224 return RHSC; 225 226 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 227 if (LHS == 0) 228 return 0; 229 return ConstantExpr::getOr(LHS, RHS); 230 } 231 case Instruction::And: { 232 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 233 if (RHS == 0) 234 return 0; 235 236 // X & 0 -> 0. 237 if (RHS->isNullValue()) 238 return RHS; 239 240 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 241 if (LHS == 0) 242 return 0; 243 return ConstantExpr::getAnd(LHS, RHS); 244 } 245 case Instruction::LShr: { 246 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 247 if (Amt == 0) 248 return 0; 249 unsigned ShAmt = Amt->getZExtValue(); 250 // Cannot analyze non-byte shifts. 251 if ((ShAmt & 7) != 0) 252 return 0; 253 ShAmt >>= 3; 254 255 // If the extract is known to be all zeros, return zero. 256 if (ByteStart >= CSize-ShAmt) 257 return Constant::getNullValue(IntegerType::get(CE->getContext(), 258 ByteSize*8)); 259 // If the extract is known to be fully in the input, extract it. 260 if (ByteStart+ByteSize+ShAmt <= CSize) 261 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 262 263 // TODO: Handle the 'partially zero' case. 264 return 0; 265 } 266 267 case Instruction::Shl: { 268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 269 if (Amt == 0) 270 return 0; 271 unsigned ShAmt = Amt->getZExtValue(); 272 // Cannot analyze non-byte shifts. 273 if ((ShAmt & 7) != 0) 274 return 0; 275 ShAmt >>= 3; 276 277 // If the extract is known to be all zeros, return zero. 278 if (ByteStart+ByteSize <= ShAmt) 279 return Constant::getNullValue(IntegerType::get(CE->getContext(), 280 ByteSize*8)); 281 // If the extract is known to be fully in the input, extract it. 282 if (ByteStart >= ShAmt) 283 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 284 285 // TODO: Handle the 'partially zero' case. 286 return 0; 287 } 288 289 case Instruction::ZExt: { 290 unsigned SrcBitSize = 291 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 292 293 // If extracting something that is completely zero, return 0. 294 if (ByteStart*8 >= SrcBitSize) 295 return Constant::getNullValue(IntegerType::get(CE->getContext(), 296 ByteSize*8)); 297 298 // If exactly extracting the input, return it. 299 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 300 return CE->getOperand(0); 301 302 // If extracting something completely in the input, if if the input is a 303 // multiple of 8 bits, recurse. 304 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 305 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 306 307 // Otherwise, if extracting a subset of the input, which is not multiple of 308 // 8 bits, do a shift and trunc to get the bits. 309 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 310 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 311 Constant *Res = CE->getOperand(0); 312 if (ByteStart) 313 Res = ConstantExpr::getLShr(Res, 314 ConstantInt::get(Res->getType(), ByteStart*8)); 315 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 316 ByteSize*8)); 317 } 318 319 // TODO: Handle the 'partially zero' case. 320 return 0; 321 } 322 } 323} 324 325/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 326/// on Ty, with any known factors factored out. If Folded is false, 327/// return null if no factoring was possible, to avoid endlessly 328/// bouncing an unfoldable expression back into the top-level folder. 329/// 330static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, 331 bool Folded) { 332 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 333 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 334 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 335 return ConstantExpr::getNUWMul(E, N); 336 } 337 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 338 Constant *N = ConstantInt::get(DestTy, VTy->getNumElements()); 339 Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true); 340 return ConstantExpr::getNUWMul(E, N); 341 } 342 if (const StructType *STy = dyn_cast<StructType>(Ty)) 343 if (!STy->isPacked()) { 344 unsigned NumElems = STy->getNumElements(); 345 // An empty struct has size zero. 346 if (NumElems == 0) 347 return ConstantExpr::getNullValue(DestTy); 348 // Check for a struct with all members having the same size. 349 Constant *MemberSize = 350 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 351 bool AllSame = true; 352 for (unsigned i = 1; i != NumElems; ++i) 353 if (MemberSize != 354 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 355 AllSame = false; 356 break; 357 } 358 if (AllSame) { 359 Constant *N = ConstantInt::get(DestTy, NumElems); 360 return ConstantExpr::getNUWMul(MemberSize, N); 361 } 362 } 363 364 // Pointer size doesn't depend on the pointee type, so canonicalize them 365 // to an arbitrary pointee. 366 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 367 if (!PTy->getElementType()->isInteger(1)) 368 return 369 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 370 PTy->getAddressSpace()), 371 DestTy, true); 372 373 // If there's no interesting folding happening, bail so that we don't create 374 // a constant that looks like it needs folding but really doesn't. 375 if (!Folded) 376 return 0; 377 378 // Base case: Get a regular sizeof expression. 379 Constant *C = ConstantExpr::getSizeOf(Ty); 380 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 381 DestTy, false), 382 C, DestTy); 383 return C; 384} 385 386/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 387/// on Ty, with any known factors factored out. If Folded is false, 388/// return null if no factoring was possible, to avoid endlessly 389/// bouncing an unfoldable expression back into the top-level folder. 390/// 391static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, 392 bool Folded) { 393 // The alignment of an array is equal to the alignment of the 394 // array element. Note that this is not always true for vectors. 395 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 396 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 397 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 398 DestTy, 399 false), 400 C, DestTy); 401 return C; 402 } 403 404 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 405 // Packed structs always have an alignment of 1. 406 if (STy->isPacked()) 407 return ConstantInt::get(DestTy, 1); 408 409 // Otherwise, struct alignment is the maximum alignment of any member. 410 // Without target data, we can't compare much, but we can check to see 411 // if all the members have the same alignment. 412 unsigned NumElems = STy->getNumElements(); 413 // An empty struct has minimal alignment. 414 if (NumElems == 0) 415 return ConstantInt::get(DestTy, 1); 416 // Check for a struct with all members having the same alignment. 417 Constant *MemberAlign = 418 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 419 bool AllSame = true; 420 for (unsigned i = 1; i != NumElems; ++i) 421 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 422 AllSame = false; 423 break; 424 } 425 if (AllSame) 426 return MemberAlign; 427 } 428 429 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 430 // to an arbitrary pointee. 431 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 432 if (!PTy->getElementType()->isInteger(1)) 433 return 434 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 435 1), 436 PTy->getAddressSpace()), 437 DestTy, true); 438 439 // If there's no interesting folding happening, bail so that we don't create 440 // a constant that looks like it needs folding but really doesn't. 441 if (!Folded) 442 return 0; 443 444 // Base case: Get a regular alignof expression. 445 Constant *C = ConstantExpr::getAlignOf(Ty); 446 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 447 DestTy, false), 448 C, DestTy); 449 return C; 450} 451 452/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 453/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 454/// return null if no factoring was possible, to avoid endlessly 455/// bouncing an unfoldable expression back into the top-level folder. 456/// 457static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, 458 const Type *DestTy, 459 bool Folded) { 460 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 461 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 462 DestTy, false), 463 FieldNo, DestTy); 464 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 465 return ConstantExpr::getNUWMul(E, N); 466 } 467 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 468 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 469 DestTy, false), 470 FieldNo, DestTy); 471 Constant *E = getFoldedSizeOf(VTy->getElementType(), DestTy, true); 472 return ConstantExpr::getNUWMul(E, N); 473 } 474 if (const StructType *STy = dyn_cast<StructType>(Ty)) 475 if (!STy->isPacked()) { 476 unsigned NumElems = STy->getNumElements(); 477 // An empty struct has no members. 478 if (NumElems == 0) 479 return 0; 480 // Check for a struct with all members having the same size. 481 Constant *MemberSize = 482 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 483 bool AllSame = true; 484 for (unsigned i = 1; i != NumElems; ++i) 485 if (MemberSize != 486 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 487 AllSame = false; 488 break; 489 } 490 if (AllSame) { 491 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 492 false, 493 DestTy, 494 false), 495 FieldNo, DestTy); 496 return ConstantExpr::getNUWMul(MemberSize, N); 497 } 498 } 499 500 // If there's no interesting folding happening, bail so that we don't create 501 // a constant that looks like it needs folding but really doesn't. 502 if (!Folded) 503 return 0; 504 505 // Base case: Get a regular offsetof expression. 506 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 507 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 508 DestTy, false), 509 C, DestTy); 510 return C; 511} 512 513Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 514 const Type *DestTy) { 515 if (isa<UndefValue>(V)) { 516 // zext(undef) = 0, because the top bits will be zero. 517 // sext(undef) = 0, because the top bits will all be the same. 518 // [us]itofp(undef) = 0, because the result value is bounded. 519 if (opc == Instruction::ZExt || opc == Instruction::SExt || 520 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 521 return Constant::getNullValue(DestTy); 522 return UndefValue::get(DestTy); 523 } 524 // No compile-time operations on this type yet. 525 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) 526 return 0; 527 528 // If the cast operand is a constant expression, there's a few things we can 529 // do to try to simplify it. 530 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 531 if (CE->isCast()) { 532 // Try hard to fold cast of cast because they are often eliminable. 533 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 534 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 535 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 536 // If all of the indexes in the GEP are null values, there is no pointer 537 // adjustment going on. We might as well cast the source pointer. 538 bool isAllNull = true; 539 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 540 if (!CE->getOperand(i)->isNullValue()) { 541 isAllNull = false; 542 break; 543 } 544 if (isAllNull) 545 // This is casting one pointer type to another, always BitCast 546 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 547 } 548 } 549 550 // If the cast operand is a constant vector, perform the cast by 551 // operating on each element. In the cast of bitcasts, the element 552 // count may be mismatched; don't attempt to handle that here. 553 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 554 if (isa<VectorType>(DestTy) && 555 cast<VectorType>(DestTy)->getNumElements() == 556 CV->getType()->getNumElements()) { 557 std::vector<Constant*> res; 558 const VectorType *DestVecTy = cast<VectorType>(DestTy); 559 const Type *DstEltTy = DestVecTy->getElementType(); 560 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 561 res.push_back(ConstantExpr::getCast(opc, 562 CV->getOperand(i), DstEltTy)); 563 return ConstantVector::get(DestVecTy, res); 564 } 565 566 // We actually have to do a cast now. Perform the cast according to the 567 // opcode specified. 568 switch (opc) { 569 default: 570 llvm_unreachable("Failed to cast constant expression"); 571 case Instruction::FPTrunc: 572 case Instruction::FPExt: 573 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 574 bool ignored; 575 APFloat Val = FPC->getValueAPF(); 576 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle : 577 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 578 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 579 DestTy->isFP128Ty() ? APFloat::IEEEquad : 580 APFloat::Bogus, 581 APFloat::rmNearestTiesToEven, &ignored); 582 return ConstantFP::get(V->getContext(), Val); 583 } 584 return 0; // Can't fold. 585 case Instruction::FPToUI: 586 case Instruction::FPToSI: 587 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 588 const APFloat &V = FPC->getValueAPF(); 589 bool ignored; 590 uint64_t x[2]; 591 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 592 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 593 APFloat::rmTowardZero, &ignored); 594 APInt Val(DestBitWidth, 2, x); 595 return ConstantInt::get(FPC->getContext(), Val); 596 } 597 return 0; // Can't fold. 598 case Instruction::IntToPtr: //always treated as unsigned 599 if (V->isNullValue()) // Is it an integral null value? 600 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 601 return 0; // Other pointer types cannot be casted 602 case Instruction::PtrToInt: // always treated as unsigned 603 // Is it a null pointer value? 604 if (V->isNullValue()) 605 return ConstantInt::get(DestTy, 0); 606 // If this is a sizeof-like expression, pull out multiplications by 607 // known factors to expose them to subsequent folding. If it's an 608 // alignof-like expression, factor out known factors. 609 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 610 if (CE->getOpcode() == Instruction::GetElementPtr && 611 CE->getOperand(0)->isNullValue()) { 612 const Type *Ty = 613 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 614 if (CE->getNumOperands() == 2) { 615 // Handle a sizeof-like expression. 616 Constant *Idx = CE->getOperand(1); 617 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 618 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 619 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 620 DestTy, false), 621 Idx, DestTy); 622 return ConstantExpr::getMul(C, Idx); 623 } 624 } else if (CE->getNumOperands() == 3 && 625 CE->getOperand(1)->isNullValue()) { 626 // Handle an alignof-like expression. 627 if (const StructType *STy = dyn_cast<StructType>(Ty)) 628 if (!STy->isPacked()) { 629 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 630 if (CI->isOne() && 631 STy->getNumElements() == 2 && 632 STy->getElementType(0)->isInteger(1)) { 633 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 634 } 635 } 636 // Handle an offsetof-like expression. 637 if (isa<StructType>(Ty) || isa<ArrayType>(Ty) || isa<VectorType>(Ty)){ 638 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 639 DestTy, false)) 640 return C; 641 } 642 } 643 } 644 // Other pointer types cannot be casted 645 return 0; 646 case Instruction::UIToFP: 647 case Instruction::SIToFP: 648 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 649 APInt api = CI->getValue(); 650 const uint64_t zero[] = {0, 0}; 651 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), 652 2, zero)); 653 (void)apf.convertFromAPInt(api, 654 opc==Instruction::SIToFP, 655 APFloat::rmNearestTiesToEven); 656 return ConstantFP::get(V->getContext(), apf); 657 } 658 return 0; 659 case Instruction::ZExt: 660 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 661 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 662 APInt Result(CI->getValue()); 663 Result.zext(BitWidth); 664 return ConstantInt::get(V->getContext(), Result); 665 } 666 return 0; 667 case Instruction::SExt: 668 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 669 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 670 APInt Result(CI->getValue()); 671 Result.sext(BitWidth); 672 return ConstantInt::get(V->getContext(), Result); 673 } 674 return 0; 675 case Instruction::Trunc: { 676 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 677 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 678 APInt Result(CI->getValue()); 679 Result.trunc(DestBitWidth); 680 return ConstantInt::get(V->getContext(), Result); 681 } 682 683 // The input must be a constantexpr. See if we can simplify this based on 684 // the bytes we are demanding. Only do this if the source and dest are an 685 // even multiple of a byte. 686 if ((DestBitWidth & 7) == 0 && 687 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 688 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 689 return Res; 690 691 return 0; 692 } 693 case Instruction::BitCast: 694 return FoldBitCast(V, DestTy); 695 } 696} 697 698Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 699 Constant *V1, Constant *V2) { 700 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 701 return CB->getZExtValue() ? V1 : V2; 702 703 if (isa<UndefValue>(V1)) return V2; 704 if (isa<UndefValue>(V2)) return V1; 705 if (isa<UndefValue>(Cond)) return V1; 706 if (V1 == V2) return V1; 707 return 0; 708} 709 710Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 711 Constant *Idx) { 712 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 713 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 714 if (Val->isNullValue()) // ee(zero, x) -> zero 715 return Constant::getNullValue( 716 cast<VectorType>(Val->getType())->getElementType()); 717 718 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 719 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 720 return CVal->getOperand(CIdx->getZExtValue()); 721 } else if (isa<UndefValue>(Idx)) { 722 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 723 return CVal->getOperand(0); 724 } 725 } 726 return 0; 727} 728 729Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 730 Constant *Elt, 731 Constant *Idx) { 732 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 733 if (!CIdx) return 0; 734 APInt idxVal = CIdx->getValue(); 735 if (isa<UndefValue>(Val)) { 736 // Insertion of scalar constant into vector undef 737 // Optimize away insertion of undef 738 if (isa<UndefValue>(Elt)) 739 return Val; 740 // Otherwise break the aggregate undef into multiple undefs and do 741 // the insertion 742 unsigned numOps = 743 cast<VectorType>(Val->getType())->getNumElements(); 744 std::vector<Constant*> Ops; 745 Ops.reserve(numOps); 746 for (unsigned i = 0; i < numOps; ++i) { 747 Constant *Op = 748 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 749 Ops.push_back(Op); 750 } 751 return ConstantVector::get(Ops); 752 } 753 if (isa<ConstantAggregateZero>(Val)) { 754 // Insertion of scalar constant into vector aggregate zero 755 // Optimize away insertion of zero 756 if (Elt->isNullValue()) 757 return Val; 758 // Otherwise break the aggregate zero into multiple zeros and do 759 // the insertion 760 unsigned numOps = 761 cast<VectorType>(Val->getType())->getNumElements(); 762 std::vector<Constant*> Ops; 763 Ops.reserve(numOps); 764 for (unsigned i = 0; i < numOps; ++i) { 765 Constant *Op = 766 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 767 Ops.push_back(Op); 768 } 769 return ConstantVector::get(Ops); 770 } 771 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 772 // Insertion of scalar constant into vector constant 773 std::vector<Constant*> Ops; 774 Ops.reserve(CVal->getNumOperands()); 775 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 776 Constant *Op = 777 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 778 Ops.push_back(Op); 779 } 780 return ConstantVector::get(Ops); 781 } 782 783 return 0; 784} 785 786/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 787/// return the specified element value. Otherwise return null. 788static Constant *GetVectorElement(Constant *C, unsigned EltNo) { 789 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) 790 return CV->getOperand(EltNo); 791 792 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 793 if (isa<ConstantAggregateZero>(C)) 794 return Constant::getNullValue(EltTy); 795 if (isa<UndefValue>(C)) 796 return UndefValue::get(EltTy); 797 return 0; 798} 799 800Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 801 Constant *V2, 802 Constant *Mask) { 803 // Undefined shuffle mask -> undefined value. 804 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 805 806 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 807 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 808 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 809 810 // Loop over the shuffle mask, evaluating each element. 811 SmallVector<Constant*, 32> Result; 812 for (unsigned i = 0; i != MaskNumElts; ++i) { 813 Constant *InElt = GetVectorElement(Mask, i); 814 if (InElt == 0) return 0; 815 816 if (isa<UndefValue>(InElt)) 817 InElt = UndefValue::get(EltTy); 818 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 819 unsigned Elt = CI->getZExtValue(); 820 if (Elt >= SrcNumElts*2) 821 InElt = UndefValue::get(EltTy); 822 else if (Elt >= SrcNumElts) 823 InElt = GetVectorElement(V2, Elt - SrcNumElts); 824 else 825 InElt = GetVectorElement(V1, Elt); 826 if (InElt == 0) return 0; 827 } else { 828 // Unknown value. 829 return 0; 830 } 831 Result.push_back(InElt); 832 } 833 834 return ConstantVector::get(&Result[0], Result.size()); 835} 836 837Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 838 const unsigned *Idxs, 839 unsigned NumIdx) { 840 // Base case: no indices, so return the entire value. 841 if (NumIdx == 0) 842 return Agg; 843 844 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 845 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 846 Idxs, 847 Idxs + NumIdx)); 848 849 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 850 return 851 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 852 Idxs, 853 Idxs + NumIdx)); 854 855 // Otherwise recurse. 856 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) 857 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs), 858 Idxs+1, NumIdx-1); 859 860 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) 861 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs), 862 Idxs+1, NumIdx-1); 863 ConstantVector *CV = cast<ConstantVector>(Agg); 864 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs), 865 Idxs+1, NumIdx-1); 866} 867 868Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 869 Constant *Val, 870 const unsigned *Idxs, 871 unsigned NumIdx) { 872 // Base case: no indices, so replace the entire value. 873 if (NumIdx == 0) 874 return Val; 875 876 if (isa<UndefValue>(Agg)) { 877 // Insertion of constant into aggregate undef 878 // Optimize away insertion of undef. 879 if (isa<UndefValue>(Val)) 880 return Agg; 881 882 // Otherwise break the aggregate undef into multiple undefs and do 883 // the insertion. 884 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 885 unsigned numOps; 886 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 887 numOps = AR->getNumElements(); 888 else if (isa<UnionType>(AggTy)) 889 numOps = 1; 890 else 891 numOps = cast<StructType>(AggTy)->getNumElements(); 892 893 std::vector<Constant*> Ops(numOps); 894 for (unsigned i = 0; i < numOps; ++i) { 895 const Type *MemberTy = AggTy->getTypeAtIndex(i); 896 Constant *Op = 897 (*Idxs == i) ? 898 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), 899 Val, Idxs+1, NumIdx-1) : 900 UndefValue::get(MemberTy); 901 Ops[i] = Op; 902 } 903 904 if (const StructType* ST = dyn_cast<StructType>(AggTy)) 905 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 906 if (const UnionType* UT = dyn_cast<UnionType>(AggTy)) { 907 assert(Ops.size() == 1 && "Union can only contain a single value!"); 908 return ConstantUnion::get(UT, Ops[0]); 909 } 910 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 911 } 912 913 if (isa<ConstantAggregateZero>(Agg)) { 914 // Insertion of constant into aggregate zero 915 // Optimize away insertion of zero. 916 if (Val->isNullValue()) 917 return Agg; 918 919 // Otherwise break the aggregate zero into multiple zeros and do 920 // the insertion. 921 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 922 unsigned numOps; 923 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 924 numOps = AR->getNumElements(); 925 else 926 numOps = cast<StructType>(AggTy)->getNumElements(); 927 928 std::vector<Constant*> Ops(numOps); 929 for (unsigned i = 0; i < numOps; ++i) { 930 const Type *MemberTy = AggTy->getTypeAtIndex(i); 931 Constant *Op = 932 (*Idxs == i) ? 933 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), 934 Val, Idxs+1, NumIdx-1) : 935 Constant::getNullValue(MemberTy); 936 Ops[i] = Op; 937 } 938 939 if (const StructType *ST = dyn_cast<StructType>(AggTy)) 940 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 941 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 942 } 943 944 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 945 // Insertion of constant into aggregate constant. 946 std::vector<Constant*> Ops(Agg->getNumOperands()); 947 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 948 Constant *Op = cast<Constant>(Agg->getOperand(i)); 949 if (*Idxs == i) 950 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1); 951 Ops[i] = Op; 952 } 953 954 if (const StructType* ST = dyn_cast<StructType>(Agg->getType())) 955 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 956 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 957 } 958 959 return 0; 960} 961 962 963Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 964 Constant *C1, Constant *C2) { 965 // No compile-time operations on this type yet. 966 if (C1->getType()->isPPC_FP128Ty()) 967 return 0; 968 969 // Handle UndefValue up front. 970 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 971 switch (Opcode) { 972 case Instruction::Xor: 973 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 974 // Handle undef ^ undef -> 0 special case. This is a common 975 // idiom (misuse). 976 return Constant::getNullValue(C1->getType()); 977 // Fallthrough 978 case Instruction::Add: 979 case Instruction::Sub: 980 return UndefValue::get(C1->getType()); 981 case Instruction::Mul: 982 case Instruction::And: 983 return Constant::getNullValue(C1->getType()); 984 case Instruction::UDiv: 985 case Instruction::SDiv: 986 case Instruction::URem: 987 case Instruction::SRem: 988 if (!isa<UndefValue>(C2)) // undef / X -> 0 989 return Constant::getNullValue(C1->getType()); 990 return C2; // X / undef -> undef 991 case Instruction::Or: // X | undef -> -1 992 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType())) 993 return Constant::getAllOnesValue(PTy); 994 return Constant::getAllOnesValue(C1->getType()); 995 case Instruction::LShr: 996 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 997 return C1; // undef lshr undef -> undef 998 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 999 // undef lshr X -> 0 1000 case Instruction::AShr: 1001 if (!isa<UndefValue>(C2)) 1002 return C1; // undef ashr X --> undef 1003 else if (isa<UndefValue>(C1)) 1004 return C1; // undef ashr undef -> undef 1005 else 1006 return C1; // X ashr undef --> X 1007 case Instruction::Shl: 1008 // undef << X -> 0 or X << undef -> 0 1009 return Constant::getNullValue(C1->getType()); 1010 } 1011 } 1012 1013 // Handle simplifications when the RHS is a constant int. 1014 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1015 switch (Opcode) { 1016 case Instruction::Add: 1017 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1018 break; 1019 case Instruction::Sub: 1020 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1021 break; 1022 case Instruction::Mul: 1023 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1024 if (CI2->equalsInt(1)) 1025 return C1; // X * 1 == X 1026 break; 1027 case Instruction::UDiv: 1028 case Instruction::SDiv: 1029 if (CI2->equalsInt(1)) 1030 return C1; // X / 1 == X 1031 if (CI2->equalsInt(0)) 1032 return UndefValue::get(CI2->getType()); // X / 0 == undef 1033 break; 1034 case Instruction::URem: 1035 case Instruction::SRem: 1036 if (CI2->equalsInt(1)) 1037 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1038 if (CI2->equalsInt(0)) 1039 return UndefValue::get(CI2->getType()); // X % 0 == undef 1040 break; 1041 case Instruction::And: 1042 if (CI2->isZero()) return C2; // X & 0 == 0 1043 if (CI2->isAllOnesValue()) 1044 return C1; // X & -1 == X 1045 1046 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1047 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1048 if (CE1->getOpcode() == Instruction::ZExt) { 1049 unsigned DstWidth = CI2->getType()->getBitWidth(); 1050 unsigned SrcWidth = 1051 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1052 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1053 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1054 return C1; 1055 } 1056 1057 // If and'ing the address of a global with a constant, fold it. 1058 if (CE1->getOpcode() == Instruction::PtrToInt && 1059 isa<GlobalValue>(CE1->getOperand(0))) { 1060 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1061 1062 // Functions are at least 4-byte aligned. 1063 unsigned GVAlign = GV->getAlignment(); 1064 if (isa<Function>(GV)) 1065 GVAlign = std::max(GVAlign, 4U); 1066 1067 if (GVAlign > 1) { 1068 unsigned DstWidth = CI2->getType()->getBitWidth(); 1069 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1070 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1071 1072 // If checking bits we know are clear, return zero. 1073 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1074 return Constant::getNullValue(CI2->getType()); 1075 } 1076 } 1077 } 1078 break; 1079 case Instruction::Or: 1080 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1081 if (CI2->isAllOnesValue()) 1082 return C2; // X | -1 == -1 1083 break; 1084 case Instruction::Xor: 1085 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1086 1087 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1088 switch (CE1->getOpcode()) { 1089 default: break; 1090 case Instruction::ICmp: 1091 case Instruction::FCmp: 1092 // cmp pred ^ true -> cmp !pred 1093 assert(CI2->equalsInt(1)); 1094 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1095 pred = CmpInst::getInversePredicate(pred); 1096 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1097 CE1->getOperand(1)); 1098 } 1099 } 1100 break; 1101 case Instruction::AShr: 1102 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1103 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1104 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1105 return ConstantExpr::getLShr(C1, C2); 1106 break; 1107 } 1108 } 1109 1110 // At this point we know neither constant is an UndefValue. 1111 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1112 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1113 using namespace APIntOps; 1114 const APInt &C1V = CI1->getValue(); 1115 const APInt &C2V = CI2->getValue(); 1116 switch (Opcode) { 1117 default: 1118 break; 1119 case Instruction::Add: 1120 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1121 case Instruction::Sub: 1122 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1123 case Instruction::Mul: 1124 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1125 case Instruction::UDiv: 1126 assert(!CI2->isNullValue() && "Div by zero handled above"); 1127 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1128 case Instruction::SDiv: 1129 assert(!CI2->isNullValue() && "Div by zero handled above"); 1130 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1131 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1132 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1133 case Instruction::URem: 1134 assert(!CI2->isNullValue() && "Div by zero handled above"); 1135 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1136 case Instruction::SRem: 1137 assert(!CI2->isNullValue() && "Div by zero handled above"); 1138 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1139 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1140 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1141 case Instruction::And: 1142 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1143 case Instruction::Or: 1144 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1145 case Instruction::Xor: 1146 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1147 case Instruction::Shl: { 1148 uint32_t shiftAmt = C2V.getZExtValue(); 1149 if (shiftAmt < C1V.getBitWidth()) 1150 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1151 else 1152 return UndefValue::get(C1->getType()); // too big shift is undef 1153 } 1154 case Instruction::LShr: { 1155 uint32_t shiftAmt = C2V.getZExtValue(); 1156 if (shiftAmt < C1V.getBitWidth()) 1157 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1158 else 1159 return UndefValue::get(C1->getType()); // too big shift is undef 1160 } 1161 case Instruction::AShr: { 1162 uint32_t shiftAmt = C2V.getZExtValue(); 1163 if (shiftAmt < C1V.getBitWidth()) 1164 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1165 else 1166 return UndefValue::get(C1->getType()); // too big shift is undef 1167 } 1168 } 1169 } 1170 1171 switch (Opcode) { 1172 case Instruction::SDiv: 1173 case Instruction::UDiv: 1174 case Instruction::URem: 1175 case Instruction::SRem: 1176 case Instruction::LShr: 1177 case Instruction::AShr: 1178 case Instruction::Shl: 1179 if (CI1->equalsInt(0)) return C1; 1180 break; 1181 default: 1182 break; 1183 } 1184 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1185 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1186 APFloat C1V = CFP1->getValueAPF(); 1187 APFloat C2V = CFP2->getValueAPF(); 1188 APFloat C3V = C1V; // copy for modification 1189 switch (Opcode) { 1190 default: 1191 break; 1192 case Instruction::FAdd: 1193 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1194 return ConstantFP::get(C1->getContext(), C3V); 1195 case Instruction::FSub: 1196 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1197 return ConstantFP::get(C1->getContext(), C3V); 1198 case Instruction::FMul: 1199 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1200 return ConstantFP::get(C1->getContext(), C3V); 1201 case Instruction::FDiv: 1202 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1203 return ConstantFP::get(C1->getContext(), C3V); 1204 case Instruction::FRem: 1205 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1206 return ConstantFP::get(C1->getContext(), C3V); 1207 } 1208 } 1209 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1210 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 1211 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 1212 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 1213 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 1214 std::vector<Constant*> Res; 1215 const Type* EltTy = VTy->getElementType(); 1216 Constant *C1 = 0; 1217 Constant *C2 = 0; 1218 switch (Opcode) { 1219 default: 1220 break; 1221 case Instruction::Add: 1222 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1223 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1224 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1225 Res.push_back(ConstantExpr::getAdd(C1, C2)); 1226 } 1227 return ConstantVector::get(Res); 1228 case Instruction::FAdd: 1229 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1230 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1231 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1232 Res.push_back(ConstantExpr::getFAdd(C1, C2)); 1233 } 1234 return ConstantVector::get(Res); 1235 case Instruction::Sub: 1236 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1237 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1238 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1239 Res.push_back(ConstantExpr::getSub(C1, C2)); 1240 } 1241 return ConstantVector::get(Res); 1242 case Instruction::FSub: 1243 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1244 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1245 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1246 Res.push_back(ConstantExpr::getFSub(C1, C2)); 1247 } 1248 return ConstantVector::get(Res); 1249 case Instruction::Mul: 1250 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1251 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1252 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1253 Res.push_back(ConstantExpr::getMul(C1, C2)); 1254 } 1255 return ConstantVector::get(Res); 1256 case Instruction::FMul: 1257 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1258 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1259 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1260 Res.push_back(ConstantExpr::getFMul(C1, C2)); 1261 } 1262 return ConstantVector::get(Res); 1263 case Instruction::UDiv: 1264 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1265 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1266 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1267 Res.push_back(ConstantExpr::getUDiv(C1, C2)); 1268 } 1269 return ConstantVector::get(Res); 1270 case Instruction::SDiv: 1271 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1272 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1273 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1274 Res.push_back(ConstantExpr::getSDiv(C1, C2)); 1275 } 1276 return ConstantVector::get(Res); 1277 case Instruction::FDiv: 1278 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1279 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1280 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1281 Res.push_back(ConstantExpr::getFDiv(C1, C2)); 1282 } 1283 return ConstantVector::get(Res); 1284 case Instruction::URem: 1285 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1286 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1287 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1288 Res.push_back(ConstantExpr::getURem(C1, C2)); 1289 } 1290 return ConstantVector::get(Res); 1291 case Instruction::SRem: 1292 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1293 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1294 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1295 Res.push_back(ConstantExpr::getSRem(C1, C2)); 1296 } 1297 return ConstantVector::get(Res); 1298 case Instruction::FRem: 1299 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1300 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1301 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1302 Res.push_back(ConstantExpr::getFRem(C1, C2)); 1303 } 1304 return ConstantVector::get(Res); 1305 case Instruction::And: 1306 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1307 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1308 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1309 Res.push_back(ConstantExpr::getAnd(C1, C2)); 1310 } 1311 return ConstantVector::get(Res); 1312 case Instruction::Or: 1313 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1314 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1315 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1316 Res.push_back(ConstantExpr::getOr(C1, C2)); 1317 } 1318 return ConstantVector::get(Res); 1319 case Instruction::Xor: 1320 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1321 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1322 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1323 Res.push_back(ConstantExpr::getXor(C1, C2)); 1324 } 1325 return ConstantVector::get(Res); 1326 case Instruction::LShr: 1327 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1328 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1329 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1330 Res.push_back(ConstantExpr::getLShr(C1, C2)); 1331 } 1332 return ConstantVector::get(Res); 1333 case Instruction::AShr: 1334 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1335 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1336 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1337 Res.push_back(ConstantExpr::getAShr(C1, C2)); 1338 } 1339 return ConstantVector::get(Res); 1340 case Instruction::Shl: 1341 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1342 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1343 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1344 Res.push_back(ConstantExpr::getShl(C1, C2)); 1345 } 1346 return ConstantVector::get(Res); 1347 } 1348 } 1349 } 1350 1351 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1352 // There are many possible foldings we could do here. We should probably 1353 // at least fold add of a pointer with an integer into the appropriate 1354 // getelementptr. This will improve alias analysis a bit. 1355 1356 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1357 // (a + (b + c)). 1358 if (Instruction::isAssociative(Opcode, C1->getType()) && 1359 CE1->getOpcode() == Opcode) { 1360 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1361 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1362 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1363 } 1364 } else if (isa<ConstantExpr>(C2)) { 1365 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1366 // other way if possible. 1367 switch (Opcode) { 1368 case Instruction::Add: 1369 case Instruction::FAdd: 1370 case Instruction::Mul: 1371 case Instruction::FMul: 1372 case Instruction::And: 1373 case Instruction::Or: 1374 case Instruction::Xor: 1375 // No change of opcode required. 1376 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1377 1378 case Instruction::Shl: 1379 case Instruction::LShr: 1380 case Instruction::AShr: 1381 case Instruction::Sub: 1382 case Instruction::FSub: 1383 case Instruction::SDiv: 1384 case Instruction::UDiv: 1385 case Instruction::FDiv: 1386 case Instruction::URem: 1387 case Instruction::SRem: 1388 case Instruction::FRem: 1389 default: // These instructions cannot be flopped around. 1390 break; 1391 } 1392 } 1393 1394 // i1 can be simplified in many cases. 1395 if (C1->getType()->isInteger(1)) { 1396 switch (Opcode) { 1397 case Instruction::Add: 1398 case Instruction::Sub: 1399 return ConstantExpr::getXor(C1, C2); 1400 case Instruction::Mul: 1401 return ConstantExpr::getAnd(C1, C2); 1402 case Instruction::Shl: 1403 case Instruction::LShr: 1404 case Instruction::AShr: 1405 // We can assume that C2 == 0. If it were one the result would be 1406 // undefined because the shift value is as large as the bitwidth. 1407 return C1; 1408 case Instruction::SDiv: 1409 case Instruction::UDiv: 1410 // We can assume that C2 == 1. If it were zero the result would be 1411 // undefined through division by zero. 1412 return C1; 1413 case Instruction::URem: 1414 case Instruction::SRem: 1415 // We can assume that C2 == 1. If it were zero the result would be 1416 // undefined through division by zero. 1417 return ConstantInt::getFalse(C1->getContext()); 1418 default: 1419 break; 1420 } 1421 } 1422 1423 // We don't know how to fold this. 1424 return 0; 1425} 1426 1427/// isZeroSizedType - This type is zero sized if its an array or structure of 1428/// zero sized types. The only leaf zero sized type is an empty structure. 1429static bool isMaybeZeroSizedType(const Type *Ty) { 1430 if (isa<OpaqueType>(Ty)) return true; // Can't say. 1431 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1432 1433 // If all of elements have zero size, this does too. 1434 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1435 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1436 return true; 1437 1438 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1439 return isMaybeZeroSizedType(ATy->getElementType()); 1440 } 1441 return false; 1442} 1443 1444/// IdxCompare - Compare the two constants as though they were getelementptr 1445/// indices. This allows coersion of the types to be the same thing. 1446/// 1447/// If the two constants are the "same" (after coersion), return 0. If the 1448/// first is less than the second, return -1, if the second is less than the 1449/// first, return 1. If the constants are not integral, return -2. 1450/// 1451static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) { 1452 if (C1 == C2) return 0; 1453 1454 // Ok, we found a different index. If they are not ConstantInt, we can't do 1455 // anything with them. 1456 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1457 return -2; // don't know! 1458 1459 // Ok, we have two differing integer indices. Sign extend them to be the same 1460 // type. Long is always big enough, so we use it. 1461 if (!C1->getType()->isInteger(64)) 1462 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1463 1464 if (!C2->getType()->isInteger(64)) 1465 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1466 1467 if (C1 == C2) return 0; // They are equal 1468 1469 // If the type being indexed over is really just a zero sized type, there is 1470 // no pointer difference being made here. 1471 if (isMaybeZeroSizedType(ElTy)) 1472 return -2; // dunno. 1473 1474 // If they are really different, now that they are the same type, then we 1475 // found a difference! 1476 if (cast<ConstantInt>(C1)->getSExtValue() < 1477 cast<ConstantInt>(C2)->getSExtValue()) 1478 return -1; 1479 else 1480 return 1; 1481} 1482 1483/// evaluateFCmpRelation - This function determines if there is anything we can 1484/// decide about the two constants provided. This doesn't need to handle simple 1485/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1486/// If we can determine that the two constants have a particular relation to 1487/// each other, we should return the corresponding FCmpInst predicate, 1488/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1489/// ConstantFoldCompareInstruction. 1490/// 1491/// To simplify this code we canonicalize the relation so that the first 1492/// operand is always the most "complex" of the two. We consider ConstantFP 1493/// to be the simplest, and ConstantExprs to be the most complex. 1494static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1495 assert(V1->getType() == V2->getType() && 1496 "Cannot compare values of different types!"); 1497 1498 // No compile-time operations on this type yet. 1499 if (V1->getType()->isPPC_FP128Ty()) 1500 return FCmpInst::BAD_FCMP_PREDICATE; 1501 1502 // Handle degenerate case quickly 1503 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1504 1505 if (!isa<ConstantExpr>(V1)) { 1506 if (!isa<ConstantExpr>(V2)) { 1507 // We distilled thisUse the standard constant folder for a few cases 1508 ConstantInt *R = 0; 1509 R = dyn_cast<ConstantInt>( 1510 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1511 if (R && !R->isZero()) 1512 return FCmpInst::FCMP_OEQ; 1513 R = dyn_cast<ConstantInt>( 1514 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1515 if (R && !R->isZero()) 1516 return FCmpInst::FCMP_OLT; 1517 R = dyn_cast<ConstantInt>( 1518 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1519 if (R && !R->isZero()) 1520 return FCmpInst::FCMP_OGT; 1521 1522 // Nothing more we can do 1523 return FCmpInst::BAD_FCMP_PREDICATE; 1524 } 1525 1526 // If the first operand is simple and second is ConstantExpr, swap operands. 1527 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1528 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1529 return FCmpInst::getSwappedPredicate(SwappedRelation); 1530 } else { 1531 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1532 // constantexpr or a simple constant. 1533 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1534 switch (CE1->getOpcode()) { 1535 case Instruction::FPTrunc: 1536 case Instruction::FPExt: 1537 case Instruction::UIToFP: 1538 case Instruction::SIToFP: 1539 // We might be able to do something with these but we don't right now. 1540 break; 1541 default: 1542 break; 1543 } 1544 } 1545 // There are MANY other foldings that we could perform here. They will 1546 // probably be added on demand, as they seem needed. 1547 return FCmpInst::BAD_FCMP_PREDICATE; 1548} 1549 1550/// evaluateICmpRelation - This function determines if there is anything we can 1551/// decide about the two constants provided. This doesn't need to handle simple 1552/// things like integer comparisons, but should instead handle ConstantExprs 1553/// and GlobalValues. If we can determine that the two constants have a 1554/// particular relation to each other, we should return the corresponding ICmp 1555/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1556/// 1557/// To simplify this code we canonicalize the relation so that the first 1558/// operand is always the most "complex" of the two. We consider simple 1559/// constants (like ConstantInt) to be the simplest, followed by 1560/// GlobalValues, followed by ConstantExpr's (the most complex). 1561/// 1562static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1563 bool isSigned) { 1564 assert(V1->getType() == V2->getType() && 1565 "Cannot compare different types of values!"); 1566 if (V1 == V2) return ICmpInst::ICMP_EQ; 1567 1568 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1569 !isa<BlockAddress>(V1)) { 1570 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1571 !isa<BlockAddress>(V2)) { 1572 // We distilled this down to a simple case, use the standard constant 1573 // folder. 1574 ConstantInt *R = 0; 1575 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1576 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1577 if (R && !R->isZero()) 1578 return pred; 1579 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1580 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1581 if (R && !R->isZero()) 1582 return pred; 1583 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1584 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1585 if (R && !R->isZero()) 1586 return pred; 1587 1588 // If we couldn't figure it out, bail. 1589 return ICmpInst::BAD_ICMP_PREDICATE; 1590 } 1591 1592 // If the first operand is simple, swap operands. 1593 ICmpInst::Predicate SwappedRelation = 1594 evaluateICmpRelation(V2, V1, isSigned); 1595 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1596 return ICmpInst::getSwappedPredicate(SwappedRelation); 1597 1598 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1599 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1600 ICmpInst::Predicate SwappedRelation = 1601 evaluateICmpRelation(V2, V1, isSigned); 1602 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1603 return ICmpInst::getSwappedPredicate(SwappedRelation); 1604 return ICmpInst::BAD_ICMP_PREDICATE; 1605 } 1606 1607 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1608 // constant (which, since the types must match, means that it's a 1609 // ConstantPointerNull). 1610 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1611 // Don't try to decide equality of aliases. 1612 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1613 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1614 return ICmpInst::ICMP_NE; 1615 } else if (isa<BlockAddress>(V2)) { 1616 return ICmpInst::ICMP_NE; // Globals never equal labels. 1617 } else { 1618 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1619 // GlobalVals can never be null unless they have external weak linkage. 1620 // We don't try to evaluate aliases here. 1621 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1622 return ICmpInst::ICMP_NE; 1623 } 1624 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1625 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1626 ICmpInst::Predicate SwappedRelation = 1627 evaluateICmpRelation(V2, V1, isSigned); 1628 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1629 return ICmpInst::getSwappedPredicate(SwappedRelation); 1630 return ICmpInst::BAD_ICMP_PREDICATE; 1631 } 1632 1633 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1634 // constant (which, since the types must match, means that it is a 1635 // ConstantPointerNull). 1636 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1637 // Block address in another function can't equal this one, but block 1638 // addresses in the current function might be the same if blocks are 1639 // empty. 1640 if (BA2->getFunction() != BA->getFunction()) 1641 return ICmpInst::ICMP_NE; 1642 } else { 1643 // Block addresses aren't null, don't equal the address of globals. 1644 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1645 "Canonicalization guarantee!"); 1646 return ICmpInst::ICMP_NE; 1647 } 1648 } else { 1649 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1650 // constantexpr, a global, block address, or a simple constant. 1651 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1652 Constant *CE1Op0 = CE1->getOperand(0); 1653 1654 switch (CE1->getOpcode()) { 1655 case Instruction::Trunc: 1656 case Instruction::FPTrunc: 1657 case Instruction::FPExt: 1658 case Instruction::FPToUI: 1659 case Instruction::FPToSI: 1660 break; // We can't evaluate floating point casts or truncations. 1661 1662 case Instruction::UIToFP: 1663 case Instruction::SIToFP: 1664 case Instruction::BitCast: 1665 case Instruction::ZExt: 1666 case Instruction::SExt: 1667 // If the cast is not actually changing bits, and the second operand is a 1668 // null pointer, do the comparison with the pre-casted value. 1669 if (V2->isNullValue() && 1670 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) { 1671 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1672 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1673 return evaluateICmpRelation(CE1Op0, 1674 Constant::getNullValue(CE1Op0->getType()), 1675 isSigned); 1676 } 1677 break; 1678 1679 case Instruction::GetElementPtr: 1680 // Ok, since this is a getelementptr, we know that the constant has a 1681 // pointer type. Check the various cases. 1682 if (isa<ConstantPointerNull>(V2)) { 1683 // If we are comparing a GEP to a null pointer, check to see if the base 1684 // of the GEP equals the null pointer. 1685 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1686 if (GV->hasExternalWeakLinkage()) 1687 // Weak linkage GVals could be zero or not. We're comparing that 1688 // to null pointer so its greater-or-equal 1689 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1690 else 1691 // If its not weak linkage, the GVal must have a non-zero address 1692 // so the result is greater-than 1693 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1694 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1695 // If we are indexing from a null pointer, check to see if we have any 1696 // non-zero indices. 1697 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1698 if (!CE1->getOperand(i)->isNullValue()) 1699 // Offsetting from null, must not be equal. 1700 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1701 // Only zero indexes from null, must still be zero. 1702 return ICmpInst::ICMP_EQ; 1703 } 1704 // Otherwise, we can't really say if the first operand is null or not. 1705 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1706 if (isa<ConstantPointerNull>(CE1Op0)) { 1707 if (GV2->hasExternalWeakLinkage()) 1708 // Weak linkage GVals could be zero or not. We're comparing it to 1709 // a null pointer, so its less-or-equal 1710 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1711 else 1712 // If its not weak linkage, the GVal must have a non-zero address 1713 // so the result is less-than 1714 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1715 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1716 if (GV == GV2) { 1717 // If this is a getelementptr of the same global, then it must be 1718 // different. Because the types must match, the getelementptr could 1719 // only have at most one index, and because we fold getelementptr's 1720 // with a single zero index, it must be nonzero. 1721 assert(CE1->getNumOperands() == 2 && 1722 !CE1->getOperand(1)->isNullValue() && 1723 "Suprising getelementptr!"); 1724 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1725 } else { 1726 // If they are different globals, we don't know what the value is, 1727 // but they can't be equal. 1728 return ICmpInst::ICMP_NE; 1729 } 1730 } 1731 } else { 1732 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1733 Constant *CE2Op0 = CE2->getOperand(0); 1734 1735 // There are MANY other foldings that we could perform here. They will 1736 // probably be added on demand, as they seem needed. 1737 switch (CE2->getOpcode()) { 1738 default: break; 1739 case Instruction::GetElementPtr: 1740 // By far the most common case to handle is when the base pointers are 1741 // obviously to the same or different globals. 1742 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1743 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1744 return ICmpInst::ICMP_NE; 1745 // Ok, we know that both getelementptr instructions are based on the 1746 // same global. From this, we can precisely determine the relative 1747 // ordering of the resultant pointers. 1748 unsigned i = 1; 1749 1750 // The logic below assumes that the result of the comparison 1751 // can be determined by finding the first index that differs. 1752 // This doesn't work if there is over-indexing in any 1753 // subsequent indices, so check for that case first. 1754 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1755 !CE2->isGEPWithNoNotionalOverIndexing()) 1756 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1757 1758 // Compare all of the operands the GEP's have in common. 1759 gep_type_iterator GTI = gep_type_begin(CE1); 1760 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1761 ++i, ++GTI) 1762 switch (IdxCompare(CE1->getOperand(i), 1763 CE2->getOperand(i), GTI.getIndexedType())) { 1764 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1765 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1766 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1767 } 1768 1769 // Ok, we ran out of things they have in common. If any leftovers 1770 // are non-zero then we have a difference, otherwise we are equal. 1771 for (; i < CE1->getNumOperands(); ++i) 1772 if (!CE1->getOperand(i)->isNullValue()) { 1773 if (isa<ConstantInt>(CE1->getOperand(i))) 1774 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1775 else 1776 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1777 } 1778 1779 for (; i < CE2->getNumOperands(); ++i) 1780 if (!CE2->getOperand(i)->isNullValue()) { 1781 if (isa<ConstantInt>(CE2->getOperand(i))) 1782 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1783 else 1784 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1785 } 1786 return ICmpInst::ICMP_EQ; 1787 } 1788 } 1789 } 1790 default: 1791 break; 1792 } 1793 } 1794 1795 return ICmpInst::BAD_ICMP_PREDICATE; 1796} 1797 1798Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1799 Constant *C1, Constant *C2) { 1800 const Type *ResultTy; 1801 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1802 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1803 VT->getNumElements()); 1804 else 1805 ResultTy = Type::getInt1Ty(C1->getContext()); 1806 1807 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1808 if (pred == FCmpInst::FCMP_FALSE) 1809 return Constant::getNullValue(ResultTy); 1810 1811 if (pred == FCmpInst::FCMP_TRUE) 1812 return Constant::getAllOnesValue(ResultTy); 1813 1814 // Handle some degenerate cases first 1815 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) 1816 return UndefValue::get(ResultTy); 1817 1818 // No compile-time operations on this type yet. 1819 if (C1->getType()->isPPC_FP128Ty()) 1820 return 0; 1821 1822 // icmp eq/ne(null,GV) -> false/true 1823 if (C1->isNullValue()) { 1824 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1825 // Don't try to evaluate aliases. External weak GV can be null. 1826 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1827 if (pred == ICmpInst::ICMP_EQ) 1828 return ConstantInt::getFalse(C1->getContext()); 1829 else if (pred == ICmpInst::ICMP_NE) 1830 return ConstantInt::getTrue(C1->getContext()); 1831 } 1832 // icmp eq/ne(GV,null) -> false/true 1833 } else if (C2->isNullValue()) { 1834 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1835 // Don't try to evaluate aliases. External weak GV can be null. 1836 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1837 if (pred == ICmpInst::ICMP_EQ) 1838 return ConstantInt::getFalse(C1->getContext()); 1839 else if (pred == ICmpInst::ICMP_NE) 1840 return ConstantInt::getTrue(C1->getContext()); 1841 } 1842 } 1843 1844 // If the comparison is a comparison between two i1's, simplify it. 1845 if (C1->getType()->isInteger(1)) { 1846 switch(pred) { 1847 case ICmpInst::ICMP_EQ: 1848 if (isa<ConstantInt>(C2)) 1849 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1850 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1851 case ICmpInst::ICMP_NE: 1852 return ConstantExpr::getXor(C1, C2); 1853 default: 1854 break; 1855 } 1856 } 1857 1858 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1859 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1860 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1861 switch (pred) { 1862 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1863 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1864 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1865 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1866 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1867 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1868 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1869 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1870 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1871 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1872 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1873 } 1874 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1875 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1876 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1877 APFloat::cmpResult R = C1V.compare(C2V); 1878 switch (pred) { 1879 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1880 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1881 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1882 case FCmpInst::FCMP_UNO: 1883 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1884 case FCmpInst::FCMP_ORD: 1885 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1886 case FCmpInst::FCMP_UEQ: 1887 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1888 R==APFloat::cmpEqual); 1889 case FCmpInst::FCMP_OEQ: 1890 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1891 case FCmpInst::FCMP_UNE: 1892 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1893 case FCmpInst::FCMP_ONE: 1894 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1895 R==APFloat::cmpGreaterThan); 1896 case FCmpInst::FCMP_ULT: 1897 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1898 R==APFloat::cmpLessThan); 1899 case FCmpInst::FCMP_OLT: 1900 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1901 case FCmpInst::FCMP_UGT: 1902 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1903 R==APFloat::cmpGreaterThan); 1904 case FCmpInst::FCMP_OGT: 1905 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1906 case FCmpInst::FCMP_ULE: 1907 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1908 case FCmpInst::FCMP_OLE: 1909 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1910 R==APFloat::cmpEqual); 1911 case FCmpInst::FCMP_UGE: 1912 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1913 case FCmpInst::FCMP_OGE: 1914 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1915 R==APFloat::cmpEqual); 1916 } 1917 } else if (isa<VectorType>(C1->getType())) { 1918 SmallVector<Constant*, 16> C1Elts, C2Elts; 1919 C1->getVectorElements(C1Elts); 1920 C2->getVectorElements(C2Elts); 1921 if (C1Elts.empty() || C2Elts.empty()) 1922 return 0; 1923 1924 // If we can constant fold the comparison of each element, constant fold 1925 // the whole vector comparison. 1926 SmallVector<Constant*, 4> ResElts; 1927 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { 1928 // Compare the elements, producing an i1 result or constant expr. 1929 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1930 } 1931 return ConstantVector::get(&ResElts[0], ResElts.size()); 1932 } 1933 1934 if (C1->getType()->isFloatingPoint()) { 1935 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1936 switch (evaluateFCmpRelation(C1, C2)) { 1937 default: llvm_unreachable("Unknown relation!"); 1938 case FCmpInst::FCMP_UNO: 1939 case FCmpInst::FCMP_ORD: 1940 case FCmpInst::FCMP_UEQ: 1941 case FCmpInst::FCMP_UNE: 1942 case FCmpInst::FCMP_ULT: 1943 case FCmpInst::FCMP_UGT: 1944 case FCmpInst::FCMP_ULE: 1945 case FCmpInst::FCMP_UGE: 1946 case FCmpInst::FCMP_TRUE: 1947 case FCmpInst::FCMP_FALSE: 1948 case FCmpInst::BAD_FCMP_PREDICATE: 1949 break; // Couldn't determine anything about these constants. 1950 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1951 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1952 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1953 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1954 break; 1955 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1956 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1957 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1958 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1959 break; 1960 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1961 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1962 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1963 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1964 break; 1965 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1966 // We can only partially decide this relation. 1967 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1968 Result = 0; 1969 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1970 Result = 1; 1971 break; 1972 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1973 // We can only partially decide this relation. 1974 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1975 Result = 0; 1976 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1977 Result = 1; 1978 break; 1979 case ICmpInst::ICMP_NE: // We know that C1 != C2 1980 // We can only partially decide this relation. 1981 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1982 Result = 0; 1983 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1984 Result = 1; 1985 break; 1986 } 1987 1988 // If we evaluated the result, return it now. 1989 if (Result != -1) 1990 return ConstantInt::get(ResultTy, Result); 1991 1992 } else { 1993 // Evaluate the relation between the two constants, per the predicate. 1994 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1995 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 1996 default: llvm_unreachable("Unknown relational!"); 1997 case ICmpInst::BAD_ICMP_PREDICATE: 1998 break; // Couldn't determine anything about these constants. 1999 case ICmpInst::ICMP_EQ: // We know the constants are equal! 2000 // If we know the constants are equal, we can decide the result of this 2001 // computation precisely. 2002 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 2003 break; 2004 case ICmpInst::ICMP_ULT: 2005 switch (pred) { 2006 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 2007 Result = 1; break; 2008 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 2009 Result = 0; break; 2010 } 2011 break; 2012 case ICmpInst::ICMP_SLT: 2013 switch (pred) { 2014 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 2015 Result = 1; break; 2016 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 2017 Result = 0; break; 2018 } 2019 break; 2020 case ICmpInst::ICMP_UGT: 2021 switch (pred) { 2022 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 2023 Result = 1; break; 2024 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 2025 Result = 0; break; 2026 } 2027 break; 2028 case ICmpInst::ICMP_SGT: 2029 switch (pred) { 2030 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 2031 Result = 1; break; 2032 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 2033 Result = 0; break; 2034 } 2035 break; 2036 case ICmpInst::ICMP_ULE: 2037 if (pred == ICmpInst::ICMP_UGT) Result = 0; 2038 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 2039 break; 2040 case ICmpInst::ICMP_SLE: 2041 if (pred == ICmpInst::ICMP_SGT) Result = 0; 2042 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 2043 break; 2044 case ICmpInst::ICMP_UGE: 2045 if (pred == ICmpInst::ICMP_ULT) Result = 0; 2046 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 2047 break; 2048 case ICmpInst::ICMP_SGE: 2049 if (pred == ICmpInst::ICMP_SLT) Result = 0; 2050 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 2051 break; 2052 case ICmpInst::ICMP_NE: 2053 if (pred == ICmpInst::ICMP_EQ) Result = 0; 2054 if (pred == ICmpInst::ICMP_NE) Result = 1; 2055 break; 2056 } 2057 2058 // If we evaluated the result, return it now. 2059 if (Result != -1) 2060 return ConstantInt::get(ResultTy, Result); 2061 2062 // If the right hand side is a bitcast, try using its inverse to simplify 2063 // it by moving it to the left hand side. We can't do this if it would turn 2064 // a vector compare into a scalar compare or visa versa. 2065 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 2066 Constant *CE2Op0 = CE2->getOperand(0); 2067 if (CE2->getOpcode() == Instruction::BitCast && 2068 isa<VectorType>(CE2->getType())==isa<VectorType>(CE2Op0->getType())) { 2069 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 2070 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 2071 } 2072 } 2073 2074 // If the left hand side is an extension, try eliminating it. 2075 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 2076 if (CE1->getOpcode() == Instruction::SExt || 2077 CE1->getOpcode() == Instruction::ZExt) { 2078 Constant *CE1Op0 = CE1->getOperand(0); 2079 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2080 if (CE1Inverse == CE1Op0) { 2081 // Check whether we can safely truncate the right hand side. 2082 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2083 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { 2084 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2085 } 2086 } 2087 } 2088 } 2089 2090 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2091 (C1->isNullValue() && !C2->isNullValue())) { 2092 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2093 // other way if possible. 2094 // Also, if C1 is null and C2 isn't, flip them around. 2095 switch (pred) { 2096 case ICmpInst::ICMP_EQ: 2097 case ICmpInst::ICMP_NE: 2098 // No change of predicate required. 2099 return ConstantExpr::getICmp(pred, C2, C1); 2100 2101 case ICmpInst::ICMP_ULT: 2102 case ICmpInst::ICMP_SLT: 2103 case ICmpInst::ICMP_UGT: 2104 case ICmpInst::ICMP_SGT: 2105 case ICmpInst::ICMP_ULE: 2106 case ICmpInst::ICMP_SLE: 2107 case ICmpInst::ICMP_UGE: 2108 case ICmpInst::ICMP_SGE: 2109 // Change the predicate as necessary to swap the operands. 2110 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2111 return ConstantExpr::getICmp(pred, C2, C1); 2112 2113 default: // These predicates cannot be flopped around. 2114 break; 2115 } 2116 } 2117 } 2118 return 0; 2119} 2120 2121/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 2122/// is "inbounds". 2123static bool isInBoundsIndices(Constant *const *Idxs, size_t NumIdx) { 2124 // No indices means nothing that could be out of bounds. 2125 if (NumIdx == 0) return true; 2126 2127 // If the first index is zero, it's in bounds. 2128 if (Idxs[0]->isNullValue()) return true; 2129 2130 // If the first index is one and all the rest are zero, it's in bounds, 2131 // by the one-past-the-end rule. 2132 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2133 return false; 2134 for (unsigned i = 1, e = NumIdx; i != e; ++i) 2135 if (!Idxs[i]->isNullValue()) 2136 return false; 2137 return true; 2138} 2139 2140Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2141 bool inBounds, 2142 Constant* const *Idxs, 2143 unsigned NumIdx) { 2144 if (NumIdx == 0 || 2145 (NumIdx == 1 && Idxs[0]->isNullValue())) 2146 return C; 2147 2148 if (isa<UndefValue>(C)) { 2149 const PointerType *Ptr = cast<PointerType>(C->getType()); 2150 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 2151 (Value **)Idxs, 2152 (Value **)Idxs+NumIdx); 2153 assert(Ty != 0 && "Invalid indices for GEP!"); 2154 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2155 } 2156 2157 Constant *Idx0 = Idxs[0]; 2158 if (C->isNullValue()) { 2159 bool isNull = true; 2160 for (unsigned i = 0, e = NumIdx; i != e; ++i) 2161 if (!Idxs[i]->isNullValue()) { 2162 isNull = false; 2163 break; 2164 } 2165 if (isNull) { 2166 const PointerType *Ptr = cast<PointerType>(C->getType()); 2167 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 2168 (Value**)Idxs, 2169 (Value**)Idxs+NumIdx); 2170 assert(Ty != 0 && "Invalid indices for GEP!"); 2171 return ConstantPointerNull::get( 2172 PointerType::get(Ty,Ptr->getAddressSpace())); 2173 } 2174 } 2175 2176 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2177 // Combine Indices - If the source pointer to this getelementptr instruction 2178 // is a getelementptr instruction, combine the indices of the two 2179 // getelementptr instructions into a single instruction. 2180 // 2181 if (CE->getOpcode() == Instruction::GetElementPtr) { 2182 const Type *LastTy = 0; 2183 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2184 I != E; ++I) 2185 LastTy = *I; 2186 2187 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) { 2188 SmallVector<Value*, 16> NewIndices; 2189 NewIndices.reserve(NumIdx + CE->getNumOperands()); 2190 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2191 NewIndices.push_back(CE->getOperand(i)); 2192 2193 // Add the last index of the source with the first index of the new GEP. 2194 // Make sure to handle the case when they are actually different types. 2195 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2196 // Otherwise it must be an array. 2197 if (!Idx0->isNullValue()) { 2198 const Type *IdxTy = Combined->getType(); 2199 if (IdxTy != Idx0->getType()) { 2200 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2201 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2202 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2203 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2204 } else { 2205 Combined = 2206 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2207 } 2208 } 2209 2210 NewIndices.push_back(Combined); 2211 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); 2212 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 2213 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 2214 &NewIndices[0], 2215 NewIndices.size()) : 2216 ConstantExpr::getGetElementPtr(CE->getOperand(0), 2217 &NewIndices[0], 2218 NewIndices.size()); 2219 } 2220 } 2221 2222 // Implement folding of: 2223 // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*), 2224 // long 0, long 0) 2225 // To: int* getelementptr ([3 x int]* %X, long 0, long 0) 2226 // 2227 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 2228 if (const PointerType *SPT = 2229 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 2230 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 2231 if (const ArrayType *CAT = 2232 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 2233 if (CAT->getElementType() == SAT->getElementType()) 2234 return inBounds ? 2235 ConstantExpr::getInBoundsGetElementPtr( 2236 (Constant*)CE->getOperand(0), Idxs, NumIdx) : 2237 ConstantExpr::getGetElementPtr( 2238 (Constant*)CE->getOperand(0), Idxs, NumIdx); 2239 } 2240 } 2241 2242 // Check to see if any array indices are not within the corresponding 2243 // notional array bounds. If so, try to determine if they can be factored 2244 // out into preceding dimensions. 2245 bool Unknown = false; 2246 SmallVector<Constant *, 8> NewIdxs; 2247 const Type *Ty = C->getType(); 2248 const Type *Prev = 0; 2249 for (unsigned i = 0; i != NumIdx; 2250 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2251 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2252 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2253 if (ATy->getNumElements() <= INT64_MAX && 2254 ATy->getNumElements() != 0 && 2255 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 2256 if (isa<SequentialType>(Prev)) { 2257 // It's out of range, but we can factor it into the prior 2258 // dimension. 2259 NewIdxs.resize(NumIdx); 2260 ConstantInt *Factor = ConstantInt::get(CI->getType(), 2261 ATy->getNumElements()); 2262 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2263 2264 Constant *PrevIdx = Idxs[i-1]; 2265 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2266 2267 // Before adding, extend both operands to i64 to avoid 2268 // overflow trouble. 2269 if (!PrevIdx->getType()->isInteger(64)) 2270 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2271 Type::getInt64Ty(Div->getContext())); 2272 if (!Div->getType()->isInteger(64)) 2273 Div = ConstantExpr::getSExt(Div, 2274 Type::getInt64Ty(Div->getContext())); 2275 2276 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2277 } else { 2278 // It's out of range, but the prior dimension is a struct 2279 // so we can't do anything about it. 2280 Unknown = true; 2281 } 2282 } 2283 } else { 2284 // We don't know if it's in range or not. 2285 Unknown = true; 2286 } 2287 } 2288 2289 // If we did any factoring, start over with the adjusted indices. 2290 if (!NewIdxs.empty()) { 2291 for (unsigned i = 0; i != NumIdx; ++i) 2292 if (!NewIdxs[i]) NewIdxs[i] = Idxs[i]; 2293 return inBounds ? 2294 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), 2295 NewIdxs.size()) : 2296 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); 2297 } 2298 2299 // If all indices are known integers and normalized, we can do a simple 2300 // check for the "inbounds" property. 2301 if (!Unknown && !inBounds && 2302 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx)) 2303 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); 2304 2305 return 0; 2306} 2307