ConstantFold.cpp revision d59ae907eea28285ece6696d6f3271b4ca578c0d
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> 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// ConstantFold*Instruction Implementations 39//===----------------------------------------------------------------------===// 40 41/// BitCastConstantVector - Convert the specified vector Constant 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. 44static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 45 46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 47 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 48 49 // If this cast changes element count then we can't handle it here: 50 // doing so requires endianness information. This should be handled by 51 // Analysis/ConstantFolding.cpp 52 unsigned NumElts = DstTy->getNumElements(); 53 if (NumElts != CV->getType()->getVectorNumElements()) 54 return 0; 55 56 Type *DstEltTy = DstTy->getElementType(); 57 58 // Check to verify that all elements of the input are simple. 59 SmallVector<Constant*, 16> Result; 60 for (unsigned i = 0; i != NumElts; ++i) { 61 Constant *C = CV->getAggregateElement(i); 62 if (C == 0) return 0; 63 C = ConstantExpr::getBitCast(C, DstEltTy); 64 if (isa<ConstantExpr>(C)) return 0; 65 Result.push_back(C); 66 } 67 68 return ConstantVector::get(Result); 69} 70 71/// This function determines which opcode to use to fold two constant cast 72/// expressions together. It uses CastInst::isEliminableCastPair to determine 73/// the opcode. Consequently its just a wrapper around that function. 74/// @brief Determine if it is valid to fold a cast of a cast 75static unsigned 76foldConstantCastPair( 77 unsigned opc, ///< opcode of the second cast constant expression 78 ConstantExpr *Op, ///< the first cast constant expression 79 Type *DstTy ///< desintation type of the first cast 80) { 81 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 82 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 83 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 84 85 // The the types and opcodes for the two Cast constant expressions 86 Type *SrcTy = Op->getOperand(0)->getType(); 87 Type *MidTy = Op->getType(); 88 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 89 Instruction::CastOps secondOp = Instruction::CastOps(opc); 90 91 // Let CastInst::isEliminableCastPair do the heavy lifting. 92 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 93 Type::getInt64Ty(DstTy->getContext())); 94} 95 96static Constant *FoldBitCast(Constant *V, Type *DestTy) { 97 Type *SrcTy = V->getType(); 98 if (SrcTy == DestTy) 99 return V; // no-op cast 100 101 // Check to see if we are casting a pointer to an aggregate to a pointer to 102 // the first element. If so, return the appropriate GEP instruction. 103 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 104 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 105 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 106 && DPTy->getElementType()->isSized()) { 107 SmallVector<Value*, 8> IdxList; 108 Value *Zero = 109 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 110 IdxList.push_back(Zero); 111 Type *ElTy = PTy->getElementType(); 112 while (ElTy != DPTy->getElementType()) { 113 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 114 if (STy->getNumElements() == 0) break; 115 ElTy = STy->getElementType(0); 116 IdxList.push_back(Zero); 117 } else if (SequentialType *STy = 118 dyn_cast<SequentialType>(ElTy)) { 119 if (ElTy->isPointerTy()) break; // Can't index into pointers! 120 ElTy = STy->getElementType(); 121 IdxList.push_back(Zero); 122 } else { 123 break; 124 } 125 } 126 127 if (ElTy == DPTy->getElementType()) 128 // This GEP is inbounds because all indices are zero. 129 return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); 130 } 131 132 // Handle casts from one vector constant to another. We know that the src 133 // and dest type have the same size (otherwise its an illegal cast). 134 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 135 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 136 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 137 "Not cast between same sized vectors!"); 138 SrcTy = NULL; 139 // First, check for null. Undef is already handled. 140 if (isa<ConstantAggregateZero>(V)) 141 return Constant::getNullValue(DestTy); 142 143 // Handle ConstantVector and ConstantAggregateVector. 144 return BitCastConstantVector(V, DestPTy); 145 } 146 147 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 148 // This allows for other simplifications (although some of them 149 // can only be handled by Analysis/ConstantFolding.cpp). 150 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 151 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 152 } 153 154 // Finally, implement bitcast folding now. The code below doesn't handle 155 // bitcast right. 156 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 157 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 158 159 // Handle integral constant input. 160 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 161 if (DestTy->isIntegerTy()) 162 // Integral -> Integral. This is a no-op because the bit widths must 163 // be the same. Consequently, we just fold to V. 164 return V; 165 166 if (DestTy->isFloatingPointTy()) 167 return ConstantFP::get(DestTy->getContext(), 168 APFloat(CI->getValue(), 169 !DestTy->isPPC_FP128Ty())); 170 171 // Otherwise, can't fold this (vector?) 172 return 0; 173 } 174 175 // Handle ConstantFP input: FP -> Integral. 176 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 177 return ConstantInt::get(FP->getContext(), 178 FP->getValueAPF().bitcastToAPInt()); 179 180 return 0; 181} 182 183 184/// ExtractConstantBytes - V is an integer constant which only has a subset of 185/// its bytes used. The bytes used are indicated by ByteStart (which is the 186/// first byte used, counting from the least significant byte) and ByteSize, 187/// which is the number of bytes used. 188/// 189/// This function analyzes the specified constant to see if the specified byte 190/// range can be returned as a simplified constant. If so, the constant is 191/// returned, otherwise null is returned. 192/// 193static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 194 unsigned ByteSize) { 195 assert(C->getType()->isIntegerTy() && 196 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 197 "Non-byte sized integer input"); 198 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 199 assert(ByteSize && "Must be accessing some piece"); 200 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 201 assert(ByteSize != CSize && "Should not extract everything"); 202 203 // Constant Integers are simple. 204 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 205 APInt V = CI->getValue(); 206 if (ByteStart) 207 V = V.lshr(ByteStart*8); 208 V = V.trunc(ByteSize*8); 209 return ConstantInt::get(CI->getContext(), V); 210 } 211 212 // In the input is a constant expr, we might be able to recursively simplify. 213 // If not, we definitely can't do anything. 214 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 215 if (CE == 0) return 0; 216 217 switch (CE->getOpcode()) { 218 default: return 0; 219 case Instruction::Or: { 220 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 221 if (RHS == 0) 222 return 0; 223 224 // X | -1 -> -1. 225 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 226 if (RHSC->isAllOnesValue()) 227 return RHSC; 228 229 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 230 if (LHS == 0) 231 return 0; 232 return ConstantExpr::getOr(LHS, RHS); 233 } 234 case Instruction::And: { 235 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 236 if (RHS == 0) 237 return 0; 238 239 // X & 0 -> 0. 240 if (RHS->isNullValue()) 241 return RHS; 242 243 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 244 if (LHS == 0) 245 return 0; 246 return ConstantExpr::getAnd(LHS, RHS); 247 } 248 case Instruction::LShr: { 249 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 250 if (Amt == 0) 251 return 0; 252 unsigned ShAmt = Amt->getZExtValue(); 253 // Cannot analyze non-byte shifts. 254 if ((ShAmt & 7) != 0) 255 return 0; 256 ShAmt >>= 3; 257 258 // If the extract is known to be all zeros, return zero. 259 if (ByteStart >= CSize-ShAmt) 260 return Constant::getNullValue(IntegerType::get(CE->getContext(), 261 ByteSize*8)); 262 // If the extract is known to be fully in the input, extract it. 263 if (ByteStart+ByteSize+ShAmt <= CSize) 264 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 265 266 // TODO: Handle the 'partially zero' case. 267 return 0; 268 } 269 270 case Instruction::Shl: { 271 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 272 if (Amt == 0) 273 return 0; 274 unsigned ShAmt = Amt->getZExtValue(); 275 // Cannot analyze non-byte shifts. 276 if ((ShAmt & 7) != 0) 277 return 0; 278 ShAmt >>= 3; 279 280 // If the extract is known to be all zeros, return zero. 281 if (ByteStart+ByteSize <= ShAmt) 282 return Constant::getNullValue(IntegerType::get(CE->getContext(), 283 ByteSize*8)); 284 // If the extract is known to be fully in the input, extract it. 285 if (ByteStart >= ShAmt) 286 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 287 288 // TODO: Handle the 'partially zero' case. 289 return 0; 290 } 291 292 case Instruction::ZExt: { 293 unsigned SrcBitSize = 294 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 295 296 // If extracting something that is completely zero, return 0. 297 if (ByteStart*8 >= SrcBitSize) 298 return Constant::getNullValue(IntegerType::get(CE->getContext(), 299 ByteSize*8)); 300 301 // If exactly extracting the input, return it. 302 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 303 return CE->getOperand(0); 304 305 // If extracting something completely in the input, if if the input is a 306 // multiple of 8 bits, recurse. 307 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 308 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 309 310 // Otherwise, if extracting a subset of the input, which is not multiple of 311 // 8 bits, do a shift and trunc to get the bits. 312 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 313 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 314 Constant *Res = CE->getOperand(0); 315 if (ByteStart) 316 Res = ConstantExpr::getLShr(Res, 317 ConstantInt::get(Res->getType(), ByteStart*8)); 318 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 319 ByteSize*8)); 320 } 321 322 // TODO: Handle the 'partially zero' case. 323 return 0; 324 } 325 } 326} 327 328/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 329/// on Ty, with any known factors factored out. If Folded is false, 330/// return null if no factoring was possible, to avoid endlessly 331/// bouncing an unfoldable expression back into the top-level folder. 332/// 333static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 334 bool Folded) { 335 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 336 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 337 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 338 return ConstantExpr::getNUWMul(E, N); 339 } 340 341 if (StructType *STy = dyn_cast<StructType>(Ty)) 342 if (!STy->isPacked()) { 343 unsigned NumElems = STy->getNumElements(); 344 // An empty struct has size zero. 345 if (NumElems == 0) 346 return ConstantExpr::getNullValue(DestTy); 347 // Check for a struct with all members having the same size. 348 Constant *MemberSize = 349 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 350 bool AllSame = true; 351 for (unsigned i = 1; i != NumElems; ++i) 352 if (MemberSize != 353 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 354 AllSame = false; 355 break; 356 } 357 if (AllSame) { 358 Constant *N = ConstantInt::get(DestTy, NumElems); 359 return ConstantExpr::getNUWMul(MemberSize, N); 360 } 361 } 362 363 // Pointer size doesn't depend on the pointee type, so canonicalize them 364 // to an arbitrary pointee. 365 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 366 if (!PTy->getElementType()->isIntegerTy(1)) 367 return 368 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 369 PTy->getAddressSpace()), 370 DestTy, true); 371 372 // If there's no interesting folding happening, bail so that we don't create 373 // a constant that looks like it needs folding but really doesn't. 374 if (!Folded) 375 return 0; 376 377 // Base case: Get a regular sizeof expression. 378 Constant *C = ConstantExpr::getSizeOf(Ty); 379 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 380 DestTy, false), 381 C, DestTy); 382 return C; 383} 384 385/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 386/// on Ty, with any known factors factored out. If Folded is false, 387/// return null if no factoring was possible, to avoid endlessly 388/// bouncing an unfoldable expression back into the top-level folder. 389/// 390static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 391 bool Folded) { 392 // The alignment of an array is equal to the alignment of the 393 // array element. Note that this is not always true for vectors. 394 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 395 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 396 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 397 DestTy, 398 false), 399 C, DestTy); 400 return C; 401 } 402 403 if (StructType *STy = dyn_cast<StructType>(Ty)) { 404 // Packed structs always have an alignment of 1. 405 if (STy->isPacked()) 406 return ConstantInt::get(DestTy, 1); 407 408 // Otherwise, struct alignment is the maximum alignment of any member. 409 // Without target data, we can't compare much, but we can check to see 410 // if all the members have the same alignment. 411 unsigned NumElems = STy->getNumElements(); 412 // An empty struct has minimal alignment. 413 if (NumElems == 0) 414 return ConstantInt::get(DestTy, 1); 415 // Check for a struct with all members having the same alignment. 416 Constant *MemberAlign = 417 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 418 bool AllSame = true; 419 for (unsigned i = 1; i != NumElems; ++i) 420 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 421 AllSame = false; 422 break; 423 } 424 if (AllSame) 425 return MemberAlign; 426 } 427 428 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 429 // to an arbitrary pointee. 430 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 431 if (!PTy->getElementType()->isIntegerTy(1)) 432 return 433 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 434 1), 435 PTy->getAddressSpace()), 436 DestTy, true); 437 438 // If there's no interesting folding happening, bail so that we don't create 439 // a constant that looks like it needs folding but really doesn't. 440 if (!Folded) 441 return 0; 442 443 // Base case: Get a regular alignof expression. 444 Constant *C = ConstantExpr::getAlignOf(Ty); 445 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 446 DestTy, false), 447 C, DestTy); 448 return C; 449} 450 451/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 452/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 453/// return null if no factoring was possible, to avoid endlessly 454/// bouncing an unfoldable expression back into the top-level folder. 455/// 456static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 457 Type *DestTy, 458 bool Folded) { 459 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 460 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 461 DestTy, false), 462 FieldNo, DestTy); 463 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 464 return ConstantExpr::getNUWMul(E, N); 465 } 466 467 if (StructType *STy = dyn_cast<StructType>(Ty)) 468 if (!STy->isPacked()) { 469 unsigned NumElems = STy->getNumElements(); 470 // An empty struct has no members. 471 if (NumElems == 0) 472 return 0; 473 // Check for a struct with all members having the same size. 474 Constant *MemberSize = 475 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 476 bool AllSame = true; 477 for (unsigned i = 1; i != NumElems; ++i) 478 if (MemberSize != 479 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 480 AllSame = false; 481 break; 482 } 483 if (AllSame) { 484 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 485 false, 486 DestTy, 487 false), 488 FieldNo, DestTy); 489 return ConstantExpr::getNUWMul(MemberSize, N); 490 } 491 } 492 493 // If there's no interesting folding happening, bail so that we don't create 494 // a constant that looks like it needs folding but really doesn't. 495 if (!Folded) 496 return 0; 497 498 // Base case: Get a regular offsetof expression. 499 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 500 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 501 DestTy, false), 502 C, DestTy); 503 return C; 504} 505 506Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 507 Type *DestTy) { 508 if (isa<UndefValue>(V)) { 509 // zext(undef) = 0, because the top bits will be zero. 510 // sext(undef) = 0, because the top bits will all be the same. 511 // [us]itofp(undef) = 0, because the result value is bounded. 512 if (opc == Instruction::ZExt || opc == Instruction::SExt || 513 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 514 return Constant::getNullValue(DestTy); 515 return UndefValue::get(DestTy); 516 } 517 518 // No compile-time operations on this type yet. 519 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) 520 return 0; 521 522 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 523 return Constant::getNullValue(DestTy); 524 525 // If the cast operand is a constant expression, there's a few things we can 526 // do to try to simplify it. 527 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 528 if (CE->isCast()) { 529 // Try hard to fold cast of cast because they are often eliminable. 530 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 531 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 532 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 533 // If all of the indexes in the GEP are null values, there is no pointer 534 // adjustment going on. We might as well cast the source pointer. 535 bool isAllNull = true; 536 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 537 if (!CE->getOperand(i)->isNullValue()) { 538 isAllNull = false; 539 break; 540 } 541 if (isAllNull) 542 // This is casting one pointer type to another, always BitCast 543 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 544 } 545 } 546 547 // If the cast operand is a constant vector, perform the cast by 548 // operating on each element. In the cast of bitcasts, the element 549 // count may be mismatched; don't attempt to handle that here. 550 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 551 if (DestTy->isVectorTy() && 552 cast<VectorType>(DestTy)->getNumElements() == 553 CV->getType()->getNumElements()) { 554 std::vector<Constant*> res; 555 VectorType *DestVecTy = cast<VectorType>(DestTy); 556 Type *DstEltTy = DestVecTy->getElementType(); 557 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 558 res.push_back(ConstantExpr::getCast(opc, 559 CV->getOperand(i), DstEltTy)); 560 return ConstantVector::get(res); 561 } 562 563 // We actually have to do a cast now. Perform the cast according to the 564 // opcode specified. 565 switch (opc) { 566 default: 567 llvm_unreachable("Failed to cast constant expression"); 568 case Instruction::FPTrunc: 569 case Instruction::FPExt: 570 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 571 bool ignored; 572 APFloat Val = FPC->getValueAPF(); 573 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 574 DestTy->isFloatTy() ? APFloat::IEEEsingle : 575 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 576 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 577 DestTy->isFP128Ty() ? APFloat::IEEEquad : 578 APFloat::Bogus, 579 APFloat::rmNearestTiesToEven, &ignored); 580 return ConstantFP::get(V->getContext(), Val); 581 } 582 return 0; // Can't fold. 583 case Instruction::FPToUI: 584 case Instruction::FPToSI: 585 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 586 const APFloat &V = FPC->getValueAPF(); 587 bool ignored; 588 uint64_t x[2]; 589 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 590 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 591 APFloat::rmTowardZero, &ignored); 592 APInt Val(DestBitWidth, x); 593 return ConstantInt::get(FPC->getContext(), Val); 594 } 595 return 0; // Can't fold. 596 case Instruction::IntToPtr: //always treated as unsigned 597 if (V->isNullValue()) // Is it an integral null value? 598 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 599 return 0; // Other pointer types cannot be casted 600 case Instruction::PtrToInt: // always treated as unsigned 601 // Is it a null pointer value? 602 if (V->isNullValue()) 603 return ConstantInt::get(DestTy, 0); 604 // If this is a sizeof-like expression, pull out multiplications by 605 // known factors to expose them to subsequent folding. If it's an 606 // alignof-like expression, factor out known factors. 607 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 608 if (CE->getOpcode() == Instruction::GetElementPtr && 609 CE->getOperand(0)->isNullValue()) { 610 Type *Ty = 611 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 612 if (CE->getNumOperands() == 2) { 613 // Handle a sizeof-like expression. 614 Constant *Idx = CE->getOperand(1); 615 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 616 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 617 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 618 DestTy, false), 619 Idx, DestTy); 620 return ConstantExpr::getMul(C, Idx); 621 } 622 } else if (CE->getNumOperands() == 3 && 623 CE->getOperand(1)->isNullValue()) { 624 // Handle an alignof-like expression. 625 if (StructType *STy = dyn_cast<StructType>(Ty)) 626 if (!STy->isPacked()) { 627 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 628 if (CI->isOne() && 629 STy->getNumElements() == 2 && 630 STy->getElementType(0)->isIntegerTy(1)) { 631 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 632 } 633 } 634 // Handle an offsetof-like expression. 635 if (Ty->isStructTy() || Ty->isArrayTy()) { 636 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 637 DestTy, false)) 638 return C; 639 } 640 } 641 } 642 // Other pointer types cannot be casted 643 return 0; 644 case Instruction::UIToFP: 645 case Instruction::SIToFP: 646 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 647 APInt api = CI->getValue(); 648 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); 649 (void)apf.convertFromAPInt(api, 650 opc==Instruction::SIToFP, 651 APFloat::rmNearestTiesToEven); 652 return ConstantFP::get(V->getContext(), apf); 653 } 654 return 0; 655 case Instruction::ZExt: 656 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 657 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 658 return ConstantInt::get(V->getContext(), 659 CI->getValue().zext(BitWidth)); 660 } 661 return 0; 662 case Instruction::SExt: 663 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 664 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 665 return ConstantInt::get(V->getContext(), 666 CI->getValue().sext(BitWidth)); 667 } 668 return 0; 669 case Instruction::Trunc: { 670 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 671 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 672 return ConstantInt::get(V->getContext(), 673 CI->getValue().trunc(DestBitWidth)); 674 } 675 676 // The input must be a constantexpr. See if we can simplify this based on 677 // the bytes we are demanding. Only do this if the source and dest are an 678 // even multiple of a byte. 679 if ((DestBitWidth & 7) == 0 && 680 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 681 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 682 return Res; 683 684 return 0; 685 } 686 case Instruction::BitCast: 687 return FoldBitCast(V, DestTy); 688 } 689} 690 691Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 692 Constant *V1, Constant *V2) { 693 // Check for i1 and vector true/false conditions. 694 if (Cond->isNullValue()) return V2; 695 if (Cond->isAllOnesValue()) return V1; 696 697 // FIXME: CDV Condition. 698 // If the condition is a vector constant, fold the result elementwise. 699 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 700 SmallVector<Constant*, 16> Result; 701 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 702 ConstantInt *Cond = dyn_cast<ConstantInt>(CondV->getOperand(i)); 703 if (Cond == 0) break; 704 705 Constant *Res = (Cond->getZExtValue() ? V2 : V1)->getAggregateElement(i); 706 if (Res == 0) break; 707 Result.push_back(Res); 708 } 709 710 // If we were able to build the vector, return it. 711 if (Result.size() == V1->getType()->getVectorNumElements()) 712 return ConstantVector::get(Result); 713 } 714 715 716 if (isa<UndefValue>(Cond)) { 717 if (isa<UndefValue>(V1)) return V1; 718 return V2; 719 } 720 if (isa<UndefValue>(V1)) return V2; 721 if (isa<UndefValue>(V2)) return V1; 722 if (V1 == V2) return V1; 723 724 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 725 if (TrueVal->getOpcode() == Instruction::Select) 726 if (TrueVal->getOperand(0) == Cond) 727 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 728 } 729 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 730 if (FalseVal->getOpcode() == Instruction::Select) 731 if (FalseVal->getOperand(0) == Cond) 732 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 733 } 734 735 return 0; 736} 737 738Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 739 Constant *Idx) { 740 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 741 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 742 if (Val->isNullValue()) // ee(zero, x) -> zero 743 return Constant::getNullValue( 744 cast<VectorType>(Val->getType())->getElementType()); 745 746 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 747 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 748 uint64_t Index = CIdx->getZExtValue(); 749 if (Index >= CVal->getNumOperands()) 750 // ee({w,x,y,z}, wrong_value) -> undef 751 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 752 return CVal->getOperand(CIdx->getZExtValue()); 753 } else if (isa<UndefValue>(Idx)) { 754 // ee({w,x,y,z}, undef) -> undef 755 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 756 } 757 } 758 return 0; 759} 760 761Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 762 Constant *Elt, 763 Constant *Idx) { 764 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 765 if (!CIdx) return 0; 766 const APInt &IdxVal = CIdx->getValue(); 767 768 SmallVector<Constant*, 16> Result; 769 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ 770 if (i == IdxVal) { 771 Result.push_back(Elt); 772 continue; 773 } 774 775 if (Constant *C = Val->getAggregateElement(i)) 776 Result.push_back(C); 777 else 778 return 0; 779 } 780 781 return ConstantVector::get(Result); 782} 783 784Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 785 Constant *V2, 786 Constant *Mask) { 787 // Undefined shuffle mask -> undefined value. 788 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 789 790 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 791 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 792 Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 793 794 // Loop over the shuffle mask, evaluating each element. 795 SmallVector<Constant*, 32> Result; 796 for (unsigned i = 0; i != MaskNumElts; ++i) { 797 Constant *InElt = Mask->getAggregateElement(i); 798 if (InElt == 0) return 0; 799 800 if (isa<UndefValue>(InElt)) { 801 Result.push_back(UndefValue::get(EltTy)); 802 continue; 803 } 804 unsigned Elt = cast<ConstantInt>(InElt)->getZExtValue(); 805 if (Elt >= SrcNumElts*2) 806 InElt = UndefValue::get(EltTy); 807 else if (Elt >= SrcNumElts) 808 InElt = V2->getAggregateElement(Elt - SrcNumElts); 809 else 810 InElt = V1->getAggregateElement(Elt); 811 if (InElt == 0) return 0; 812 Result.push_back(InElt); 813 } 814 815 return ConstantVector::get(Result); 816} 817 818Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 819 ArrayRef<unsigned> Idxs) { 820 // Base case: no indices, so return the entire value. 821 if (Idxs.empty()) 822 return Agg; 823 824 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 825 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 826 827 return 0; 828} 829 830Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 831 Constant *Val, 832 ArrayRef<unsigned> Idxs) { 833 // Base case: no indices, so replace the entire value. 834 if (Idxs.empty()) 835 return Val; 836 837 unsigned NumElts; 838 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 839 NumElts = ST->getNumElements(); 840 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 841 NumElts = AT->getNumElements(); 842 else 843 NumElts = AT->getVectorNumElements(); 844 845 SmallVector<Constant*, 32> Result; 846 for (unsigned i = 0; i != NumElts; ++i) { 847 Constant *C = Agg->getAggregateElement(i); 848 if (C == 0) return 0; 849 850 if (Idxs[0] == i) 851 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 852 853 Result.push_back(C); 854 } 855 856 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 857 return ConstantStruct::get(ST, Result); 858 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 859 return ConstantArray::get(AT, Result); 860 return ConstantVector::get(Result); 861} 862 863 864Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 865 Constant *C1, Constant *C2) { 866 // No compile-time operations on this type yet. 867 if (C1->getType()->isPPC_FP128Ty()) 868 return 0; 869 870 // Handle UndefValue up front. 871 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 872 switch (Opcode) { 873 case Instruction::Xor: 874 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 875 // Handle undef ^ undef -> 0 special case. This is a common 876 // idiom (misuse). 877 return Constant::getNullValue(C1->getType()); 878 // Fallthrough 879 case Instruction::Add: 880 case Instruction::Sub: 881 return UndefValue::get(C1->getType()); 882 case Instruction::And: 883 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 884 return C1; 885 return Constant::getNullValue(C1->getType()); // undef & X -> 0 886 case Instruction::Mul: { 887 ConstantInt *CI; 888 // X * undef -> undef if X is odd or undef 889 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) || 890 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) || 891 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 892 return UndefValue::get(C1->getType()); 893 894 // X * undef -> 0 otherwise 895 return Constant::getNullValue(C1->getType()); 896 } 897 case Instruction::UDiv: 898 case Instruction::SDiv: 899 // undef / 1 -> undef 900 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) 901 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) 902 if (CI2->isOne()) 903 return C1; 904 // FALL THROUGH 905 case Instruction::URem: 906 case Instruction::SRem: 907 if (!isa<UndefValue>(C2)) // undef / X -> 0 908 return Constant::getNullValue(C1->getType()); 909 return C2; // X / undef -> undef 910 case Instruction::Or: // X | undef -> -1 911 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 912 return C1; 913 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 914 case Instruction::LShr: 915 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 916 return C1; // undef lshr undef -> undef 917 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 918 // undef lshr X -> 0 919 case Instruction::AShr: 920 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 921 return Constant::getAllOnesValue(C1->getType()); 922 else if (isa<UndefValue>(C1)) 923 return C1; // undef ashr undef -> undef 924 else 925 return C1; // X ashr undef --> X 926 case Instruction::Shl: 927 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 928 return C1; // undef shl undef -> undef 929 // undef << X -> 0 or X << undef -> 0 930 return Constant::getNullValue(C1->getType()); 931 } 932 } 933 934 // Handle simplifications when the RHS is a constant int. 935 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 936 switch (Opcode) { 937 case Instruction::Add: 938 if (CI2->equalsInt(0)) return C1; // X + 0 == X 939 break; 940 case Instruction::Sub: 941 if (CI2->equalsInt(0)) return C1; // X - 0 == X 942 break; 943 case Instruction::Mul: 944 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 945 if (CI2->equalsInt(1)) 946 return C1; // X * 1 == X 947 break; 948 case Instruction::UDiv: 949 case Instruction::SDiv: 950 if (CI2->equalsInt(1)) 951 return C1; // X / 1 == X 952 if (CI2->equalsInt(0)) 953 return UndefValue::get(CI2->getType()); // X / 0 == undef 954 break; 955 case Instruction::URem: 956 case Instruction::SRem: 957 if (CI2->equalsInt(1)) 958 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 959 if (CI2->equalsInt(0)) 960 return UndefValue::get(CI2->getType()); // X % 0 == undef 961 break; 962 case Instruction::And: 963 if (CI2->isZero()) return C2; // X & 0 == 0 964 if (CI2->isAllOnesValue()) 965 return C1; // X & -1 == X 966 967 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 968 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 969 if (CE1->getOpcode() == Instruction::ZExt) { 970 unsigned DstWidth = CI2->getType()->getBitWidth(); 971 unsigned SrcWidth = 972 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 973 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 974 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 975 return C1; 976 } 977 978 // If and'ing the address of a global with a constant, fold it. 979 if (CE1->getOpcode() == Instruction::PtrToInt && 980 isa<GlobalValue>(CE1->getOperand(0))) { 981 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 982 983 // Functions are at least 4-byte aligned. 984 unsigned GVAlign = GV->getAlignment(); 985 if (isa<Function>(GV)) 986 GVAlign = std::max(GVAlign, 4U); 987 988 if (GVAlign > 1) { 989 unsigned DstWidth = CI2->getType()->getBitWidth(); 990 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 991 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 992 993 // If checking bits we know are clear, return zero. 994 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 995 return Constant::getNullValue(CI2->getType()); 996 } 997 } 998 } 999 break; 1000 case Instruction::Or: 1001 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1002 if (CI2->isAllOnesValue()) 1003 return C2; // X | -1 == -1 1004 break; 1005 case Instruction::Xor: 1006 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1007 1008 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1009 switch (CE1->getOpcode()) { 1010 default: break; 1011 case Instruction::ICmp: 1012 case Instruction::FCmp: 1013 // cmp pred ^ true -> cmp !pred 1014 assert(CI2->equalsInt(1)); 1015 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1016 pred = CmpInst::getInversePredicate(pred); 1017 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1018 CE1->getOperand(1)); 1019 } 1020 } 1021 break; 1022 case Instruction::AShr: 1023 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1024 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1025 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1026 return ConstantExpr::getLShr(C1, C2); 1027 break; 1028 } 1029 } else if (isa<ConstantInt>(C1)) { 1030 // If C1 is a ConstantInt and C2 is not, swap the operands. 1031 if (Instruction::isCommutative(Opcode)) 1032 return ConstantExpr::get(Opcode, C2, C1); 1033 } 1034 1035 // At this point we know neither constant is an UndefValue. 1036 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1037 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1038 const APInt &C1V = CI1->getValue(); 1039 const APInt &C2V = CI2->getValue(); 1040 switch (Opcode) { 1041 default: 1042 break; 1043 case Instruction::Add: 1044 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1045 case Instruction::Sub: 1046 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1047 case Instruction::Mul: 1048 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1049 case Instruction::UDiv: 1050 assert(!CI2->isNullValue() && "Div by zero handled above"); 1051 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1052 case Instruction::SDiv: 1053 assert(!CI2->isNullValue() && "Div by zero handled above"); 1054 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1055 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1056 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1057 case Instruction::URem: 1058 assert(!CI2->isNullValue() && "Div by zero handled above"); 1059 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1060 case Instruction::SRem: 1061 assert(!CI2->isNullValue() && "Div by zero handled above"); 1062 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1063 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1064 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1065 case Instruction::And: 1066 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1067 case Instruction::Or: 1068 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1069 case Instruction::Xor: 1070 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1071 case Instruction::Shl: { 1072 uint32_t shiftAmt = C2V.getZExtValue(); 1073 if (shiftAmt < C1V.getBitWidth()) 1074 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1075 else 1076 return UndefValue::get(C1->getType()); // too big shift is undef 1077 } 1078 case Instruction::LShr: { 1079 uint32_t shiftAmt = C2V.getZExtValue(); 1080 if (shiftAmt < C1V.getBitWidth()) 1081 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1082 else 1083 return UndefValue::get(C1->getType()); // too big shift is undef 1084 } 1085 case Instruction::AShr: { 1086 uint32_t shiftAmt = C2V.getZExtValue(); 1087 if (shiftAmt < C1V.getBitWidth()) 1088 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1089 else 1090 return UndefValue::get(C1->getType()); // too big shift is undef 1091 } 1092 } 1093 } 1094 1095 switch (Opcode) { 1096 case Instruction::SDiv: 1097 case Instruction::UDiv: 1098 case Instruction::URem: 1099 case Instruction::SRem: 1100 case Instruction::LShr: 1101 case Instruction::AShr: 1102 case Instruction::Shl: 1103 if (CI1->equalsInt(0)) return C1; 1104 break; 1105 default: 1106 break; 1107 } 1108 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1109 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1110 APFloat C1V = CFP1->getValueAPF(); 1111 APFloat C2V = CFP2->getValueAPF(); 1112 APFloat C3V = C1V; // copy for modification 1113 switch (Opcode) { 1114 default: 1115 break; 1116 case Instruction::FAdd: 1117 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1118 return ConstantFP::get(C1->getContext(), C3V); 1119 case Instruction::FSub: 1120 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1121 return ConstantFP::get(C1->getContext(), C3V); 1122 case Instruction::FMul: 1123 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1124 return ConstantFP::get(C1->getContext(), C3V); 1125 case Instruction::FDiv: 1126 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1127 return ConstantFP::get(C1->getContext(), C3V); 1128 case Instruction::FRem: 1129 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1130 return ConstantFP::get(C1->getContext(), C3V); 1131 } 1132 } 1133 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1134 // Perform elementwise folding. 1135 SmallVector<Constant*, 16> Result; 1136 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1137 Constant *LHS = C1->getAggregateElement(i); 1138 Constant *RHS = C2->getAggregateElement(i); 1139 if (LHS == 0 || RHS == 0) break; 1140 1141 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1142 } 1143 1144 if (Result.size() == VTy->getNumElements()) 1145 return ConstantVector::get(Result); 1146 } 1147 1148 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1149 // There are many possible foldings we could do here. We should probably 1150 // at least fold add of a pointer with an integer into the appropriate 1151 // getelementptr. This will improve alias analysis a bit. 1152 1153 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1154 // (a + (b + c)). 1155 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1156 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1157 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1158 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1159 } 1160 } else if (isa<ConstantExpr>(C2)) { 1161 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1162 // other way if possible. 1163 if (Instruction::isCommutative(Opcode)) 1164 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1165 } 1166 1167 // i1 can be simplified in many cases. 1168 if (C1->getType()->isIntegerTy(1)) { 1169 switch (Opcode) { 1170 case Instruction::Add: 1171 case Instruction::Sub: 1172 return ConstantExpr::getXor(C1, C2); 1173 case Instruction::Mul: 1174 return ConstantExpr::getAnd(C1, C2); 1175 case Instruction::Shl: 1176 case Instruction::LShr: 1177 case Instruction::AShr: 1178 // We can assume that C2 == 0. If it were one the result would be 1179 // undefined because the shift value is as large as the bitwidth. 1180 return C1; 1181 case Instruction::SDiv: 1182 case Instruction::UDiv: 1183 // We can assume that C2 == 1. If it were zero the result would be 1184 // undefined through division by zero. 1185 return C1; 1186 case Instruction::URem: 1187 case Instruction::SRem: 1188 // We can assume that C2 == 1. If it were zero the result would be 1189 // undefined through division by zero. 1190 return ConstantInt::getFalse(C1->getContext()); 1191 default: 1192 break; 1193 } 1194 } 1195 1196 // We don't know how to fold this. 1197 return 0; 1198} 1199 1200/// isZeroSizedType - This type is zero sized if its an array or structure of 1201/// zero sized types. The only leaf zero sized type is an empty structure. 1202static bool isMaybeZeroSizedType(Type *Ty) { 1203 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1204 if (STy->isOpaque()) return true; // Can't say. 1205 1206 // If all of elements have zero size, this does too. 1207 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1208 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1209 return true; 1210 1211 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1212 return isMaybeZeroSizedType(ATy->getElementType()); 1213 } 1214 return false; 1215} 1216 1217/// IdxCompare - Compare the two constants as though they were getelementptr 1218/// indices. This allows coersion of the types to be the same thing. 1219/// 1220/// If the two constants are the "same" (after coersion), return 0. If the 1221/// first is less than the second, return -1, if the second is less than the 1222/// first, return 1. If the constants are not integral, return -2. 1223/// 1224static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1225 if (C1 == C2) return 0; 1226 1227 // Ok, we found a different index. If they are not ConstantInt, we can't do 1228 // anything with them. 1229 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1230 return -2; // don't know! 1231 1232 // Ok, we have two differing integer indices. Sign extend them to be the same 1233 // type. Long is always big enough, so we use it. 1234 if (!C1->getType()->isIntegerTy(64)) 1235 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1236 1237 if (!C2->getType()->isIntegerTy(64)) 1238 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1239 1240 if (C1 == C2) return 0; // They are equal 1241 1242 // If the type being indexed over is really just a zero sized type, there is 1243 // no pointer difference being made here. 1244 if (isMaybeZeroSizedType(ElTy)) 1245 return -2; // dunno. 1246 1247 // If they are really different, now that they are the same type, then we 1248 // found a difference! 1249 if (cast<ConstantInt>(C1)->getSExtValue() < 1250 cast<ConstantInt>(C2)->getSExtValue()) 1251 return -1; 1252 else 1253 return 1; 1254} 1255 1256/// evaluateFCmpRelation - This function determines if there is anything we can 1257/// decide about the two constants provided. This doesn't need to handle simple 1258/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1259/// If we can determine that the two constants have a particular relation to 1260/// each other, we should return the corresponding FCmpInst predicate, 1261/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1262/// ConstantFoldCompareInstruction. 1263/// 1264/// To simplify this code we canonicalize the relation so that the first 1265/// operand is always the most "complex" of the two. We consider ConstantFP 1266/// to be the simplest, and ConstantExprs to be the most complex. 1267static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1268 assert(V1->getType() == V2->getType() && 1269 "Cannot compare values of different types!"); 1270 1271 // No compile-time operations on this type yet. 1272 if (V1->getType()->isPPC_FP128Ty()) 1273 return FCmpInst::BAD_FCMP_PREDICATE; 1274 1275 // Handle degenerate case quickly 1276 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1277 1278 if (!isa<ConstantExpr>(V1)) { 1279 if (!isa<ConstantExpr>(V2)) { 1280 // We distilled thisUse the standard constant folder for a few cases 1281 ConstantInt *R = 0; 1282 R = dyn_cast<ConstantInt>( 1283 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1284 if (R && !R->isZero()) 1285 return FCmpInst::FCMP_OEQ; 1286 R = dyn_cast<ConstantInt>( 1287 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1288 if (R && !R->isZero()) 1289 return FCmpInst::FCMP_OLT; 1290 R = dyn_cast<ConstantInt>( 1291 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1292 if (R && !R->isZero()) 1293 return FCmpInst::FCMP_OGT; 1294 1295 // Nothing more we can do 1296 return FCmpInst::BAD_FCMP_PREDICATE; 1297 } 1298 1299 // If the first operand is simple and second is ConstantExpr, swap operands. 1300 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1301 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1302 return FCmpInst::getSwappedPredicate(SwappedRelation); 1303 } else { 1304 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1305 // constantexpr or a simple constant. 1306 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1307 switch (CE1->getOpcode()) { 1308 case Instruction::FPTrunc: 1309 case Instruction::FPExt: 1310 case Instruction::UIToFP: 1311 case Instruction::SIToFP: 1312 // We might be able to do something with these but we don't right now. 1313 break; 1314 default: 1315 break; 1316 } 1317 } 1318 // There are MANY other foldings that we could perform here. They will 1319 // probably be added on demand, as they seem needed. 1320 return FCmpInst::BAD_FCMP_PREDICATE; 1321} 1322 1323/// evaluateICmpRelation - This function determines if there is anything we can 1324/// decide about the two constants provided. This doesn't need to handle simple 1325/// things like integer comparisons, but should instead handle ConstantExprs 1326/// and GlobalValues. If we can determine that the two constants have a 1327/// particular relation to each other, we should return the corresponding ICmp 1328/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1329/// 1330/// To simplify this code we canonicalize the relation so that the first 1331/// operand is always the most "complex" of the two. We consider simple 1332/// constants (like ConstantInt) to be the simplest, followed by 1333/// GlobalValues, followed by ConstantExpr's (the most complex). 1334/// 1335static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1336 bool isSigned) { 1337 assert(V1->getType() == V2->getType() && 1338 "Cannot compare different types of values!"); 1339 if (V1 == V2) return ICmpInst::ICMP_EQ; 1340 1341 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1342 !isa<BlockAddress>(V1)) { 1343 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1344 !isa<BlockAddress>(V2)) { 1345 // We distilled this down to a simple case, use the standard constant 1346 // folder. 1347 ConstantInt *R = 0; 1348 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1349 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1350 if (R && !R->isZero()) 1351 return pred; 1352 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1353 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1354 if (R && !R->isZero()) 1355 return pred; 1356 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1357 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1358 if (R && !R->isZero()) 1359 return pred; 1360 1361 // If we couldn't figure it out, bail. 1362 return ICmpInst::BAD_ICMP_PREDICATE; 1363 } 1364 1365 // If the first operand is simple, swap operands. 1366 ICmpInst::Predicate SwappedRelation = 1367 evaluateICmpRelation(V2, V1, isSigned); 1368 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1369 return ICmpInst::getSwappedPredicate(SwappedRelation); 1370 1371 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1372 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1373 ICmpInst::Predicate SwappedRelation = 1374 evaluateICmpRelation(V2, V1, isSigned); 1375 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1376 return ICmpInst::getSwappedPredicate(SwappedRelation); 1377 return ICmpInst::BAD_ICMP_PREDICATE; 1378 } 1379 1380 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1381 // constant (which, since the types must match, means that it's a 1382 // ConstantPointerNull). 1383 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1384 // Don't try to decide equality of aliases. 1385 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1386 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1387 return ICmpInst::ICMP_NE; 1388 } else if (isa<BlockAddress>(V2)) { 1389 return ICmpInst::ICMP_NE; // Globals never equal labels. 1390 } else { 1391 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1392 // GlobalVals can never be null unless they have external weak linkage. 1393 // We don't try to evaluate aliases here. 1394 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1395 return ICmpInst::ICMP_NE; 1396 } 1397 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1398 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1399 ICmpInst::Predicate SwappedRelation = 1400 evaluateICmpRelation(V2, V1, isSigned); 1401 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1402 return ICmpInst::getSwappedPredicate(SwappedRelation); 1403 return ICmpInst::BAD_ICMP_PREDICATE; 1404 } 1405 1406 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1407 // constant (which, since the types must match, means that it is a 1408 // ConstantPointerNull). 1409 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1410 // Block address in another function can't equal this one, but block 1411 // addresses in the current function might be the same if blocks are 1412 // empty. 1413 if (BA2->getFunction() != BA->getFunction()) 1414 return ICmpInst::ICMP_NE; 1415 } else { 1416 // Block addresses aren't null, don't equal the address of globals. 1417 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1418 "Canonicalization guarantee!"); 1419 return ICmpInst::ICMP_NE; 1420 } 1421 } else { 1422 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1423 // constantexpr, a global, block address, or a simple constant. 1424 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1425 Constant *CE1Op0 = CE1->getOperand(0); 1426 1427 switch (CE1->getOpcode()) { 1428 case Instruction::Trunc: 1429 case Instruction::FPTrunc: 1430 case Instruction::FPExt: 1431 case Instruction::FPToUI: 1432 case Instruction::FPToSI: 1433 break; // We can't evaluate floating point casts or truncations. 1434 1435 case Instruction::UIToFP: 1436 case Instruction::SIToFP: 1437 case Instruction::BitCast: 1438 case Instruction::ZExt: 1439 case Instruction::SExt: 1440 // If the cast is not actually changing bits, and the second operand is a 1441 // null pointer, do the comparison with the pre-casted value. 1442 if (V2->isNullValue() && 1443 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1444 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1445 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1446 return evaluateICmpRelation(CE1Op0, 1447 Constant::getNullValue(CE1Op0->getType()), 1448 isSigned); 1449 } 1450 break; 1451 1452 case Instruction::GetElementPtr: 1453 // Ok, since this is a getelementptr, we know that the constant has a 1454 // pointer type. Check the various cases. 1455 if (isa<ConstantPointerNull>(V2)) { 1456 // If we are comparing a GEP to a null pointer, check to see if the base 1457 // of the GEP equals the null pointer. 1458 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1459 if (GV->hasExternalWeakLinkage()) 1460 // Weak linkage GVals could be zero or not. We're comparing that 1461 // to null pointer so its greater-or-equal 1462 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1463 else 1464 // If its not weak linkage, the GVal must have a non-zero address 1465 // so the result is greater-than 1466 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1467 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1468 // If we are indexing from a null pointer, check to see if we have any 1469 // non-zero indices. 1470 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1471 if (!CE1->getOperand(i)->isNullValue()) 1472 // Offsetting from null, must not be equal. 1473 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1474 // Only zero indexes from null, must still be zero. 1475 return ICmpInst::ICMP_EQ; 1476 } 1477 // Otherwise, we can't really say if the first operand is null or not. 1478 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1479 if (isa<ConstantPointerNull>(CE1Op0)) { 1480 if (GV2->hasExternalWeakLinkage()) 1481 // Weak linkage GVals could be zero or not. We're comparing it to 1482 // a null pointer, so its less-or-equal 1483 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1484 else 1485 // If its not weak linkage, the GVal must have a non-zero address 1486 // so the result is less-than 1487 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1488 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1489 if (GV == GV2) { 1490 // If this is a getelementptr of the same global, then it must be 1491 // different. Because the types must match, the getelementptr could 1492 // only have at most one index, and because we fold getelementptr's 1493 // with a single zero index, it must be nonzero. 1494 assert(CE1->getNumOperands() == 2 && 1495 !CE1->getOperand(1)->isNullValue() && 1496 "Surprising getelementptr!"); 1497 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1498 } else { 1499 // If they are different globals, we don't know what the value is, 1500 // but they can't be equal. 1501 return ICmpInst::ICMP_NE; 1502 } 1503 } 1504 } else { 1505 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1506 Constant *CE2Op0 = CE2->getOperand(0); 1507 1508 // There are MANY other foldings that we could perform here. They will 1509 // probably be added on demand, as they seem needed. 1510 switch (CE2->getOpcode()) { 1511 default: break; 1512 case Instruction::GetElementPtr: 1513 // By far the most common case to handle is when the base pointers are 1514 // obviously to the same or different globals. 1515 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1516 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1517 return ICmpInst::ICMP_NE; 1518 // Ok, we know that both getelementptr instructions are based on the 1519 // same global. From this, we can precisely determine the relative 1520 // ordering of the resultant pointers. 1521 unsigned i = 1; 1522 1523 // The logic below assumes that the result of the comparison 1524 // can be determined by finding the first index that differs. 1525 // This doesn't work if there is over-indexing in any 1526 // subsequent indices, so check for that case first. 1527 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1528 !CE2->isGEPWithNoNotionalOverIndexing()) 1529 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1530 1531 // Compare all of the operands the GEP's have in common. 1532 gep_type_iterator GTI = gep_type_begin(CE1); 1533 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1534 ++i, ++GTI) 1535 switch (IdxCompare(CE1->getOperand(i), 1536 CE2->getOperand(i), GTI.getIndexedType())) { 1537 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1538 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1539 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1540 } 1541 1542 // Ok, we ran out of things they have in common. If any leftovers 1543 // are non-zero then we have a difference, otherwise we are equal. 1544 for (; i < CE1->getNumOperands(); ++i) 1545 if (!CE1->getOperand(i)->isNullValue()) { 1546 if (isa<ConstantInt>(CE1->getOperand(i))) 1547 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1548 else 1549 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1550 } 1551 1552 for (; i < CE2->getNumOperands(); ++i) 1553 if (!CE2->getOperand(i)->isNullValue()) { 1554 if (isa<ConstantInt>(CE2->getOperand(i))) 1555 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1556 else 1557 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1558 } 1559 return ICmpInst::ICMP_EQ; 1560 } 1561 } 1562 } 1563 default: 1564 break; 1565 } 1566 } 1567 1568 return ICmpInst::BAD_ICMP_PREDICATE; 1569} 1570 1571Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1572 Constant *C1, Constant *C2) { 1573 Type *ResultTy; 1574 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1575 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1576 VT->getNumElements()); 1577 else 1578 ResultTy = Type::getInt1Ty(C1->getContext()); 1579 1580 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1581 if (pred == FCmpInst::FCMP_FALSE) 1582 return Constant::getNullValue(ResultTy); 1583 1584 if (pred == FCmpInst::FCMP_TRUE) 1585 return Constant::getAllOnesValue(ResultTy); 1586 1587 // Handle some degenerate cases first 1588 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1589 // For EQ and NE, we can always pick a value for the undef to make the 1590 // predicate pass or fail, so we can return undef. 1591 // Also, if both operands are undef, we can return undef. 1592 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || 1593 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1594 return UndefValue::get(ResultTy); 1595 // Otherwise, pick the same value as the non-undef operand, and fold 1596 // it to true or false. 1597 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1598 } 1599 1600 // No compile-time operations on this type yet. 1601 if (C1->getType()->isPPC_FP128Ty()) 1602 return 0; 1603 1604 // icmp eq/ne(null,GV) -> false/true 1605 if (C1->isNullValue()) { 1606 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1607 // Don't try to evaluate aliases. External weak GV can be null. 1608 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1609 if (pred == ICmpInst::ICMP_EQ) 1610 return ConstantInt::getFalse(C1->getContext()); 1611 else if (pred == ICmpInst::ICMP_NE) 1612 return ConstantInt::getTrue(C1->getContext()); 1613 } 1614 // icmp eq/ne(GV,null) -> false/true 1615 } else if (C2->isNullValue()) { 1616 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1617 // Don't try to evaluate aliases. External weak GV can be null. 1618 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1619 if (pred == ICmpInst::ICMP_EQ) 1620 return ConstantInt::getFalse(C1->getContext()); 1621 else if (pred == ICmpInst::ICMP_NE) 1622 return ConstantInt::getTrue(C1->getContext()); 1623 } 1624 } 1625 1626 // If the comparison is a comparison between two i1's, simplify it. 1627 if (C1->getType()->isIntegerTy(1)) { 1628 switch(pred) { 1629 case ICmpInst::ICMP_EQ: 1630 if (isa<ConstantInt>(C2)) 1631 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1632 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1633 case ICmpInst::ICMP_NE: 1634 return ConstantExpr::getXor(C1, C2); 1635 default: 1636 break; 1637 } 1638 } 1639 1640 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1641 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1642 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1643 switch (pred) { 1644 default: llvm_unreachable("Invalid ICmp Predicate"); 1645 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1646 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1647 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1648 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1649 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1650 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1651 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1652 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1653 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1654 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1655 } 1656 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1657 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1658 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1659 APFloat::cmpResult R = C1V.compare(C2V); 1660 switch (pred) { 1661 default: llvm_unreachable("Invalid FCmp Predicate"); 1662 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1663 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1664 case FCmpInst::FCMP_UNO: 1665 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1666 case FCmpInst::FCMP_ORD: 1667 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1668 case FCmpInst::FCMP_UEQ: 1669 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1670 R==APFloat::cmpEqual); 1671 case FCmpInst::FCMP_OEQ: 1672 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1673 case FCmpInst::FCMP_UNE: 1674 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1675 case FCmpInst::FCMP_ONE: 1676 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1677 R==APFloat::cmpGreaterThan); 1678 case FCmpInst::FCMP_ULT: 1679 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1680 R==APFloat::cmpLessThan); 1681 case FCmpInst::FCMP_OLT: 1682 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1683 case FCmpInst::FCMP_UGT: 1684 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1685 R==APFloat::cmpGreaterThan); 1686 case FCmpInst::FCMP_OGT: 1687 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1688 case FCmpInst::FCMP_ULE: 1689 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1690 case FCmpInst::FCMP_OLE: 1691 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1692 R==APFloat::cmpEqual); 1693 case FCmpInst::FCMP_UGE: 1694 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1695 case FCmpInst::FCMP_OGE: 1696 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1697 R==APFloat::cmpEqual); 1698 } 1699 } else if (C1->getType()->isVectorTy()) { 1700 SmallVector<Constant*, 16> C1Elts, C2Elts; 1701 C1->getVectorElements(C1Elts); 1702 C2->getVectorElements(C2Elts); 1703 if (C1Elts.empty() || C2Elts.empty()) 1704 return 0; 1705 1706 // If we can constant fold the comparison of each element, constant fold 1707 // the whole vector comparison. 1708 SmallVector<Constant*, 4> ResElts; 1709 // Compare the elements, producing an i1 result or constant expr. 1710 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) 1711 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1712 1713 return ConstantVector::get(ResElts); 1714 } 1715 1716 if (C1->getType()->isFloatingPointTy()) { 1717 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1718 switch (evaluateFCmpRelation(C1, C2)) { 1719 default: llvm_unreachable("Unknown relation!"); 1720 case FCmpInst::FCMP_UNO: 1721 case FCmpInst::FCMP_ORD: 1722 case FCmpInst::FCMP_UEQ: 1723 case FCmpInst::FCMP_UNE: 1724 case FCmpInst::FCMP_ULT: 1725 case FCmpInst::FCMP_UGT: 1726 case FCmpInst::FCMP_ULE: 1727 case FCmpInst::FCMP_UGE: 1728 case FCmpInst::FCMP_TRUE: 1729 case FCmpInst::FCMP_FALSE: 1730 case FCmpInst::BAD_FCMP_PREDICATE: 1731 break; // Couldn't determine anything about these constants. 1732 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1733 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1734 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1735 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1736 break; 1737 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1738 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1739 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1740 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1741 break; 1742 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1743 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1744 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1745 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1746 break; 1747 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1748 // We can only partially decide this relation. 1749 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1750 Result = 0; 1751 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1752 Result = 1; 1753 break; 1754 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1755 // We can only partially decide this relation. 1756 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1757 Result = 0; 1758 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1759 Result = 1; 1760 break; 1761 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1762 // We can only partially decide this relation. 1763 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1764 Result = 0; 1765 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1766 Result = 1; 1767 break; 1768 } 1769 1770 // If we evaluated the result, return it now. 1771 if (Result != -1) 1772 return ConstantInt::get(ResultTy, Result); 1773 1774 } else { 1775 // Evaluate the relation between the two constants, per the predicate. 1776 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1777 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 1778 default: llvm_unreachable("Unknown relational!"); 1779 case ICmpInst::BAD_ICMP_PREDICATE: 1780 break; // Couldn't determine anything about these constants. 1781 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1782 // If we know the constants are equal, we can decide the result of this 1783 // computation precisely. 1784 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1785 break; 1786 case ICmpInst::ICMP_ULT: 1787 switch (pred) { 1788 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1789 Result = 1; break; 1790 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1791 Result = 0; break; 1792 } 1793 break; 1794 case ICmpInst::ICMP_SLT: 1795 switch (pred) { 1796 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1797 Result = 1; break; 1798 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1799 Result = 0; break; 1800 } 1801 break; 1802 case ICmpInst::ICMP_UGT: 1803 switch (pred) { 1804 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1805 Result = 1; break; 1806 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1807 Result = 0; break; 1808 } 1809 break; 1810 case ICmpInst::ICMP_SGT: 1811 switch (pred) { 1812 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1813 Result = 1; break; 1814 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1815 Result = 0; break; 1816 } 1817 break; 1818 case ICmpInst::ICMP_ULE: 1819 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1820 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1821 break; 1822 case ICmpInst::ICMP_SLE: 1823 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1824 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1825 break; 1826 case ICmpInst::ICMP_UGE: 1827 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1828 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1829 break; 1830 case ICmpInst::ICMP_SGE: 1831 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1832 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1833 break; 1834 case ICmpInst::ICMP_NE: 1835 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1836 if (pred == ICmpInst::ICMP_NE) Result = 1; 1837 break; 1838 } 1839 1840 // If we evaluated the result, return it now. 1841 if (Result != -1) 1842 return ConstantInt::get(ResultTy, Result); 1843 1844 // If the right hand side is a bitcast, try using its inverse to simplify 1845 // it by moving it to the left hand side. We can't do this if it would turn 1846 // a vector compare into a scalar compare or visa versa. 1847 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1848 Constant *CE2Op0 = CE2->getOperand(0); 1849 if (CE2->getOpcode() == Instruction::BitCast && 1850 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1851 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1852 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1853 } 1854 } 1855 1856 // If the left hand side is an extension, try eliminating it. 1857 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1858 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 1859 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 1860 Constant *CE1Op0 = CE1->getOperand(0); 1861 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1862 if (CE1Inverse == CE1Op0) { 1863 // Check whether we can safely truncate the right hand side. 1864 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1865 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { 1866 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1867 } 1868 } 1869 } 1870 } 1871 1872 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1873 (C1->isNullValue() && !C2->isNullValue())) { 1874 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1875 // other way if possible. 1876 // Also, if C1 is null and C2 isn't, flip them around. 1877 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1878 return ConstantExpr::getICmp(pred, C2, C1); 1879 } 1880 } 1881 return 0; 1882} 1883 1884/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1885/// is "inbounds". 1886template<typename IndexTy> 1887static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1888 // No indices means nothing that could be out of bounds. 1889 if (Idxs.empty()) return true; 1890 1891 // If the first index is zero, it's in bounds. 1892 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1893 1894 // If the first index is one and all the rest are zero, it's in bounds, 1895 // by the one-past-the-end rule. 1896 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1897 return false; 1898 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1899 if (!cast<Constant>(Idxs[i])->isNullValue()) 1900 return false; 1901 return true; 1902} 1903 1904template<typename IndexTy> 1905static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 1906 bool inBounds, 1907 ArrayRef<IndexTy> Idxs) { 1908 if (Idxs.empty()) return C; 1909 Constant *Idx0 = cast<Constant>(Idxs[0]); 1910 if ((Idxs.size() == 1 && Idx0->isNullValue())) 1911 return C; 1912 1913 if (isa<UndefValue>(C)) { 1914 PointerType *Ptr = cast<PointerType>(C->getType()); 1915 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 1916 assert(Ty != 0 && "Invalid indices for GEP!"); 1917 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 1918 } 1919 1920 if (C->isNullValue()) { 1921 bool isNull = true; 1922 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 1923 if (!cast<Constant>(Idxs[i])->isNullValue()) { 1924 isNull = false; 1925 break; 1926 } 1927 if (isNull) { 1928 PointerType *Ptr = cast<PointerType>(C->getType()); 1929 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 1930 assert(Ty != 0 && "Invalid indices for GEP!"); 1931 return ConstantPointerNull::get(PointerType::get(Ty, 1932 Ptr->getAddressSpace())); 1933 } 1934 } 1935 1936 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 1937 // Combine Indices - If the source pointer to this getelementptr instruction 1938 // is a getelementptr instruction, combine the indices of the two 1939 // getelementptr instructions into a single instruction. 1940 // 1941 if (CE->getOpcode() == Instruction::GetElementPtr) { 1942 Type *LastTy = 0; 1943 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 1944 I != E; ++I) 1945 LastTy = *I; 1946 1947 if ((LastTy && isa<SequentialType>(LastTy)) || Idx0->isNullValue()) { 1948 SmallVector<Value*, 16> NewIndices; 1949 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 1950 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 1951 NewIndices.push_back(CE->getOperand(i)); 1952 1953 // Add the last index of the source with the first index of the new GEP. 1954 // Make sure to handle the case when they are actually different types. 1955 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 1956 // Otherwise it must be an array. 1957 if (!Idx0->isNullValue()) { 1958 Type *IdxTy = Combined->getType(); 1959 if (IdxTy != Idx0->getType()) { 1960 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 1961 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 1962 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 1963 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 1964 } else { 1965 Combined = 1966 ConstantExpr::get(Instruction::Add, Idx0, Combined); 1967 } 1968 } 1969 1970 NewIndices.push_back(Combined); 1971 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 1972 return 1973 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, 1974 inBounds && 1975 cast<GEPOperator>(CE)->isInBounds()); 1976 } 1977 } 1978 1979 // Implement folding of: 1980 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 1981 // i64 0, i64 0) 1982 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 1983 // 1984 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 1985 if (PointerType *SPT = 1986 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 1987 if (ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 1988 if (ArrayType *CAT = 1989 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 1990 if (CAT->getElementType() == SAT->getElementType()) 1991 return 1992 ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), 1993 Idxs, inBounds); 1994 } 1995 } 1996 1997 // Check to see if any array indices are not within the corresponding 1998 // notional array bounds. If so, try to determine if they can be factored 1999 // out into preceding dimensions. 2000 bool Unknown = false; 2001 SmallVector<Constant *, 8> NewIdxs; 2002 Type *Ty = C->getType(); 2003 Type *Prev = 0; 2004 for (unsigned i = 0, e = Idxs.size(); i != e; 2005 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2006 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2007 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2008 if (ATy->getNumElements() <= INT64_MAX && 2009 ATy->getNumElements() != 0 && 2010 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 2011 if (isa<SequentialType>(Prev)) { 2012 // It's out of range, but we can factor it into the prior 2013 // dimension. 2014 NewIdxs.resize(Idxs.size()); 2015 ConstantInt *Factor = ConstantInt::get(CI->getType(), 2016 ATy->getNumElements()); 2017 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2018 2019 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2020 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2021 2022 // Before adding, extend both operands to i64 to avoid 2023 // overflow trouble. 2024 if (!PrevIdx->getType()->isIntegerTy(64)) 2025 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2026 Type::getInt64Ty(Div->getContext())); 2027 if (!Div->getType()->isIntegerTy(64)) 2028 Div = ConstantExpr::getSExt(Div, 2029 Type::getInt64Ty(Div->getContext())); 2030 2031 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2032 } else { 2033 // It's out of range, but the prior dimension is a struct 2034 // so we can't do anything about it. 2035 Unknown = true; 2036 } 2037 } 2038 } else { 2039 // We don't know if it's in range or not. 2040 Unknown = true; 2041 } 2042 } 2043 2044 // If we did any factoring, start over with the adjusted indices. 2045 if (!NewIdxs.empty()) { 2046 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2047 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2048 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); 2049 } 2050 2051 // If all indices are known integers and normalized, we can do a simple 2052 // check for the "inbounds" property. 2053 if (!Unknown && !inBounds && 2054 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs)) 2055 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); 2056 2057 return 0; 2058} 2059 2060Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2061 bool inBounds, 2062 ArrayRef<Constant *> Idxs) { 2063 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2064} 2065 2066Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2067 bool inBounds, 2068 ArrayRef<Value *> Idxs) { 2069 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2070} 2071