ConstantFold.cpp revision baf3c404409d5e47b13984a7f95bfbd6d1f2e79e
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// template-based folder for simple primitive constants like ConstantInt, and 16// the special case hackery that we use to symbolically evaluate expressions 17// that use ConstantExprs. 18// 19//===----------------------------------------------------------------------===// 20 21#include "ConstantFold.h" 22#include "llvm/Constants.h" 23#include "llvm/Instructions.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Function.h" 26#include "llvm/GlobalAlias.h" 27#include "llvm/LLVMContext.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 ConstantVector node to the 42/// specified vector type. At this point, we know that the elements of the 43/// input vector constant are all simple integer or FP values. 44static Constant *BitCastConstantVector(LLVMContext &Context, ConstantVector *CV, 45 const VectorType *DstTy) { 46 // If this cast changes element count then we can't handle it here: 47 // doing so requires endianness information. This should be handled by 48 // Analysis/ConstantFolding.cpp 49 unsigned NumElts = DstTy->getNumElements(); 50 if (NumElts != CV->getNumOperands()) 51 return 0; 52 53 // Check to verify that all elements of the input are simple. 54 for (unsigned i = 0; i != NumElts; ++i) { 55 if (!isa<ConstantInt>(CV->getOperand(i)) && 56 !isa<ConstantFP>(CV->getOperand(i))) 57 return 0; 58 } 59 60 // Bitcast each element now. 61 std::vector<Constant*> Result; 62 const Type *DstEltTy = DstTy->getElementType(); 63 for (unsigned i = 0; i != NumElts; ++i) 64 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 65 DstEltTy)); 66 return ConstantVector::get(Result); 67} 68 69/// This function determines which opcode to use to fold two constant cast 70/// expressions together. It uses CastInst::isEliminableCastPair to determine 71/// the opcode. Consequently its just a wrapper around that function. 72/// @brief Determine if it is valid to fold a cast of a cast 73static unsigned 74foldConstantCastPair( 75 unsigned opc, ///< opcode of the second cast constant expression 76 const ConstantExpr*Op, ///< the first cast constant expression 77 const Type *DstTy ///< desintation type of the first cast 78) { 79 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 80 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 81 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 82 83 // The the types and opcodes for the two Cast constant expressions 84 const Type *SrcTy = Op->getOperand(0)->getType(); 85 const Type *MidTy = Op->getType(); 86 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 87 Instruction::CastOps secondOp = Instruction::CastOps(opc); 88 89 // Let CastInst::isEliminableCastPair do the heavy lifting. 90 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 91 Type::Int64Ty); 92} 93 94static Constant *FoldBitCast(LLVMContext &Context, 95 Constant *V, const Type *DestTy) { 96 const Type *SrcTy = V->getType(); 97 if (SrcTy == DestTy) 98 return V; // no-op cast 99 100 // Check to see if we are casting a pointer to an aggregate to a pointer to 101 // the first element. If so, return the appropriate GEP instruction. 102 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) 103 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 104 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 105 SmallVector<Value*, 8> IdxList; 106 IdxList.push_back(Context.getNullValue(Type::Int32Ty)); 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(Context.getNullValue(Type::Int32Ty)); 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(IdxList[0]); 118 } else { 119 break; 120 } 121 } 122 123 if (ElTy == DPTy->getElementType()) 124 return ConstantExpr::getGetElementPtr(V, &IdxList[0], 125 IdxList.size()); 126 } 127 128 // Handle casts from one vector constant to another. We know that the src 129 // and dest type have the same size (otherwise its an illegal cast). 130 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 131 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 132 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 133 "Not cast between same sized vectors!"); 134 SrcTy = NULL; 135 // First, check for null. Undef is already handled. 136 if (isa<ConstantAggregateZero>(V)) 137 return Context.getNullValue(DestTy); 138 139 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 140 return BitCastConstantVector(Context, CV, DestPTy); 141 } 142 143 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 144 // This allows for other simplifications (although some of them 145 // can only be handled by Analysis/ConstantFolding.cpp). 146 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 147 return ConstantExpr::getBitCast( 148 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 Context.getConstantPointerNull(cast<PointerType>(DestTy)); 155 156 // Handle integral constant input. 157 if (const 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(Context, APFloat(CI->getValue(), 165 DestTy != Type::PPC_FP128Ty)); 166 167 // Otherwise, can't fold this (vector?) 168 return 0; 169 } 170 171 // Handle ConstantFP input. 172 if (const ConstantFP *FP = dyn_cast<ConstantFP>(V)) 173 // FP -> Integral. 174 return ConstantInt::get(Context, FP->getValueAPF().bitcastToAPInt()); 175 176 return 0; 177} 178 179 180Constant *llvm::ConstantFoldCastInstruction(LLVMContext &Context, 181 unsigned opc, const Constant *V, 182 const Type *DestTy) { 183 if (isa<UndefValue>(V)) { 184 // zext(undef) = 0, because the top bits will be zero. 185 // sext(undef) = 0, because the top bits will all be the same. 186 // [us]itofp(undef) = 0, because the result value is bounded. 187 if (opc == Instruction::ZExt || opc == Instruction::SExt || 188 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 189 return Context.getNullValue(DestTy); 190 return Context.getUndef(DestTy); 191 } 192 // No compile-time operations on this type yet. 193 if (V->getType() == Type::PPC_FP128Ty || DestTy == Type::PPC_FP128Ty) 194 return 0; 195 196 // If the cast operand is a constant expression, there's a few things we can 197 // do to try to simplify it. 198 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 199 if (CE->isCast()) { 200 // Try hard to fold cast of cast because they are often eliminable. 201 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 202 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 203 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 204 // If all of the indexes in the GEP are null values, there is no pointer 205 // adjustment going on. We might as well cast the source pointer. 206 bool isAllNull = true; 207 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 208 if (!CE->getOperand(i)->isNullValue()) { 209 isAllNull = false; 210 break; 211 } 212 if (isAllNull) 213 // This is casting one pointer type to another, always BitCast 214 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 215 } 216 } 217 218 // If the cast operand is a constant vector, perform the cast by 219 // operating on each element. In the cast of bitcasts, the element 220 // count may be mismatched; don't attempt to handle that here. 221 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) 222 if (isa<VectorType>(DestTy) && 223 cast<VectorType>(DestTy)->getNumElements() == 224 CV->getType()->getNumElements()) { 225 std::vector<Constant*> res; 226 const VectorType *DestVecTy = cast<VectorType>(DestTy); 227 const Type *DstEltTy = DestVecTy->getElementType(); 228 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 229 res.push_back(ConstantExpr::getCast(opc, 230 CV->getOperand(i), DstEltTy)); 231 return ConstantVector::get(DestVecTy, res); 232 } 233 234 // We actually have to do a cast now. Perform the cast according to the 235 // opcode specified. 236 switch (opc) { 237 case Instruction::FPTrunc: 238 case Instruction::FPExt: 239 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 240 bool ignored; 241 APFloat Val = FPC->getValueAPF(); 242 Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle : 243 DestTy == Type::DoubleTy ? APFloat::IEEEdouble : 244 DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended : 245 DestTy == Type::FP128Ty ? APFloat::IEEEquad : 246 APFloat::Bogus, 247 APFloat::rmNearestTiesToEven, &ignored); 248 return ConstantFP::get(Context, Val); 249 } 250 return 0; // Can't fold. 251 case Instruction::FPToUI: 252 case Instruction::FPToSI: 253 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 254 const APFloat &V = FPC->getValueAPF(); 255 bool ignored; 256 uint64_t x[2]; 257 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 258 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 259 APFloat::rmTowardZero, &ignored); 260 APInt Val(DestBitWidth, 2, x); 261 return ConstantInt::get(Context, Val); 262 } 263 return 0; // Can't fold. 264 case Instruction::IntToPtr: //always treated as unsigned 265 if (V->isNullValue()) // Is it an integral null value? 266 return Context.getConstantPointerNull(cast<PointerType>(DestTy)); 267 return 0; // Other pointer types cannot be casted 268 case Instruction::PtrToInt: // always treated as unsigned 269 if (V->isNullValue()) // is it a null pointer value? 270 return ConstantInt::get(DestTy, 0); 271 return 0; // Other pointer types cannot be casted 272 case Instruction::UIToFP: 273 case Instruction::SIToFP: 274 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 275 APInt api = CI->getValue(); 276 const uint64_t zero[] = {0, 0}; 277 APFloat apf = APFloat(APInt(DestTy->getPrimitiveSizeInBits(), 278 2, zero)); 279 (void)apf.convertFromAPInt(api, 280 opc==Instruction::SIToFP, 281 APFloat::rmNearestTiesToEven); 282 return ConstantFP::get(Context, apf); 283 } 284 return 0; 285 case Instruction::ZExt: 286 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 287 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 288 APInt Result(CI->getValue()); 289 Result.zext(BitWidth); 290 return ConstantInt::get(Context, Result); 291 } 292 return 0; 293 case Instruction::SExt: 294 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 295 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 296 APInt Result(CI->getValue()); 297 Result.sext(BitWidth); 298 return ConstantInt::get(Context, Result); 299 } 300 return 0; 301 case Instruction::Trunc: 302 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 303 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 304 APInt Result(CI->getValue()); 305 Result.trunc(BitWidth); 306 return ConstantInt::get(Context, Result); 307 } 308 return 0; 309 case Instruction::BitCast: 310 return FoldBitCast(Context, const_cast<Constant*>(V), DestTy); 311 default: 312 assert(!"Invalid CE CastInst opcode"); 313 break; 314 } 315 316 llvm_unreachable("Failed to cast constant expression"); 317 return 0; 318} 319 320Constant *llvm::ConstantFoldSelectInstruction(LLVMContext&, 321 const Constant *Cond, 322 const Constant *V1, 323 const Constant *V2) { 324 if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 325 return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2); 326 327 if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2); 328 if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1); 329 if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1); 330 if (V1 == V2) return const_cast<Constant*>(V1); 331 return 0; 332} 333 334Constant *llvm::ConstantFoldExtractElementInstruction(LLVMContext &Context, 335 const Constant *Val, 336 const Constant *Idx) { 337 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 338 return Context.getUndef(cast<VectorType>(Val->getType())->getElementType()); 339 if (Val->isNullValue()) // ee(zero, x) -> zero 340 return Context.getNullValue( 341 cast<VectorType>(Val->getType())->getElementType()); 342 343 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 344 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 345 return CVal->getOperand(CIdx->getZExtValue()); 346 } else if (isa<UndefValue>(Idx)) { 347 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 348 return CVal->getOperand(0); 349 } 350 } 351 return 0; 352} 353 354Constant *llvm::ConstantFoldInsertElementInstruction(LLVMContext &Context, 355 const Constant *Val, 356 const Constant *Elt, 357 const Constant *Idx) { 358 const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 359 if (!CIdx) return 0; 360 APInt idxVal = CIdx->getValue(); 361 if (isa<UndefValue>(Val)) { 362 // Insertion of scalar constant into vector undef 363 // Optimize away insertion of undef 364 if (isa<UndefValue>(Elt)) 365 return const_cast<Constant*>(Val); 366 // Otherwise break the aggregate undef into multiple undefs and do 367 // the insertion 368 unsigned numOps = 369 cast<VectorType>(Val->getType())->getNumElements(); 370 std::vector<Constant*> Ops; 371 Ops.reserve(numOps); 372 for (unsigned i = 0; i < numOps; ++i) { 373 const Constant *Op = 374 (idxVal == i) ? Elt : Context.getUndef(Elt->getType()); 375 Ops.push_back(const_cast<Constant*>(Op)); 376 } 377 return ConstantVector::get(Ops); 378 } 379 if (isa<ConstantAggregateZero>(Val)) { 380 // Insertion of scalar constant into vector aggregate zero 381 // Optimize away insertion of zero 382 if (Elt->isNullValue()) 383 return const_cast<Constant*>(Val); 384 // Otherwise break the aggregate zero into multiple zeros and do 385 // the insertion 386 unsigned numOps = 387 cast<VectorType>(Val->getType())->getNumElements(); 388 std::vector<Constant*> Ops; 389 Ops.reserve(numOps); 390 for (unsigned i = 0; i < numOps; ++i) { 391 const Constant *Op = 392 (idxVal == i) ? Elt : Context.getNullValue(Elt->getType()); 393 Ops.push_back(const_cast<Constant*>(Op)); 394 } 395 return ConstantVector::get(Ops); 396 } 397 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 398 // Insertion of scalar constant into vector constant 399 std::vector<Constant*> Ops; 400 Ops.reserve(CVal->getNumOperands()); 401 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 402 const Constant *Op = 403 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 404 Ops.push_back(const_cast<Constant*>(Op)); 405 } 406 return ConstantVector::get(Ops); 407 } 408 409 return 0; 410} 411 412/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 413/// return the specified element value. Otherwise return null. 414static Constant *GetVectorElement(LLVMContext &Context, const Constant *C, 415 unsigned EltNo) { 416 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) 417 return CV->getOperand(EltNo); 418 419 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 420 if (isa<ConstantAggregateZero>(C)) 421 return Context.getNullValue(EltTy); 422 if (isa<UndefValue>(C)) 423 return Context.getUndef(EltTy); 424 return 0; 425} 426 427Constant *llvm::ConstantFoldShuffleVectorInstruction(LLVMContext &Context, 428 const Constant *V1, 429 const Constant *V2, 430 const Constant *Mask) { 431 // Undefined shuffle mask -> undefined value. 432 if (isa<UndefValue>(Mask)) return Context.getUndef(V1->getType()); 433 434 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 435 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 436 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 437 438 // Loop over the shuffle mask, evaluating each element. 439 SmallVector<Constant*, 32> Result; 440 for (unsigned i = 0; i != MaskNumElts; ++i) { 441 Constant *InElt = GetVectorElement(Context, Mask, i); 442 if (InElt == 0) return 0; 443 444 if (isa<UndefValue>(InElt)) 445 InElt = Context.getUndef(EltTy); 446 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 447 unsigned Elt = CI->getZExtValue(); 448 if (Elt >= SrcNumElts*2) 449 InElt = Context.getUndef(EltTy); 450 else if (Elt >= SrcNumElts) 451 InElt = GetVectorElement(Context, V2, Elt - SrcNumElts); 452 else 453 InElt = GetVectorElement(Context, V1, Elt); 454 if (InElt == 0) return 0; 455 } else { 456 // Unknown value. 457 return 0; 458 } 459 Result.push_back(InElt); 460 } 461 462 return ConstantVector::get(&Result[0], Result.size()); 463} 464 465Constant *llvm::ConstantFoldExtractValueInstruction(LLVMContext &Context, 466 const Constant *Agg, 467 const unsigned *Idxs, 468 unsigned NumIdx) { 469 // Base case: no indices, so return the entire value. 470 if (NumIdx == 0) 471 return const_cast<Constant *>(Agg); 472 473 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 474 return Context.getUndef(ExtractValueInst::getIndexedType(Agg->getType(), 475 Idxs, 476 Idxs + NumIdx)); 477 478 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 479 return 480 Context.getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 481 Idxs, 482 Idxs + NumIdx)); 483 484 // Otherwise recurse. 485 return ConstantFoldExtractValueInstruction(Context, Agg->getOperand(*Idxs), 486 Idxs+1, NumIdx-1); 487} 488 489Constant *llvm::ConstantFoldInsertValueInstruction(LLVMContext &Context, 490 const Constant *Agg, 491 const Constant *Val, 492 const unsigned *Idxs, 493 unsigned NumIdx) { 494 // Base case: no indices, so replace the entire value. 495 if (NumIdx == 0) 496 return const_cast<Constant *>(Val); 497 498 if (isa<UndefValue>(Agg)) { 499 // Insertion of constant into aggregate undef 500 // Optimize away insertion of undef 501 if (isa<UndefValue>(Val)) 502 return const_cast<Constant*>(Agg); 503 // Otherwise break the aggregate undef into multiple undefs and do 504 // the insertion 505 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 506 unsigned numOps; 507 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 508 numOps = AR->getNumElements(); 509 else 510 numOps = cast<StructType>(AggTy)->getNumElements(); 511 std::vector<Constant*> Ops(numOps); 512 for (unsigned i = 0; i < numOps; ++i) { 513 const Type *MemberTy = AggTy->getTypeAtIndex(i); 514 const Constant *Op = 515 (*Idxs == i) ? 516 ConstantFoldInsertValueInstruction(Context, Context.getUndef(MemberTy), 517 Val, Idxs+1, NumIdx-1) : 518 Context.getUndef(MemberTy); 519 Ops[i] = const_cast<Constant*>(Op); 520 } 521 if (isa<StructType>(AggTy)) 522 return ConstantStruct::get(Ops); 523 else 524 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 525 } 526 if (isa<ConstantAggregateZero>(Agg)) { 527 // Insertion of constant into aggregate zero 528 // Optimize away insertion of zero 529 if (Val->isNullValue()) 530 return const_cast<Constant*>(Agg); 531 // Otherwise break the aggregate zero into multiple zeros and do 532 // the insertion 533 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 534 unsigned numOps; 535 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 536 numOps = AR->getNumElements(); 537 else 538 numOps = cast<StructType>(AggTy)->getNumElements(); 539 std::vector<Constant*> Ops(numOps); 540 for (unsigned i = 0; i < numOps; ++i) { 541 const Type *MemberTy = AggTy->getTypeAtIndex(i); 542 const Constant *Op = 543 (*Idxs == i) ? 544 ConstantFoldInsertValueInstruction(Context, 545 Context.getNullValue(MemberTy), 546 Val, Idxs+1, NumIdx-1) : 547 Context.getNullValue(MemberTy); 548 Ops[i] = const_cast<Constant*>(Op); 549 } 550 if (isa<StructType>(AggTy)) 551 return ConstantStruct::get(Ops); 552 else 553 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 554 } 555 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 556 // Insertion of constant into aggregate constant 557 std::vector<Constant*> Ops(Agg->getNumOperands()); 558 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 559 const Constant *Op = 560 (*Idxs == i) ? 561 ConstantFoldInsertValueInstruction(Context, Agg->getOperand(i), 562 Val, Idxs+1, NumIdx-1) : 563 Agg->getOperand(i); 564 Ops[i] = const_cast<Constant*>(Op); 565 } 566 Constant *C; 567 if (isa<StructType>(Agg->getType())) 568 C = ConstantStruct::get(Ops); 569 else 570 C = ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 571 return C; 572 } 573 574 return 0; 575} 576 577 578Constant *llvm::ConstantFoldBinaryInstruction(LLVMContext &Context, 579 unsigned Opcode, 580 const Constant *C1, 581 const Constant *C2) { 582 // No compile-time operations on this type yet. 583 if (C1->getType() == Type::PPC_FP128Ty) 584 return 0; 585 586 // Handle UndefValue up front 587 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 588 switch (Opcode) { 589 case Instruction::Xor: 590 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 591 // Handle undef ^ undef -> 0 special case. This is a common 592 // idiom (misuse). 593 return Context.getNullValue(C1->getType()); 594 // Fallthrough 595 case Instruction::Add: 596 case Instruction::Sub: 597 return Context.getUndef(C1->getType()); 598 case Instruction::Mul: 599 case Instruction::And: 600 return Context.getNullValue(C1->getType()); 601 case Instruction::UDiv: 602 case Instruction::SDiv: 603 case Instruction::URem: 604 case Instruction::SRem: 605 if (!isa<UndefValue>(C2)) // undef / X -> 0 606 return Context.getNullValue(C1->getType()); 607 return const_cast<Constant*>(C2); // X / undef -> undef 608 case Instruction::Or: // X | undef -> -1 609 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType())) 610 return Context.getAllOnesValue(PTy); 611 return Context.getAllOnesValue(C1->getType()); 612 case Instruction::LShr: 613 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 614 return const_cast<Constant*>(C1); // undef lshr undef -> undef 615 return Context.getNullValue(C1->getType()); // X lshr undef -> 0 616 // undef lshr X -> 0 617 case Instruction::AShr: 618 if (!isa<UndefValue>(C2)) 619 return const_cast<Constant*>(C1); // undef ashr X --> undef 620 else if (isa<UndefValue>(C1)) 621 return const_cast<Constant*>(C1); // undef ashr undef -> undef 622 else 623 return const_cast<Constant*>(C1); // X ashr undef --> X 624 case Instruction::Shl: 625 // undef << X -> 0 or X << undef -> 0 626 return Context.getNullValue(C1->getType()); 627 } 628 } 629 630 // Handle simplifications when the RHS is a constant int. 631 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 632 switch (Opcode) { 633 case Instruction::Add: 634 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X + 0 == X 635 break; 636 case Instruction::Sub: 637 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X - 0 == X 638 break; 639 case Instruction::Mul: 640 if (CI2->equalsInt(0)) return const_cast<Constant*>(C2); // X * 0 == 0 641 if (CI2->equalsInt(1)) 642 return const_cast<Constant*>(C1); // X * 1 == X 643 break; 644 case Instruction::UDiv: 645 case Instruction::SDiv: 646 if (CI2->equalsInt(1)) 647 return const_cast<Constant*>(C1); // X / 1 == X 648 if (CI2->equalsInt(0)) 649 return Context.getUndef(CI2->getType()); // X / 0 == undef 650 break; 651 case Instruction::URem: 652 case Instruction::SRem: 653 if (CI2->equalsInt(1)) 654 return Context.getNullValue(CI2->getType()); // X % 1 == 0 655 if (CI2->equalsInt(0)) 656 return Context.getUndef(CI2->getType()); // X % 0 == undef 657 break; 658 case Instruction::And: 659 if (CI2->isZero()) return const_cast<Constant*>(C2); // X & 0 == 0 660 if (CI2->isAllOnesValue()) 661 return const_cast<Constant*>(C1); // X & -1 == X 662 663 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 664 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 665 if (CE1->getOpcode() == Instruction::ZExt) { 666 unsigned DstWidth = CI2->getType()->getBitWidth(); 667 unsigned SrcWidth = 668 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 669 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 670 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 671 return const_cast<Constant*>(C1); 672 } 673 674 // If and'ing the address of a global with a constant, fold it. 675 if (CE1->getOpcode() == Instruction::PtrToInt && 676 isa<GlobalValue>(CE1->getOperand(0))) { 677 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 678 679 // Functions are at least 4-byte aligned. 680 unsigned GVAlign = GV->getAlignment(); 681 if (isa<Function>(GV)) 682 GVAlign = std::max(GVAlign, 4U); 683 684 if (GVAlign > 1) { 685 unsigned DstWidth = CI2->getType()->getBitWidth(); 686 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 687 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 688 689 // If checking bits we know are clear, return zero. 690 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 691 return Context.getNullValue(CI2->getType()); 692 } 693 } 694 } 695 break; 696 case Instruction::Or: 697 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X | 0 == X 698 if (CI2->isAllOnesValue()) 699 return const_cast<Constant*>(C2); // X | -1 == -1 700 break; 701 case Instruction::Xor: 702 if (CI2->equalsInt(0)) return const_cast<Constant*>(C1); // X ^ 0 == X 703 break; 704 case Instruction::AShr: 705 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 706 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 707 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 708 return ConstantExpr::getLShr(const_cast<Constant*>(C1), 709 const_cast<Constant*>(C2)); 710 break; 711 } 712 } 713 714 // At this point we know neither constant is an UndefValue. 715 if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 716 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 717 using namespace APIntOps; 718 const APInt &C1V = CI1->getValue(); 719 const APInt &C2V = CI2->getValue(); 720 switch (Opcode) { 721 default: 722 break; 723 case Instruction::Add: 724 return ConstantInt::get(Context, C1V + C2V); 725 case Instruction::Sub: 726 return ConstantInt::get(Context, C1V - C2V); 727 case Instruction::Mul: 728 return ConstantInt::get(Context, C1V * C2V); 729 case Instruction::UDiv: 730 assert(!CI2->isNullValue() && "Div by zero handled above"); 731 return ConstantInt::get(Context, C1V.udiv(C2V)); 732 case Instruction::SDiv: 733 assert(!CI2->isNullValue() && "Div by zero handled above"); 734 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 735 return Context.getUndef(CI1->getType()); // MIN_INT / -1 -> undef 736 return ConstantInt::get(Context, C1V.sdiv(C2V)); 737 case Instruction::URem: 738 assert(!CI2->isNullValue() && "Div by zero handled above"); 739 return ConstantInt::get(Context, C1V.urem(C2V)); 740 case Instruction::SRem: 741 assert(!CI2->isNullValue() && "Div by zero handled above"); 742 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 743 return Context.getUndef(CI1->getType()); // MIN_INT % -1 -> undef 744 return ConstantInt::get(Context, C1V.srem(C2V)); 745 case Instruction::And: 746 return ConstantInt::get(Context, C1V & C2V); 747 case Instruction::Or: 748 return ConstantInt::get(Context, C1V | C2V); 749 case Instruction::Xor: 750 return ConstantInt::get(Context, C1V ^ C2V); 751 case Instruction::Shl: { 752 uint32_t shiftAmt = C2V.getZExtValue(); 753 if (shiftAmt < C1V.getBitWidth()) 754 return ConstantInt::get(Context, C1V.shl(shiftAmt)); 755 else 756 return Context.getUndef(C1->getType()); // too big shift is undef 757 } 758 case Instruction::LShr: { 759 uint32_t shiftAmt = C2V.getZExtValue(); 760 if (shiftAmt < C1V.getBitWidth()) 761 return ConstantInt::get(Context, C1V.lshr(shiftAmt)); 762 else 763 return Context.getUndef(C1->getType()); // too big shift is undef 764 } 765 case Instruction::AShr: { 766 uint32_t shiftAmt = C2V.getZExtValue(); 767 if (shiftAmt < C1V.getBitWidth()) 768 return ConstantInt::get(Context, C1V.ashr(shiftAmt)); 769 else 770 return Context.getUndef(C1->getType()); // too big shift is undef 771 } 772 } 773 } 774 775 switch (Opcode) { 776 case Instruction::SDiv: 777 case Instruction::UDiv: 778 case Instruction::URem: 779 case Instruction::SRem: 780 case Instruction::LShr: 781 case Instruction::AShr: 782 case Instruction::Shl: 783 if (CI1->equalsInt(0)) return const_cast<Constant*>(C1); 784 break; 785 default: 786 break; 787 } 788 } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 789 if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 790 APFloat C1V = CFP1->getValueAPF(); 791 APFloat C2V = CFP2->getValueAPF(); 792 APFloat C3V = C1V; // copy for modification 793 switch (Opcode) { 794 default: 795 break; 796 case Instruction::FAdd: 797 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 798 return ConstantFP::get(Context, C3V); 799 case Instruction::FSub: 800 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 801 return ConstantFP::get(Context, C3V); 802 case Instruction::FMul: 803 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 804 return ConstantFP::get(Context, C3V); 805 case Instruction::FDiv: 806 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 807 return ConstantFP::get(Context, C3V); 808 case Instruction::FRem: 809 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 810 return ConstantFP::get(Context, C3V); 811 } 812 } 813 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 814 const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 815 const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 816 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 817 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 818 std::vector<Constant*> Res; 819 const Type* EltTy = VTy->getElementType(); 820 const Constant *C1 = 0; 821 const Constant *C2 = 0; 822 switch (Opcode) { 823 default: 824 break; 825 case Instruction::Add: 826 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 827 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 828 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 829 Res.push_back(ConstantExpr::getAdd(const_cast<Constant*>(C1), 830 const_cast<Constant*>(C2))); 831 } 832 return ConstantVector::get(Res); 833 case Instruction::FAdd: 834 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 835 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 836 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 837 Res.push_back(ConstantExpr::getFAdd(const_cast<Constant*>(C1), 838 const_cast<Constant*>(C2))); 839 } 840 return ConstantVector::get(Res); 841 case Instruction::Sub: 842 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 843 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 844 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 845 Res.push_back(ConstantExpr::getSub(const_cast<Constant*>(C1), 846 const_cast<Constant*>(C2))); 847 } 848 return ConstantVector::get(Res); 849 case Instruction::FSub: 850 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 851 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 852 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 853 Res.push_back(ConstantExpr::getFSub(const_cast<Constant*>(C1), 854 const_cast<Constant*>(C2))); 855 } 856 return ConstantVector::get(Res); 857 case Instruction::Mul: 858 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 859 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 860 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 861 Res.push_back(ConstantExpr::getMul(const_cast<Constant*>(C1), 862 const_cast<Constant*>(C2))); 863 } 864 return ConstantVector::get(Res); 865 case Instruction::FMul: 866 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 867 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 868 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 869 Res.push_back(ConstantExpr::getFMul(const_cast<Constant*>(C1), 870 const_cast<Constant*>(C2))); 871 } 872 return ConstantVector::get(Res); 873 case Instruction::UDiv: 874 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 875 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 876 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 877 Res.push_back(ConstantExpr::getUDiv(const_cast<Constant*>(C1), 878 const_cast<Constant*>(C2))); 879 } 880 return ConstantVector::get(Res); 881 case Instruction::SDiv: 882 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 883 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 884 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 885 Res.push_back(ConstantExpr::getSDiv(const_cast<Constant*>(C1), 886 const_cast<Constant*>(C2))); 887 } 888 return ConstantVector::get(Res); 889 case Instruction::FDiv: 890 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 891 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 892 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 893 Res.push_back(ConstantExpr::getFDiv(const_cast<Constant*>(C1), 894 const_cast<Constant*>(C2))); 895 } 896 return ConstantVector::get(Res); 897 case Instruction::URem: 898 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 899 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 900 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 901 Res.push_back(ConstantExpr::getURem(const_cast<Constant*>(C1), 902 const_cast<Constant*>(C2))); 903 } 904 return ConstantVector::get(Res); 905 case Instruction::SRem: 906 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 907 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 908 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 909 Res.push_back(ConstantExpr::getSRem(const_cast<Constant*>(C1), 910 const_cast<Constant*>(C2))); 911 } 912 return ConstantVector::get(Res); 913 case Instruction::FRem: 914 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 915 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 916 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 917 Res.push_back(ConstantExpr::getFRem(const_cast<Constant*>(C1), 918 const_cast<Constant*>(C2))); 919 } 920 return ConstantVector::get(Res); 921 case Instruction::And: 922 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 923 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 924 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 925 Res.push_back(ConstantExpr::getAnd(const_cast<Constant*>(C1), 926 const_cast<Constant*>(C2))); 927 } 928 return ConstantVector::get(Res); 929 case Instruction::Or: 930 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 931 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 932 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 933 Res.push_back(ConstantExpr::getOr(const_cast<Constant*>(C1), 934 const_cast<Constant*>(C2))); 935 } 936 return ConstantVector::get(Res); 937 case Instruction::Xor: 938 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 939 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 940 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 941 Res.push_back(ConstantExpr::getXor(const_cast<Constant*>(C1), 942 const_cast<Constant*>(C2))); 943 } 944 return ConstantVector::get(Res); 945 case Instruction::LShr: 946 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 947 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 948 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 949 Res.push_back(ConstantExpr::getLShr(const_cast<Constant*>(C1), 950 const_cast<Constant*>(C2))); 951 } 952 return ConstantVector::get(Res); 953 case Instruction::AShr: 954 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 955 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 956 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 957 Res.push_back(ConstantExpr::getAShr(const_cast<Constant*>(C1), 958 const_cast<Constant*>(C2))); 959 } 960 return ConstantVector::get(Res); 961 case Instruction::Shl: 962 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 963 C1 = CP1 ? CP1->getOperand(i) : Context.getNullValue(EltTy); 964 C2 = CP2 ? CP2->getOperand(i) : Context.getNullValue(EltTy); 965 Res.push_back(ConstantExpr::getShl(const_cast<Constant*>(C1), 966 const_cast<Constant*>(C2))); 967 } 968 return ConstantVector::get(Res); 969 } 970 } 971 } 972 973 if (isa<ConstantExpr>(C1)) { 974 // There are many possible foldings we could do here. We should probably 975 // at least fold add of a pointer with an integer into the appropriate 976 // getelementptr. This will improve alias analysis a bit. 977 } else if (isa<ConstantExpr>(C2)) { 978 // If C2 is a constant expr and C1 isn't, flop them around and fold the 979 // other way if possible. 980 switch (Opcode) { 981 case Instruction::Add: 982 case Instruction::FAdd: 983 case Instruction::Mul: 984 case Instruction::FMul: 985 case Instruction::And: 986 case Instruction::Or: 987 case Instruction::Xor: 988 // No change of opcode required. 989 return ConstantFoldBinaryInstruction(Context, Opcode, C2, C1); 990 991 case Instruction::Shl: 992 case Instruction::LShr: 993 case Instruction::AShr: 994 case Instruction::Sub: 995 case Instruction::FSub: 996 case Instruction::SDiv: 997 case Instruction::UDiv: 998 case Instruction::FDiv: 999 case Instruction::URem: 1000 case Instruction::SRem: 1001 case Instruction::FRem: 1002 default: // These instructions cannot be flopped around. 1003 break; 1004 } 1005 } 1006 1007 // We don't know how to fold this. 1008 return 0; 1009} 1010 1011/// isZeroSizedType - This type is zero sized if its an array or structure of 1012/// zero sized types. The only leaf zero sized type is an empty structure. 1013static bool isMaybeZeroSizedType(const Type *Ty) { 1014 if (isa<OpaqueType>(Ty)) return true; // Can't say. 1015 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1016 1017 // If all of elements have zero size, this does too. 1018 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1019 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1020 return true; 1021 1022 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1023 return isMaybeZeroSizedType(ATy->getElementType()); 1024 } 1025 return false; 1026} 1027 1028/// IdxCompare - Compare the two constants as though they were getelementptr 1029/// indices. This allows coersion of the types to be the same thing. 1030/// 1031/// If the two constants are the "same" (after coersion), return 0. If the 1032/// first is less than the second, return -1, if the second is less than the 1033/// first, return 1. If the constants are not integral, return -2. 1034/// 1035static int IdxCompare(LLVMContext &Context, Constant *C1, Constant *C2, 1036 const Type *ElTy) { 1037 if (C1 == C2) return 0; 1038 1039 // Ok, we found a different index. If they are not ConstantInt, we can't do 1040 // anything with them. 1041 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1042 return -2; // don't know! 1043 1044 // Ok, we have two differing integer indices. Sign extend them to be the same 1045 // type. Long is always big enough, so we use it. 1046 if (C1->getType() != Type::Int64Ty) 1047 C1 = ConstantExpr::getSExt(C1, Type::Int64Ty); 1048 1049 if (C2->getType() != Type::Int64Ty) 1050 C2 = ConstantExpr::getSExt(C2, Type::Int64Ty); 1051 1052 if (C1 == C2) return 0; // They are equal 1053 1054 // If the type being indexed over is really just a zero sized type, there is 1055 // no pointer difference being made here. 1056 if (isMaybeZeroSizedType(ElTy)) 1057 return -2; // dunno. 1058 1059 // If they are really different, now that they are the same type, then we 1060 // found a difference! 1061 if (cast<ConstantInt>(C1)->getSExtValue() < 1062 cast<ConstantInt>(C2)->getSExtValue()) 1063 return -1; 1064 else 1065 return 1; 1066} 1067 1068/// evaluateFCmpRelation - This function determines if there is anything we can 1069/// decide about the two constants provided. This doesn't need to handle simple 1070/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1071/// If we can determine that the two constants have a particular relation to 1072/// each other, we should return the corresponding FCmpInst predicate, 1073/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1074/// ConstantFoldCompareInstruction. 1075/// 1076/// To simplify this code we canonicalize the relation so that the first 1077/// operand is always the most "complex" of the two. We consider ConstantFP 1078/// to be the simplest, and ConstantExprs to be the most complex. 1079static FCmpInst::Predicate evaluateFCmpRelation(LLVMContext &Context, 1080 const Constant *V1, 1081 const Constant *V2) { 1082 assert(V1->getType() == V2->getType() && 1083 "Cannot compare values of different types!"); 1084 1085 // No compile-time operations on this type yet. 1086 if (V1->getType() == Type::PPC_FP128Ty) 1087 return FCmpInst::BAD_FCMP_PREDICATE; 1088 1089 // Handle degenerate case quickly 1090 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1091 1092 if (!isa<ConstantExpr>(V1)) { 1093 if (!isa<ConstantExpr>(V2)) { 1094 // We distilled thisUse the standard constant folder for a few cases 1095 ConstantInt *R = 0; 1096 Constant *C1 = const_cast<Constant*>(V1); 1097 Constant *C2 = const_cast<Constant*>(V2); 1098 R = dyn_cast<ConstantInt>( 1099 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2)); 1100 if (R && !R->isZero()) 1101 return FCmpInst::FCMP_OEQ; 1102 R = dyn_cast<ConstantInt>( 1103 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2)); 1104 if (R && !R->isZero()) 1105 return FCmpInst::FCMP_OLT; 1106 R = dyn_cast<ConstantInt>( 1107 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2)); 1108 if (R && !R->isZero()) 1109 return FCmpInst::FCMP_OGT; 1110 1111 // Nothing more we can do 1112 return FCmpInst::BAD_FCMP_PREDICATE; 1113 } 1114 1115 // If the first operand is simple and second is ConstantExpr, swap operands. 1116 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(Context, V2, V1); 1117 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1118 return FCmpInst::getSwappedPredicate(SwappedRelation); 1119 } else { 1120 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1121 // constantexpr or a simple constant. 1122 const ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1123 switch (CE1->getOpcode()) { 1124 case Instruction::FPTrunc: 1125 case Instruction::FPExt: 1126 case Instruction::UIToFP: 1127 case Instruction::SIToFP: 1128 // We might be able to do something with these but we don't right now. 1129 break; 1130 default: 1131 break; 1132 } 1133 } 1134 // There are MANY other foldings that we could perform here. They will 1135 // probably be added on demand, as they seem needed. 1136 return FCmpInst::BAD_FCMP_PREDICATE; 1137} 1138 1139/// evaluateICmpRelation - This function determines if there is anything we can 1140/// decide about the two constants provided. This doesn't need to handle simple 1141/// things like integer comparisons, but should instead handle ConstantExprs 1142/// and GlobalValues. If we can determine that the two constants have a 1143/// particular relation to each other, we should return the corresponding ICmp 1144/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1145/// 1146/// To simplify this code we canonicalize the relation so that the first 1147/// operand is always the most "complex" of the two. We consider simple 1148/// constants (like ConstantInt) to be the simplest, followed by 1149/// GlobalValues, followed by ConstantExpr's (the most complex). 1150/// 1151static ICmpInst::Predicate evaluateICmpRelation(LLVMContext &Context, 1152 const Constant *V1, 1153 const Constant *V2, 1154 bool isSigned) { 1155 assert(V1->getType() == V2->getType() && 1156 "Cannot compare different types of values!"); 1157 if (V1 == V2) return ICmpInst::ICMP_EQ; 1158 1159 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) { 1160 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) { 1161 // We distilled this down to a simple case, use the standard constant 1162 // folder. 1163 ConstantInt *R = 0; 1164 Constant *C1 = const_cast<Constant*>(V1); 1165 Constant *C2 = const_cast<Constant*>(V2); 1166 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1167 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1168 if (R && !R->isZero()) 1169 return pred; 1170 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1171 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1172 if (R && !R->isZero()) 1173 return pred; 1174 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1175 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2)); 1176 if (R && !R->isZero()) 1177 return pred; 1178 1179 // If we couldn't figure it out, bail. 1180 return ICmpInst::BAD_ICMP_PREDICATE; 1181 } 1182 1183 // If the first operand is simple, swap operands. 1184 ICmpInst::Predicate SwappedRelation = 1185 evaluateICmpRelation(Context, V2, V1, isSigned); 1186 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1187 return ICmpInst::getSwappedPredicate(SwappedRelation); 1188 1189 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) { 1190 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1191 ICmpInst::Predicate SwappedRelation = 1192 evaluateICmpRelation(Context, V2, V1, isSigned); 1193 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1194 return ICmpInst::getSwappedPredicate(SwappedRelation); 1195 else 1196 return ICmpInst::BAD_ICMP_PREDICATE; 1197 } 1198 1199 // Now we know that the RHS is a GlobalValue or simple constant, 1200 // which (since the types must match) means that it's a ConstantPointerNull. 1201 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1202 // Don't try to decide equality of aliases. 1203 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2)) 1204 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage()) 1205 return ICmpInst::ICMP_NE; 1206 } else { 1207 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1208 // GlobalVals can never be null. Don't try to evaluate aliases. 1209 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1)) 1210 return ICmpInst::ICMP_NE; 1211 } 1212 } else { 1213 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1214 // constantexpr, a CPR, or a simple constant. 1215 const ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1216 const Constant *CE1Op0 = CE1->getOperand(0); 1217 1218 switch (CE1->getOpcode()) { 1219 case Instruction::Trunc: 1220 case Instruction::FPTrunc: 1221 case Instruction::FPExt: 1222 case Instruction::FPToUI: 1223 case Instruction::FPToSI: 1224 break; // We can't evaluate floating point casts or truncations. 1225 1226 case Instruction::UIToFP: 1227 case Instruction::SIToFP: 1228 case Instruction::BitCast: 1229 case Instruction::ZExt: 1230 case Instruction::SExt: 1231 // If the cast is not actually changing bits, and the second operand is a 1232 // null pointer, do the comparison with the pre-casted value. 1233 if (V2->isNullValue() && 1234 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) { 1235 bool sgnd = isSigned; 1236 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1237 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1238 return evaluateICmpRelation(Context, CE1Op0, 1239 Context.getNullValue(CE1Op0->getType()), 1240 sgnd); 1241 } 1242 1243 // If the dest type is a pointer type, and the RHS is a constantexpr cast 1244 // from the same type as the src of the LHS, evaluate the inputs. This is 1245 // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)", 1246 // which happens a lot in compilers with tagged integers. 1247 if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2)) 1248 if (CE2->isCast() && isa<PointerType>(CE1->getType()) && 1249 CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() && 1250 CE1->getOperand(0)->getType()->isInteger()) { 1251 bool sgnd = isSigned; 1252 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1253 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1254 return evaluateICmpRelation(Context, CE1->getOperand(0), 1255 CE2->getOperand(0), sgnd); 1256 } 1257 break; 1258 1259 case Instruction::GetElementPtr: 1260 // Ok, since this is a getelementptr, we know that the constant has a 1261 // pointer type. Check the various cases. 1262 if (isa<ConstantPointerNull>(V2)) { 1263 // If we are comparing a GEP to a null pointer, check to see if the base 1264 // of the GEP equals the null pointer. 1265 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1266 if (GV->hasExternalWeakLinkage()) 1267 // Weak linkage GVals could be zero or not. We're comparing that 1268 // to null pointer so its greater-or-equal 1269 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1270 else 1271 // If its not weak linkage, the GVal must have a non-zero address 1272 // so the result is greater-than 1273 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1274 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1275 // If we are indexing from a null pointer, check to see if we have any 1276 // non-zero indices. 1277 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1278 if (!CE1->getOperand(i)->isNullValue()) 1279 // Offsetting from null, must not be equal. 1280 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1281 // Only zero indexes from null, must still be zero. 1282 return ICmpInst::ICMP_EQ; 1283 } 1284 // Otherwise, we can't really say if the first operand is null or not. 1285 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) { 1286 if (isa<ConstantPointerNull>(CE1Op0)) { 1287 if (CPR2->hasExternalWeakLinkage()) 1288 // Weak linkage GVals could be zero or not. We're comparing it to 1289 // a null pointer, so its less-or-equal 1290 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1291 else 1292 // If its not weak linkage, the GVal must have a non-zero address 1293 // so the result is less-than 1294 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1295 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) { 1296 if (CPR1 == CPR2) { 1297 // If this is a getelementptr of the same global, then it must be 1298 // different. Because the types must match, the getelementptr could 1299 // only have at most one index, and because we fold getelementptr's 1300 // with a single zero index, it must be nonzero. 1301 assert(CE1->getNumOperands() == 2 && 1302 !CE1->getOperand(1)->isNullValue() && 1303 "Suprising getelementptr!"); 1304 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1305 } else { 1306 // If they are different globals, we don't know what the value is, 1307 // but they can't be equal. 1308 return ICmpInst::ICMP_NE; 1309 } 1310 } 1311 } else { 1312 const ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1313 const Constant *CE2Op0 = CE2->getOperand(0); 1314 1315 // There are MANY other foldings that we could perform here. They will 1316 // probably be added on demand, as they seem needed. 1317 switch (CE2->getOpcode()) { 1318 default: break; 1319 case Instruction::GetElementPtr: 1320 // By far the most common case to handle is when the base pointers are 1321 // obviously to the same or different globals. 1322 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1323 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1324 return ICmpInst::ICMP_NE; 1325 // Ok, we know that both getelementptr instructions are based on the 1326 // same global. From this, we can precisely determine the relative 1327 // ordering of the resultant pointers. 1328 unsigned i = 1; 1329 1330 // Compare all of the operands the GEP's have in common. 1331 gep_type_iterator GTI = gep_type_begin(CE1); 1332 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1333 ++i, ++GTI) 1334 switch (IdxCompare(Context, CE1->getOperand(i), 1335 CE2->getOperand(i), GTI.getIndexedType())) { 1336 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1337 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1338 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1339 } 1340 1341 // Ok, we ran out of things they have in common. If any leftovers 1342 // are non-zero then we have a difference, otherwise we are equal. 1343 for (; i < CE1->getNumOperands(); ++i) 1344 if (!CE1->getOperand(i)->isNullValue()) { 1345 if (isa<ConstantInt>(CE1->getOperand(i))) 1346 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1347 else 1348 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1349 } 1350 1351 for (; i < CE2->getNumOperands(); ++i) 1352 if (!CE2->getOperand(i)->isNullValue()) { 1353 if (isa<ConstantInt>(CE2->getOperand(i))) 1354 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1355 else 1356 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1357 } 1358 return ICmpInst::ICMP_EQ; 1359 } 1360 } 1361 } 1362 default: 1363 break; 1364 } 1365 } 1366 1367 return ICmpInst::BAD_ICMP_PREDICATE; 1368} 1369 1370Constant *llvm::ConstantFoldCompareInstruction(LLVMContext &Context, 1371 unsigned short pred, 1372 const Constant *C1, 1373 const Constant *C2) { 1374 const Type *ResultTy; 1375 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1376 ResultTy = Context.getVectorType(Type::Int1Ty, VT->getNumElements()); 1377 else 1378 ResultTy = Type::Int1Ty; 1379 1380 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1381 if (pred == FCmpInst::FCMP_FALSE) 1382 return Context.getNullValue(ResultTy); 1383 1384 if (pred == FCmpInst::FCMP_TRUE) 1385 return Context.getAllOnesValue(ResultTy); 1386 1387 // Handle some degenerate cases first 1388 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) 1389 return Context.getUndef(ResultTy); 1390 1391 // No compile-time operations on this type yet. 1392 if (C1->getType() == Type::PPC_FP128Ty) 1393 return 0; 1394 1395 // icmp eq/ne(null,GV) -> false/true 1396 if (C1->isNullValue()) { 1397 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1398 // Don't try to evaluate aliases. External weak GV can be null. 1399 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1400 if (pred == ICmpInst::ICMP_EQ) 1401 return Context.getFalse(); 1402 else if (pred == ICmpInst::ICMP_NE) 1403 return Context.getTrue(); 1404 } 1405 // icmp eq/ne(GV,null) -> false/true 1406 } else if (C2->isNullValue()) { 1407 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1408 // Don't try to evaluate aliases. External weak GV can be null. 1409 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1410 if (pred == ICmpInst::ICMP_EQ) 1411 return Context.getFalse(); 1412 else if (pred == ICmpInst::ICMP_NE) 1413 return Context.getTrue(); 1414 } 1415 } 1416 1417 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1418 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1419 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1420 switch (pred) { 1421 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1422 case ICmpInst::ICMP_EQ: 1423 return ConstantInt::get(Type::Int1Ty, V1 == V2); 1424 case ICmpInst::ICMP_NE: 1425 return ConstantInt::get(Type::Int1Ty, V1 != V2); 1426 case ICmpInst::ICMP_SLT: 1427 return ConstantInt::get(Type::Int1Ty, V1.slt(V2)); 1428 case ICmpInst::ICMP_SGT: 1429 return ConstantInt::get(Type::Int1Ty, V1.sgt(V2)); 1430 case ICmpInst::ICMP_SLE: 1431 return ConstantInt::get(Type::Int1Ty, V1.sle(V2)); 1432 case ICmpInst::ICMP_SGE: 1433 return ConstantInt::get(Type::Int1Ty, V1.sge(V2)); 1434 case ICmpInst::ICMP_ULT: 1435 return ConstantInt::get(Type::Int1Ty, V1.ult(V2)); 1436 case ICmpInst::ICMP_UGT: 1437 return ConstantInt::get(Type::Int1Ty, V1.ugt(V2)); 1438 case ICmpInst::ICMP_ULE: 1439 return ConstantInt::get(Type::Int1Ty, V1.ule(V2)); 1440 case ICmpInst::ICMP_UGE: 1441 return ConstantInt::get(Type::Int1Ty, V1.uge(V2)); 1442 } 1443 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1444 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1445 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1446 APFloat::cmpResult R = C1V.compare(C2V); 1447 switch (pred) { 1448 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1449 case FCmpInst::FCMP_FALSE: return Context.getFalse(); 1450 case FCmpInst::FCMP_TRUE: return Context.getTrue(); 1451 case FCmpInst::FCMP_UNO: 1452 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered); 1453 case FCmpInst::FCMP_ORD: 1454 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpUnordered); 1455 case FCmpInst::FCMP_UEQ: 1456 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || 1457 R==APFloat::cmpEqual); 1458 case FCmpInst::FCMP_OEQ: 1459 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpEqual); 1460 case FCmpInst::FCMP_UNE: 1461 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpEqual); 1462 case FCmpInst::FCMP_ONE: 1463 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan || 1464 R==APFloat::cmpGreaterThan); 1465 case FCmpInst::FCMP_ULT: 1466 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || 1467 R==APFloat::cmpLessThan); 1468 case FCmpInst::FCMP_OLT: 1469 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan); 1470 case FCmpInst::FCMP_UGT: 1471 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered || 1472 R==APFloat::cmpGreaterThan); 1473 case FCmpInst::FCMP_OGT: 1474 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan); 1475 case FCmpInst::FCMP_ULE: 1476 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpGreaterThan); 1477 case FCmpInst::FCMP_OLE: 1478 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan || 1479 R==APFloat::cmpEqual); 1480 case FCmpInst::FCMP_UGE: 1481 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpLessThan); 1482 case FCmpInst::FCMP_OGE: 1483 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan || 1484 R==APFloat::cmpEqual); 1485 } 1486 } else if (isa<VectorType>(C1->getType())) { 1487 SmallVector<Constant*, 16> C1Elts, C2Elts; 1488 C1->getVectorElements(Context, C1Elts); 1489 C2->getVectorElements(Context, C2Elts); 1490 1491 // If we can constant fold the comparison of each element, constant fold 1492 // the whole vector comparison. 1493 SmallVector<Constant*, 4> ResElts; 1494 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) { 1495 // Compare the elements, producing an i1 result or constant expr. 1496 ResElts.push_back( 1497 ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1498 } 1499 return ConstantVector::get(&ResElts[0], ResElts.size()); 1500 } 1501 1502 if (C1->getType()->isFloatingPoint()) { 1503 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1504 switch (evaluateFCmpRelation(Context, C1, C2)) { 1505 default: llvm_unreachable("Unknown relation!"); 1506 case FCmpInst::FCMP_UNO: 1507 case FCmpInst::FCMP_ORD: 1508 case FCmpInst::FCMP_UEQ: 1509 case FCmpInst::FCMP_UNE: 1510 case FCmpInst::FCMP_ULT: 1511 case FCmpInst::FCMP_UGT: 1512 case FCmpInst::FCMP_ULE: 1513 case FCmpInst::FCMP_UGE: 1514 case FCmpInst::FCMP_TRUE: 1515 case FCmpInst::FCMP_FALSE: 1516 case FCmpInst::BAD_FCMP_PREDICATE: 1517 break; // Couldn't determine anything about these constants. 1518 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1519 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1520 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1521 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1522 break; 1523 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1524 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1525 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1526 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1527 break; 1528 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1529 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1530 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1531 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1532 break; 1533 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1534 // We can only partially decide this relation. 1535 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1536 Result = 0; 1537 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1538 Result = 1; 1539 break; 1540 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1541 // We can only partially decide this relation. 1542 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1543 Result = 0; 1544 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1545 Result = 1; 1546 break; 1547 case ICmpInst::ICMP_NE: // We know that C1 != C2 1548 // We can only partially decide this relation. 1549 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1550 Result = 0; 1551 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1552 Result = 1; 1553 break; 1554 } 1555 1556 // If we evaluated the result, return it now. 1557 if (Result != -1) 1558 return ConstantInt::get(Type::Int1Ty, Result); 1559 1560 } else { 1561 // Evaluate the relation between the two constants, per the predicate. 1562 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1563 switch (evaluateICmpRelation(Context, C1, C2, CmpInst::isSigned(pred))) { 1564 default: llvm_unreachable("Unknown relational!"); 1565 case ICmpInst::BAD_ICMP_PREDICATE: 1566 break; // Couldn't determine anything about these constants. 1567 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1568 // If we know the constants are equal, we can decide the result of this 1569 // computation precisely. 1570 Result = (pred == ICmpInst::ICMP_EQ || 1571 pred == ICmpInst::ICMP_ULE || 1572 pred == ICmpInst::ICMP_SLE || 1573 pred == ICmpInst::ICMP_UGE || 1574 pred == ICmpInst::ICMP_SGE); 1575 break; 1576 case ICmpInst::ICMP_ULT: 1577 // If we know that C1 < C2, we can decide the result of this computation 1578 // precisely. 1579 Result = (pred == ICmpInst::ICMP_ULT || 1580 pred == ICmpInst::ICMP_NE || 1581 pred == ICmpInst::ICMP_ULE); 1582 break; 1583 case ICmpInst::ICMP_SLT: 1584 // If we know that C1 < C2, we can decide the result of this computation 1585 // precisely. 1586 Result = (pred == ICmpInst::ICMP_SLT || 1587 pred == ICmpInst::ICMP_NE || 1588 pred == ICmpInst::ICMP_SLE); 1589 break; 1590 case ICmpInst::ICMP_UGT: 1591 // If we know that C1 > C2, we can decide the result of this computation 1592 // precisely. 1593 Result = (pred == ICmpInst::ICMP_UGT || 1594 pred == ICmpInst::ICMP_NE || 1595 pred == ICmpInst::ICMP_UGE); 1596 break; 1597 case ICmpInst::ICMP_SGT: 1598 // If we know that C1 > C2, we can decide the result of this computation 1599 // precisely. 1600 Result = (pred == ICmpInst::ICMP_SGT || 1601 pred == ICmpInst::ICMP_NE || 1602 pred == ICmpInst::ICMP_SGE); 1603 break; 1604 case ICmpInst::ICMP_ULE: 1605 // If we know that C1 <= C2, we can only partially decide this relation. 1606 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1607 if (pred == ICmpInst::ICMP_ULT) Result = 1; 1608 break; 1609 case ICmpInst::ICMP_SLE: 1610 // If we know that C1 <= C2, we can only partially decide this relation. 1611 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1612 if (pred == ICmpInst::ICMP_SLT) Result = 1; 1613 break; 1614 1615 case ICmpInst::ICMP_UGE: 1616 // If we know that C1 >= C2, we can only partially decide this relation. 1617 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1618 if (pred == ICmpInst::ICMP_UGT) Result = 1; 1619 break; 1620 case ICmpInst::ICMP_SGE: 1621 // If we know that C1 >= C2, we can only partially decide this relation. 1622 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1623 if (pred == ICmpInst::ICMP_SGT) Result = 1; 1624 break; 1625 1626 case ICmpInst::ICMP_NE: 1627 // If we know that C1 != C2, we can only partially decide this relation. 1628 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1629 if (pred == ICmpInst::ICMP_NE) Result = 1; 1630 break; 1631 } 1632 1633 // If we evaluated the result, return it now. 1634 if (Result != -1) 1635 return ConstantInt::get(Type::Int1Ty, Result); 1636 1637 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) { 1638 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1639 // other way if possible. 1640 switch (pred) { 1641 case ICmpInst::ICMP_EQ: 1642 case ICmpInst::ICMP_NE: 1643 // No change of predicate required. 1644 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1645 1646 case ICmpInst::ICMP_ULT: 1647 case ICmpInst::ICMP_SLT: 1648 case ICmpInst::ICMP_UGT: 1649 case ICmpInst::ICMP_SGT: 1650 case ICmpInst::ICMP_ULE: 1651 case ICmpInst::ICMP_SLE: 1652 case ICmpInst::ICMP_UGE: 1653 case ICmpInst::ICMP_SGE: 1654 // Change the predicate as necessary to swap the operands. 1655 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1656 return ConstantFoldCompareInstruction(Context, pred, C2, C1); 1657 1658 default: // These predicates cannot be flopped around. 1659 break; 1660 } 1661 } 1662 } 1663 return 0; 1664 } 1665 1666Constant *llvm::ConstantFoldGetElementPtr(LLVMContext &Context, 1667 const Constant *C, 1668 Constant* const *Idxs, 1669 unsigned NumIdx) { 1670 if (NumIdx == 0 || 1671 (NumIdx == 1 && Idxs[0]->isNullValue())) 1672 return const_cast<Constant*>(C); 1673 1674 if (isa<UndefValue>(C)) { 1675 const PointerType *Ptr = cast<PointerType>(C->getType()); 1676 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 1677 (Value **)Idxs, 1678 (Value **)Idxs+NumIdx); 1679 assert(Ty != 0 && "Invalid indices for GEP!"); 1680 return Context.getUndef(Context.getPointerType(Ty, Ptr->getAddressSpace())); 1681 } 1682 1683 Constant *Idx0 = Idxs[0]; 1684 if (C->isNullValue()) { 1685 bool isNull = true; 1686 for (unsigned i = 0, e = NumIdx; i != e; ++i) 1687 if (!Idxs[i]->isNullValue()) { 1688 isNull = false; 1689 break; 1690 } 1691 if (isNull) { 1692 const PointerType *Ptr = cast<PointerType>(C->getType()); 1693 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, 1694 (Value**)Idxs, 1695 (Value**)Idxs+NumIdx); 1696 assert(Ty != 0 && "Invalid indices for GEP!"); 1697 return Context.getConstantPointerNull( 1698 Context.getPointerType(Ty,Ptr->getAddressSpace())); 1699 } 1700 } 1701 1702 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) { 1703 // Combine Indices - If the source pointer to this getelementptr instruction 1704 // is a getelementptr instruction, combine the indices of the two 1705 // getelementptr instructions into a single instruction. 1706 // 1707 if (CE->getOpcode() == Instruction::GetElementPtr) { 1708 const Type *LastTy = 0; 1709 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 1710 I != E; ++I) 1711 LastTy = *I; 1712 1713 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) { 1714 SmallVector<Value*, 16> NewIndices; 1715 NewIndices.reserve(NumIdx + CE->getNumOperands()); 1716 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 1717 NewIndices.push_back(CE->getOperand(i)); 1718 1719 // Add the last index of the source with the first index of the new GEP. 1720 // Make sure to handle the case when they are actually different types. 1721 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 1722 // Otherwise it must be an array. 1723 if (!Idx0->isNullValue()) { 1724 const Type *IdxTy = Combined->getType(); 1725 if (IdxTy != Idx0->getType()) { 1726 Constant *C1 = 1727 ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty); 1728 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, 1729 Type::Int64Ty); 1730 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 1731 } else { 1732 Combined = 1733 ConstantExpr::get(Instruction::Add, Idx0, Combined); 1734 } 1735 } 1736 1737 NewIndices.push_back(Combined); 1738 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx); 1739 return ConstantExpr::getGetElementPtr(CE->getOperand(0), 1740 &NewIndices[0], 1741 NewIndices.size()); 1742 } 1743 } 1744 1745 // Implement folding of: 1746 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*), 1747 // long 0, long 0) 1748 // To: int* getelementptr ([3 x int]* %X, long 0, long 0) 1749 // 1750 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 1751 if (const PointerType *SPT = 1752 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 1753 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 1754 if (const ArrayType *CAT = 1755 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 1756 if (CAT->getElementType() == SAT->getElementType()) 1757 return ConstantExpr::getGetElementPtr( 1758 (Constant*)CE->getOperand(0), Idxs, NumIdx); 1759 } 1760 1761 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1) 1762 // Into: inttoptr (i64 0 to i8*) 1763 // This happens with pointers to member functions in C++. 1764 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 && 1765 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) && 1766 cast<PointerType>(CE->getType())->getElementType() == Type::Int8Ty) { 1767 Constant *Base = CE->getOperand(0); 1768 Constant *Offset = Idxs[0]; 1769 1770 // Convert the smaller integer to the larger type. 1771 if (Offset->getType()->getPrimitiveSizeInBits() < 1772 Base->getType()->getPrimitiveSizeInBits()) 1773 Offset = ConstantExpr::getSExt(Offset, Base->getType()); 1774 else if (Base->getType()->getPrimitiveSizeInBits() < 1775 Offset->getType()->getPrimitiveSizeInBits()) 1776 Base = ConstantExpr::getZExt(Base, Offset->getType()); 1777 1778 Base = ConstantExpr::getAdd(Base, Offset); 1779 return ConstantExpr::getIntToPtr(Base, CE->getType()); 1780 } 1781 } 1782 return 0; 1783} 1784 1785