1 //===- InstCombineSelect.cpp ----------------------------------------------===// 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 the visitSelect function. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombine.h" 15 #include "llvm/Analysis/ConstantFolding.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/Support/PatternMatch.h" 18 using namespace llvm; 19 using namespace PatternMatch; 20 21 /// MatchSelectPattern - Pattern match integer [SU]MIN, [SU]MAX, and ABS idioms, 22 /// returning the kind and providing the out parameter results if we 23 /// successfully match. 24 static SelectPatternFlavor 25 MatchSelectPattern(Value *V, Value *&LHS, Value *&RHS) { 26 SelectInst *SI = dyn_cast<SelectInst>(V); 27 if (SI == 0) return SPF_UNKNOWN; 28 29 ICmpInst *ICI = dyn_cast<ICmpInst>(SI->getCondition()); 30 if (ICI == 0) return SPF_UNKNOWN; 31 32 LHS = ICI->getOperand(0); 33 RHS = ICI->getOperand(1); 34 35 // (icmp X, Y) ? X : Y 36 if (SI->getTrueValue() == ICI->getOperand(0) && 37 SI->getFalseValue() == ICI->getOperand(1)) { 38 switch (ICI->getPredicate()) { 39 default: return SPF_UNKNOWN; // Equality. 40 case ICmpInst::ICMP_UGT: 41 case ICmpInst::ICMP_UGE: return SPF_UMAX; 42 case ICmpInst::ICMP_SGT: 43 case ICmpInst::ICMP_SGE: return SPF_SMAX; 44 case ICmpInst::ICMP_ULT: 45 case ICmpInst::ICMP_ULE: return SPF_UMIN; 46 case ICmpInst::ICMP_SLT: 47 case ICmpInst::ICMP_SLE: return SPF_SMIN; 48 } 49 } 50 51 // (icmp X, Y) ? Y : X 52 if (SI->getTrueValue() == ICI->getOperand(1) && 53 SI->getFalseValue() == ICI->getOperand(0)) { 54 switch (ICI->getPredicate()) { 55 default: return SPF_UNKNOWN; // Equality. 56 case ICmpInst::ICMP_UGT: 57 case ICmpInst::ICMP_UGE: return SPF_UMIN; 58 case ICmpInst::ICMP_SGT: 59 case ICmpInst::ICMP_SGE: return SPF_SMIN; 60 case ICmpInst::ICMP_ULT: 61 case ICmpInst::ICMP_ULE: return SPF_UMAX; 62 case ICmpInst::ICMP_SLT: 63 case ICmpInst::ICMP_SLE: return SPF_SMAX; 64 } 65 } 66 67 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5) 68 69 return SPF_UNKNOWN; 70 } 71 72 73 /// GetSelectFoldableOperands - We want to turn code that looks like this: 74 /// %C = or %A, %B 75 /// %D = select %cond, %C, %A 76 /// into: 77 /// %C = select %cond, %B, 0 78 /// %D = or %A, %C 79 /// 80 /// Assuming that the specified instruction is an operand to the select, return 81 /// a bitmask indicating which operands of this instruction are foldable if they 82 /// equal the other incoming value of the select. 83 /// 84 static unsigned GetSelectFoldableOperands(Instruction *I) { 85 switch (I->getOpcode()) { 86 case Instruction::Add: 87 case Instruction::Mul: 88 case Instruction::And: 89 case Instruction::Or: 90 case Instruction::Xor: 91 return 3; // Can fold through either operand. 92 case Instruction::Sub: // Can only fold on the amount subtracted. 93 case Instruction::Shl: // Can only fold on the shift amount. 94 case Instruction::LShr: 95 case Instruction::AShr: 96 return 1; 97 default: 98 return 0; // Cannot fold 99 } 100 } 101 102 /// GetSelectFoldableConstant - For the same transformation as the previous 103 /// function, return the identity constant that goes into the select. 104 static Constant *GetSelectFoldableConstant(Instruction *I) { 105 switch (I->getOpcode()) { 106 default: llvm_unreachable("This cannot happen!"); 107 case Instruction::Add: 108 case Instruction::Sub: 109 case Instruction::Or: 110 case Instruction::Xor: 111 case Instruction::Shl: 112 case Instruction::LShr: 113 case Instruction::AShr: 114 return Constant::getNullValue(I->getType()); 115 case Instruction::And: 116 return Constant::getAllOnesValue(I->getType()); 117 case Instruction::Mul: 118 return ConstantInt::get(I->getType(), 1); 119 } 120 } 121 122 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI 123 /// have the same opcode and only one use each. Try to simplify this. 124 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI, 125 Instruction *FI) { 126 if (TI->getNumOperands() == 1) { 127 // If this is a non-volatile load or a cast from the same type, 128 // merge. 129 if (TI->isCast()) { 130 Type *FIOpndTy = FI->getOperand(0)->getType(); 131 if (TI->getOperand(0)->getType() != FIOpndTy) 132 return 0; 133 // The select condition may be a vector. We may only change the operand 134 // type if the vector width remains the same (and matches the condition). 135 Type *CondTy = SI.getCondition()->getType(); 136 if (CondTy->isVectorTy() && (!FIOpndTy->isVectorTy() || 137 CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())) 138 return 0; 139 } else { 140 return 0; // unknown unary op. 141 } 142 143 // Fold this by inserting a select from the input values. 144 Value *NewSI = Builder->CreateSelect(SI.getCondition(), TI->getOperand(0), 145 FI->getOperand(0), SI.getName()+".v"); 146 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 147 TI->getType()); 148 } 149 150 // Only handle binary operators here. 151 if (!isa<BinaryOperator>(TI)) 152 return 0; 153 154 // Figure out if the operations have any operands in common. 155 Value *MatchOp, *OtherOpT, *OtherOpF; 156 bool MatchIsOpZero; 157 if (TI->getOperand(0) == FI->getOperand(0)) { 158 MatchOp = TI->getOperand(0); 159 OtherOpT = TI->getOperand(1); 160 OtherOpF = FI->getOperand(1); 161 MatchIsOpZero = true; 162 } else if (TI->getOperand(1) == FI->getOperand(1)) { 163 MatchOp = TI->getOperand(1); 164 OtherOpT = TI->getOperand(0); 165 OtherOpF = FI->getOperand(0); 166 MatchIsOpZero = false; 167 } else if (!TI->isCommutative()) { 168 return 0; 169 } else if (TI->getOperand(0) == FI->getOperand(1)) { 170 MatchOp = TI->getOperand(0); 171 OtherOpT = TI->getOperand(1); 172 OtherOpF = FI->getOperand(0); 173 MatchIsOpZero = true; 174 } else if (TI->getOperand(1) == FI->getOperand(0)) { 175 MatchOp = TI->getOperand(1); 176 OtherOpT = TI->getOperand(0); 177 OtherOpF = FI->getOperand(1); 178 MatchIsOpZero = true; 179 } else { 180 return 0; 181 } 182 183 // If we reach here, they do have operations in common. 184 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, 185 OtherOpF, SI.getName()+".v"); 186 187 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) { 188 if (MatchIsOpZero) 189 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI); 190 else 191 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp); 192 } 193 llvm_unreachable("Shouldn't get here"); 194 } 195 196 static bool isSelect01(Constant *C1, Constant *C2) { 197 ConstantInt *C1I = dyn_cast<ConstantInt>(C1); 198 if (!C1I) 199 return false; 200 ConstantInt *C2I = dyn_cast<ConstantInt>(C2); 201 if (!C2I) 202 return false; 203 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero. 204 return false; 205 return C1I->isOne() || C1I->isAllOnesValue() || 206 C2I->isOne() || C2I->isAllOnesValue(); 207 } 208 209 /// FoldSelectIntoOp - Try fold the select into one of the operands to 210 /// facilitate further optimization. 211 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal, 212 Value *FalseVal) { 213 // See the comment above GetSelectFoldableOperands for a description of the 214 // transformation we are doing here. 215 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) { 216 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 && 217 !isa<Constant>(FalseVal)) { 218 if (unsigned SFO = GetSelectFoldableOperands(TVI)) { 219 unsigned OpToFold = 0; 220 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 221 OpToFold = 1; 222 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 223 OpToFold = 2; 224 } 225 226 if (OpToFold) { 227 Constant *C = GetSelectFoldableConstant(TVI); 228 Value *OOp = TVI->getOperand(2-OpToFold); 229 // Avoid creating select between 2 constants unless it's selecting 230 // between 0, 1 and -1. 231 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 232 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C); 233 NewSel->takeName(TVI); 234 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI); 235 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(), 236 FalseVal, NewSel); 237 if (isa<PossiblyExactOperator>(BO)) 238 BO->setIsExact(TVI_BO->isExact()); 239 if (isa<OverflowingBinaryOperator>(BO)) { 240 BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap()); 241 BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap()); 242 } 243 return BO; 244 } 245 } 246 } 247 } 248 } 249 250 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) { 251 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 && 252 !isa<Constant>(TrueVal)) { 253 if (unsigned SFO = GetSelectFoldableOperands(FVI)) { 254 unsigned OpToFold = 0; 255 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 256 OpToFold = 1; 257 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 258 OpToFold = 2; 259 } 260 261 if (OpToFold) { 262 Constant *C = GetSelectFoldableConstant(FVI); 263 Value *OOp = FVI->getOperand(2-OpToFold); 264 // Avoid creating select between 2 constants unless it's selecting 265 // between 0, 1 and -1. 266 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 267 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp); 268 NewSel->takeName(FVI); 269 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI); 270 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(), 271 TrueVal, NewSel); 272 if (isa<PossiblyExactOperator>(BO)) 273 BO->setIsExact(FVI_BO->isExact()); 274 if (isa<OverflowingBinaryOperator>(BO)) { 275 BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap()); 276 BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap()); 277 } 278 return BO; 279 } 280 } 281 } 282 } 283 } 284 285 return 0; 286 } 287 288 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is 289 /// replaced with RepOp. 290 static Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 291 const DataLayout *TD, 292 const TargetLibraryInfo *TLI) { 293 // Trivial replacement. 294 if (V == Op) 295 return RepOp; 296 297 Instruction *I = dyn_cast<Instruction>(V); 298 if (!I) 299 return 0; 300 301 // If this is a binary operator, try to simplify it with the replaced op. 302 if (BinaryOperator *B = dyn_cast<BinaryOperator>(I)) { 303 if (B->getOperand(0) == Op) 304 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), TD, TLI); 305 if (B->getOperand(1) == Op) 306 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, TD, TLI); 307 } 308 309 // Same for CmpInsts. 310 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 311 if (C->getOperand(0) == Op) 312 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), TD, 313 TLI); 314 if (C->getOperand(1) == Op) 315 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, TD, 316 TLI); 317 } 318 319 // TODO: We could hand off more cases to instsimplify here. 320 321 // If all operands are constant after substituting Op for RepOp then we can 322 // constant fold the instruction. 323 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 324 // Build a list of all constant operands. 325 SmallVector<Constant*, 8> ConstOps; 326 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 327 if (I->getOperand(i) == Op) 328 ConstOps.push_back(CRepOp); 329 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 330 ConstOps.push_back(COp); 331 else 332 break; 333 } 334 335 // All operands were constants, fold it. 336 if (ConstOps.size() == I->getNumOperands()) { 337 if (CmpInst *C = dyn_cast<CmpInst>(I)) 338 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 339 ConstOps[1], TD, TLI); 340 341 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 342 if (!LI->isVolatile()) 343 return ConstantFoldLoadFromConstPtr(ConstOps[0], TD); 344 345 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 346 ConstOps, TD, TLI); 347 } 348 } 349 350 return 0; 351 } 352 353 /// foldSelectICmpAndOr - We want to turn: 354 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 355 /// into: 356 /// (or (shl (and X, C1), C3), y) 357 /// iff: 358 /// C1 and C2 are both powers of 2 359 /// where: 360 /// C3 = Log(C2) - Log(C1) 361 /// 362 /// This transform handles cases where: 363 /// 1. The icmp predicate is inverted 364 /// 2. The select operands are reversed 365 /// 3. The magnitude of C2 and C1 are flipped 366 static Value *foldSelectICmpAndOr(const SelectInst &SI, Value *TrueVal, 367 Value *FalseVal, 368 InstCombiner::BuilderTy *Builder) { 369 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 370 if (!IC || !IC->isEquality()) 371 return 0; 372 373 Value *CmpLHS = IC->getOperand(0); 374 Value *CmpRHS = IC->getOperand(1); 375 376 if (!match(CmpRHS, m_Zero())) 377 return 0; 378 379 Value *X; 380 const APInt *C1; 381 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1)))) 382 return 0; 383 384 const APInt *C2; 385 bool OrOnTrueVal = false; 386 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 387 if (!OrOnFalseVal) 388 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 389 390 if (!OrOnFalseVal && !OrOnTrueVal) 391 return 0; 392 393 Value *V = CmpLHS; 394 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 395 396 unsigned C1Log = C1->logBase2(); 397 unsigned C2Log = C2->logBase2(); 398 if (C2Log > C1Log) { 399 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 400 V = Builder->CreateShl(V, C2Log - C1Log); 401 } else if (C1Log > C2Log) { 402 V = Builder->CreateLShr(V, C1Log - C2Log); 403 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 404 } else 405 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 406 407 ICmpInst::Predicate Pred = IC->getPredicate(); 408 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) || 409 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal)) 410 V = Builder->CreateXor(V, *C2); 411 412 return Builder->CreateOr(V, Y); 413 } 414 415 /// visitSelectInstWithICmp - Visit a SelectInst that has an 416 /// ICmpInst as its first operand. 417 /// 418 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, 419 ICmpInst *ICI) { 420 bool Changed = false; 421 ICmpInst::Predicate Pred = ICI->getPredicate(); 422 Value *CmpLHS = ICI->getOperand(0); 423 Value *CmpRHS = ICI->getOperand(1); 424 Value *TrueVal = SI.getTrueValue(); 425 Value *FalseVal = SI.getFalseValue(); 426 427 // Check cases where the comparison is with a constant that 428 // can be adjusted to fit the min/max idiom. We may move or edit ICI 429 // here, so make sure the select is the only user. 430 if (ICI->hasOneUse()) 431 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) { 432 // X < MIN ? T : F --> F 433 if ((Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT) 434 && CI->isMinValue(Pred == ICmpInst::ICMP_SLT)) 435 return ReplaceInstUsesWith(SI, FalseVal); 436 // X > MAX ? T : F --> F 437 else if ((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT) 438 && CI->isMaxValue(Pred == ICmpInst::ICMP_SGT)) 439 return ReplaceInstUsesWith(SI, FalseVal); 440 switch (Pred) { 441 default: break; 442 case ICmpInst::ICMP_ULT: 443 case ICmpInst::ICMP_SLT: 444 case ICmpInst::ICMP_UGT: 445 case ICmpInst::ICMP_SGT: { 446 // These transformations only work for selects over integers. 447 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType()); 448 if (!SelectTy) 449 break; 450 451 Constant *AdjustedRHS; 452 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 453 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1); 454 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 455 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1); 456 457 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 458 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 459 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 460 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) 461 ; // Nothing to do here. Values match without any sign/zero extension. 462 463 // Types do not match. Instead of calculating this with mixed types 464 // promote all to the larger type. This enables scalar evolution to 465 // analyze this expression. 466 else if (CmpRHS->getType()->getScalarSizeInBits() 467 < SelectTy->getBitWidth()) { 468 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy); 469 470 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 471 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 472 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 473 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 474 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && 475 sextRHS == FalseVal) { 476 CmpLHS = TrueVal; 477 AdjustedRHS = sextRHS; 478 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 479 sextRHS == TrueVal) { 480 CmpLHS = FalseVal; 481 AdjustedRHS = sextRHS; 482 } else if (ICI->isUnsigned()) { 483 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy); 484 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 485 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 486 // zext + signed compare cannot be changed: 487 // 0xff <s 0x00, but 0x00ff >s 0x0000 488 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && 489 zextRHS == FalseVal) { 490 CmpLHS = TrueVal; 491 AdjustedRHS = zextRHS; 492 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 493 zextRHS == TrueVal) { 494 CmpLHS = FalseVal; 495 AdjustedRHS = zextRHS; 496 } else 497 break; 498 } else 499 break; 500 } else 501 break; 502 503 Pred = ICmpInst::getSwappedPredicate(Pred); 504 CmpRHS = AdjustedRHS; 505 std::swap(FalseVal, TrueVal); 506 ICI->setPredicate(Pred); 507 ICI->setOperand(0, CmpLHS); 508 ICI->setOperand(1, CmpRHS); 509 SI.setOperand(1, TrueVal); 510 SI.setOperand(2, FalseVal); 511 512 // Move ICI instruction right before the select instruction. Otherwise 513 // the sext/zext value may be defined after the ICI instruction uses it. 514 ICI->moveBefore(&SI); 515 516 Changed = true; 517 break; 518 } 519 } 520 } 521 522 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 523 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1 524 // FIXME: Type and constness constraints could be lifted, but we have to 525 // watch code size carefully. We should consider xor instead of 526 // sub/add when we decide to do that. 527 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) { 528 if (TrueVal->getType() == Ty) { 529 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) { 530 ConstantInt *C1 = NULL, *C2 = NULL; 531 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) { 532 C1 = dyn_cast<ConstantInt>(TrueVal); 533 C2 = dyn_cast<ConstantInt>(FalseVal); 534 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) { 535 C1 = dyn_cast<ConstantInt>(FalseVal); 536 C2 = dyn_cast<ConstantInt>(TrueVal); 537 } 538 if (C1 && C2) { 539 // This shift results in either -1 or 0. 540 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1); 541 542 // Check if we can express the operation with a single or. 543 if (C2->isAllOnesValue()) 544 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1)); 545 546 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue()); 547 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1)); 548 } 549 } 550 } 551 } 552 553 // If we have an equality comparison then we know the value in one of the 554 // arms of the select. See if substituting this value into the arm and 555 // simplifying the result yields the same value as the other arm. 556 if (Pred == ICmpInst::ICMP_EQ) { 557 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 558 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 559 return ReplaceInstUsesWith(SI, FalseVal); 560 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 561 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 562 return ReplaceInstUsesWith(SI, FalseVal); 563 } else if (Pred == ICmpInst::ICMP_NE) { 564 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 565 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 566 return ReplaceInstUsesWith(SI, TrueVal); 567 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 568 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 569 return ReplaceInstUsesWith(SI, TrueVal); 570 } 571 572 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 573 574 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 575 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 576 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 577 SI.setOperand(1, CmpRHS); 578 Changed = true; 579 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 580 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 581 SI.setOperand(2, CmpRHS); 582 Changed = true; 583 } 584 } 585 586 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder)) 587 return ReplaceInstUsesWith(SI, V); 588 589 return Changed ? &SI : 0; 590 } 591 592 593 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a 594 /// PHI node (but the two may be in different blocks). See if the true/false 595 /// values (V) are live in all of the predecessor blocks of the PHI. For 596 /// example, cases like this cannot be mapped: 597 /// 598 /// X = phi [ C1, BB1], [C2, BB2] 599 /// Y = add 600 /// Z = select X, Y, 0 601 /// 602 /// because Y is not live in BB1/BB2. 603 /// 604 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V, 605 const SelectInst &SI) { 606 // If the value is a non-instruction value like a constant or argument, it 607 // can always be mapped. 608 const Instruction *I = dyn_cast<Instruction>(V); 609 if (I == 0) return true; 610 611 // If V is a PHI node defined in the same block as the condition PHI, we can 612 // map the arguments. 613 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 614 615 if (const PHINode *VP = dyn_cast<PHINode>(I)) 616 if (VP->getParent() == CondPHI->getParent()) 617 return true; 618 619 // Otherwise, if the PHI and select are defined in the same block and if V is 620 // defined in a different block, then we can transform it. 621 if (SI.getParent() == CondPHI->getParent() && 622 I->getParent() != CondPHI->getParent()) 623 return true; 624 625 // Otherwise we have a 'hard' case and we can't tell without doing more 626 // detailed dominator based analysis, punt. 627 return false; 628 } 629 630 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form: 631 /// SPF2(SPF1(A, B), C) 632 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner, 633 SelectPatternFlavor SPF1, 634 Value *A, Value *B, 635 Instruction &Outer, 636 SelectPatternFlavor SPF2, Value *C) { 637 if (C == A || C == B) { 638 // MAX(MAX(A, B), B) -> MAX(A, B) 639 // MIN(MIN(a, b), a) -> MIN(a, b) 640 if (SPF1 == SPF2) 641 return ReplaceInstUsesWith(Outer, Inner); 642 643 // MAX(MIN(a, b), a) -> a 644 // MIN(MAX(a, b), a) -> a 645 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 646 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 647 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 648 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 649 return ReplaceInstUsesWith(Outer, C); 650 } 651 652 // TODO: MIN(MIN(A, 23), 97) 653 return 0; 654 } 655 656 657 /// foldSelectICmpAnd - If one of the constants is zero (we know they can't 658 /// both be) and we have an icmp instruction with zero, and we have an 'and' 659 /// with the non-constant value and a power of two we can turn the select 660 /// into a shift on the result of the 'and'. 661 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal, 662 ConstantInt *FalseVal, 663 InstCombiner::BuilderTy *Builder) { 664 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 665 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy()) 666 return 0; 667 668 if (!match(IC->getOperand(1), m_Zero())) 669 return 0; 670 671 ConstantInt *AndRHS; 672 Value *LHS = IC->getOperand(0); 673 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS)))) 674 return 0; 675 676 // If both select arms are non-zero see if we have a select of the form 677 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic 678 // for 'x ? 2^n : 0' and fix the thing up at the end. 679 ConstantInt *Offset = 0; 680 if (!TrueVal->isZero() && !FalseVal->isZero()) { 681 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2()) 682 Offset = FalseVal; 683 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2()) 684 Offset = TrueVal; 685 else 686 return 0; 687 688 // Adjust TrueVal and FalseVal to the offset. 689 TrueVal = ConstantInt::get(Builder->getContext(), 690 TrueVal->getValue() - Offset->getValue()); 691 FalseVal = ConstantInt::get(Builder->getContext(), 692 FalseVal->getValue() - Offset->getValue()); 693 } 694 695 // Make sure the mask in the 'and' and one of the select arms is a power of 2. 696 if (!AndRHS->getValue().isPowerOf2() || 697 (!TrueVal->getValue().isPowerOf2() && 698 !FalseVal->getValue().isPowerOf2())) 699 return 0; 700 701 // Determine which shift is needed to transform result of the 'and' into the 702 // desired result. 703 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal; 704 unsigned ValZeros = ValC->getValue().logBase2(); 705 unsigned AndZeros = AndRHS->getValue().logBase2(); 706 707 // If types don't match we can still convert the select by introducing a zext 708 // or a trunc of the 'and'. The trunc case requires that all of the truncated 709 // bits are zero, we can figure that out by looking at the 'and' mask. 710 if (AndZeros >= ValC->getBitWidth()) 711 return 0; 712 713 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType()); 714 if (ValZeros > AndZeros) 715 V = Builder->CreateShl(V, ValZeros - AndZeros); 716 else if (ValZeros < AndZeros) 717 V = Builder->CreateLShr(V, AndZeros - ValZeros); 718 719 // Okay, now we know that everything is set up, we just don't know whether we 720 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 721 bool ShouldNotVal = !TrueVal->isZero(); 722 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; 723 if (ShouldNotVal) 724 V = Builder->CreateXor(V, ValC); 725 726 // Apply an offset if needed. 727 if (Offset) 728 V = Builder->CreateAdd(V, Offset); 729 return V; 730 } 731 732 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { 733 Value *CondVal = SI.getCondition(); 734 Value *TrueVal = SI.getTrueValue(); 735 Value *FalseVal = SI.getFalseValue(); 736 737 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, TD)) 738 return ReplaceInstUsesWith(SI, V); 739 740 if (SI.getType()->isIntegerTy(1)) { 741 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { 742 if (C->getZExtValue()) { 743 // Change: A = select B, true, C --> A = or B, C 744 return BinaryOperator::CreateOr(CondVal, FalseVal); 745 } 746 // Change: A = select B, false, C --> A = and !B, C 747 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 748 return BinaryOperator::CreateAnd(NotCond, FalseVal); 749 } 750 if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) { 751 if (C->getZExtValue() == false) { 752 // Change: A = select B, C, false --> A = and B, C 753 return BinaryOperator::CreateAnd(CondVal, TrueVal); 754 } 755 // Change: A = select B, C, true --> A = or !B, C 756 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 757 return BinaryOperator::CreateOr(NotCond, TrueVal); 758 } 759 760 // select a, b, a -> a&b 761 // select a, a, b -> a|b 762 if (CondVal == TrueVal) 763 return BinaryOperator::CreateOr(CondVal, FalseVal); 764 if (CondVal == FalseVal) 765 return BinaryOperator::CreateAnd(CondVal, TrueVal); 766 767 // select a, ~a, b -> (~a)&b 768 // select a, b, ~a -> (~a)|b 769 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 770 return BinaryOperator::CreateAnd(TrueVal, FalseVal); 771 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 772 return BinaryOperator::CreateOr(TrueVal, FalseVal); 773 } 774 775 // Selecting between two integer constants? 776 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) 777 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { 778 // select C, 1, 0 -> zext C to int 779 if (FalseValC->isZero() && TrueValC->getValue() == 1) 780 return new ZExtInst(CondVal, SI.getType()); 781 782 // select C, -1, 0 -> sext C to int 783 if (FalseValC->isZero() && TrueValC->isAllOnesValue()) 784 return new SExtInst(CondVal, SI.getType()); 785 786 // select C, 0, 1 -> zext !C to int 787 if (TrueValC->isZero() && FalseValC->getValue() == 1) { 788 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 789 return new ZExtInst(NotCond, SI.getType()); 790 } 791 792 // select C, 0, -1 -> sext !C to int 793 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { 794 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 795 return new SExtInst(NotCond, SI.getType()); 796 } 797 798 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder)) 799 return ReplaceInstUsesWith(SI, V); 800 } 801 802 // See if we are selecting two values based on a comparison of the two values. 803 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 804 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) { 805 // Transform (X == Y) ? X : Y -> Y 806 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 807 // This is not safe in general for floating point: 808 // consider X== -0, Y== +0. 809 // It becomes safe if either operand is a nonzero constant. 810 ConstantFP *CFPt, *CFPf; 811 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 812 !CFPt->getValueAPF().isZero()) || 813 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 814 !CFPf->getValueAPF().isZero())) 815 return ReplaceInstUsesWith(SI, FalseVal); 816 } 817 // Transform (X une Y) ? X : Y -> X 818 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 819 // This is not safe in general for floating point: 820 // consider X== -0, Y== +0. 821 // It becomes safe if either operand is a nonzero constant. 822 ConstantFP *CFPt, *CFPf; 823 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 824 !CFPt->getValueAPF().isZero()) || 825 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 826 !CFPf->getValueAPF().isZero())) 827 return ReplaceInstUsesWith(SI, TrueVal); 828 } 829 // NOTE: if we wanted to, this is where to detect MIN/MAX 830 831 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){ 832 // Transform (X == Y) ? Y : X -> X 833 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 834 // This is not safe in general for floating point: 835 // consider X== -0, Y== +0. 836 // It becomes safe if either operand is a nonzero constant. 837 ConstantFP *CFPt, *CFPf; 838 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 839 !CFPt->getValueAPF().isZero()) || 840 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 841 !CFPf->getValueAPF().isZero())) 842 return ReplaceInstUsesWith(SI, FalseVal); 843 } 844 // Transform (X une Y) ? Y : X -> Y 845 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 846 // This is not safe in general for floating point: 847 // consider X== -0, Y== +0. 848 // It becomes safe if either operand is a nonzero constant. 849 ConstantFP *CFPt, *CFPf; 850 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 851 !CFPt->getValueAPF().isZero()) || 852 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 853 !CFPf->getValueAPF().isZero())) 854 return ReplaceInstUsesWith(SI, TrueVal); 855 } 856 // NOTE: if we wanted to, this is where to detect MIN/MAX 857 } 858 // NOTE: if we wanted to, this is where to detect ABS 859 } 860 861 // See if we are selecting two values based on a comparison of the two values. 862 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 863 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI)) 864 return Result; 865 866 if (Instruction *TI = dyn_cast<Instruction>(TrueVal)) 867 if (Instruction *FI = dyn_cast<Instruction>(FalseVal)) 868 if (TI->hasOneUse() && FI->hasOneUse()) { 869 Instruction *AddOp = 0, *SubOp = 0; 870 871 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 872 if (TI->getOpcode() == FI->getOpcode()) 873 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI)) 874 return IV; 875 876 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is 877 // even legal for FP. 878 if ((TI->getOpcode() == Instruction::Sub && 879 FI->getOpcode() == Instruction::Add) || 880 (TI->getOpcode() == Instruction::FSub && 881 FI->getOpcode() == Instruction::FAdd)) { 882 AddOp = FI; SubOp = TI; 883 } else if ((FI->getOpcode() == Instruction::Sub && 884 TI->getOpcode() == Instruction::Add) || 885 (FI->getOpcode() == Instruction::FSub && 886 TI->getOpcode() == Instruction::FAdd)) { 887 AddOp = TI; SubOp = FI; 888 } 889 890 if (AddOp) { 891 Value *OtherAddOp = 0; 892 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 893 OtherAddOp = AddOp->getOperand(1); 894 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 895 OtherAddOp = AddOp->getOperand(0); 896 } 897 898 if (OtherAddOp) { 899 // So at this point we know we have (Y -> OtherAddOp): 900 // select C, (add X, Y), (sub X, Z) 901 Value *NegVal; // Compute -Z 902 if (SI.getType()->isFPOrFPVectorTy()) { 903 NegVal = Builder->CreateFNeg(SubOp->getOperand(1)); 904 } else { 905 NegVal = Builder->CreateNeg(SubOp->getOperand(1)); 906 } 907 908 Value *NewTrueOp = OtherAddOp; 909 Value *NewFalseOp = NegVal; 910 if (AddOp != TI) 911 std::swap(NewTrueOp, NewFalseOp); 912 Value *NewSel = 913 Builder->CreateSelect(CondVal, NewTrueOp, 914 NewFalseOp, SI.getName() + ".p"); 915 916 if (SI.getType()->isFPOrFPVectorTy()) 917 return BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 918 else 919 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 920 } 921 } 922 } 923 924 // See if we can fold the select into one of our operands. 925 if (SI.getType()->isIntegerTy()) { 926 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) 927 return FoldI; 928 929 // MAX(MAX(a, b), a) -> MAX(a, b) 930 // MIN(MIN(a, b), a) -> MIN(a, b) 931 // MAX(MIN(a, b), a) -> a 932 // MIN(MAX(a, b), a) -> a 933 Value *LHS, *RHS, *LHS2, *RHS2; 934 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) { 935 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2)) 936 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2, 937 SI, SPF, RHS)) 938 return R; 939 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2)) 940 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2, 941 SI, SPF, LHS)) 942 return R; 943 } 944 945 // TODO. 946 // ABS(-X) -> ABS(X) 947 // ABS(ABS(X)) -> ABS(X) 948 } 949 950 // See if we can fold the select into a phi node if the condition is a select. 951 if (isa<PHINode>(SI.getCondition())) 952 // The true/false values have to be live in the PHI predecessor's blocks. 953 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 954 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 955 if (Instruction *NV = FoldOpIntoPhi(SI)) 956 return NV; 957 958 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 959 if (TrueSI->getCondition() == CondVal) { 960 if (SI.getTrueValue() == TrueSI->getTrueValue()) 961 return 0; 962 SI.setOperand(1, TrueSI->getTrueValue()); 963 return &SI; 964 } 965 } 966 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 967 if (FalseSI->getCondition() == CondVal) { 968 if (SI.getFalseValue() == FalseSI->getFalseValue()) 969 return 0; 970 SI.setOperand(2, FalseSI->getFalseValue()); 971 return &SI; 972 } 973 } 974 975 if (BinaryOperator::isNot(CondVal)) { 976 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal)); 977 SI.setOperand(1, FalseVal); 978 SI.setOperand(2, TrueVal); 979 return &SI; 980 } 981 982 if (VectorType* VecTy = dyn_cast<VectorType>(SI.getType())) { 983 unsigned VWidth = VecTy->getNumElements(); 984 APInt UndefElts(VWidth, 0); 985 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 986 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) { 987 if (V != &SI) 988 return ReplaceInstUsesWith(SI, V); 989 return &SI; 990 } 991 992 if (isa<ConstantAggregateZero>(CondVal)) { 993 return ReplaceInstUsesWith(SI, FalseVal); 994 } 995 } 996 997 return 0; 998 } 999
