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 "InstCombineInternal.h" 15 #include "llvm/Analysis/ConstantFolding.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/Analysis/ValueTracking.h" 18 #include "llvm/IR/PatternMatch.h" 19 using namespace llvm; 20 using namespace PatternMatch; 21 22 #define DEBUG_TYPE "instcombine" 23 24 static SelectPatternFlavor 25 getInverseMinMaxSelectPattern(SelectPatternFlavor SPF) { 26 switch (SPF) { 27 default: 28 llvm_unreachable("unhandled!"); 29 30 case SPF_SMIN: 31 return SPF_SMAX; 32 case SPF_UMIN: 33 return SPF_UMAX; 34 case SPF_SMAX: 35 return SPF_SMIN; 36 case SPF_UMAX: 37 return SPF_UMIN; 38 } 39 } 40 41 static CmpInst::Predicate getICmpPredicateForMinMax(SelectPatternFlavor SPF) { 42 switch (SPF) { 43 default: 44 llvm_unreachable("unhandled!"); 45 46 case SPF_SMIN: 47 return ICmpInst::ICMP_SLT; 48 case SPF_UMIN: 49 return ICmpInst::ICMP_ULT; 50 case SPF_SMAX: 51 return ICmpInst::ICMP_SGT; 52 case SPF_UMAX: 53 return ICmpInst::ICMP_UGT; 54 } 55 } 56 57 static Value *generateMinMaxSelectPattern(InstCombiner::BuilderTy *Builder, 58 SelectPatternFlavor SPF, Value *A, 59 Value *B) { 60 CmpInst::Predicate Pred = getICmpPredicateForMinMax(SPF); 61 return Builder->CreateSelect(Builder->CreateICmp(Pred, A, B), A, B); 62 } 63 64 /// GetSelectFoldableOperands - We want to turn code that looks like this: 65 /// %C = or %A, %B 66 /// %D = select %cond, %C, %A 67 /// into: 68 /// %C = select %cond, %B, 0 69 /// %D = or %A, %C 70 /// 71 /// Assuming that the specified instruction is an operand to the select, return 72 /// a bitmask indicating which operands of this instruction are foldable if they 73 /// equal the other incoming value of the select. 74 /// 75 static unsigned GetSelectFoldableOperands(Instruction *I) { 76 switch (I->getOpcode()) { 77 case Instruction::Add: 78 case Instruction::Mul: 79 case Instruction::And: 80 case Instruction::Or: 81 case Instruction::Xor: 82 return 3; // Can fold through either operand. 83 case Instruction::Sub: // Can only fold on the amount subtracted. 84 case Instruction::Shl: // Can only fold on the shift amount. 85 case Instruction::LShr: 86 case Instruction::AShr: 87 return 1; 88 default: 89 return 0; // Cannot fold 90 } 91 } 92 93 /// GetSelectFoldableConstant - For the same transformation as the previous 94 /// function, return the identity constant that goes into the select. 95 static Constant *GetSelectFoldableConstant(Instruction *I) { 96 switch (I->getOpcode()) { 97 default: llvm_unreachable("This cannot happen!"); 98 case Instruction::Add: 99 case Instruction::Sub: 100 case Instruction::Or: 101 case Instruction::Xor: 102 case Instruction::Shl: 103 case Instruction::LShr: 104 case Instruction::AShr: 105 return Constant::getNullValue(I->getType()); 106 case Instruction::And: 107 return Constant::getAllOnesValue(I->getType()); 108 case Instruction::Mul: 109 return ConstantInt::get(I->getType(), 1); 110 } 111 } 112 113 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI 114 /// have the same opcode and only one use each. Try to simplify this. 115 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI, 116 Instruction *FI) { 117 if (TI->getNumOperands() == 1) { 118 // If this is a non-volatile load or a cast from the same type, 119 // merge. 120 if (TI->isCast()) { 121 Type *FIOpndTy = FI->getOperand(0)->getType(); 122 if (TI->getOperand(0)->getType() != FIOpndTy) 123 return nullptr; 124 // The select condition may be a vector. We may only change the operand 125 // type if the vector width remains the same (and matches the condition). 126 Type *CondTy = SI.getCondition()->getType(); 127 if (CondTy->isVectorTy() && (!FIOpndTy->isVectorTy() || 128 CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())) 129 return nullptr; 130 } else { 131 return nullptr; // unknown unary op. 132 } 133 134 // Fold this by inserting a select from the input values. 135 Value *NewSI = Builder->CreateSelect(SI.getCondition(), TI->getOperand(0), 136 FI->getOperand(0), SI.getName()+".v"); 137 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 138 TI->getType()); 139 } 140 141 // Only handle binary operators here. 142 if (!isa<BinaryOperator>(TI)) 143 return nullptr; 144 145 // Figure out if the operations have any operands in common. 146 Value *MatchOp, *OtherOpT, *OtherOpF; 147 bool MatchIsOpZero; 148 if (TI->getOperand(0) == FI->getOperand(0)) { 149 MatchOp = TI->getOperand(0); 150 OtherOpT = TI->getOperand(1); 151 OtherOpF = FI->getOperand(1); 152 MatchIsOpZero = true; 153 } else if (TI->getOperand(1) == FI->getOperand(1)) { 154 MatchOp = TI->getOperand(1); 155 OtherOpT = TI->getOperand(0); 156 OtherOpF = FI->getOperand(0); 157 MatchIsOpZero = false; 158 } else if (!TI->isCommutative()) { 159 return nullptr; 160 } else if (TI->getOperand(0) == FI->getOperand(1)) { 161 MatchOp = TI->getOperand(0); 162 OtherOpT = TI->getOperand(1); 163 OtherOpF = FI->getOperand(0); 164 MatchIsOpZero = true; 165 } else if (TI->getOperand(1) == FI->getOperand(0)) { 166 MatchOp = TI->getOperand(1); 167 OtherOpT = TI->getOperand(0); 168 OtherOpF = FI->getOperand(1); 169 MatchIsOpZero = true; 170 } else { 171 return nullptr; 172 } 173 174 // If we reach here, they do have operations in common. 175 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, 176 OtherOpF, SI.getName()+".v"); 177 178 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) { 179 if (MatchIsOpZero) 180 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI); 181 else 182 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp); 183 } 184 llvm_unreachable("Shouldn't get here"); 185 } 186 187 static bool isSelect01(Constant *C1, Constant *C2) { 188 ConstantInt *C1I = dyn_cast<ConstantInt>(C1); 189 if (!C1I) 190 return false; 191 ConstantInt *C2I = dyn_cast<ConstantInt>(C2); 192 if (!C2I) 193 return false; 194 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero. 195 return false; 196 return C1I->isOne() || C1I->isAllOnesValue() || 197 C2I->isOne() || C2I->isAllOnesValue(); 198 } 199 200 /// FoldSelectIntoOp - Try fold the select into one of the operands to 201 /// facilitate further optimization. 202 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal, 203 Value *FalseVal) { 204 // See the comment above GetSelectFoldableOperands for a description of the 205 // transformation we are doing here. 206 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) { 207 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 && 208 !isa<Constant>(FalseVal)) { 209 if (unsigned SFO = GetSelectFoldableOperands(TVI)) { 210 unsigned OpToFold = 0; 211 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 212 OpToFold = 1; 213 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 214 OpToFold = 2; 215 } 216 217 if (OpToFold) { 218 Constant *C = GetSelectFoldableConstant(TVI); 219 Value *OOp = TVI->getOperand(2-OpToFold); 220 // Avoid creating select between 2 constants unless it's selecting 221 // between 0, 1 and -1. 222 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 223 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C); 224 NewSel->takeName(TVI); 225 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI); 226 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(), 227 FalseVal, NewSel); 228 if (isa<PossiblyExactOperator>(BO)) 229 BO->setIsExact(TVI_BO->isExact()); 230 if (isa<OverflowingBinaryOperator>(BO)) { 231 BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap()); 232 BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap()); 233 } 234 return BO; 235 } 236 } 237 } 238 } 239 } 240 241 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) { 242 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 && 243 !isa<Constant>(TrueVal)) { 244 if (unsigned SFO = GetSelectFoldableOperands(FVI)) { 245 unsigned OpToFold = 0; 246 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 247 OpToFold = 1; 248 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 249 OpToFold = 2; 250 } 251 252 if (OpToFold) { 253 Constant *C = GetSelectFoldableConstant(FVI); 254 Value *OOp = FVI->getOperand(2-OpToFold); 255 // Avoid creating select between 2 constants unless it's selecting 256 // between 0, 1 and -1. 257 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 258 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp); 259 NewSel->takeName(FVI); 260 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI); 261 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(), 262 TrueVal, NewSel); 263 if (isa<PossiblyExactOperator>(BO)) 264 BO->setIsExact(FVI_BO->isExact()); 265 if (isa<OverflowingBinaryOperator>(BO)) { 266 BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap()); 267 BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap()); 268 } 269 return BO; 270 } 271 } 272 } 273 } 274 } 275 276 return nullptr; 277 } 278 279 /// foldSelectICmpAndOr - We want to turn: 280 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 281 /// into: 282 /// (or (shl (and X, C1), C3), y) 283 /// iff: 284 /// C1 and C2 are both powers of 2 285 /// where: 286 /// C3 = Log(C2) - Log(C1) 287 /// 288 /// This transform handles cases where: 289 /// 1. The icmp predicate is inverted 290 /// 2. The select operands are reversed 291 /// 3. The magnitude of C2 and C1 are flipped 292 static Value *foldSelectICmpAndOr(const SelectInst &SI, Value *TrueVal, 293 Value *FalseVal, 294 InstCombiner::BuilderTy *Builder) { 295 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 296 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy()) 297 return nullptr; 298 299 Value *CmpLHS = IC->getOperand(0); 300 Value *CmpRHS = IC->getOperand(1); 301 302 if (!match(CmpRHS, m_Zero())) 303 return nullptr; 304 305 Value *X; 306 const APInt *C1; 307 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1)))) 308 return nullptr; 309 310 const APInt *C2; 311 bool OrOnTrueVal = false; 312 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 313 if (!OrOnFalseVal) 314 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 315 316 if (!OrOnFalseVal && !OrOnTrueVal) 317 return nullptr; 318 319 Value *V = CmpLHS; 320 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 321 322 unsigned C1Log = C1->logBase2(); 323 unsigned C2Log = C2->logBase2(); 324 if (C2Log > C1Log) { 325 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 326 V = Builder->CreateShl(V, C2Log - C1Log); 327 } else if (C1Log > C2Log) { 328 V = Builder->CreateLShr(V, C1Log - C2Log); 329 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 330 } else 331 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 332 333 ICmpInst::Predicate Pred = IC->getPredicate(); 334 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) || 335 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal)) 336 V = Builder->CreateXor(V, *C2); 337 338 return Builder->CreateOr(V, Y); 339 } 340 341 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single 342 /// call to cttz/ctlz with flag 'is_zero_undef' cleared. 343 /// 344 /// For example, we can fold the following code sequence: 345 /// \code 346 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) 347 /// %1 = icmp ne i32 %x, 0 348 /// %2 = select i1 %1, i32 %0, i32 32 349 /// \code 350 /// 351 /// into: 352 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) 353 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, 354 InstCombiner::BuilderTy *Builder) { 355 ICmpInst::Predicate Pred = ICI->getPredicate(); 356 Value *CmpLHS = ICI->getOperand(0); 357 Value *CmpRHS = ICI->getOperand(1); 358 359 // Check if the condition value compares a value for equality against zero. 360 if (!ICI->isEquality() || !match(CmpRHS, m_Zero())) 361 return nullptr; 362 363 Value *Count = FalseVal; 364 Value *ValueOnZero = TrueVal; 365 if (Pred == ICmpInst::ICMP_NE) 366 std::swap(Count, ValueOnZero); 367 368 // Skip zero extend/truncate. 369 Value *V = nullptr; 370 if (match(Count, m_ZExt(m_Value(V))) || 371 match(Count, m_Trunc(m_Value(V)))) 372 Count = V; 373 374 // Check if the value propagated on zero is a constant number equal to the 375 // sizeof in bits of 'Count'. 376 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); 377 if (!match(ValueOnZero, m_SpecificInt(SizeOfInBits))) 378 return nullptr; 379 380 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the 381 // input to the cttz/ctlz is used as LHS for the compare instruction. 382 if (match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) || 383 match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) { 384 IntrinsicInst *II = cast<IntrinsicInst>(Count); 385 IRBuilder<> Builder(II); 386 // Explicitly clear the 'undef_on_zero' flag. 387 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone()); 388 Type *Ty = NewI->getArgOperand(1)->getType(); 389 NewI->setArgOperand(1, Constant::getNullValue(Ty)); 390 Builder.Insert(NewI); 391 return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType()); 392 } 393 394 return nullptr; 395 } 396 397 /// visitSelectInstWithICmp - Visit a SelectInst that has an 398 /// ICmpInst as its first operand. 399 /// 400 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, 401 ICmpInst *ICI) { 402 bool Changed = false; 403 ICmpInst::Predicate Pred = ICI->getPredicate(); 404 Value *CmpLHS = ICI->getOperand(0); 405 Value *CmpRHS = ICI->getOperand(1); 406 Value *TrueVal = SI.getTrueValue(); 407 Value *FalseVal = SI.getFalseValue(); 408 409 // Check cases where the comparison is with a constant that 410 // can be adjusted to fit the min/max idiom. We may move or edit ICI 411 // here, so make sure the select is the only user. 412 if (ICI->hasOneUse()) 413 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) { 414 switch (Pred) { 415 default: break; 416 case ICmpInst::ICMP_ULT: 417 case ICmpInst::ICMP_SLT: 418 case ICmpInst::ICMP_UGT: 419 case ICmpInst::ICMP_SGT: { 420 // These transformations only work for selects over integers. 421 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType()); 422 if (!SelectTy) 423 break; 424 425 Constant *AdjustedRHS; 426 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 427 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1); 428 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 429 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1); 430 431 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 432 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 433 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 434 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) 435 ; // Nothing to do here. Values match without any sign/zero extension. 436 437 // Types do not match. Instead of calculating this with mixed types 438 // promote all to the larger type. This enables scalar evolution to 439 // analyze this expression. 440 else if (CmpRHS->getType()->getScalarSizeInBits() 441 < SelectTy->getBitWidth()) { 442 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy); 443 444 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 445 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 446 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 447 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 448 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && 449 sextRHS == FalseVal) { 450 CmpLHS = TrueVal; 451 AdjustedRHS = sextRHS; 452 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 453 sextRHS == TrueVal) { 454 CmpLHS = FalseVal; 455 AdjustedRHS = sextRHS; 456 } else if (ICI->isUnsigned()) { 457 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy); 458 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 459 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 460 // zext + signed compare cannot be changed: 461 // 0xff <s 0x00, but 0x00ff >s 0x0000 462 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && 463 zextRHS == FalseVal) { 464 CmpLHS = TrueVal; 465 AdjustedRHS = zextRHS; 466 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 467 zextRHS == TrueVal) { 468 CmpLHS = FalseVal; 469 AdjustedRHS = zextRHS; 470 } else 471 break; 472 } else 473 break; 474 } else 475 break; 476 477 Pred = ICmpInst::getSwappedPredicate(Pred); 478 CmpRHS = AdjustedRHS; 479 std::swap(FalseVal, TrueVal); 480 ICI->setPredicate(Pred); 481 ICI->setOperand(0, CmpLHS); 482 ICI->setOperand(1, CmpRHS); 483 SI.setOperand(1, TrueVal); 484 SI.setOperand(2, FalseVal); 485 486 // Move ICI instruction right before the select instruction. Otherwise 487 // the sext/zext value may be defined after the ICI instruction uses it. 488 ICI->moveBefore(&SI); 489 490 Changed = true; 491 break; 492 } 493 } 494 } 495 496 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 497 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1 498 // FIXME: Type and constness constraints could be lifted, but we have to 499 // watch code size carefully. We should consider xor instead of 500 // sub/add when we decide to do that. 501 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) { 502 if (TrueVal->getType() == Ty) { 503 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) { 504 ConstantInt *C1 = nullptr, *C2 = nullptr; 505 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) { 506 C1 = dyn_cast<ConstantInt>(TrueVal); 507 C2 = dyn_cast<ConstantInt>(FalseVal); 508 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) { 509 C1 = dyn_cast<ConstantInt>(FalseVal); 510 C2 = dyn_cast<ConstantInt>(TrueVal); 511 } 512 if (C1 && C2) { 513 // This shift results in either -1 or 0. 514 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1); 515 516 // Check if we can express the operation with a single or. 517 if (C2->isAllOnesValue()) 518 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1)); 519 520 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue()); 521 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1)); 522 } 523 } 524 } 525 } 526 527 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 528 529 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 530 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 531 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 532 SI.setOperand(1, CmpRHS); 533 Changed = true; 534 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 535 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 536 SI.setOperand(2, CmpRHS); 537 Changed = true; 538 } 539 } 540 541 { 542 unsigned BitWidth = DL.getTypeSizeInBits(TrueVal->getType()); 543 APInt MinSignedValue = APInt::getSignBit(BitWidth); 544 Value *X; 545 const APInt *Y, *C; 546 bool TrueWhenUnset; 547 bool IsBitTest = false; 548 if (ICmpInst::isEquality(Pred) && 549 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && 550 match(CmpRHS, m_Zero())) { 551 IsBitTest = true; 552 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 553 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 554 X = CmpLHS; 555 Y = &MinSignedValue; 556 IsBitTest = true; 557 TrueWhenUnset = false; 558 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 559 X = CmpLHS; 560 Y = &MinSignedValue; 561 IsBitTest = true; 562 TrueWhenUnset = true; 563 } 564 if (IsBitTest) { 565 Value *V = nullptr; 566 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y 567 if (TrueWhenUnset && TrueVal == X && 568 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 569 V = Builder->CreateAnd(X, ~(*Y)); 570 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y 571 else if (!TrueWhenUnset && FalseVal == X && 572 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 573 V = Builder->CreateAnd(X, ~(*Y)); 574 // (X & Y) == 0 ? X ^ Y : X --> X | Y 575 else if (TrueWhenUnset && FalseVal == X && 576 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 577 V = Builder->CreateOr(X, *Y); 578 // (X & Y) != 0 ? X : X ^ Y --> X | Y 579 else if (!TrueWhenUnset && TrueVal == X && 580 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 581 V = Builder->CreateOr(X, *Y); 582 583 if (V) 584 return ReplaceInstUsesWith(SI, V); 585 } 586 } 587 588 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder)) 589 return ReplaceInstUsesWith(SI, V); 590 591 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) 592 return ReplaceInstUsesWith(SI, V); 593 594 return Changed ? &SI : nullptr; 595 } 596 597 598 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a 599 /// PHI node (but the two may be in different blocks). See if the true/false 600 /// values (V) are live in all of the predecessor blocks of the PHI. For 601 /// example, cases like this cannot be mapped: 602 /// 603 /// X = phi [ C1, BB1], [C2, BB2] 604 /// Y = add 605 /// Z = select X, Y, 0 606 /// 607 /// because Y is not live in BB1/BB2. 608 /// 609 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V, 610 const SelectInst &SI) { 611 // If the value is a non-instruction value like a constant or argument, it 612 // can always be mapped. 613 const Instruction *I = dyn_cast<Instruction>(V); 614 if (!I) return true; 615 616 // If V is a PHI node defined in the same block as the condition PHI, we can 617 // map the arguments. 618 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 619 620 if (const PHINode *VP = dyn_cast<PHINode>(I)) 621 if (VP->getParent() == CondPHI->getParent()) 622 return true; 623 624 // Otherwise, if the PHI and select are defined in the same block and if V is 625 // defined in a different block, then we can transform it. 626 if (SI.getParent() == CondPHI->getParent() && 627 I->getParent() != CondPHI->getParent()) 628 return true; 629 630 // Otherwise we have a 'hard' case and we can't tell without doing more 631 // detailed dominator based analysis, punt. 632 return false; 633 } 634 635 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form: 636 /// SPF2(SPF1(A, B), C) 637 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner, 638 SelectPatternFlavor SPF1, 639 Value *A, Value *B, 640 Instruction &Outer, 641 SelectPatternFlavor SPF2, Value *C) { 642 if (C == A || C == B) { 643 // MAX(MAX(A, B), B) -> MAX(A, B) 644 // MIN(MIN(a, b), a) -> MIN(a, b) 645 if (SPF1 == SPF2) 646 return ReplaceInstUsesWith(Outer, Inner); 647 648 // MAX(MIN(a, b), a) -> a 649 // MIN(MAX(a, b), a) -> a 650 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 651 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 652 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 653 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 654 return ReplaceInstUsesWith(Outer, C); 655 } 656 657 if (SPF1 == SPF2) { 658 if (ConstantInt *CB = dyn_cast<ConstantInt>(B)) { 659 if (ConstantInt *CC = dyn_cast<ConstantInt>(C)) { 660 APInt ACB = CB->getValue(); 661 APInt ACC = CC->getValue(); 662 663 // MIN(MIN(A, 23), 97) -> MIN(A, 23) 664 // MAX(MAX(A, 97), 23) -> MAX(A, 97) 665 if ((SPF1 == SPF_UMIN && ACB.ule(ACC)) || 666 (SPF1 == SPF_SMIN && ACB.sle(ACC)) || 667 (SPF1 == SPF_UMAX && ACB.uge(ACC)) || 668 (SPF1 == SPF_SMAX && ACB.sge(ACC))) 669 return ReplaceInstUsesWith(Outer, Inner); 670 671 // MIN(MIN(A, 97), 23) -> MIN(A, 23) 672 // MAX(MAX(A, 23), 97) -> MAX(A, 97) 673 if ((SPF1 == SPF_UMIN && ACB.ugt(ACC)) || 674 (SPF1 == SPF_SMIN && ACB.sgt(ACC)) || 675 (SPF1 == SPF_UMAX && ACB.ult(ACC)) || 676 (SPF1 == SPF_SMAX && ACB.slt(ACC))) { 677 Outer.replaceUsesOfWith(Inner, A); 678 return &Outer; 679 } 680 } 681 } 682 } 683 684 // ABS(ABS(X)) -> ABS(X) 685 // NABS(NABS(X)) -> NABS(X) 686 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) { 687 return ReplaceInstUsesWith(Outer, Inner); 688 } 689 690 // ABS(NABS(X)) -> ABS(X) 691 // NABS(ABS(X)) -> NABS(X) 692 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) || 693 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) { 694 SelectInst *SI = cast<SelectInst>(Inner); 695 Value *NewSI = Builder->CreateSelect( 696 SI->getCondition(), SI->getFalseValue(), SI->getTrueValue()); 697 return ReplaceInstUsesWith(Outer, NewSI); 698 } 699 700 auto IsFreeOrProfitableToInvert = 701 [&](Value *V, Value *&NotV, bool &ElidesXor) { 702 if (match(V, m_Not(m_Value(NotV)))) { 703 // If V has at most 2 uses then we can get rid of the xor operation 704 // entirely. 705 ElidesXor |= !V->hasNUsesOrMore(3); 706 return true; 707 } 708 709 if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) { 710 NotV = nullptr; 711 return true; 712 } 713 714 return false; 715 }; 716 717 Value *NotA, *NotB, *NotC; 718 bool ElidesXor = false; 719 720 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C) 721 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C) 722 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C) 723 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C) 724 // 725 // This transform is performance neutral if we can elide at least one xor from 726 // the set of three operands, since we'll be tacking on an xor at the very 727 // end. 728 if (IsFreeOrProfitableToInvert(A, NotA, ElidesXor) && 729 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) && 730 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) { 731 if (!NotA) 732 NotA = Builder->CreateNot(A); 733 if (!NotB) 734 NotB = Builder->CreateNot(B); 735 if (!NotC) 736 NotC = Builder->CreateNot(C); 737 738 Value *NewInner = generateMinMaxSelectPattern( 739 Builder, getInverseMinMaxSelectPattern(SPF1), NotA, NotB); 740 Value *NewOuter = Builder->CreateNot(generateMinMaxSelectPattern( 741 Builder, getInverseMinMaxSelectPattern(SPF2), NewInner, NotC)); 742 return ReplaceInstUsesWith(Outer, NewOuter); 743 } 744 745 return nullptr; 746 } 747 748 /// foldSelectICmpAnd - If one of the constants is zero (we know they can't 749 /// both be) and we have an icmp instruction with zero, and we have an 'and' 750 /// with the non-constant value and a power of two we can turn the select 751 /// into a shift on the result of the 'and'. 752 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal, 753 ConstantInt *FalseVal, 754 InstCombiner::BuilderTy *Builder) { 755 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 756 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy()) 757 return nullptr; 758 759 if (!match(IC->getOperand(1), m_Zero())) 760 return nullptr; 761 762 ConstantInt *AndRHS; 763 Value *LHS = IC->getOperand(0); 764 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS)))) 765 return nullptr; 766 767 // If both select arms are non-zero see if we have a select of the form 768 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic 769 // for 'x ? 2^n : 0' and fix the thing up at the end. 770 ConstantInt *Offset = nullptr; 771 if (!TrueVal->isZero() && !FalseVal->isZero()) { 772 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2()) 773 Offset = FalseVal; 774 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2()) 775 Offset = TrueVal; 776 else 777 return nullptr; 778 779 // Adjust TrueVal and FalseVal to the offset. 780 TrueVal = ConstantInt::get(Builder->getContext(), 781 TrueVal->getValue() - Offset->getValue()); 782 FalseVal = ConstantInt::get(Builder->getContext(), 783 FalseVal->getValue() - Offset->getValue()); 784 } 785 786 // Make sure the mask in the 'and' and one of the select arms is a power of 2. 787 if (!AndRHS->getValue().isPowerOf2() || 788 (!TrueVal->getValue().isPowerOf2() && 789 !FalseVal->getValue().isPowerOf2())) 790 return nullptr; 791 792 // Determine which shift is needed to transform result of the 'and' into the 793 // desired result. 794 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal; 795 unsigned ValZeros = ValC->getValue().logBase2(); 796 unsigned AndZeros = AndRHS->getValue().logBase2(); 797 798 // If types don't match we can still convert the select by introducing a zext 799 // or a trunc of the 'and'. The trunc case requires that all of the truncated 800 // bits are zero, we can figure that out by looking at the 'and' mask. 801 if (AndZeros >= ValC->getBitWidth()) 802 return nullptr; 803 804 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType()); 805 if (ValZeros > AndZeros) 806 V = Builder->CreateShl(V, ValZeros - AndZeros); 807 else if (ValZeros < AndZeros) 808 V = Builder->CreateLShr(V, AndZeros - ValZeros); 809 810 // Okay, now we know that everything is set up, we just don't know whether we 811 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 812 bool ShouldNotVal = !TrueVal->isZero(); 813 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; 814 if (ShouldNotVal) 815 V = Builder->CreateXor(V, ValC); 816 817 // Apply an offset if needed. 818 if (Offset) 819 V = Builder->CreateAdd(V, Offset); 820 return V; 821 } 822 823 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { 824 Value *CondVal = SI.getCondition(); 825 Value *TrueVal = SI.getTrueValue(); 826 Value *FalseVal = SI.getFalseValue(); 827 828 if (Value *V = 829 SimplifySelectInst(CondVal, TrueVal, FalseVal, DL, TLI, DT, AC)) 830 return ReplaceInstUsesWith(SI, V); 831 832 if (SI.getType()->isIntegerTy(1)) { 833 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { 834 if (C->getZExtValue()) { 835 // Change: A = select B, true, C --> A = or B, C 836 return BinaryOperator::CreateOr(CondVal, FalseVal); 837 } 838 // Change: A = select B, false, C --> A = and !B, C 839 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 840 return BinaryOperator::CreateAnd(NotCond, FalseVal); 841 } 842 if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) { 843 if (!C->getZExtValue()) { 844 // Change: A = select B, C, false --> A = and B, C 845 return BinaryOperator::CreateAnd(CondVal, TrueVal); 846 } 847 // Change: A = select B, C, true --> A = or !B, C 848 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 849 return BinaryOperator::CreateOr(NotCond, TrueVal); 850 } 851 852 // select a, b, a -> a&b 853 // select a, a, b -> a|b 854 if (CondVal == TrueVal) 855 return BinaryOperator::CreateOr(CondVal, FalseVal); 856 if (CondVal == FalseVal) 857 return BinaryOperator::CreateAnd(CondVal, TrueVal); 858 859 // select a, ~a, b -> (~a)&b 860 // select a, b, ~a -> (~a)|b 861 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 862 return BinaryOperator::CreateAnd(TrueVal, FalseVal); 863 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 864 return BinaryOperator::CreateOr(TrueVal, FalseVal); 865 } 866 867 // Selecting between two integer constants? 868 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) 869 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { 870 // select C, 1, 0 -> zext C to int 871 if (FalseValC->isZero() && TrueValC->getValue() == 1) 872 return new ZExtInst(CondVal, SI.getType()); 873 874 // select C, -1, 0 -> sext C to int 875 if (FalseValC->isZero() && TrueValC->isAllOnesValue()) 876 return new SExtInst(CondVal, SI.getType()); 877 878 // select C, 0, 1 -> zext !C to int 879 if (TrueValC->isZero() && FalseValC->getValue() == 1) { 880 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 881 return new ZExtInst(NotCond, SI.getType()); 882 } 883 884 // select C, 0, -1 -> sext !C to int 885 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { 886 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 887 return new SExtInst(NotCond, SI.getType()); 888 } 889 890 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder)) 891 return ReplaceInstUsesWith(SI, V); 892 } 893 894 // See if we are selecting two values based on a comparison of the two values. 895 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 896 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) { 897 // Transform (X == Y) ? X : Y -> Y 898 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 899 // This is not safe in general for floating point: 900 // consider X== -0, Y== +0. 901 // It becomes safe if either operand is a nonzero constant. 902 ConstantFP *CFPt, *CFPf; 903 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 904 !CFPt->getValueAPF().isZero()) || 905 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 906 !CFPf->getValueAPF().isZero())) 907 return ReplaceInstUsesWith(SI, FalseVal); 908 } 909 // Transform (X une Y) ? X : Y -> X 910 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 911 // This is not safe in general for floating point: 912 // consider X== -0, Y== +0. 913 // It becomes safe if either operand is a nonzero constant. 914 ConstantFP *CFPt, *CFPf; 915 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 916 !CFPt->getValueAPF().isZero()) || 917 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 918 !CFPf->getValueAPF().isZero())) 919 return ReplaceInstUsesWith(SI, TrueVal); 920 } 921 922 // Canonicalize to use ordered comparisons by swapping the select 923 // operands. 924 // 925 // e.g. 926 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X 927 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) { 928 FCmpInst::Predicate InvPred = FCI->getInversePredicate(); 929 Value *NewCond = Builder->CreateFCmp(InvPred, TrueVal, FalseVal, 930 FCI->getName() + ".inv"); 931 932 return SelectInst::Create(NewCond, FalseVal, TrueVal, 933 SI.getName() + ".p"); 934 } 935 936 // NOTE: if we wanted to, this is where to detect MIN/MAX 937 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){ 938 // Transform (X == Y) ? Y : X -> X 939 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 940 // This is not safe in general for floating point: 941 // consider X== -0, Y== +0. 942 // It becomes safe if either operand is a nonzero constant. 943 ConstantFP *CFPt, *CFPf; 944 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 945 !CFPt->getValueAPF().isZero()) || 946 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 947 !CFPf->getValueAPF().isZero())) 948 return ReplaceInstUsesWith(SI, FalseVal); 949 } 950 // Transform (X une Y) ? Y : X -> Y 951 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 952 // This is not safe in general for floating point: 953 // consider X== -0, Y== +0. 954 // It becomes safe if either operand is a nonzero constant. 955 ConstantFP *CFPt, *CFPf; 956 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 957 !CFPt->getValueAPF().isZero()) || 958 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 959 !CFPf->getValueAPF().isZero())) 960 return ReplaceInstUsesWith(SI, TrueVal); 961 } 962 963 // Canonicalize to use ordered comparisons by swapping the select 964 // operands. 965 // 966 // e.g. 967 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y 968 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) { 969 FCmpInst::Predicate InvPred = FCI->getInversePredicate(); 970 Value *NewCond = Builder->CreateFCmp(InvPred, FalseVal, TrueVal, 971 FCI->getName() + ".inv"); 972 973 return SelectInst::Create(NewCond, FalseVal, TrueVal, 974 SI.getName() + ".p"); 975 } 976 977 // NOTE: if we wanted to, this is where to detect MIN/MAX 978 } 979 // NOTE: if we wanted to, this is where to detect ABS 980 } 981 982 // See if we are selecting two values based on a comparison of the two values. 983 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 984 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI)) 985 return Result; 986 987 if (Instruction *TI = dyn_cast<Instruction>(TrueVal)) 988 if (Instruction *FI = dyn_cast<Instruction>(FalseVal)) 989 if (TI->hasOneUse() && FI->hasOneUse()) { 990 Instruction *AddOp = nullptr, *SubOp = nullptr; 991 992 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 993 if (TI->getOpcode() == FI->getOpcode()) 994 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI)) 995 return IV; 996 997 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is 998 // even legal for FP. 999 if ((TI->getOpcode() == Instruction::Sub && 1000 FI->getOpcode() == Instruction::Add) || 1001 (TI->getOpcode() == Instruction::FSub && 1002 FI->getOpcode() == Instruction::FAdd)) { 1003 AddOp = FI; SubOp = TI; 1004 } else if ((FI->getOpcode() == Instruction::Sub && 1005 TI->getOpcode() == Instruction::Add) || 1006 (FI->getOpcode() == Instruction::FSub && 1007 TI->getOpcode() == Instruction::FAdd)) { 1008 AddOp = TI; SubOp = FI; 1009 } 1010 1011 if (AddOp) { 1012 Value *OtherAddOp = nullptr; 1013 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 1014 OtherAddOp = AddOp->getOperand(1); 1015 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 1016 OtherAddOp = AddOp->getOperand(0); 1017 } 1018 1019 if (OtherAddOp) { 1020 // So at this point we know we have (Y -> OtherAddOp): 1021 // select C, (add X, Y), (sub X, Z) 1022 Value *NegVal; // Compute -Z 1023 if (SI.getType()->isFPOrFPVectorTy()) { 1024 NegVal = Builder->CreateFNeg(SubOp->getOperand(1)); 1025 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { 1026 FastMathFlags Flags = AddOp->getFastMathFlags(); 1027 Flags &= SubOp->getFastMathFlags(); 1028 NegInst->setFastMathFlags(Flags); 1029 } 1030 } else { 1031 NegVal = Builder->CreateNeg(SubOp->getOperand(1)); 1032 } 1033 1034 Value *NewTrueOp = OtherAddOp; 1035 Value *NewFalseOp = NegVal; 1036 if (AddOp != TI) 1037 std::swap(NewTrueOp, NewFalseOp); 1038 Value *NewSel = 1039 Builder->CreateSelect(CondVal, NewTrueOp, 1040 NewFalseOp, SI.getName() + ".p"); 1041 1042 if (SI.getType()->isFPOrFPVectorTy()) { 1043 Instruction *RI = 1044 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 1045 1046 FastMathFlags Flags = AddOp->getFastMathFlags(); 1047 Flags &= SubOp->getFastMathFlags(); 1048 RI->setFastMathFlags(Flags); 1049 return RI; 1050 } else 1051 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 1052 } 1053 } 1054 } 1055 1056 // See if we can fold the select into one of our operands. 1057 if (SI.getType()->isIntOrIntVectorTy()) { 1058 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) 1059 return FoldI; 1060 1061 Value *LHS, *RHS, *LHS2, *RHS2; 1062 Instruction::CastOps CastOp; 1063 SelectPatternFlavor SPF = matchSelectPattern(&SI, LHS, RHS, &CastOp); 1064 1065 if (SPF) { 1066 // Canonicalize so that type casts are outside select patterns. 1067 if (LHS->getType()->getPrimitiveSizeInBits() != 1068 SI.getType()->getPrimitiveSizeInBits()) { 1069 CmpInst::Predicate Pred = getICmpPredicateForMinMax(SPF); 1070 Value *Cmp = Builder->CreateICmp(Pred, LHS, RHS); 1071 Value *NewSI = Builder->CreateCast(CastOp, 1072 Builder->CreateSelect(Cmp, LHS, RHS), 1073 SI.getType()); 1074 return ReplaceInstUsesWith(SI, NewSI); 1075 } 1076 1077 // MAX(MAX(a, b), a) -> MAX(a, b) 1078 // MIN(MIN(a, b), a) -> MIN(a, b) 1079 // MAX(MIN(a, b), a) -> a 1080 // MIN(MAX(a, b), a) -> a 1081 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2)) 1082 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2, 1083 SI, SPF, RHS)) 1084 return R; 1085 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2)) 1086 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2, 1087 SI, SPF, LHS)) 1088 return R; 1089 } 1090 1091 // MAX(~a, ~b) -> ~MIN(a, b) 1092 if (SPF == SPF_SMAX || SPF == SPF_UMAX) { 1093 if (IsFreeToInvert(LHS, LHS->hasNUses(2)) && 1094 IsFreeToInvert(RHS, RHS->hasNUses(2))) { 1095 1096 // This transform adds a xor operation and that extra cost needs to be 1097 // justified. We look for simplifications that will result from 1098 // applying this rule: 1099 1100 bool Profitable = 1101 (LHS->hasNUses(2) && match(LHS, m_Not(m_Value()))) || 1102 (RHS->hasNUses(2) && match(RHS, m_Not(m_Value()))) || 1103 (SI.hasOneUse() && match(*SI.user_begin(), m_Not(m_Value()))); 1104 1105 if (Profitable) { 1106 Value *NewLHS = Builder->CreateNot(LHS); 1107 Value *NewRHS = Builder->CreateNot(RHS); 1108 Value *NewCmp = SPF == SPF_SMAX 1109 ? Builder->CreateICmpSLT(NewLHS, NewRHS) 1110 : Builder->CreateICmpULT(NewLHS, NewRHS); 1111 Value *NewSI = 1112 Builder->CreateNot(Builder->CreateSelect(NewCmp, NewLHS, NewRHS)); 1113 return ReplaceInstUsesWith(SI, NewSI); 1114 } 1115 } 1116 } 1117 1118 // TODO. 1119 // ABS(-X) -> ABS(X) 1120 } 1121 1122 // See if we can fold the select into a phi node if the condition is a select. 1123 if (isa<PHINode>(SI.getCondition())) 1124 // The true/false values have to be live in the PHI predecessor's blocks. 1125 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 1126 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 1127 if (Instruction *NV = FoldOpIntoPhi(SI)) 1128 return NV; 1129 1130 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 1131 if (TrueSI->getCondition()->getType() == CondVal->getType()) { 1132 // select(C, select(C, a, b), c) -> select(C, a, c) 1133 if (TrueSI->getCondition() == CondVal) { 1134 if (SI.getTrueValue() == TrueSI->getTrueValue()) 1135 return nullptr; 1136 SI.setOperand(1, TrueSI->getTrueValue()); 1137 return &SI; 1138 } 1139 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) 1140 // We choose this as normal form to enable folding on the And and shortening 1141 // paths for the values (this helps GetUnderlyingObjects() for example). 1142 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { 1143 Value *And = Builder->CreateAnd(CondVal, TrueSI->getCondition()); 1144 SI.setOperand(0, And); 1145 SI.setOperand(1, TrueSI->getTrueValue()); 1146 return &SI; 1147 } 1148 } 1149 } 1150 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 1151 if (FalseSI->getCondition()->getType() == CondVal->getType()) { 1152 // select(C, a, select(C, b, c)) -> select(C, a, c) 1153 if (FalseSI->getCondition() == CondVal) { 1154 if (SI.getFalseValue() == FalseSI->getFalseValue()) 1155 return nullptr; 1156 SI.setOperand(2, FalseSI->getFalseValue()); 1157 return &SI; 1158 } 1159 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) 1160 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { 1161 Value *Or = Builder->CreateOr(CondVal, FalseSI->getCondition()); 1162 SI.setOperand(0, Or); 1163 SI.setOperand(2, FalseSI->getFalseValue()); 1164 return &SI; 1165 } 1166 } 1167 } 1168 1169 if (BinaryOperator::isNot(CondVal)) { 1170 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal)); 1171 SI.setOperand(1, FalseVal); 1172 SI.setOperand(2, TrueVal); 1173 return &SI; 1174 } 1175 1176 if (VectorType* VecTy = dyn_cast<VectorType>(SI.getType())) { 1177 unsigned VWidth = VecTy->getNumElements(); 1178 APInt UndefElts(VWidth, 0); 1179 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 1180 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) { 1181 if (V != &SI) 1182 return ReplaceInstUsesWith(SI, V); 1183 return &SI; 1184 } 1185 1186 if (isa<ConstantAggregateZero>(CondVal)) { 1187 return ReplaceInstUsesWith(SI, FalseVal); 1188 } 1189 } 1190 1191 return nullptr; 1192 } 1193