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 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType()) 131 return 0; 132 // The select condition may be a vector. We may only change the operand 133 // type if the vector width remains the same (and matches the condition). 134 Type *CondTy = SI.getCondition()->getType(); 135 if (CondTy->isVectorTy() && CondTy->getVectorNumElements() != 136 FI->getOperand(0)->getType()->getVectorNumElements()) 137 return 0; 138 } else { 139 return 0; // unknown unary op. 140 } 141 142 // Fold this by inserting a select from the input values. 143 Value *NewSI = Builder->CreateSelect(SI.getCondition(), TI->getOperand(0), 144 FI->getOperand(0), SI.getName()+".v"); 145 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 146 TI->getType()); 147 } 148 149 // Only handle binary operators here. 150 if (!isa<BinaryOperator>(TI)) 151 return 0; 152 153 // Figure out if the operations have any operands in common. 154 Value *MatchOp, *OtherOpT, *OtherOpF; 155 bool MatchIsOpZero; 156 if (TI->getOperand(0) == FI->getOperand(0)) { 157 MatchOp = TI->getOperand(0); 158 OtherOpT = TI->getOperand(1); 159 OtherOpF = FI->getOperand(1); 160 MatchIsOpZero = true; 161 } else if (TI->getOperand(1) == FI->getOperand(1)) { 162 MatchOp = TI->getOperand(1); 163 OtherOpT = TI->getOperand(0); 164 OtherOpF = FI->getOperand(0); 165 MatchIsOpZero = false; 166 } else if (!TI->isCommutative()) { 167 return 0; 168 } else if (TI->getOperand(0) == FI->getOperand(1)) { 169 MatchOp = TI->getOperand(0); 170 OtherOpT = TI->getOperand(1); 171 OtherOpF = FI->getOperand(0); 172 MatchIsOpZero = true; 173 } else if (TI->getOperand(1) == FI->getOperand(0)) { 174 MatchOp = TI->getOperand(1); 175 OtherOpT = TI->getOperand(0); 176 OtherOpF = FI->getOperand(1); 177 MatchIsOpZero = true; 178 } else { 179 return 0; 180 } 181 182 // If we reach here, they do have operations in common. 183 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, 184 OtherOpF, SI.getName()+".v"); 185 186 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) { 187 if (MatchIsOpZero) 188 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI); 189 else 190 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp); 191 } 192 llvm_unreachable("Shouldn't get here"); 193 } 194 195 static bool isSelect01(Constant *C1, Constant *C2) { 196 ConstantInt *C1I = dyn_cast<ConstantInt>(C1); 197 if (!C1I) 198 return false; 199 ConstantInt *C2I = dyn_cast<ConstantInt>(C2); 200 if (!C2I) 201 return false; 202 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero. 203 return false; 204 return C1I->isOne() || C1I->isAllOnesValue() || 205 C2I->isOne() || C2I->isAllOnesValue(); 206 } 207 208 /// FoldSelectIntoOp - Try fold the select into one of the operands to 209 /// facilitate further optimization. 210 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal, 211 Value *FalseVal) { 212 // See the comment above GetSelectFoldableOperands for a description of the 213 // transformation we are doing here. 214 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) { 215 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 && 216 !isa<Constant>(FalseVal)) { 217 if (unsigned SFO = GetSelectFoldableOperands(TVI)) { 218 unsigned OpToFold = 0; 219 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 220 OpToFold = 1; 221 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 222 OpToFold = 2; 223 } 224 225 if (OpToFold) { 226 Constant *C = GetSelectFoldableConstant(TVI); 227 Value *OOp = TVI->getOperand(2-OpToFold); 228 // Avoid creating select between 2 constants unless it's selecting 229 // between 0, 1 and -1. 230 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 231 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C); 232 NewSel->takeName(TVI); 233 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI); 234 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(), 235 FalseVal, NewSel); 236 if (isa<PossiblyExactOperator>(BO)) 237 BO->setIsExact(TVI_BO->isExact()); 238 if (isa<OverflowingBinaryOperator>(BO)) { 239 BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap()); 240 BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap()); 241 } 242 return BO; 243 } 244 } 245 } 246 } 247 } 248 249 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) { 250 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 && 251 !isa<Constant>(TrueVal)) { 252 if (unsigned SFO = GetSelectFoldableOperands(FVI)) { 253 unsigned OpToFold = 0; 254 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 255 OpToFold = 1; 256 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 257 OpToFold = 2; 258 } 259 260 if (OpToFold) { 261 Constant *C = GetSelectFoldableConstant(FVI); 262 Value *OOp = FVI->getOperand(2-OpToFold); 263 // Avoid creating select between 2 constants unless it's selecting 264 // between 0, 1 and -1. 265 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 266 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp); 267 NewSel->takeName(FVI); 268 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI); 269 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(), 270 TrueVal, NewSel); 271 if (isa<PossiblyExactOperator>(BO)) 272 BO->setIsExact(FVI_BO->isExact()); 273 if (isa<OverflowingBinaryOperator>(BO)) { 274 BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap()); 275 BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap()); 276 } 277 return BO; 278 } 279 } 280 } 281 } 282 } 283 284 return 0; 285 } 286 287 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is 288 /// replaced with RepOp. 289 static Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, 290 const DataLayout *TD, 291 const TargetLibraryInfo *TLI) { 292 // Trivial replacement. 293 if (V == Op) 294 return RepOp; 295 296 Instruction *I = dyn_cast<Instruction>(V); 297 if (!I) 298 return 0; 299 300 // If this is a binary operator, try to simplify it with the replaced op. 301 if (BinaryOperator *B = dyn_cast<BinaryOperator>(I)) { 302 if (B->getOperand(0) == Op) 303 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), TD, TLI); 304 if (B->getOperand(1) == Op) 305 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, TD, TLI); 306 } 307 308 // Same for CmpInsts. 309 if (CmpInst *C = dyn_cast<CmpInst>(I)) { 310 if (C->getOperand(0) == Op) 311 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), TD, 312 TLI); 313 if (C->getOperand(1) == Op) 314 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, TD, 315 TLI); 316 } 317 318 // TODO: We could hand off more cases to instsimplify here. 319 320 // If all operands are constant after substituting Op for RepOp then we can 321 // constant fold the instruction. 322 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { 323 // Build a list of all constant operands. 324 SmallVector<Constant*, 8> ConstOps; 325 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 326 if (I->getOperand(i) == Op) 327 ConstOps.push_back(CRepOp); 328 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) 329 ConstOps.push_back(COp); 330 else 331 break; 332 } 333 334 // All operands were constants, fold it. 335 if (ConstOps.size() == I->getNumOperands()) { 336 if (CmpInst *C = dyn_cast<CmpInst>(I)) 337 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], 338 ConstOps[1], TD, TLI); 339 340 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 341 if (!LI->isVolatile()) 342 return ConstantFoldLoadFromConstPtr(ConstOps[0], TD); 343 344 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 345 ConstOps, TD, TLI); 346 } 347 } 348 349 return 0; 350 } 351 352 /// visitSelectInstWithICmp - Visit a SelectInst that has an 353 /// ICmpInst as its first operand. 354 /// 355 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, 356 ICmpInst *ICI) { 357 bool Changed = false; 358 ICmpInst::Predicate Pred = ICI->getPredicate(); 359 Value *CmpLHS = ICI->getOperand(0); 360 Value *CmpRHS = ICI->getOperand(1); 361 Value *TrueVal = SI.getTrueValue(); 362 Value *FalseVal = SI.getFalseValue(); 363 364 // Check cases where the comparison is with a constant that 365 // can be adjusted to fit the min/max idiom. We may move or edit ICI 366 // here, so make sure the select is the only user. 367 if (ICI->hasOneUse()) 368 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) { 369 // X < MIN ? T : F --> F 370 if ((Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT) 371 && CI->isMinValue(Pred == ICmpInst::ICMP_SLT)) 372 return ReplaceInstUsesWith(SI, FalseVal); 373 // X > MAX ? T : F --> F 374 else if ((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT) 375 && CI->isMaxValue(Pred == ICmpInst::ICMP_SGT)) 376 return ReplaceInstUsesWith(SI, FalseVal); 377 switch (Pred) { 378 default: break; 379 case ICmpInst::ICMP_ULT: 380 case ICmpInst::ICMP_SLT: 381 case ICmpInst::ICMP_UGT: 382 case ICmpInst::ICMP_SGT: { 383 // These transformations only work for selects over integers. 384 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType()); 385 if (!SelectTy) 386 break; 387 388 Constant *AdjustedRHS; 389 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 390 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1); 391 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 392 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1); 393 394 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 395 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 396 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 397 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) 398 ; // Nothing to do here. Values match without any sign/zero extension. 399 400 // Types do not match. Instead of calculating this with mixed types 401 // promote all to the larger type. This enables scalar evolution to 402 // analyze this expression. 403 else if (CmpRHS->getType()->getScalarSizeInBits() 404 < SelectTy->getBitWidth()) { 405 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy); 406 407 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 408 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 409 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 410 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 411 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && 412 sextRHS == FalseVal) { 413 CmpLHS = TrueVal; 414 AdjustedRHS = sextRHS; 415 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 416 sextRHS == TrueVal) { 417 CmpLHS = FalseVal; 418 AdjustedRHS = sextRHS; 419 } else if (ICI->isUnsigned()) { 420 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy); 421 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 422 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 423 // zext + signed compare cannot be changed: 424 // 0xff <s 0x00, but 0x00ff >s 0x0000 425 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && 426 zextRHS == FalseVal) { 427 CmpLHS = TrueVal; 428 AdjustedRHS = zextRHS; 429 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 430 zextRHS == TrueVal) { 431 CmpLHS = FalseVal; 432 AdjustedRHS = zextRHS; 433 } else 434 break; 435 } else 436 break; 437 } else 438 break; 439 440 Pred = ICmpInst::getSwappedPredicate(Pred); 441 CmpRHS = AdjustedRHS; 442 std::swap(FalseVal, TrueVal); 443 ICI->setPredicate(Pred); 444 ICI->setOperand(0, CmpLHS); 445 ICI->setOperand(1, CmpRHS); 446 SI.setOperand(1, TrueVal); 447 SI.setOperand(2, FalseVal); 448 449 // Move ICI instruction right before the select instruction. Otherwise 450 // the sext/zext value may be defined after the ICI instruction uses it. 451 ICI->moveBefore(&SI); 452 453 Changed = true; 454 break; 455 } 456 } 457 } 458 459 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 460 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1 461 // FIXME: Type and constness constraints could be lifted, but we have to 462 // watch code size carefully. We should consider xor instead of 463 // sub/add when we decide to do that. 464 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) { 465 if (TrueVal->getType() == Ty) { 466 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) { 467 ConstantInt *C1 = NULL, *C2 = NULL; 468 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) { 469 C1 = dyn_cast<ConstantInt>(TrueVal); 470 C2 = dyn_cast<ConstantInt>(FalseVal); 471 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) { 472 C1 = dyn_cast<ConstantInt>(FalseVal); 473 C2 = dyn_cast<ConstantInt>(TrueVal); 474 } 475 if (C1 && C2) { 476 // This shift results in either -1 or 0. 477 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1); 478 479 // Check if we can express the operation with a single or. 480 if (C2->isAllOnesValue()) 481 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1)); 482 483 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue()); 484 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1)); 485 } 486 } 487 } 488 } 489 490 // If we have an equality comparison then we know the value in one of the 491 // arms of the select. See if substituting this value into the arm and 492 // simplifying the result yields the same value as the other arm. 493 if (Pred == ICmpInst::ICMP_EQ) { 494 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 495 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 496 return ReplaceInstUsesWith(SI, FalseVal); 497 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 498 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 499 return ReplaceInstUsesWith(SI, FalseVal); 500 } else if (Pred == ICmpInst::ICMP_NE) { 501 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 502 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 503 return ReplaceInstUsesWith(SI, TrueVal); 504 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 505 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 506 return ReplaceInstUsesWith(SI, TrueVal); 507 } 508 509 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 510 511 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 512 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 513 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 514 SI.setOperand(1, CmpRHS); 515 Changed = true; 516 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 517 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 518 SI.setOperand(2, CmpRHS); 519 Changed = true; 520 } 521 } 522 523 return Changed ? &SI : 0; 524 } 525 526 527 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a 528 /// PHI node (but the two may be in different blocks). See if the true/false 529 /// values (V) are live in all of the predecessor blocks of the PHI. For 530 /// example, cases like this cannot be mapped: 531 /// 532 /// X = phi [ C1, BB1], [C2, BB2] 533 /// Y = add 534 /// Z = select X, Y, 0 535 /// 536 /// because Y is not live in BB1/BB2. 537 /// 538 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V, 539 const SelectInst &SI) { 540 // If the value is a non-instruction value like a constant or argument, it 541 // can always be mapped. 542 const Instruction *I = dyn_cast<Instruction>(V); 543 if (I == 0) return true; 544 545 // If V is a PHI node defined in the same block as the condition PHI, we can 546 // map the arguments. 547 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 548 549 if (const PHINode *VP = dyn_cast<PHINode>(I)) 550 if (VP->getParent() == CondPHI->getParent()) 551 return true; 552 553 // Otherwise, if the PHI and select are defined in the same block and if V is 554 // defined in a different block, then we can transform it. 555 if (SI.getParent() == CondPHI->getParent() && 556 I->getParent() != CondPHI->getParent()) 557 return true; 558 559 // Otherwise we have a 'hard' case and we can't tell without doing more 560 // detailed dominator based analysis, punt. 561 return false; 562 } 563 564 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form: 565 /// SPF2(SPF1(A, B), C) 566 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner, 567 SelectPatternFlavor SPF1, 568 Value *A, Value *B, 569 Instruction &Outer, 570 SelectPatternFlavor SPF2, Value *C) { 571 if (C == A || C == B) { 572 // MAX(MAX(A, B), B) -> MAX(A, B) 573 // MIN(MIN(a, b), a) -> MIN(a, b) 574 if (SPF1 == SPF2) 575 return ReplaceInstUsesWith(Outer, Inner); 576 577 // MAX(MIN(a, b), a) -> a 578 // MIN(MAX(a, b), a) -> a 579 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 580 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 581 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 582 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 583 return ReplaceInstUsesWith(Outer, C); 584 } 585 586 // TODO: MIN(MIN(A, 23), 97) 587 return 0; 588 } 589 590 591 /// foldSelectICmpAnd - If one of the constants is zero (we know they can't 592 /// both be) and we have an icmp instruction with zero, and we have an 'and' 593 /// with the non-constant value and a power of two we can turn the select 594 /// into a shift on the result of the 'and'. 595 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal, 596 ConstantInt *FalseVal, 597 InstCombiner::BuilderTy *Builder) { 598 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 599 if (!IC || !IC->isEquality()) 600 return 0; 601 602 if (!match(IC->getOperand(1), m_Zero())) 603 return 0; 604 605 ConstantInt *AndRHS; 606 Value *LHS = IC->getOperand(0); 607 if (LHS->getType() != SI.getType() || 608 !match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS)))) 609 return 0; 610 611 // If both select arms are non-zero see if we have a select of the form 612 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic 613 // for 'x ? 2^n : 0' and fix the thing up at the end. 614 ConstantInt *Offset = 0; 615 if (!TrueVal->isZero() && !FalseVal->isZero()) { 616 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2()) 617 Offset = FalseVal; 618 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2()) 619 Offset = TrueVal; 620 else 621 return 0; 622 623 // Adjust TrueVal and FalseVal to the offset. 624 TrueVal = ConstantInt::get(Builder->getContext(), 625 TrueVal->getValue() - Offset->getValue()); 626 FalseVal = ConstantInt::get(Builder->getContext(), 627 FalseVal->getValue() - Offset->getValue()); 628 } 629 630 // Make sure the mask in the 'and' and one of the select arms is a power of 2. 631 if (!AndRHS->getValue().isPowerOf2() || 632 (!TrueVal->getValue().isPowerOf2() && 633 !FalseVal->getValue().isPowerOf2())) 634 return 0; 635 636 // Determine which shift is needed to transform result of the 'and' into the 637 // desired result. 638 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal; 639 unsigned ValZeros = ValC->getValue().logBase2(); 640 unsigned AndZeros = AndRHS->getValue().logBase2(); 641 642 Value *V = LHS; 643 if (ValZeros > AndZeros) 644 V = Builder->CreateShl(V, ValZeros - AndZeros); 645 else if (ValZeros < AndZeros) 646 V = Builder->CreateLShr(V, AndZeros - ValZeros); 647 648 // Okay, now we know that everything is set up, we just don't know whether we 649 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 650 bool ShouldNotVal = !TrueVal->isZero(); 651 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; 652 if (ShouldNotVal) 653 V = Builder->CreateXor(V, ValC); 654 655 // Apply an offset if needed. 656 if (Offset) 657 V = Builder->CreateAdd(V, Offset); 658 return V; 659 } 660 661 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { 662 Value *CondVal = SI.getCondition(); 663 Value *TrueVal = SI.getTrueValue(); 664 Value *FalseVal = SI.getFalseValue(); 665 666 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, TD)) 667 return ReplaceInstUsesWith(SI, V); 668 669 if (SI.getType()->isIntegerTy(1)) { 670 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { 671 if (C->getZExtValue()) { 672 // Change: A = select B, true, C --> A = or B, C 673 return BinaryOperator::CreateOr(CondVal, FalseVal); 674 } 675 // Change: A = select B, false, C --> A = and !B, C 676 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 677 return BinaryOperator::CreateAnd(NotCond, FalseVal); 678 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) { 679 if (C->getZExtValue() == false) { 680 // Change: A = select B, C, false --> A = and B, C 681 return BinaryOperator::CreateAnd(CondVal, TrueVal); 682 } 683 // Change: A = select B, C, true --> A = or !B, C 684 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 685 return BinaryOperator::CreateOr(NotCond, TrueVal); 686 } 687 688 // select a, b, a -> a&b 689 // select a, a, b -> a|b 690 if (CondVal == TrueVal) 691 return BinaryOperator::CreateOr(CondVal, FalseVal); 692 else if (CondVal == FalseVal) 693 return BinaryOperator::CreateAnd(CondVal, TrueVal); 694 695 // select a, ~a, b -> (~a)&b 696 // select a, b, ~a -> (~a)|b 697 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 698 return BinaryOperator::CreateAnd(TrueVal, FalseVal); 699 else if (match(FalseVal, m_Not(m_Specific(CondVal)))) 700 return BinaryOperator::CreateOr(TrueVal, FalseVal); 701 } 702 703 // Selecting between two integer constants? 704 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) 705 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { 706 // select C, 1, 0 -> zext C to int 707 if (FalseValC->isZero() && TrueValC->getValue() == 1) 708 return new ZExtInst(CondVal, SI.getType()); 709 710 // select C, -1, 0 -> sext C to int 711 if (FalseValC->isZero() && TrueValC->isAllOnesValue()) 712 return new SExtInst(CondVal, SI.getType()); 713 714 // select C, 0, 1 -> zext !C to int 715 if (TrueValC->isZero() && FalseValC->getValue() == 1) { 716 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 717 return new ZExtInst(NotCond, SI.getType()); 718 } 719 720 // select C, 0, -1 -> sext !C to int 721 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { 722 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 723 return new SExtInst(NotCond, SI.getType()); 724 } 725 726 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder)) 727 return ReplaceInstUsesWith(SI, V); 728 } 729 730 // See if we are selecting two values based on a comparison of the two values. 731 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 732 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) { 733 // Transform (X == Y) ? X : Y -> Y 734 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 735 // This is not safe in general for floating point: 736 // consider X== -0, Y== +0. 737 // It becomes safe if either operand is a nonzero constant. 738 ConstantFP *CFPt, *CFPf; 739 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 740 !CFPt->getValueAPF().isZero()) || 741 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 742 !CFPf->getValueAPF().isZero())) 743 return ReplaceInstUsesWith(SI, FalseVal); 744 } 745 // Transform (X une Y) ? X : Y -> X 746 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 747 // This is not safe in general for floating point: 748 // consider X== -0, Y== +0. 749 // It becomes safe if either operand is a nonzero constant. 750 ConstantFP *CFPt, *CFPf; 751 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 752 !CFPt->getValueAPF().isZero()) || 753 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 754 !CFPf->getValueAPF().isZero())) 755 return ReplaceInstUsesWith(SI, TrueVal); 756 } 757 // NOTE: if we wanted to, this is where to detect MIN/MAX 758 759 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){ 760 // Transform (X == Y) ? Y : X -> X 761 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 762 // This is not safe in general for floating point: 763 // consider X== -0, Y== +0. 764 // It becomes safe if either operand is a nonzero constant. 765 ConstantFP *CFPt, *CFPf; 766 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 767 !CFPt->getValueAPF().isZero()) || 768 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 769 !CFPf->getValueAPF().isZero())) 770 return ReplaceInstUsesWith(SI, FalseVal); 771 } 772 // Transform (X une Y) ? Y : X -> Y 773 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 774 // This is not safe in general for floating point: 775 // consider X== -0, Y== +0. 776 // It becomes safe if either operand is a nonzero constant. 777 ConstantFP *CFPt, *CFPf; 778 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 779 !CFPt->getValueAPF().isZero()) || 780 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 781 !CFPf->getValueAPF().isZero())) 782 return ReplaceInstUsesWith(SI, TrueVal); 783 } 784 // NOTE: if we wanted to, this is where to detect MIN/MAX 785 } 786 // NOTE: if we wanted to, this is where to detect ABS 787 } 788 789 // See if we are selecting two values based on a comparison of the two values. 790 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 791 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI)) 792 return Result; 793 794 if (Instruction *TI = dyn_cast<Instruction>(TrueVal)) 795 if (Instruction *FI = dyn_cast<Instruction>(FalseVal)) 796 if (TI->hasOneUse() && FI->hasOneUse()) { 797 Instruction *AddOp = 0, *SubOp = 0; 798 799 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 800 if (TI->getOpcode() == FI->getOpcode()) 801 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI)) 802 return IV; 803 804 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is 805 // even legal for FP. 806 if ((TI->getOpcode() == Instruction::Sub && 807 FI->getOpcode() == Instruction::Add) || 808 (TI->getOpcode() == Instruction::FSub && 809 FI->getOpcode() == Instruction::FAdd)) { 810 AddOp = FI; SubOp = TI; 811 } else if ((FI->getOpcode() == Instruction::Sub && 812 TI->getOpcode() == Instruction::Add) || 813 (FI->getOpcode() == Instruction::FSub && 814 TI->getOpcode() == Instruction::FAdd)) { 815 AddOp = TI; SubOp = FI; 816 } 817 818 if (AddOp) { 819 Value *OtherAddOp = 0; 820 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 821 OtherAddOp = AddOp->getOperand(1); 822 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 823 OtherAddOp = AddOp->getOperand(0); 824 } 825 826 if (OtherAddOp) { 827 // So at this point we know we have (Y -> OtherAddOp): 828 // select C, (add X, Y), (sub X, Z) 829 Value *NegVal; // Compute -Z 830 if (SI.getType()->isFPOrFPVectorTy()) { 831 NegVal = Builder->CreateFNeg(SubOp->getOperand(1)); 832 } else { 833 NegVal = Builder->CreateNeg(SubOp->getOperand(1)); 834 } 835 836 Value *NewTrueOp = OtherAddOp; 837 Value *NewFalseOp = NegVal; 838 if (AddOp != TI) 839 std::swap(NewTrueOp, NewFalseOp); 840 Value *NewSel = 841 Builder->CreateSelect(CondVal, NewTrueOp, 842 NewFalseOp, SI.getName() + ".p"); 843 844 if (SI.getType()->isFPOrFPVectorTy()) 845 return BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 846 else 847 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 848 } 849 } 850 } 851 852 // See if we can fold the select into one of our operands. 853 if (SI.getType()->isIntegerTy()) { 854 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) 855 return FoldI; 856 857 // MAX(MAX(a, b), a) -> MAX(a, b) 858 // MIN(MIN(a, b), a) -> MIN(a, b) 859 // MAX(MIN(a, b), a) -> a 860 // MIN(MAX(a, b), a) -> a 861 Value *LHS, *RHS, *LHS2, *RHS2; 862 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) { 863 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2)) 864 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2, 865 SI, SPF, RHS)) 866 return R; 867 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2)) 868 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2, 869 SI, SPF, LHS)) 870 return R; 871 } 872 873 // TODO. 874 // ABS(-X) -> ABS(X) 875 // ABS(ABS(X)) -> ABS(X) 876 } 877 878 // See if we can fold the select into a phi node if the condition is a select. 879 if (isa<PHINode>(SI.getCondition())) 880 // The true/false values have to be live in the PHI predecessor's blocks. 881 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 882 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 883 if (Instruction *NV = FoldOpIntoPhi(SI)) 884 return NV; 885 886 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 887 if (TrueSI->getCondition() == CondVal) { 888 if (SI.getTrueValue() == TrueSI->getTrueValue()) 889 return 0; 890 SI.setOperand(1, TrueSI->getTrueValue()); 891 return &SI; 892 } 893 } 894 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 895 if (FalseSI->getCondition() == CondVal) { 896 if (SI.getFalseValue() == FalseSI->getFalseValue()) 897 return 0; 898 SI.setOperand(2, FalseSI->getFalseValue()); 899 return &SI; 900 } 901 } 902 903 if (BinaryOperator::isNot(CondVal)) { 904 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal)); 905 SI.setOperand(1, FalseVal); 906 SI.setOperand(2, TrueVal); 907 return &SI; 908 } 909 910 if (VectorType *VecTy = dyn_cast<VectorType>(SI.getType())) { 911 unsigned VWidth = VecTy->getNumElements(); 912 APInt UndefElts(VWidth, 0); 913 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 914 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) { 915 if (V != &SI) 916 return ReplaceInstUsesWith(SI, V); 917 return &SI; 918 } 919 920 if (ConstantVector *CV = dyn_cast<ConstantVector>(CondVal)) { 921 // Form a shufflevector instruction. 922 SmallVector<Constant *, 8> Mask(VWidth); 923 Type *Int32Ty = Type::getInt32Ty(CV->getContext()); 924 for (unsigned i = 0; i != VWidth; ++i) { 925 Constant *Elem = cast<Constant>(CV->getOperand(i)); 926 if (ConstantInt *E = dyn_cast<ConstantInt>(Elem)) 927 Mask[i] = ConstantInt::get(Int32Ty, i + (E->isZero() ? VWidth : 0)); 928 else if (isa<UndefValue>(Elem)) 929 Mask[i] = UndefValue::get(Int32Ty); 930 else 931 return 0; 932 } 933 Constant *MaskVal = ConstantVector::get(Mask); 934 Value *V = Builder->CreateShuffleVector(TrueVal, FalseVal, MaskVal); 935 return ReplaceInstUsesWith(SI, V); 936 } 937 938 if (isa<ConstantAggregateZero>(CondVal)) { 939 return ReplaceInstUsesWith(SI, FalseVal); 940 } 941 } 942 943 return 0; 944 } 945
