InstCombineSelect.cpp revision 9affd163611b90113406b8729cc591eaad4778fa
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" 18using namespace llvm; 19using 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. 24static SelectPatternFlavor 25MatchSelectPattern(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/// 84static 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. 104static 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. 124Instruction *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 196static 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. 211Instruction *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. 290static 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/// visitSelectInstWithICmp - Visit a SelectInst that has an 354/// ICmpInst as its first operand. 355/// 356Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, 357 ICmpInst *ICI) { 358 bool Changed = false; 359 ICmpInst::Predicate Pred = ICI->getPredicate(); 360 Value *CmpLHS = ICI->getOperand(0); 361 Value *CmpRHS = ICI->getOperand(1); 362 Value *TrueVal = SI.getTrueValue(); 363 Value *FalseVal = SI.getFalseValue(); 364 365 // Check cases where the comparison is with a constant that 366 // can be adjusted to fit the min/max idiom. We may move or edit ICI 367 // here, so make sure the select is the only user. 368 if (ICI->hasOneUse()) 369 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) { 370 // X < MIN ? T : F --> F 371 if ((Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT) 372 && CI->isMinValue(Pred == ICmpInst::ICMP_SLT)) 373 return ReplaceInstUsesWith(SI, FalseVal); 374 // X > MAX ? T : F --> F 375 else if ((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT) 376 && CI->isMaxValue(Pred == ICmpInst::ICMP_SGT)) 377 return ReplaceInstUsesWith(SI, FalseVal); 378 switch (Pred) { 379 default: break; 380 case ICmpInst::ICMP_ULT: 381 case ICmpInst::ICMP_SLT: 382 case ICmpInst::ICMP_UGT: 383 case ICmpInst::ICMP_SGT: { 384 // These transformations only work for selects over integers. 385 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType()); 386 if (!SelectTy) 387 break; 388 389 Constant *AdjustedRHS; 390 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 391 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1); 392 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 393 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1); 394 395 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 396 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 397 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 398 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) 399 ; // Nothing to do here. Values match without any sign/zero extension. 400 401 // Types do not match. Instead of calculating this with mixed types 402 // promote all to the larger type. This enables scalar evolution to 403 // analyze this expression. 404 else if (CmpRHS->getType()->getScalarSizeInBits() 405 < SelectTy->getBitWidth()) { 406 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy); 407 408 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 409 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s 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 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 412 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && 413 sextRHS == FalseVal) { 414 CmpLHS = TrueVal; 415 AdjustedRHS = sextRHS; 416 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 417 sextRHS == TrueVal) { 418 CmpLHS = FalseVal; 419 AdjustedRHS = sextRHS; 420 } else if (ICI->isUnsigned()) { 421 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy); 422 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 423 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 424 // zext + signed compare cannot be changed: 425 // 0xff <s 0x00, but 0x00ff >s 0x0000 426 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && 427 zextRHS == FalseVal) { 428 CmpLHS = TrueVal; 429 AdjustedRHS = zextRHS; 430 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 431 zextRHS == TrueVal) { 432 CmpLHS = FalseVal; 433 AdjustedRHS = zextRHS; 434 } else 435 break; 436 } else 437 break; 438 } else 439 break; 440 441 Pred = ICmpInst::getSwappedPredicate(Pred); 442 CmpRHS = AdjustedRHS; 443 std::swap(FalseVal, TrueVal); 444 ICI->setPredicate(Pred); 445 ICI->setOperand(0, CmpLHS); 446 ICI->setOperand(1, CmpRHS); 447 SI.setOperand(1, TrueVal); 448 SI.setOperand(2, FalseVal); 449 450 // Move ICI instruction right before the select instruction. Otherwise 451 // the sext/zext value may be defined after the ICI instruction uses it. 452 ICI->moveBefore(&SI); 453 454 Changed = true; 455 break; 456 } 457 } 458 } 459 460 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 461 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1 462 // FIXME: Type and constness constraints could be lifted, but we have to 463 // watch code size carefully. We should consider xor instead of 464 // sub/add when we decide to do that. 465 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) { 466 if (TrueVal->getType() == Ty) { 467 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) { 468 ConstantInt *C1 = NULL, *C2 = NULL; 469 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) { 470 C1 = dyn_cast<ConstantInt>(TrueVal); 471 C2 = dyn_cast<ConstantInt>(FalseVal); 472 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) { 473 C1 = dyn_cast<ConstantInt>(FalseVal); 474 C2 = dyn_cast<ConstantInt>(TrueVal); 475 } 476 if (C1 && C2) { 477 // This shift results in either -1 or 0. 478 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1); 479 480 // Check if we can express the operation with a single or. 481 if (C2->isAllOnesValue()) 482 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1)); 483 484 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue()); 485 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1)); 486 } 487 } 488 } 489 } 490 491 // If we have an equality comparison then we know the value in one of the 492 // arms of the select. See if substituting this value into the arm and 493 // simplifying the result yields the same value as the other arm. 494 if (Pred == ICmpInst::ICMP_EQ) { 495 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 496 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 497 return ReplaceInstUsesWith(SI, FalseVal); 498 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 499 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 500 return ReplaceInstUsesWith(SI, FalseVal); 501 } else if (Pred == ICmpInst::ICMP_NE) { 502 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, TD, TLI) == FalseVal || 503 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, TD, TLI) == FalseVal) 504 return ReplaceInstUsesWith(SI, TrueVal); 505 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, TD, TLI) == TrueVal || 506 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, TD, TLI) == TrueVal) 507 return ReplaceInstUsesWith(SI, TrueVal); 508 } 509 510 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 511 512 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 513 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 514 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 515 SI.setOperand(1, CmpRHS); 516 Changed = true; 517 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 518 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 519 SI.setOperand(2, CmpRHS); 520 Changed = true; 521 } 522 } 523 524 return Changed ? &SI : 0; 525} 526 527 528/// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a 529/// PHI node (but the two may be in different blocks). See if the true/false 530/// values (V) are live in all of the predecessor blocks of the PHI. For 531/// example, cases like this cannot be mapped: 532/// 533/// X = phi [ C1, BB1], [C2, BB2] 534/// Y = add 535/// Z = select X, Y, 0 536/// 537/// because Y is not live in BB1/BB2. 538/// 539static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V, 540 const SelectInst &SI) { 541 // If the value is a non-instruction value like a constant or argument, it 542 // can always be mapped. 543 const Instruction *I = dyn_cast<Instruction>(V); 544 if (I == 0) return true; 545 546 // If V is a PHI node defined in the same block as the condition PHI, we can 547 // map the arguments. 548 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 549 550 if (const PHINode *VP = dyn_cast<PHINode>(I)) 551 if (VP->getParent() == CondPHI->getParent()) 552 return true; 553 554 // Otherwise, if the PHI and select are defined in the same block and if V is 555 // defined in a different block, then we can transform it. 556 if (SI.getParent() == CondPHI->getParent() && 557 I->getParent() != CondPHI->getParent()) 558 return true; 559 560 // Otherwise we have a 'hard' case and we can't tell without doing more 561 // detailed dominator based analysis, punt. 562 return false; 563} 564 565/// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form: 566/// SPF2(SPF1(A, B), C) 567Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner, 568 SelectPatternFlavor SPF1, 569 Value *A, Value *B, 570 Instruction &Outer, 571 SelectPatternFlavor SPF2, Value *C) { 572 if (C == A || C == B) { 573 // MAX(MAX(A, B), B) -> MAX(A, B) 574 // MIN(MIN(a, b), a) -> MIN(a, b) 575 if (SPF1 == SPF2) 576 return ReplaceInstUsesWith(Outer, Inner); 577 578 // MAX(MIN(a, b), a) -> a 579 // MIN(MAX(a, b), a) -> a 580 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 581 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 582 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 583 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 584 return ReplaceInstUsesWith(Outer, C); 585 } 586 587 // TODO: MIN(MIN(A, 23), 97) 588 return 0; 589} 590 591 592/// foldSelectICmpAnd - If one of the constants is zero (we know they can't 593/// both be) and we have an icmp instruction with zero, and we have an 'and' 594/// with the non-constant value and a power of two we can turn the select 595/// into a shift on the result of the 'and'. 596static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal, 597 ConstantInt *FalseVal, 598 InstCombiner::BuilderTy *Builder) { 599 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 600 if (!IC || !IC->isEquality()) 601 return 0; 602 603 if (!match(IC->getOperand(1), m_Zero())) 604 return 0; 605 606 ConstantInt *AndRHS; 607 Value *LHS = IC->getOperand(0); 608 if (LHS->getType() != SI.getType() || 609 !match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS)))) 610 return 0; 611 612 // If both select arms are non-zero see if we have a select of the form 613 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic 614 // for 'x ? 2^n : 0' and fix the thing up at the end. 615 ConstantInt *Offset = 0; 616 if (!TrueVal->isZero() && !FalseVal->isZero()) { 617 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2()) 618 Offset = FalseVal; 619 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2()) 620 Offset = TrueVal; 621 else 622 return 0; 623 624 // Adjust TrueVal and FalseVal to the offset. 625 TrueVal = ConstantInt::get(Builder->getContext(), 626 TrueVal->getValue() - Offset->getValue()); 627 FalseVal = ConstantInt::get(Builder->getContext(), 628 FalseVal->getValue() - Offset->getValue()); 629 } 630 631 // Make sure the mask in the 'and' and one of the select arms is a power of 2. 632 if (!AndRHS->getValue().isPowerOf2() || 633 (!TrueVal->getValue().isPowerOf2() && 634 !FalseVal->getValue().isPowerOf2())) 635 return 0; 636 637 // Determine which shift is needed to transform result of the 'and' into the 638 // desired result. 639 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal; 640 unsigned ValZeros = ValC->getValue().logBase2(); 641 unsigned AndZeros = AndRHS->getValue().logBase2(); 642 643 Value *V = LHS; 644 if (ValZeros > AndZeros) 645 V = Builder->CreateShl(V, ValZeros - AndZeros); 646 else if (ValZeros < AndZeros) 647 V = Builder->CreateLShr(V, AndZeros - ValZeros); 648 649 // Okay, now we know that everything is set up, we just don't know whether we 650 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 651 bool ShouldNotVal = !TrueVal->isZero(); 652 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; 653 if (ShouldNotVal) 654 V = Builder->CreateXor(V, ValC); 655 656 // Apply an offset if needed. 657 if (Offset) 658 V = Builder->CreateAdd(V, Offset); 659 return V; 660} 661 662Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { 663 Value *CondVal = SI.getCondition(); 664 Value *TrueVal = SI.getTrueValue(); 665 Value *FalseVal = SI.getFalseValue(); 666 667 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, TD)) 668 return ReplaceInstUsesWith(SI, V); 669 670 if (SI.getType()->isIntegerTy(1)) { 671 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { 672 if (C->getZExtValue()) { 673 // Change: A = select B, true, C --> A = or B, C 674 return BinaryOperator::CreateOr(CondVal, FalseVal); 675 } 676 // Change: A = select B, false, C --> A = and !B, C 677 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 678 return BinaryOperator::CreateAnd(NotCond, FalseVal); 679 } 680 if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) { 681 if (C->getZExtValue() == false) { 682 // Change: A = select B, C, false --> A = and B, C 683 return BinaryOperator::CreateAnd(CondVal, TrueVal); 684 } 685 // Change: A = select B, C, true --> A = or !B, C 686 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 687 return BinaryOperator::CreateOr(NotCond, TrueVal); 688 } 689 690 // select a, b, a -> a&b 691 // select a, a, b -> a|b 692 if (CondVal == TrueVal) 693 return BinaryOperator::CreateOr(CondVal, FalseVal); 694 if (CondVal == FalseVal) 695 return BinaryOperator::CreateAnd(CondVal, TrueVal); 696 697 // select a, ~a, b -> (~a)&b 698 // select a, b, ~a -> (~a)|b 699 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 700 return BinaryOperator::CreateAnd(TrueVal, FalseVal); 701 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 702 return BinaryOperator::CreateOr(TrueVal, FalseVal); 703 } 704 705 // Selecting between two integer constants? 706 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) 707 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { 708 // select C, 1, 0 -> zext C to int 709 if (FalseValC->isZero() && TrueValC->getValue() == 1) 710 return new ZExtInst(CondVal, SI.getType()); 711 712 // select C, -1, 0 -> sext C to int 713 if (FalseValC->isZero() && TrueValC->isAllOnesValue()) 714 return new SExtInst(CondVal, SI.getType()); 715 716 // select C, 0, 1 -> zext !C to int 717 if (TrueValC->isZero() && FalseValC->getValue() == 1) { 718 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 719 return new ZExtInst(NotCond, SI.getType()); 720 } 721 722 // select C, 0, -1 -> sext !C to int 723 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { 724 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 725 return new SExtInst(NotCond, SI.getType()); 726 } 727 728 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder)) 729 return ReplaceInstUsesWith(SI, V); 730 } 731 732 // See if we are selecting two values based on a comparison of the two values. 733 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 734 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) { 735 // Transform (X == Y) ? X : Y -> Y 736 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 737 // This is not safe in general for floating point: 738 // consider X== -0, Y== +0. 739 // It becomes safe if either operand is a nonzero constant. 740 ConstantFP *CFPt, *CFPf; 741 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 742 !CFPt->getValueAPF().isZero()) || 743 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 744 !CFPf->getValueAPF().isZero())) 745 return ReplaceInstUsesWith(SI, FalseVal); 746 } 747 // Transform (X une Y) ? X : Y -> X 748 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 749 // This is not safe in general for floating point: 750 // consider X== -0, Y== +0. 751 // It becomes safe if either operand is a nonzero constant. 752 ConstantFP *CFPt, *CFPf; 753 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 754 !CFPt->getValueAPF().isZero()) || 755 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 756 !CFPf->getValueAPF().isZero())) 757 return ReplaceInstUsesWith(SI, TrueVal); 758 } 759 // NOTE: if we wanted to, this is where to detect MIN/MAX 760 761 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){ 762 // Transform (X == Y) ? Y : X -> X 763 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 764 // This is not safe in general for floating point: 765 // consider X== -0, Y== +0. 766 // It becomes safe if either operand is a nonzero constant. 767 ConstantFP *CFPt, *CFPf; 768 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 769 !CFPt->getValueAPF().isZero()) || 770 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 771 !CFPf->getValueAPF().isZero())) 772 return ReplaceInstUsesWith(SI, FalseVal); 773 } 774 // Transform (X une Y) ? Y : X -> Y 775 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 776 // This is not safe in general for floating point: 777 // consider X== -0, Y== +0. 778 // It becomes safe if either operand is a nonzero constant. 779 ConstantFP *CFPt, *CFPf; 780 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 781 !CFPt->getValueAPF().isZero()) || 782 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 783 !CFPf->getValueAPF().isZero())) 784 return ReplaceInstUsesWith(SI, TrueVal); 785 } 786 // NOTE: if we wanted to, this is where to detect MIN/MAX 787 } 788 // NOTE: if we wanted to, this is where to detect ABS 789 } 790 791 // See if we are selecting two values based on a comparison of the two values. 792 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 793 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI)) 794 return Result; 795 796 if (Instruction *TI = dyn_cast<Instruction>(TrueVal)) 797 if (Instruction *FI = dyn_cast<Instruction>(FalseVal)) 798 if (TI->hasOneUse() && FI->hasOneUse()) { 799 Instruction *AddOp = 0, *SubOp = 0; 800 801 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 802 if (TI->getOpcode() == FI->getOpcode()) 803 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI)) 804 return IV; 805 806 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is 807 // even legal for FP. 808 if ((TI->getOpcode() == Instruction::Sub && 809 FI->getOpcode() == Instruction::Add) || 810 (TI->getOpcode() == Instruction::FSub && 811 FI->getOpcode() == Instruction::FAdd)) { 812 AddOp = FI; SubOp = TI; 813 } else if ((FI->getOpcode() == Instruction::Sub && 814 TI->getOpcode() == Instruction::Add) || 815 (FI->getOpcode() == Instruction::FSub && 816 TI->getOpcode() == Instruction::FAdd)) { 817 AddOp = TI; SubOp = FI; 818 } 819 820 if (AddOp) { 821 Value *OtherAddOp = 0; 822 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 823 OtherAddOp = AddOp->getOperand(1); 824 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 825 OtherAddOp = AddOp->getOperand(0); 826 } 827 828 if (OtherAddOp) { 829 // So at this point we know we have (Y -> OtherAddOp): 830 // select C, (add X, Y), (sub X, Z) 831 Value *NegVal; // Compute -Z 832 if (SI.getType()->isFPOrFPVectorTy()) { 833 NegVal = Builder->CreateFNeg(SubOp->getOperand(1)); 834 } else { 835 NegVal = Builder->CreateNeg(SubOp->getOperand(1)); 836 } 837 838 Value *NewTrueOp = OtherAddOp; 839 Value *NewFalseOp = NegVal; 840 if (AddOp != TI) 841 std::swap(NewTrueOp, NewFalseOp); 842 Value *NewSel = 843 Builder->CreateSelect(CondVal, NewTrueOp, 844 NewFalseOp, SI.getName() + ".p"); 845 846 if (SI.getType()->isFPOrFPVectorTy()) 847 return BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 848 else 849 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 850 } 851 } 852 } 853 854 // See if we can fold the select into one of our operands. 855 if (SI.getType()->isIntegerTy()) { 856 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) 857 return FoldI; 858 859 // MAX(MAX(a, b), a) -> MAX(a, b) 860 // MIN(MIN(a, b), a) -> MIN(a, b) 861 // MAX(MIN(a, b), a) -> a 862 // MIN(MAX(a, b), a) -> a 863 Value *LHS, *RHS, *LHS2, *RHS2; 864 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) { 865 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2)) 866 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2, 867 SI, SPF, RHS)) 868 return R; 869 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2)) 870 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2, 871 SI, SPF, LHS)) 872 return R; 873 } 874 875 // TODO. 876 // ABS(-X) -> ABS(X) 877 // ABS(ABS(X)) -> ABS(X) 878 } 879 880 // See if we can fold the select into a phi node if the condition is a select. 881 if (isa<PHINode>(SI.getCondition())) 882 // The true/false values have to be live in the PHI predecessor's blocks. 883 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 884 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 885 if (Instruction *NV = FoldOpIntoPhi(SI)) 886 return NV; 887 888 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 889 if (TrueSI->getCondition() == CondVal) { 890 if (SI.getTrueValue() == TrueSI->getTrueValue()) 891 return 0; 892 SI.setOperand(1, TrueSI->getTrueValue()); 893 return &SI; 894 } 895 } 896 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 897 if (FalseSI->getCondition() == CondVal) { 898 if (SI.getFalseValue() == FalseSI->getFalseValue()) 899 return 0; 900 SI.setOperand(2, FalseSI->getFalseValue()); 901 return &SI; 902 } 903 } 904 905 if (BinaryOperator::isNot(CondVal)) { 906 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal)); 907 SI.setOperand(1, FalseVal); 908 SI.setOperand(2, TrueVal); 909 return &SI; 910 } 911 912 if (VectorType *VecTy = dyn_cast<VectorType>(SI.getType())) { 913 unsigned VWidth = VecTy->getNumElements(); 914 APInt UndefElts(VWidth, 0); 915 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 916 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) { 917 if (V != &SI) 918 return ReplaceInstUsesWith(SI, V); 919 return &SI; 920 } 921 922 if (ConstantVector *CV = dyn_cast<ConstantVector>(CondVal)) { 923 // Form a shufflevector instruction. 924 SmallVector<Constant *, 8> Mask(VWidth); 925 Type *Int32Ty = Type::getInt32Ty(CV->getContext()); 926 for (unsigned i = 0; i != VWidth; ++i) { 927 Constant *Elem = cast<Constant>(CV->getOperand(i)); 928 if (ConstantInt *E = dyn_cast<ConstantInt>(Elem)) 929 Mask[i] = ConstantInt::get(Int32Ty, i + (E->isZero() ? VWidth : 0)); 930 else if (isa<UndefValue>(Elem)) 931 Mask[i] = UndefValue::get(Int32Ty); 932 else 933 return 0; 934 } 935 Constant *MaskVal = ConstantVector::get(Mask); 936 Value *V = Builder->CreateShuffleVector(TrueVal, FalseVal, MaskVal); 937 return ReplaceInstUsesWith(SI, V); 938 } 939 940 if (isa<ConstantAggregateZero>(CondVal)) { 941 return ReplaceInstUsesWith(SI, FalseVal); 942 } 943 } 944 945 return 0; 946} 947