1 //===- llvm/Analysis/IVDescriptors.cpp - IndVar Descriptors -----*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file "describes" induction and recurrence variables. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Analysis/IVDescriptors.h" 14 #include "llvm/ADT/ScopeExit.h" 15 #include "llvm/Analysis/BasicAliasAnalysis.h" 16 #include "llvm/Analysis/DemandedBits.h" 17 #include "llvm/Analysis/DomTreeUpdater.h" 18 #include "llvm/Analysis/GlobalsModRef.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/LoopPass.h" 22 #include "llvm/Analysis/MustExecute.h" 23 #include "llvm/Analysis/ScalarEvolution.h" 24 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 25 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 26 #include "llvm/Analysis/TargetTransformInfo.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/Instructions.h" 30 #include "llvm/IR/Module.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/IR/ValueHandle.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/KnownBits.h" 36 37 using namespace llvm; 38 using namespace llvm::PatternMatch; 39 40 #define DEBUG_TYPE "iv-descriptors" 41 42 bool RecurrenceDescriptor::areAllUsesIn(Instruction *I, 43 SmallPtrSetImpl<Instruction *> &Set) { 44 for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) 45 if (!Set.count(dyn_cast<Instruction>(*Use))) 46 return false; 47 return true; 48 } 49 50 bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) { 51 switch (Kind) { 52 default: 53 break; 54 case RecurKind::Add: 55 case RecurKind::Mul: 56 case RecurKind::Or: 57 case RecurKind::And: 58 case RecurKind::Xor: 59 case RecurKind::SMax: 60 case RecurKind::SMin: 61 case RecurKind::UMax: 62 case RecurKind::UMin: 63 return true; 64 } 65 return false; 66 } 67 68 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurKind Kind) { 69 return (Kind != RecurKind::None) && !isIntegerRecurrenceKind(Kind); 70 } 71 72 bool RecurrenceDescriptor::isArithmeticRecurrenceKind(RecurKind Kind) { 73 switch (Kind) { 74 default: 75 break; 76 case RecurKind::Add: 77 case RecurKind::Mul: 78 case RecurKind::FAdd: 79 case RecurKind::FMul: 80 return true; 81 } 82 return false; 83 } 84 85 /// Determines if Phi may have been type-promoted. If Phi has a single user 86 /// that ANDs the Phi with a type mask, return the user. RT is updated to 87 /// account for the narrower bit width represented by the mask, and the AND 88 /// instruction is added to CI. 89 static Instruction *lookThroughAnd(PHINode *Phi, Type *&RT, 90 SmallPtrSetImpl<Instruction *> &Visited, 91 SmallPtrSetImpl<Instruction *> &CI) { 92 if (!Phi->hasOneUse()) 93 return Phi; 94 95 const APInt *M = nullptr; 96 Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser()); 97 98 // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT 99 // with a new integer type of the corresponding bit width. 100 if (match(J, m_c_And(m_Instruction(I), m_APInt(M)))) { 101 int32_t Bits = (*M + 1).exactLogBase2(); 102 if (Bits > 0) { 103 RT = IntegerType::get(Phi->getContext(), Bits); 104 Visited.insert(Phi); 105 CI.insert(J); 106 return J; 107 } 108 } 109 return Phi; 110 } 111 112 /// Compute the minimal bit width needed to represent a reduction whose exit 113 /// instruction is given by Exit. 114 static std::pair<Type *, bool> computeRecurrenceType(Instruction *Exit, 115 DemandedBits *DB, 116 AssumptionCache *AC, 117 DominatorTree *DT) { 118 bool IsSigned = false; 119 const DataLayout &DL = Exit->getModule()->getDataLayout(); 120 uint64_t MaxBitWidth = DL.getTypeSizeInBits(Exit->getType()); 121 122 if (DB) { 123 // Use the demanded bits analysis to determine the bits that are live out 124 // of the exit instruction, rounding up to the nearest power of two. If the 125 // use of demanded bits results in a smaller bit width, we know the value 126 // must be positive (i.e., IsSigned = false), because if this were not the 127 // case, the sign bit would have been demanded. 128 auto Mask = DB->getDemandedBits(Exit); 129 MaxBitWidth = Mask.getBitWidth() - Mask.countLeadingZeros(); 130 } 131 132 if (MaxBitWidth == DL.getTypeSizeInBits(Exit->getType()) && AC && DT) { 133 // If demanded bits wasn't able to limit the bit width, we can try to use 134 // value tracking instead. This can be the case, for example, if the value 135 // may be negative. 136 auto NumSignBits = ComputeNumSignBits(Exit, DL, 0, AC, nullptr, DT); 137 auto NumTypeBits = DL.getTypeSizeInBits(Exit->getType()); 138 MaxBitWidth = NumTypeBits - NumSignBits; 139 KnownBits Bits = computeKnownBits(Exit, DL); 140 if (!Bits.isNonNegative()) { 141 // If the value is not known to be non-negative, we set IsSigned to true, 142 // meaning that we will use sext instructions instead of zext 143 // instructions to restore the original type. 144 IsSigned = true; 145 if (!Bits.isNegative()) 146 // If the value is not known to be negative, we don't known what the 147 // upper bit is, and therefore, we don't know what kind of extend we 148 // will need. In this case, just increase the bit width by one bit and 149 // use sext. 150 ++MaxBitWidth; 151 } 152 } 153 if (!isPowerOf2_64(MaxBitWidth)) 154 MaxBitWidth = NextPowerOf2(MaxBitWidth); 155 156 return std::make_pair(Type::getIntNTy(Exit->getContext(), MaxBitWidth), 157 IsSigned); 158 } 159 160 /// Collect cast instructions that can be ignored in the vectorizer's cost 161 /// model, given a reduction exit value and the minimal type in which the 162 /// reduction can be represented. 163 static void collectCastsToIgnore(Loop *TheLoop, Instruction *Exit, 164 Type *RecurrenceType, 165 SmallPtrSetImpl<Instruction *> &Casts) { 166 167 SmallVector<Instruction *, 8> Worklist; 168 SmallPtrSet<Instruction *, 8> Visited; 169 Worklist.push_back(Exit); 170 171 while (!Worklist.empty()) { 172 Instruction *Val = Worklist.pop_back_val(); 173 Visited.insert(Val); 174 if (auto *Cast = dyn_cast<CastInst>(Val)) 175 if (Cast->getSrcTy() == RecurrenceType) { 176 // If the source type of a cast instruction is equal to the recurrence 177 // type, it will be eliminated, and should be ignored in the vectorizer 178 // cost model. 179 Casts.insert(Cast); 180 continue; 181 } 182 183 // Add all operands to the work list if they are loop-varying values that 184 // we haven't yet visited. 185 for (Value *O : cast<User>(Val)->operands()) 186 if (auto *I = dyn_cast<Instruction>(O)) 187 if (TheLoop->contains(I) && !Visited.count(I)) 188 Worklist.push_back(I); 189 } 190 } 191 192 bool RecurrenceDescriptor::AddReductionVar(PHINode *Phi, RecurKind Kind, 193 Loop *TheLoop, bool HasFunNoNaNAttr, 194 RecurrenceDescriptor &RedDes, 195 DemandedBits *DB, 196 AssumptionCache *AC, 197 DominatorTree *DT) { 198 if (Phi->getNumIncomingValues() != 2) 199 return false; 200 201 // Reduction variables are only found in the loop header block. 202 if (Phi->getParent() != TheLoop->getHeader()) 203 return false; 204 205 // Obtain the reduction start value from the value that comes from the loop 206 // preheader. 207 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 208 209 // ExitInstruction is the single value which is used outside the loop. 210 // We only allow for a single reduction value to be used outside the loop. 211 // This includes users of the reduction, variables (which form a cycle 212 // which ends in the phi node). 213 Instruction *ExitInstruction = nullptr; 214 // Indicates that we found a reduction operation in our scan. 215 bool FoundReduxOp = false; 216 217 // We start with the PHI node and scan for all of the users of this 218 // instruction. All users must be instructions that can be used as reduction 219 // variables (such as ADD). We must have a single out-of-block user. The cycle 220 // must include the original PHI. 221 bool FoundStartPHI = false; 222 223 // To recognize min/max patterns formed by a icmp select sequence, we store 224 // the number of instruction we saw from the recognized min/max pattern, 225 // to make sure we only see exactly the two instructions. 226 unsigned NumCmpSelectPatternInst = 0; 227 InstDesc ReduxDesc(false, nullptr); 228 229 // Data used for determining if the recurrence has been type-promoted. 230 Type *RecurrenceType = Phi->getType(); 231 SmallPtrSet<Instruction *, 4> CastInsts; 232 Instruction *Start = Phi; 233 bool IsSigned = false; 234 235 SmallPtrSet<Instruction *, 8> VisitedInsts; 236 SmallVector<Instruction *, 8> Worklist; 237 238 // Return early if the recurrence kind does not match the type of Phi. If the 239 // recurrence kind is arithmetic, we attempt to look through AND operations 240 // resulting from the type promotion performed by InstCombine. Vector 241 // operations are not limited to the legal integer widths, so we may be able 242 // to evaluate the reduction in the narrower width. 243 if (RecurrenceType->isFloatingPointTy()) { 244 if (!isFloatingPointRecurrenceKind(Kind)) 245 return false; 246 } else { 247 if (!isIntegerRecurrenceKind(Kind)) 248 return false; 249 if (isArithmeticRecurrenceKind(Kind)) 250 Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts); 251 } 252 253 Worklist.push_back(Start); 254 VisitedInsts.insert(Start); 255 256 // Start with all flags set because we will intersect this with the reduction 257 // flags from all the reduction operations. 258 FastMathFlags FMF = FastMathFlags::getFast(); 259 260 // A value in the reduction can be used: 261 // - By the reduction: 262 // - Reduction operation: 263 // - One use of reduction value (safe). 264 // - Multiple use of reduction value (not safe). 265 // - PHI: 266 // - All uses of the PHI must be the reduction (safe). 267 // - Otherwise, not safe. 268 // - By instructions outside of the loop (safe). 269 // * One value may have several outside users, but all outside 270 // uses must be of the same value. 271 // - By an instruction that is not part of the reduction (not safe). 272 // This is either: 273 // * An instruction type other than PHI or the reduction operation. 274 // * A PHI in the header other than the initial PHI. 275 while (!Worklist.empty()) { 276 Instruction *Cur = Worklist.pop_back_val(); 277 278 // No Users. 279 // If the instruction has no users then this is a broken chain and can't be 280 // a reduction variable. 281 if (Cur->use_empty()) 282 return false; 283 284 bool IsAPhi = isa<PHINode>(Cur); 285 286 // A header PHI use other than the original PHI. 287 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 288 return false; 289 290 // Reductions of instructions such as Div, and Sub is only possible if the 291 // LHS is the reduction variable. 292 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 293 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 294 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 295 return false; 296 297 // Any reduction instruction must be of one of the allowed kinds. We ignore 298 // the starting value (the Phi or an AND instruction if the Phi has been 299 // type-promoted). 300 if (Cur != Start) { 301 ReduxDesc = isRecurrenceInstr(Cur, Kind, ReduxDesc, HasFunNoNaNAttr); 302 if (!ReduxDesc.isRecurrence()) 303 return false; 304 // FIXME: FMF is allowed on phi, but propagation is not handled correctly. 305 if (isa<FPMathOperator>(ReduxDesc.getPatternInst()) && !IsAPhi) { 306 FastMathFlags CurFMF = ReduxDesc.getPatternInst()->getFastMathFlags(); 307 if (auto *Sel = dyn_cast<SelectInst>(ReduxDesc.getPatternInst())) { 308 // Accept FMF on either fcmp or select of a min/max idiom. 309 // TODO: This is a hack to work-around the fact that FMF may not be 310 // assigned/propagated correctly. If that problem is fixed or we 311 // standardize on fmin/fmax via intrinsics, this can be removed. 312 if (auto *FCmp = dyn_cast<FCmpInst>(Sel->getCondition())) 313 CurFMF |= FCmp->getFastMathFlags(); 314 } 315 FMF &= CurFMF; 316 } 317 // Update this reduction kind if we matched a new instruction. 318 // TODO: Can we eliminate the need for a 2nd InstDesc by keeping 'Kind' 319 // state accurate while processing the worklist? 320 if (ReduxDesc.getRecKind() != RecurKind::None) 321 Kind = ReduxDesc.getRecKind(); 322 } 323 324 bool IsASelect = isa<SelectInst>(Cur); 325 326 // A conditional reduction operation must only have 2 or less uses in 327 // VisitedInsts. 328 if (IsASelect && (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) && 329 hasMultipleUsesOf(Cur, VisitedInsts, 2)) 330 return false; 331 332 // A reduction operation must only have one use of the reduction value. 333 if (!IsAPhi && !IsASelect && !isMinMaxRecurrenceKind(Kind) && 334 hasMultipleUsesOf(Cur, VisitedInsts, 1)) 335 return false; 336 337 // All inputs to a PHI node must be a reduction value. 338 if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 339 return false; 340 341 if (isIntMinMaxRecurrenceKind(Kind) && 342 (isa<ICmpInst>(Cur) || isa<SelectInst>(Cur))) 343 ++NumCmpSelectPatternInst; 344 if (isFPMinMaxRecurrenceKind(Kind) && 345 (isa<FCmpInst>(Cur) || isa<SelectInst>(Cur))) 346 ++NumCmpSelectPatternInst; 347 348 // Check whether we found a reduction operator. 349 FoundReduxOp |= !IsAPhi && Cur != Start; 350 351 // Process users of current instruction. Push non-PHI nodes after PHI nodes 352 // onto the stack. This way we are going to have seen all inputs to PHI 353 // nodes once we get to them. 354 SmallVector<Instruction *, 8> NonPHIs; 355 SmallVector<Instruction *, 8> PHIs; 356 for (User *U : Cur->users()) { 357 Instruction *UI = cast<Instruction>(U); 358 359 // Check if we found the exit user. 360 BasicBlock *Parent = UI->getParent(); 361 if (!TheLoop->contains(Parent)) { 362 // If we already know this instruction is used externally, move on to 363 // the next user. 364 if (ExitInstruction == Cur) 365 continue; 366 367 // Exit if you find multiple values used outside or if the header phi 368 // node is being used. In this case the user uses the value of the 369 // previous iteration, in which case we would loose "VF-1" iterations of 370 // the reduction operation if we vectorize. 371 if (ExitInstruction != nullptr || Cur == Phi) 372 return false; 373 374 // The instruction used by an outside user must be the last instruction 375 // before we feed back to the reduction phi. Otherwise, we loose VF-1 376 // operations on the value. 377 if (!is_contained(Phi->operands(), Cur)) 378 return false; 379 380 ExitInstruction = Cur; 381 continue; 382 } 383 384 // Process instructions only once (termination). Each reduction cycle 385 // value must only be used once, except by phi nodes and min/max 386 // reductions which are represented as a cmp followed by a select. 387 InstDesc IgnoredVal(false, nullptr); 388 if (VisitedInsts.insert(UI).second) { 389 if (isa<PHINode>(UI)) 390 PHIs.push_back(UI); 391 else 392 NonPHIs.push_back(UI); 393 } else if (!isa<PHINode>(UI) && 394 ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) && 395 !isa<SelectInst>(UI)) || 396 (!isConditionalRdxPattern(Kind, UI).isRecurrence() && 397 !isMinMaxSelectCmpPattern(UI, IgnoredVal).isRecurrence()))) 398 return false; 399 400 // Remember that we completed the cycle. 401 if (UI == Phi) 402 FoundStartPHI = true; 403 } 404 Worklist.append(PHIs.begin(), PHIs.end()); 405 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 406 } 407 408 // This means we have seen one but not the other instruction of the 409 // pattern or more than just a select and cmp. 410 if (isMinMaxRecurrenceKind(Kind) && NumCmpSelectPatternInst != 2) 411 return false; 412 413 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 414 return false; 415 416 if (Start != Phi) { 417 // If the starting value is not the same as the phi node, we speculatively 418 // looked through an 'and' instruction when evaluating a potential 419 // arithmetic reduction to determine if it may have been type-promoted. 420 // 421 // We now compute the minimal bit width that is required to represent the 422 // reduction. If this is the same width that was indicated by the 'and', we 423 // can represent the reduction in the smaller type. The 'and' instruction 424 // will be eliminated since it will essentially be a cast instruction that 425 // can be ignore in the cost model. If we compute a different type than we 426 // did when evaluating the 'and', the 'and' will not be eliminated, and we 427 // will end up with different kinds of operations in the recurrence 428 // expression (e.g., IntegerAND, IntegerADD). We give up if this is 429 // the case. 430 // 431 // The vectorizer relies on InstCombine to perform the actual 432 // type-shrinking. It does this by inserting instructions to truncate the 433 // exit value of the reduction to the width indicated by RecurrenceType and 434 // then extend this value back to the original width. If IsSigned is false, 435 // a 'zext' instruction will be generated; otherwise, a 'sext' will be 436 // used. 437 // 438 // TODO: We should not rely on InstCombine to rewrite the reduction in the 439 // smaller type. We should just generate a correctly typed expression 440 // to begin with. 441 Type *ComputedType; 442 std::tie(ComputedType, IsSigned) = 443 computeRecurrenceType(ExitInstruction, DB, AC, DT); 444 if (ComputedType != RecurrenceType) 445 return false; 446 447 // The recurrence expression will be represented in a narrower type. If 448 // there are any cast instructions that will be unnecessary, collect them 449 // in CastInsts. Note that the 'and' instruction was already included in 450 // this list. 451 // 452 // TODO: A better way to represent this may be to tag in some way all the 453 // instructions that are a part of the reduction. The vectorizer cost 454 // model could then apply the recurrence type to these instructions, 455 // without needing a white list of instructions to ignore. 456 // This may also be useful for the inloop reductions, if it can be 457 // kept simple enough. 458 collectCastsToIgnore(TheLoop, ExitInstruction, RecurrenceType, CastInsts); 459 } 460 461 // We found a reduction var if we have reached the original phi node and we 462 // only have a single instruction with out-of-loop users. 463 464 // The ExitInstruction(Instruction which is allowed to have out-of-loop users) 465 // is saved as part of the RecurrenceDescriptor. 466 467 // Save the description of this reduction variable. 468 RecurrenceDescriptor RD(RdxStart, ExitInstruction, Kind, FMF, 469 ReduxDesc.getUnsafeAlgebraInst(), RecurrenceType, 470 IsSigned, CastInsts); 471 RedDes = RD; 472 473 return true; 474 } 475 476 RecurrenceDescriptor::InstDesc 477 RecurrenceDescriptor::isMinMaxSelectCmpPattern(Instruction *I, 478 const InstDesc &Prev) { 479 assert((isa<CmpInst>(I) || isa<SelectInst>(I)) && 480 "Expected a cmp or select instruction"); 481 482 // We must handle the select(cmp()) as a single instruction. Advance to the 483 // select. 484 CmpInst::Predicate Pred; 485 if (match(I, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) { 486 if (auto *Select = dyn_cast<SelectInst>(*I->user_begin())) 487 return InstDesc(Select, Prev.getRecKind()); 488 } 489 490 // Only match select with single use cmp condition. 491 if (!match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(), 492 m_Value()))) 493 return InstDesc(false, I); 494 495 // Look for a min/max pattern. 496 if (match(I, m_UMin(m_Value(), m_Value()))) 497 return InstDesc(I, RecurKind::UMin); 498 if (match(I, m_UMax(m_Value(), m_Value()))) 499 return InstDesc(I, RecurKind::UMax); 500 if (match(I, m_SMax(m_Value(), m_Value()))) 501 return InstDesc(I, RecurKind::SMax); 502 if (match(I, m_SMin(m_Value(), m_Value()))) 503 return InstDesc(I, RecurKind::SMin); 504 if (match(I, m_OrdFMin(m_Value(), m_Value()))) 505 return InstDesc(I, RecurKind::FMin); 506 if (match(I, m_OrdFMax(m_Value(), m_Value()))) 507 return InstDesc(I, RecurKind::FMax); 508 if (match(I, m_UnordFMin(m_Value(), m_Value()))) 509 return InstDesc(I, RecurKind::FMin); 510 if (match(I, m_UnordFMax(m_Value(), m_Value()))) 511 return InstDesc(I, RecurKind::FMax); 512 513 return InstDesc(false, I); 514 } 515 516 /// Returns true if the select instruction has users in the compare-and-add 517 /// reduction pattern below. The select instruction argument is the last one 518 /// in the sequence. 519 /// 520 /// %sum.1 = phi ... 521 /// ... 522 /// %cmp = fcmp pred %0, %CFP 523 /// %add = fadd %0, %sum.1 524 /// %sum.2 = select %cmp, %add, %sum.1 525 RecurrenceDescriptor::InstDesc 526 RecurrenceDescriptor::isConditionalRdxPattern(RecurKind Kind, Instruction *I) { 527 SelectInst *SI = dyn_cast<SelectInst>(I); 528 if (!SI) 529 return InstDesc(false, I); 530 531 CmpInst *CI = dyn_cast<CmpInst>(SI->getCondition()); 532 // Only handle single use cases for now. 533 if (!CI || !CI->hasOneUse()) 534 return InstDesc(false, I); 535 536 Value *TrueVal = SI->getTrueValue(); 537 Value *FalseVal = SI->getFalseValue(); 538 // Handle only when either of operands of select instruction is a PHI 539 // node for now. 540 if ((isa<PHINode>(*TrueVal) && isa<PHINode>(*FalseVal)) || 541 (!isa<PHINode>(*TrueVal) && !isa<PHINode>(*FalseVal))) 542 return InstDesc(false, I); 543 544 Instruction *I1 = 545 isa<PHINode>(*TrueVal) ? dyn_cast<Instruction>(FalseVal) 546 : dyn_cast<Instruction>(TrueVal); 547 if (!I1 || !I1->isBinaryOp()) 548 return InstDesc(false, I); 549 550 Value *Op1, *Op2; 551 if ((m_FAdd(m_Value(Op1), m_Value(Op2)).match(I1) || 552 m_FSub(m_Value(Op1), m_Value(Op2)).match(I1)) && 553 I1->isFast()) 554 return InstDesc(Kind == RecurKind::FAdd, SI); 555 556 if (m_FMul(m_Value(Op1), m_Value(Op2)).match(I1) && (I1->isFast())) 557 return InstDesc(Kind == RecurKind::FMul, SI); 558 559 return InstDesc(false, I); 560 } 561 562 RecurrenceDescriptor::InstDesc 563 RecurrenceDescriptor::isRecurrenceInstr(Instruction *I, RecurKind Kind, 564 InstDesc &Prev, bool HasFunNoNaNAttr) { 565 Instruction *UAI = Prev.getUnsafeAlgebraInst(); 566 if (!UAI && isa<FPMathOperator>(I) && !I->hasAllowReassoc()) 567 UAI = I; // Found an unsafe (unvectorizable) algebra instruction. 568 569 switch (I->getOpcode()) { 570 default: 571 return InstDesc(false, I); 572 case Instruction::PHI: 573 return InstDesc(I, Prev.getRecKind(), Prev.getUnsafeAlgebraInst()); 574 case Instruction::Sub: 575 case Instruction::Add: 576 return InstDesc(Kind == RecurKind::Add, I); 577 case Instruction::Mul: 578 return InstDesc(Kind == RecurKind::Mul, I); 579 case Instruction::And: 580 return InstDesc(Kind == RecurKind::And, I); 581 case Instruction::Or: 582 return InstDesc(Kind == RecurKind::Or, I); 583 case Instruction::Xor: 584 return InstDesc(Kind == RecurKind::Xor, I); 585 case Instruction::FDiv: 586 case Instruction::FMul: 587 return InstDesc(Kind == RecurKind::FMul, I, UAI); 588 case Instruction::FSub: 589 case Instruction::FAdd: 590 return InstDesc(Kind == RecurKind::FAdd, I, UAI); 591 case Instruction::Select: 592 if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) 593 return isConditionalRdxPattern(Kind, I); 594 LLVM_FALLTHROUGH; 595 case Instruction::FCmp: 596 case Instruction::ICmp: 597 if (!isIntMinMaxRecurrenceKind(Kind) && 598 (!HasFunNoNaNAttr || !isFPMinMaxRecurrenceKind(Kind))) 599 return InstDesc(false, I); 600 return isMinMaxSelectCmpPattern(I, Prev); 601 } 602 } 603 604 bool RecurrenceDescriptor::hasMultipleUsesOf( 605 Instruction *I, SmallPtrSetImpl<Instruction *> &Insts, 606 unsigned MaxNumUses) { 607 unsigned NumUses = 0; 608 for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; 609 ++Use) { 610 if (Insts.count(dyn_cast<Instruction>(*Use))) 611 ++NumUses; 612 if (NumUses > MaxNumUses) 613 return true; 614 } 615 616 return false; 617 } 618 bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop, 619 RecurrenceDescriptor &RedDes, 620 DemandedBits *DB, AssumptionCache *AC, 621 DominatorTree *DT) { 622 623 BasicBlock *Header = TheLoop->getHeader(); 624 Function &F = *Header->getParent(); 625 bool HasFunNoNaNAttr = 626 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true"; 627 628 if (AddReductionVar(Phi, RecurKind::Add, TheLoop, HasFunNoNaNAttr, RedDes, DB, 629 AC, DT)) { 630 LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n"); 631 return true; 632 } 633 if (AddReductionVar(Phi, RecurKind::Mul, TheLoop, HasFunNoNaNAttr, RedDes, DB, 634 AC, DT)) { 635 LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n"); 636 return true; 637 } 638 if (AddReductionVar(Phi, RecurKind::Or, TheLoop, HasFunNoNaNAttr, RedDes, DB, 639 AC, DT)) { 640 LLVM_DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n"); 641 return true; 642 } 643 if (AddReductionVar(Phi, RecurKind::And, TheLoop, HasFunNoNaNAttr, RedDes, DB, 644 AC, DT)) { 645 LLVM_DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n"); 646 return true; 647 } 648 if (AddReductionVar(Phi, RecurKind::Xor, TheLoop, HasFunNoNaNAttr, RedDes, DB, 649 AC, DT)) { 650 LLVM_DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n"); 651 return true; 652 } 653 if (AddReductionVar(Phi, RecurKind::SMax, TheLoop, HasFunNoNaNAttr, RedDes, 654 DB, AC, DT)) { 655 LLVM_DEBUG(dbgs() << "Found a SMAX reduction PHI." << *Phi << "\n"); 656 return true; 657 } 658 if (AddReductionVar(Phi, RecurKind::SMin, TheLoop, HasFunNoNaNAttr, RedDes, 659 DB, AC, DT)) { 660 LLVM_DEBUG(dbgs() << "Found a SMIN reduction PHI." << *Phi << "\n"); 661 return true; 662 } 663 if (AddReductionVar(Phi, RecurKind::UMax, TheLoop, HasFunNoNaNAttr, RedDes, 664 DB, AC, DT)) { 665 LLVM_DEBUG(dbgs() << "Found a UMAX reduction PHI." << *Phi << "\n"); 666 return true; 667 } 668 if (AddReductionVar(Phi, RecurKind::UMin, TheLoop, HasFunNoNaNAttr, RedDes, 669 DB, AC, DT)) { 670 LLVM_DEBUG(dbgs() << "Found a UMIN reduction PHI." << *Phi << "\n"); 671 return true; 672 } 673 if (AddReductionVar(Phi, RecurKind::FMul, TheLoop, HasFunNoNaNAttr, RedDes, 674 DB, AC, DT)) { 675 LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n"); 676 return true; 677 } 678 if (AddReductionVar(Phi, RecurKind::FAdd, TheLoop, HasFunNoNaNAttr, RedDes, 679 DB, AC, DT)) { 680 LLVM_DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n"); 681 return true; 682 } 683 if (AddReductionVar(Phi, RecurKind::FMax, TheLoop, HasFunNoNaNAttr, RedDes, 684 DB, AC, DT)) { 685 LLVM_DEBUG(dbgs() << "Found a float MAX reduction PHI." << *Phi << "\n"); 686 return true; 687 } 688 if (AddReductionVar(Phi, RecurKind::FMin, TheLoop, HasFunNoNaNAttr, RedDes, 689 DB, AC, DT)) { 690 LLVM_DEBUG(dbgs() << "Found a float MIN reduction PHI." << *Phi << "\n"); 691 return true; 692 } 693 // Not a reduction of known type. 694 return false; 695 } 696 697 bool RecurrenceDescriptor::isFirstOrderRecurrence( 698 PHINode *Phi, Loop *TheLoop, 699 DenseMap<Instruction *, Instruction *> &SinkAfter, DominatorTree *DT) { 700 701 // Ensure the phi node is in the loop header and has two incoming values. 702 if (Phi->getParent() != TheLoop->getHeader() || 703 Phi->getNumIncomingValues() != 2) 704 return false; 705 706 // Ensure the loop has a preheader and a single latch block. The loop 707 // vectorizer will need the latch to set up the next iteration of the loop. 708 auto *Preheader = TheLoop->getLoopPreheader(); 709 auto *Latch = TheLoop->getLoopLatch(); 710 if (!Preheader || !Latch) 711 return false; 712 713 // Ensure the phi node's incoming blocks are the loop preheader and latch. 714 if (Phi->getBasicBlockIndex(Preheader) < 0 || 715 Phi->getBasicBlockIndex(Latch) < 0) 716 return false; 717 718 // Get the previous value. The previous value comes from the latch edge while 719 // the initial value comes form the preheader edge. 720 auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch)); 721 if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) || 722 SinkAfter.count(Previous)) // Cannot rely on dominance due to motion. 723 return false; 724 725 // Ensure every user of the phi node is dominated by the previous value. 726 // The dominance requirement ensures the loop vectorizer will not need to 727 // vectorize the initial value prior to the first iteration of the loop. 728 // TODO: Consider extending this sinking to handle memory instructions and 729 // phis with multiple users. 730 731 // Returns true, if all users of I are dominated by DominatedBy. 732 auto allUsesDominatedBy = [DT](Instruction *I, Instruction *DominatedBy) { 733 return all_of(I->uses(), [DT, DominatedBy](Use &U) { 734 return DT->dominates(DominatedBy, U); 735 }); 736 }; 737 738 if (Phi->hasOneUse()) { 739 Instruction *I = Phi->user_back(); 740 741 // If the user of the PHI is also the incoming value, we potentially have a 742 // reduction and which cannot be handled by sinking. 743 if (Previous == I) 744 return false; 745 746 // We cannot sink terminator instructions. 747 if (I->getParent()->getTerminator() == I) 748 return false; 749 750 // Do not try to sink an instruction multiple times (if multiple operands 751 // are first order recurrences). 752 // TODO: We can support this case, by sinking the instruction after the 753 // 'deepest' previous instruction. 754 if (SinkAfter.find(I) != SinkAfter.end()) 755 return false; 756 757 if (DT->dominates(Previous, I)) // We already are good w/o sinking. 758 return true; 759 760 // We can sink any instruction without side effects, as long as all users 761 // are dominated by the instruction we are sinking after. 762 if (I->getParent() == Phi->getParent() && !I->mayHaveSideEffects() && 763 allUsesDominatedBy(I, Previous)) { 764 SinkAfter[I] = Previous; 765 return true; 766 } 767 } 768 769 return allUsesDominatedBy(Phi, Previous); 770 } 771 772 /// This function returns the identity element (or neutral element) for 773 /// the operation K. 774 Constant *RecurrenceDescriptor::getRecurrenceIdentity(RecurKind K, Type *Tp) { 775 switch (K) { 776 case RecurKind::Xor: 777 case RecurKind::Add: 778 case RecurKind::Or: 779 // Adding, Xoring, Oring zero to a number does not change it. 780 return ConstantInt::get(Tp, 0); 781 case RecurKind::Mul: 782 // Multiplying a number by 1 does not change it. 783 return ConstantInt::get(Tp, 1); 784 case RecurKind::And: 785 // AND-ing a number with an all-1 value does not change it. 786 return ConstantInt::get(Tp, -1, true); 787 case RecurKind::FMul: 788 // Multiplying a number by 1 does not change it. 789 return ConstantFP::get(Tp, 1.0L); 790 case RecurKind::FAdd: 791 // Adding zero to a number does not change it. 792 return ConstantFP::get(Tp, 0.0L); 793 case RecurKind::UMin: 794 return ConstantInt::get(Tp, -1); 795 case RecurKind::UMax: 796 return ConstantInt::get(Tp, 0); 797 case RecurKind::SMin: 798 return ConstantInt::get(Tp, 799 APInt::getSignedMaxValue(Tp->getIntegerBitWidth())); 800 case RecurKind::SMax: 801 return ConstantInt::get(Tp, 802 APInt::getSignedMinValue(Tp->getIntegerBitWidth())); 803 case RecurKind::FMin: 804 return ConstantFP::getInfinity(Tp, true); 805 case RecurKind::FMax: 806 return ConstantFP::getInfinity(Tp, false); 807 default: 808 llvm_unreachable("Unknown recurrence kind"); 809 } 810 } 811 812 unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) { 813 switch (Kind) { 814 case RecurKind::Add: 815 return Instruction::Add; 816 case RecurKind::Mul: 817 return Instruction::Mul; 818 case RecurKind::Or: 819 return Instruction::Or; 820 case RecurKind::And: 821 return Instruction::And; 822 case RecurKind::Xor: 823 return Instruction::Xor; 824 case RecurKind::FMul: 825 return Instruction::FMul; 826 case RecurKind::FAdd: 827 return Instruction::FAdd; 828 case RecurKind::SMax: 829 case RecurKind::SMin: 830 case RecurKind::UMax: 831 case RecurKind::UMin: 832 return Instruction::ICmp; 833 case RecurKind::FMax: 834 case RecurKind::FMin: 835 return Instruction::FCmp; 836 default: 837 llvm_unreachable("Unknown recurrence operation"); 838 } 839 } 840 841 SmallVector<Instruction *, 4> 842 RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const { 843 SmallVector<Instruction *, 4> ReductionOperations; 844 unsigned RedOp = getOpcode(Kind); 845 846 // Search down from the Phi to the LoopExitInstr, looking for instructions 847 // with a single user of the correct type for the reduction. 848 849 // Note that we check that the type of the operand is correct for each item in 850 // the chain, including the last (the loop exit value). This can come up from 851 // sub, which would otherwise be treated as an add reduction. MinMax also need 852 // to check for a pair of icmp/select, for which we use getNextInstruction and 853 // isCorrectOpcode functions to step the right number of instruction, and 854 // check the icmp/select pair. 855 // FIXME: We also do not attempt to look through Phi/Select's yet, which might 856 // be part of the reduction chain, or attempt to looks through And's to find a 857 // smaller bitwidth. Subs are also currently not allowed (which are usually 858 // treated as part of a add reduction) as they are expected to generally be 859 // more expensive than out-of-loop reductions, and need to be costed more 860 // carefully. 861 unsigned ExpectedUses = 1; 862 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) 863 ExpectedUses = 2; 864 865 auto getNextInstruction = [&](Instruction *Cur) { 866 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) { 867 // We are expecting a icmp/select pair, which we go to the next select 868 // instruction if we can. We already know that Cur has 2 uses. 869 if (isa<SelectInst>(*Cur->user_begin())) 870 return cast<Instruction>(*Cur->user_begin()); 871 else 872 return cast<Instruction>(*std::next(Cur->user_begin())); 873 } 874 return cast<Instruction>(*Cur->user_begin()); 875 }; 876 auto isCorrectOpcode = [&](Instruction *Cur) { 877 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) { 878 Value *LHS, *RHS; 879 return SelectPatternResult::isMinOrMax( 880 matchSelectPattern(Cur, LHS, RHS).Flavor); 881 } 882 return Cur->getOpcode() == RedOp; 883 }; 884 885 // The loop exit instruction we check first (as a quick test) but add last. We 886 // check the opcode is correct (and dont allow them to be Subs) and that they 887 // have expected to have the expected number of uses. They will have one use 888 // from the phi and one from a LCSSA value, no matter the type. 889 if (!isCorrectOpcode(LoopExitInstr) || !LoopExitInstr->hasNUses(2)) 890 return {}; 891 892 // Check that the Phi has one (or two for min/max) uses. 893 if (!Phi->hasNUses(ExpectedUses)) 894 return {}; 895 Instruction *Cur = getNextInstruction(Phi); 896 897 // Each other instruction in the chain should have the expected number of uses 898 // and be the correct opcode. 899 while (Cur != LoopExitInstr) { 900 if (!isCorrectOpcode(Cur) || !Cur->hasNUses(ExpectedUses)) 901 return {}; 902 903 ReductionOperations.push_back(Cur); 904 Cur = getNextInstruction(Cur); 905 } 906 907 ReductionOperations.push_back(Cur); 908 return ReductionOperations; 909 } 910 911 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K, 912 const SCEV *Step, BinaryOperator *BOp, 913 SmallVectorImpl<Instruction *> *Casts) 914 : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) { 915 assert(IK != IK_NoInduction && "Not an induction"); 916 917 // Start value type should match the induction kind and the value 918 // itself should not be null. 919 assert(StartValue && "StartValue is null"); 920 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) && 921 "StartValue is not a pointer for pointer induction"); 922 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) && 923 "StartValue is not an integer for integer induction"); 924 925 // Check the Step Value. It should be non-zero integer value. 926 assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) && 927 "Step value is zero"); 928 929 assert((IK != IK_PtrInduction || getConstIntStepValue()) && 930 "Step value should be constant for pointer induction"); 931 assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) && 932 "StepValue is not an integer"); 933 934 assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) && 935 "StepValue is not FP for FpInduction"); 936 assert((IK != IK_FpInduction || 937 (InductionBinOp && 938 (InductionBinOp->getOpcode() == Instruction::FAdd || 939 InductionBinOp->getOpcode() == Instruction::FSub))) && 940 "Binary opcode should be specified for FP induction"); 941 942 if (Casts) { 943 for (auto &Inst : *Casts) { 944 RedundantCasts.push_back(Inst); 945 } 946 } 947 } 948 949 ConstantInt *InductionDescriptor::getConstIntStepValue() const { 950 if (isa<SCEVConstant>(Step)) 951 return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue()); 952 return nullptr; 953 } 954 955 bool InductionDescriptor::isFPInductionPHI(PHINode *Phi, const Loop *TheLoop, 956 ScalarEvolution *SE, 957 InductionDescriptor &D) { 958 959 // Here we only handle FP induction variables. 960 assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type"); 961 962 if (TheLoop->getHeader() != Phi->getParent()) 963 return false; 964 965 // The loop may have multiple entrances or multiple exits; we can analyze 966 // this phi if it has a unique entry value and a unique backedge value. 967 if (Phi->getNumIncomingValues() != 2) 968 return false; 969 Value *BEValue = nullptr, *StartValue = nullptr; 970 if (TheLoop->contains(Phi->getIncomingBlock(0))) { 971 BEValue = Phi->getIncomingValue(0); 972 StartValue = Phi->getIncomingValue(1); 973 } else { 974 assert(TheLoop->contains(Phi->getIncomingBlock(1)) && 975 "Unexpected Phi node in the loop"); 976 BEValue = Phi->getIncomingValue(1); 977 StartValue = Phi->getIncomingValue(0); 978 } 979 980 BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue); 981 if (!BOp) 982 return false; 983 984 Value *Addend = nullptr; 985 if (BOp->getOpcode() == Instruction::FAdd) { 986 if (BOp->getOperand(0) == Phi) 987 Addend = BOp->getOperand(1); 988 else if (BOp->getOperand(1) == Phi) 989 Addend = BOp->getOperand(0); 990 } else if (BOp->getOpcode() == Instruction::FSub) 991 if (BOp->getOperand(0) == Phi) 992 Addend = BOp->getOperand(1); 993 994 if (!Addend) 995 return false; 996 997 // The addend should be loop invariant 998 if (auto *I = dyn_cast<Instruction>(Addend)) 999 if (TheLoop->contains(I)) 1000 return false; 1001 1002 // FP Step has unknown SCEV 1003 const SCEV *Step = SE->getUnknown(Addend); 1004 D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp); 1005 return true; 1006 } 1007 1008 /// This function is called when we suspect that the update-chain of a phi node 1009 /// (whose symbolic SCEV expression sin \p PhiScev) contains redundant casts, 1010 /// that can be ignored. (This can happen when the PSCEV rewriter adds a runtime 1011 /// predicate P under which the SCEV expression for the phi can be the 1012 /// AddRecurrence \p AR; See createAddRecFromPHIWithCast). We want to find the 1013 /// cast instructions that are involved in the update-chain of this induction. 1014 /// A caller that adds the required runtime predicate can be free to drop these 1015 /// cast instructions, and compute the phi using \p AR (instead of some scev 1016 /// expression with casts). 1017 /// 1018 /// For example, without a predicate the scev expression can take the following 1019 /// form: 1020 /// (Ext ix (Trunc iy ( Start + i*Step ) to ix) to iy) 1021 /// 1022 /// It corresponds to the following IR sequence: 1023 /// %for.body: 1024 /// %x = phi i64 [ 0, %ph ], [ %add, %for.body ] 1025 /// %casted_phi = "ExtTrunc i64 %x" 1026 /// %add = add i64 %casted_phi, %step 1027 /// 1028 /// where %x is given in \p PN, 1029 /// PSE.getSCEV(%x) is equal to PSE.getSCEV(%casted_phi) under a predicate, 1030 /// and the IR sequence that "ExtTrunc i64 %x" represents can take one of 1031 /// several forms, for example, such as: 1032 /// ExtTrunc1: %casted_phi = and %x, 2^n-1 1033 /// or: 1034 /// ExtTrunc2: %t = shl %x, m 1035 /// %casted_phi = ashr %t, m 1036 /// 1037 /// If we are able to find such sequence, we return the instructions 1038 /// we found, namely %casted_phi and the instructions on its use-def chain up 1039 /// to the phi (not including the phi). 1040 static bool getCastsForInductionPHI(PredicatedScalarEvolution &PSE, 1041 const SCEVUnknown *PhiScev, 1042 const SCEVAddRecExpr *AR, 1043 SmallVectorImpl<Instruction *> &CastInsts) { 1044 1045 assert(CastInsts.empty() && "CastInsts is expected to be empty."); 1046 auto *PN = cast<PHINode>(PhiScev->getValue()); 1047 assert(PSE.getSCEV(PN) == AR && "Unexpected phi node SCEV expression"); 1048 const Loop *L = AR->getLoop(); 1049 1050 // Find any cast instructions that participate in the def-use chain of 1051 // PhiScev in the loop. 1052 // FORNOW/TODO: We currently expect the def-use chain to include only 1053 // two-operand instructions, where one of the operands is an invariant. 1054 // createAddRecFromPHIWithCasts() currently does not support anything more 1055 // involved than that, so we keep the search simple. This can be 1056 // extended/generalized as needed. 1057 1058 auto getDef = [&](const Value *Val) -> Value * { 1059 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val); 1060 if (!BinOp) 1061 return nullptr; 1062 Value *Op0 = BinOp->getOperand(0); 1063 Value *Op1 = BinOp->getOperand(1); 1064 Value *Def = nullptr; 1065 if (L->isLoopInvariant(Op0)) 1066 Def = Op1; 1067 else if (L->isLoopInvariant(Op1)) 1068 Def = Op0; 1069 return Def; 1070 }; 1071 1072 // Look for the instruction that defines the induction via the 1073 // loop backedge. 1074 BasicBlock *Latch = L->getLoopLatch(); 1075 if (!Latch) 1076 return false; 1077 Value *Val = PN->getIncomingValueForBlock(Latch); 1078 if (!Val) 1079 return false; 1080 1081 // Follow the def-use chain until the induction phi is reached. 1082 // If on the way we encounter a Value that has the same SCEV Expr as the 1083 // phi node, we can consider the instructions we visit from that point 1084 // as part of the cast-sequence that can be ignored. 1085 bool InCastSequence = false; 1086 auto *Inst = dyn_cast<Instruction>(Val); 1087 while (Val != PN) { 1088 // If we encountered a phi node other than PN, or if we left the loop, 1089 // we bail out. 1090 if (!Inst || !L->contains(Inst)) { 1091 return false; 1092 } 1093 auto *AddRec = dyn_cast<SCEVAddRecExpr>(PSE.getSCEV(Val)); 1094 if (AddRec && PSE.areAddRecsEqualWithPreds(AddRec, AR)) 1095 InCastSequence = true; 1096 if (InCastSequence) { 1097 // Only the last instruction in the cast sequence is expected to have 1098 // uses outside the induction def-use chain. 1099 if (!CastInsts.empty()) 1100 if (!Inst->hasOneUse()) 1101 return false; 1102 CastInsts.push_back(Inst); 1103 } 1104 Val = getDef(Val); 1105 if (!Val) 1106 return false; 1107 Inst = dyn_cast<Instruction>(Val); 1108 } 1109 1110 return InCastSequence; 1111 } 1112 1113 bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop, 1114 PredicatedScalarEvolution &PSE, 1115 InductionDescriptor &D, bool Assume) { 1116 Type *PhiTy = Phi->getType(); 1117 1118 // Handle integer and pointer inductions variables. 1119 // Now we handle also FP induction but not trying to make a 1120 // recurrent expression from the PHI node in-place. 1121 1122 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() && !PhiTy->isFloatTy() && 1123 !PhiTy->isDoubleTy() && !PhiTy->isHalfTy()) 1124 return false; 1125 1126 if (PhiTy->isFloatingPointTy()) 1127 return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D); 1128 1129 const SCEV *PhiScev = PSE.getSCEV(Phi); 1130 const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 1131 1132 // We need this expression to be an AddRecExpr. 1133 if (Assume && !AR) 1134 AR = PSE.getAsAddRec(Phi); 1135 1136 if (!AR) { 1137 LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 1138 return false; 1139 } 1140 1141 // Record any Cast instructions that participate in the induction update 1142 const auto *SymbolicPhi = dyn_cast<SCEVUnknown>(PhiScev); 1143 // If we started from an UnknownSCEV, and managed to build an addRecurrence 1144 // only after enabling Assume with PSCEV, this means we may have encountered 1145 // cast instructions that required adding a runtime check in order to 1146 // guarantee the correctness of the AddRecurrence respresentation of the 1147 // induction. 1148 if (PhiScev != AR && SymbolicPhi) { 1149 SmallVector<Instruction *, 2> Casts; 1150 if (getCastsForInductionPHI(PSE, SymbolicPhi, AR, Casts)) 1151 return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR, &Casts); 1152 } 1153 1154 return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR); 1155 } 1156 1157 bool InductionDescriptor::isInductionPHI( 1158 PHINode *Phi, const Loop *TheLoop, ScalarEvolution *SE, 1159 InductionDescriptor &D, const SCEV *Expr, 1160 SmallVectorImpl<Instruction *> *CastsToIgnore) { 1161 Type *PhiTy = Phi->getType(); 1162 // We only handle integer and pointer inductions variables. 1163 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 1164 return false; 1165 1166 // Check that the PHI is consecutive. 1167 const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi); 1168 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 1169 1170 if (!AR) { 1171 LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 1172 return false; 1173 } 1174 1175 if (AR->getLoop() != TheLoop) { 1176 // FIXME: We should treat this as a uniform. Unfortunately, we 1177 // don't currently know how to handled uniform PHIs. 1178 LLVM_DEBUG( 1179 dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n"); 1180 return false; 1181 } 1182 1183 Value *StartValue = 1184 Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader()); 1185 1186 BasicBlock *Latch = AR->getLoop()->getLoopLatch(); 1187 if (!Latch) 1188 return false; 1189 BinaryOperator *BOp = 1190 dyn_cast<BinaryOperator>(Phi->getIncomingValueForBlock(Latch)); 1191 1192 const SCEV *Step = AR->getStepRecurrence(*SE); 1193 // Calculate the pointer stride and check if it is consecutive. 1194 // The stride may be a constant or a loop invariant integer value. 1195 const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step); 1196 if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop)) 1197 return false; 1198 1199 if (PhiTy->isIntegerTy()) { 1200 D = InductionDescriptor(StartValue, IK_IntInduction, Step, BOp, 1201 CastsToIgnore); 1202 return true; 1203 } 1204 1205 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 1206 // Pointer induction should be a constant. 1207 if (!ConstStep) 1208 return false; 1209 1210 ConstantInt *CV = ConstStep->getValue(); 1211 Type *PointerElementType = PhiTy->getPointerElementType(); 1212 // The pointer stride cannot be determined if the pointer element type is not 1213 // sized. 1214 if (!PointerElementType->isSized()) 1215 return false; 1216 1217 const DataLayout &DL = Phi->getModule()->getDataLayout(); 1218 int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType)); 1219 if (!Size) 1220 return false; 1221 1222 int64_t CVSize = CV->getSExtValue(); 1223 if (CVSize % Size) 1224 return false; 1225 auto *StepValue = 1226 SE->getConstant(CV->getType(), CVSize / Size, true /* signed */); 1227 D = InductionDescriptor(StartValue, IK_PtrInduction, StepValue, BOp); 1228 return true; 1229 } 1230