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