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/Analysis/DemandedBits.h" 15 #include "llvm/Analysis/LoopInfo.h" 16 #include "llvm/Analysis/ScalarEvolution.h" 17 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 18 #include "llvm/Analysis/ValueTracking.h" 19 #include "llvm/IR/Dominators.h" 20 #include "llvm/IR/Instructions.h" 21 #include "llvm/IR/Module.h" 22 #include "llvm/IR/PatternMatch.h" 23 #include "llvm/IR/ValueHandle.h" 24 #include "llvm/Support/Debug.h" 25 #include "llvm/Support/KnownBits.h" 26 27 #include <set> 28 29 using namespace llvm; 30 using namespace llvm::PatternMatch; 31 32 #define DEBUG_TYPE "iv-descriptors" 33 34 bool RecurrenceDescriptor::areAllUsesIn(Instruction *I, 35 SmallPtrSetImpl<Instruction *> &Set) { 36 for (const Use &Use : I->operands()) 37 if (!Set.count(dyn_cast<Instruction>(Use))) 38 return false; 39 return true; 40 } 41 42 bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) { 43 switch (Kind) { 44 default: 45 break; 46 case RecurKind::Add: 47 case RecurKind::Mul: 48 case RecurKind::Or: 49 case RecurKind::And: 50 case RecurKind::Xor: 51 case RecurKind::SMax: 52 case RecurKind::SMin: 53 case RecurKind::UMax: 54 case RecurKind::UMin: 55 case RecurKind::SelectICmp: 56 case RecurKind::SelectFCmp: 57 return true; 58 } 59 return false; 60 } 61 62 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurKind Kind) { 63 return (Kind != RecurKind::None) && !isIntegerRecurrenceKind(Kind); 64 } 65 66 /// Determines if Phi may have been type-promoted. If Phi has a single user 67 /// that ANDs the Phi with a type mask, return the user. RT is updated to 68 /// account for the narrower bit width represented by the mask, and the AND 69 /// instruction is added to CI. 70 static Instruction *lookThroughAnd(PHINode *Phi, Type *&RT, 71 SmallPtrSetImpl<Instruction *> &Visited, 72 SmallPtrSetImpl<Instruction *> &CI) { 73 if (!Phi->hasOneUse()) 74 return Phi; 75 76 const APInt *M = nullptr; 77 Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser()); 78 79 // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT 80 // with a new integer type of the corresponding bit width. 81 if (match(J, m_c_And(m_Instruction(I), m_APInt(M)))) { 82 int32_t Bits = (*M + 1).exactLogBase2(); 83 if (Bits > 0) { 84 RT = IntegerType::get(Phi->getContext(), Bits); 85 Visited.insert(Phi); 86 CI.insert(J); 87 return J; 88 } 89 } 90 return Phi; 91 } 92 93 /// Compute the minimal bit width needed to represent a reduction whose exit 94 /// instruction is given by Exit. 95 static std::pair<Type *, bool> computeRecurrenceType(Instruction *Exit, 96 DemandedBits *DB, 97 AssumptionCache *AC, 98 DominatorTree *DT) { 99 bool IsSigned = false; 100 const DataLayout &DL = Exit->getModule()->getDataLayout(); 101 uint64_t MaxBitWidth = DL.getTypeSizeInBits(Exit->getType()); 102 103 if (DB) { 104 // Use the demanded bits analysis to determine the bits that are live out 105 // of the exit instruction, rounding up to the nearest power of two. If the 106 // use of demanded bits results in a smaller bit width, we know the value 107 // must be positive (i.e., IsSigned = false), because if this were not the 108 // case, the sign bit would have been demanded. 109 auto Mask = DB->getDemandedBits(Exit); 110 MaxBitWidth = Mask.getBitWidth() - Mask.countLeadingZeros(); 111 } 112 113 if (MaxBitWidth == DL.getTypeSizeInBits(Exit->getType()) && AC && DT) { 114 // If demanded bits wasn't able to limit the bit width, we can try to use 115 // value tracking instead. This can be the case, for example, if the value 116 // may be negative. 117 auto NumSignBits = ComputeNumSignBits(Exit, DL, 0, AC, nullptr, DT); 118 auto NumTypeBits = DL.getTypeSizeInBits(Exit->getType()); 119 MaxBitWidth = NumTypeBits - NumSignBits; 120 KnownBits Bits = computeKnownBits(Exit, DL); 121 if (!Bits.isNonNegative()) { 122 // If the value is not known to be non-negative, we set IsSigned to true, 123 // meaning that we will use sext instructions instead of zext 124 // instructions to restore the original type. 125 IsSigned = true; 126 // Make sure at at least one sign bit is included in the result, so it 127 // will get properly sign-extended. 128 ++MaxBitWidth; 129 } 130 } 131 if (!isPowerOf2_64(MaxBitWidth)) 132 MaxBitWidth = NextPowerOf2(MaxBitWidth); 133 134 return std::make_pair(Type::getIntNTy(Exit->getContext(), MaxBitWidth), 135 IsSigned); 136 } 137 138 /// Collect cast instructions that can be ignored in the vectorizer's cost 139 /// model, given a reduction exit value and the minimal type in which the 140 // reduction can be represented. Also search casts to the recurrence type 141 // to find the minimum width used by the recurrence. 142 static void collectCastInstrs(Loop *TheLoop, Instruction *Exit, 143 Type *RecurrenceType, 144 SmallPtrSetImpl<Instruction *> &Casts, 145 unsigned &MinWidthCastToRecurTy) { 146 147 SmallVector<Instruction *, 8> Worklist; 148 SmallPtrSet<Instruction *, 8> Visited; 149 Worklist.push_back(Exit); 150 MinWidthCastToRecurTy = -1U; 151 152 while (!Worklist.empty()) { 153 Instruction *Val = Worklist.pop_back_val(); 154 Visited.insert(Val); 155 if (auto *Cast = dyn_cast<CastInst>(Val)) { 156 if (Cast->getSrcTy() == RecurrenceType) { 157 // If the source type of a cast instruction is equal to the recurrence 158 // type, it will be eliminated, and should be ignored in the vectorizer 159 // cost model. 160 Casts.insert(Cast); 161 continue; 162 } 163 if (Cast->getDestTy() == RecurrenceType) { 164 // The minimum width used by the recurrence is found by checking for 165 // casts on its operands. The minimum width is used by the vectorizer 166 // when finding the widest type for in-loop reductions without any 167 // loads/stores. 168 MinWidthCastToRecurTy = std::min<unsigned>( 169 MinWidthCastToRecurTy, Cast->getSrcTy()->getScalarSizeInBits()); 170 continue; 171 } 172 } 173 // Add all operands to the work list if they are loop-varying values that 174 // we haven't yet visited. 175 for (Value *O : cast<User>(Val)->operands()) 176 if (auto *I = dyn_cast<Instruction>(O)) 177 if (TheLoop->contains(I) && !Visited.count(I)) 178 Worklist.push_back(I); 179 } 180 } 181 182 // Check if a given Phi node can be recognized as an ordered reduction for 183 // vectorizing floating point operations without unsafe math. 184 static bool checkOrderedReduction(RecurKind Kind, Instruction *ExactFPMathInst, 185 Instruction *Exit, PHINode *Phi) { 186 // Currently only FAdd and FMulAdd are supported. 187 if (Kind != RecurKind::FAdd && Kind != RecurKind::FMulAdd) 188 return false; 189 190 if (Kind == RecurKind::FAdd && Exit->getOpcode() != Instruction::FAdd) 191 return false; 192 193 if (Kind == RecurKind::FMulAdd && 194 !RecurrenceDescriptor::isFMulAddIntrinsic(Exit)) 195 return false; 196 197 // Ensure the exit instruction has only one user other than the reduction PHI 198 if (Exit != ExactFPMathInst || Exit->hasNUsesOrMore(3)) 199 return false; 200 201 // The only pattern accepted is the one in which the reduction PHI 202 // is used as one of the operands of the exit instruction 203 auto *Op0 = Exit->getOperand(0); 204 auto *Op1 = Exit->getOperand(1); 205 if (Kind == RecurKind::FAdd && Op0 != Phi && Op1 != Phi) 206 return false; 207 if (Kind == RecurKind::FMulAdd && Exit->getOperand(2) != Phi) 208 return false; 209 210 LLVM_DEBUG(dbgs() << "LV: Found an ordered reduction: Phi: " << *Phi 211 << ", ExitInst: " << *Exit << "\n"); 212 213 return true; 214 } 215 216 bool RecurrenceDescriptor::AddReductionVar( 217 PHINode *Phi, RecurKind Kind, Loop *TheLoop, FastMathFlags FuncFMF, 218 RecurrenceDescriptor &RedDes, DemandedBits *DB, AssumptionCache *AC, 219 DominatorTree *DT, ScalarEvolution *SE) { 220 if (Phi->getNumIncomingValues() != 2) 221 return false; 222 223 // Reduction variables are only found in the loop header block. 224 if (Phi->getParent() != TheLoop->getHeader()) 225 return false; 226 227 // Obtain the reduction start value from the value that comes from the loop 228 // preheader. 229 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 230 231 // ExitInstruction is the single value which is used outside the loop. 232 // We only allow for a single reduction value to be used outside the loop. 233 // This includes users of the reduction, variables (which form a cycle 234 // which ends in the phi node). 235 Instruction *ExitInstruction = nullptr; 236 237 // Variable to keep last visited store instruction. By the end of the 238 // algorithm this variable will be either empty or having intermediate 239 // reduction value stored in invariant address. 240 StoreInst *IntermediateStore = nullptr; 241 242 // Indicates that we found a reduction operation in our scan. 243 bool FoundReduxOp = false; 244 245 // We start with the PHI node and scan for all of the users of this 246 // instruction. All users must be instructions that can be used as reduction 247 // variables (such as ADD). We must have a single out-of-block user. The cycle 248 // must include the original PHI. 249 bool FoundStartPHI = false; 250 251 // To recognize min/max patterns formed by a icmp select sequence, we store 252 // the number of instruction we saw from the recognized min/max pattern, 253 // to make sure we only see exactly the two instructions. 254 unsigned NumCmpSelectPatternInst = 0; 255 InstDesc ReduxDesc(false, nullptr); 256 257 // Data used for determining if the recurrence has been type-promoted. 258 Type *RecurrenceType = Phi->getType(); 259 SmallPtrSet<Instruction *, 4> CastInsts; 260 unsigned MinWidthCastToRecurrenceType; 261 Instruction *Start = Phi; 262 bool IsSigned = false; 263 264 SmallPtrSet<Instruction *, 8> VisitedInsts; 265 SmallVector<Instruction *, 8> Worklist; 266 267 // Return early if the recurrence kind does not match the type of Phi. If the 268 // recurrence kind is arithmetic, we attempt to look through AND operations 269 // resulting from the type promotion performed by InstCombine. Vector 270 // operations are not limited to the legal integer widths, so we may be able 271 // to evaluate the reduction in the narrower width. 272 if (RecurrenceType->isFloatingPointTy()) { 273 if (!isFloatingPointRecurrenceKind(Kind)) 274 return false; 275 } else if (RecurrenceType->isIntegerTy()) { 276 if (!isIntegerRecurrenceKind(Kind)) 277 return false; 278 if (!isMinMaxRecurrenceKind(Kind)) 279 Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts); 280 } else { 281 // Pointer min/max may exist, but it is not supported as a reduction op. 282 return false; 283 } 284 285 Worklist.push_back(Start); 286 VisitedInsts.insert(Start); 287 288 // Start with all flags set because we will intersect this with the reduction 289 // flags from all the reduction operations. 290 FastMathFlags FMF = FastMathFlags::getFast(); 291 292 // The first instruction in the use-def chain of the Phi node that requires 293 // exact floating point operations. 294 Instruction *ExactFPMathInst = nullptr; 295 296 // A value in the reduction can be used: 297 // - By the reduction: 298 // - Reduction operation: 299 // - One use of reduction value (safe). 300 // - Multiple use of reduction value (not safe). 301 // - PHI: 302 // - All uses of the PHI must be the reduction (safe). 303 // - Otherwise, not safe. 304 // - By instructions outside of the loop (safe). 305 // * One value may have several outside users, but all outside 306 // uses must be of the same value. 307 // - By store instructions with a loop invariant address (safe with 308 // the following restrictions): 309 // * If there are several stores, all must have the same address. 310 // * Final value should be stored in that loop invariant address. 311 // - By an instruction that is not part of the reduction (not safe). 312 // This is either: 313 // * An instruction type other than PHI or the reduction operation. 314 // * A PHI in the header other than the initial PHI. 315 while (!Worklist.empty()) { 316 Instruction *Cur = Worklist.pop_back_val(); 317 318 // Store instructions are allowed iff it is the store of the reduction 319 // value to the same loop invariant memory location. 320 if (auto *SI = dyn_cast<StoreInst>(Cur)) { 321 if (!SE) { 322 LLVM_DEBUG(dbgs() << "Store instructions are not processed without " 323 << "Scalar Evolution Analysis\n"); 324 return false; 325 } 326 327 const SCEV *PtrScev = SE->getSCEV(SI->getPointerOperand()); 328 // Check it is the same address as previous stores 329 if (IntermediateStore) { 330 const SCEV *OtherScev = 331 SE->getSCEV(IntermediateStore->getPointerOperand()); 332 333 if (OtherScev != PtrScev) { 334 LLVM_DEBUG(dbgs() << "Storing reduction value to different addresses " 335 << "inside the loop: " << *SI->getPointerOperand() 336 << " and " 337 << *IntermediateStore->getPointerOperand() << '\n'); 338 return false; 339 } 340 } 341 342 // Check the pointer is loop invariant 343 if (!SE->isLoopInvariant(PtrScev, TheLoop)) { 344 LLVM_DEBUG(dbgs() << "Storing reduction value to non-uniform address " 345 << "inside the loop: " << *SI->getPointerOperand() 346 << '\n'); 347 return false; 348 } 349 350 // IntermediateStore is always the last store in the loop. 351 IntermediateStore = SI; 352 continue; 353 } 354 355 // No Users. 356 // If the instruction has no users then this is a broken chain and can't be 357 // a reduction variable. 358 if (Cur->use_empty()) 359 return false; 360 361 bool IsAPhi = isa<PHINode>(Cur); 362 363 // A header PHI use other than the original PHI. 364 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 365 return false; 366 367 // Reductions of instructions such as Div, and Sub is only possible if the 368 // LHS is the reduction variable. 369 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 370 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 371 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 372 return false; 373 374 // Any reduction instruction must be of one of the allowed kinds. We ignore 375 // the starting value (the Phi or an AND instruction if the Phi has been 376 // type-promoted). 377 if (Cur != Start) { 378 ReduxDesc = 379 isRecurrenceInstr(TheLoop, Phi, Cur, Kind, ReduxDesc, FuncFMF); 380 ExactFPMathInst = ExactFPMathInst == nullptr 381 ? ReduxDesc.getExactFPMathInst() 382 : ExactFPMathInst; 383 if (!ReduxDesc.isRecurrence()) 384 return false; 385 // FIXME: FMF is allowed on phi, but propagation is not handled correctly. 386 if (isa<FPMathOperator>(ReduxDesc.getPatternInst()) && !IsAPhi) { 387 FastMathFlags CurFMF = ReduxDesc.getPatternInst()->getFastMathFlags(); 388 if (auto *Sel = dyn_cast<SelectInst>(ReduxDesc.getPatternInst())) { 389 // Accept FMF on either fcmp or select of a min/max idiom. 390 // TODO: This is a hack to work-around the fact that FMF may not be 391 // assigned/propagated correctly. If that problem is fixed or we 392 // standardize on fmin/fmax via intrinsics, this can be removed. 393 if (auto *FCmp = dyn_cast<FCmpInst>(Sel->getCondition())) 394 CurFMF |= FCmp->getFastMathFlags(); 395 } 396 FMF &= CurFMF; 397 } 398 // Update this reduction kind if we matched a new instruction. 399 // TODO: Can we eliminate the need for a 2nd InstDesc by keeping 'Kind' 400 // state accurate while processing the worklist? 401 if (ReduxDesc.getRecKind() != RecurKind::None) 402 Kind = ReduxDesc.getRecKind(); 403 } 404 405 bool IsASelect = isa<SelectInst>(Cur); 406 407 // A conditional reduction operation must only have 2 or less uses in 408 // VisitedInsts. 409 if (IsASelect && (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) && 410 hasMultipleUsesOf(Cur, VisitedInsts, 2)) 411 return false; 412 413 // A reduction operation must only have one use of the reduction value. 414 if (!IsAPhi && !IsASelect && !isMinMaxRecurrenceKind(Kind) && 415 !isSelectCmpRecurrenceKind(Kind) && 416 hasMultipleUsesOf(Cur, VisitedInsts, 1)) 417 return false; 418 419 // All inputs to a PHI node must be a reduction value. 420 if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 421 return false; 422 423 if ((isIntMinMaxRecurrenceKind(Kind) || Kind == RecurKind::SelectICmp) && 424 (isa<ICmpInst>(Cur) || isa<SelectInst>(Cur))) 425 ++NumCmpSelectPatternInst; 426 if ((isFPMinMaxRecurrenceKind(Kind) || Kind == RecurKind::SelectFCmp) && 427 (isa<FCmpInst>(Cur) || isa<SelectInst>(Cur))) 428 ++NumCmpSelectPatternInst; 429 430 // Check whether we found a reduction operator. 431 FoundReduxOp |= !IsAPhi && Cur != Start; 432 433 // Process users of current instruction. Push non-PHI nodes after PHI nodes 434 // onto the stack. This way we are going to have seen all inputs to PHI 435 // nodes once we get to them. 436 SmallVector<Instruction *, 8> NonPHIs; 437 SmallVector<Instruction *, 8> PHIs; 438 for (User *U : Cur->users()) { 439 Instruction *UI = cast<Instruction>(U); 440 441 // If the user is a call to llvm.fmuladd then the instruction can only be 442 // the final operand. 443 if (isFMulAddIntrinsic(UI)) 444 if (Cur == UI->getOperand(0) || Cur == UI->getOperand(1)) 445 return false; 446 447 // Check if we found the exit user. 448 BasicBlock *Parent = UI->getParent(); 449 if (!TheLoop->contains(Parent)) { 450 // If we already know this instruction is used externally, move on to 451 // the next user. 452 if (ExitInstruction == Cur) 453 continue; 454 455 // Exit if you find multiple values used outside or if the header phi 456 // node is being used. In this case the user uses the value of the 457 // previous iteration, in which case we would loose "VF-1" iterations of 458 // the reduction operation if we vectorize. 459 if (ExitInstruction != nullptr || Cur == Phi) 460 return false; 461 462 // The instruction used by an outside user must be the last instruction 463 // before we feed back to the reduction phi. Otherwise, we loose VF-1 464 // operations on the value. 465 if (!is_contained(Phi->operands(), Cur)) 466 return false; 467 468 ExitInstruction = Cur; 469 continue; 470 } 471 472 // Process instructions only once (termination). Each reduction cycle 473 // value must only be used once, except by phi nodes and min/max 474 // reductions which are represented as a cmp followed by a select. 475 InstDesc IgnoredVal(false, nullptr); 476 if (VisitedInsts.insert(UI).second) { 477 if (isa<PHINode>(UI)) { 478 PHIs.push_back(UI); 479 } else { 480 StoreInst *SI = dyn_cast<StoreInst>(UI); 481 if (SI && SI->getPointerOperand() == Cur) { 482 // Reduction variable chain can only be stored somewhere but it 483 // can't be used as an address. 484 return false; 485 } 486 NonPHIs.push_back(UI); 487 } 488 } else if (!isa<PHINode>(UI) && 489 ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) && 490 !isa<SelectInst>(UI)) || 491 (!isConditionalRdxPattern(Kind, UI).isRecurrence() && 492 !isSelectCmpPattern(TheLoop, Phi, UI, IgnoredVal) 493 .isRecurrence() && 494 !isMinMaxPattern(UI, Kind, IgnoredVal).isRecurrence()))) 495 return false; 496 497 // Remember that we completed the cycle. 498 if (UI == Phi) 499 FoundStartPHI = true; 500 } 501 Worklist.append(PHIs.begin(), PHIs.end()); 502 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 503 } 504 505 // This means we have seen one but not the other instruction of the 506 // pattern or more than just a select and cmp. Zero implies that we saw a 507 // llvm.min/max intrinsic, which is always OK. 508 if (isMinMaxRecurrenceKind(Kind) && NumCmpSelectPatternInst != 2 && 509 NumCmpSelectPatternInst != 0) 510 return false; 511 512 if (isSelectCmpRecurrenceKind(Kind) && NumCmpSelectPatternInst != 1) 513 return false; 514 515 if (IntermediateStore) { 516 // Check that stored value goes to the phi node again. This way we make sure 517 // that the value stored in IntermediateStore is indeed the final reduction 518 // value. 519 if (!is_contained(Phi->operands(), IntermediateStore->getValueOperand())) { 520 LLVM_DEBUG(dbgs() << "Not a final reduction value stored: " 521 << *IntermediateStore << '\n'); 522 return false; 523 } 524 525 // If there is an exit instruction it's value should be stored in 526 // IntermediateStore 527 if (ExitInstruction && 528 IntermediateStore->getValueOperand() != ExitInstruction) { 529 LLVM_DEBUG(dbgs() << "Last store Instruction of reduction value does not " 530 "store last calculated value of the reduction: " 531 << *IntermediateStore << '\n'); 532 return false; 533 } 534 535 // If all uses are inside the loop (intermediate stores), then the 536 // reduction value after the loop will be the one used in the last store. 537 if (!ExitInstruction) 538 ExitInstruction = cast<Instruction>(IntermediateStore->getValueOperand()); 539 } 540 541 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 542 return false; 543 544 const bool IsOrdered = 545 checkOrderedReduction(Kind, ExactFPMathInst, ExitInstruction, Phi); 546 547 if (Start != Phi) { 548 // If the starting value is not the same as the phi node, we speculatively 549 // looked through an 'and' instruction when evaluating a potential 550 // arithmetic reduction to determine if it may have been type-promoted. 551 // 552 // We now compute the minimal bit width that is required to represent the 553 // reduction. If this is the same width that was indicated by the 'and', we 554 // can represent the reduction in the smaller type. The 'and' instruction 555 // will be eliminated since it will essentially be a cast instruction that 556 // can be ignore in the cost model. If we compute a different type than we 557 // did when evaluating the 'and', the 'and' will not be eliminated, and we 558 // will end up with different kinds of operations in the recurrence 559 // expression (e.g., IntegerAND, IntegerADD). We give up if this is 560 // the case. 561 // 562 // The vectorizer relies on InstCombine to perform the actual 563 // type-shrinking. It does this by inserting instructions to truncate the 564 // exit value of the reduction to the width indicated by RecurrenceType and 565 // then extend this value back to the original width. If IsSigned is false, 566 // a 'zext' instruction will be generated; otherwise, a 'sext' will be 567 // used. 568 // 569 // TODO: We should not rely on InstCombine to rewrite the reduction in the 570 // smaller type. We should just generate a correctly typed expression 571 // to begin with. 572 Type *ComputedType; 573 std::tie(ComputedType, IsSigned) = 574 computeRecurrenceType(ExitInstruction, DB, AC, DT); 575 if (ComputedType != RecurrenceType) 576 return false; 577 } 578 579 // Collect cast instructions and the minimum width used by the recurrence. 580 // If the starting value is not the same as the phi node and the computed 581 // recurrence type is equal to the recurrence type, the recurrence expression 582 // will be represented in a narrower or wider type. If there are any cast 583 // instructions that will be unnecessary, collect them in CastsFromRecurTy. 584 // Note that the 'and' instruction was already included in this list. 585 // 586 // TODO: A better way to represent this may be to tag in some way all the 587 // instructions that are a part of the reduction. The vectorizer cost 588 // model could then apply the recurrence type to these instructions, 589 // without needing a white list of instructions to ignore. 590 // This may also be useful for the inloop reductions, if it can be 591 // kept simple enough. 592 collectCastInstrs(TheLoop, ExitInstruction, RecurrenceType, CastInsts, 593 MinWidthCastToRecurrenceType); 594 595 // We found a reduction var if we have reached the original phi node and we 596 // only have a single instruction with out-of-loop users. 597 598 // The ExitInstruction(Instruction which is allowed to have out-of-loop users) 599 // is saved as part of the RecurrenceDescriptor. 600 601 // Save the description of this reduction variable. 602 RecurrenceDescriptor RD(RdxStart, ExitInstruction, IntermediateStore, Kind, 603 FMF, ExactFPMathInst, RecurrenceType, IsSigned, 604 IsOrdered, CastInsts, MinWidthCastToRecurrenceType); 605 RedDes = RD; 606 607 return true; 608 } 609 610 // We are looking for loops that do something like this: 611 // int r = 0; 612 // for (int i = 0; i < n; i++) { 613 // if (src[i] > 3) 614 // r = 3; 615 // } 616 // where the reduction value (r) only has two states, in this example 0 or 3. 617 // The generated LLVM IR for this type of loop will be like this: 618 // for.body: 619 // %r = phi i32 [ %spec.select, %for.body ], [ 0, %entry ] 620 // ... 621 // %cmp = icmp sgt i32 %5, 3 622 // %spec.select = select i1 %cmp, i32 3, i32 %r 623 // ... 624 // In general we can support vectorization of loops where 'r' flips between 625 // any two non-constants, provided they are loop invariant. The only thing 626 // we actually care about at the end of the loop is whether or not any lane 627 // in the selected vector is different from the start value. The final 628 // across-vector reduction after the loop simply involves choosing the start 629 // value if nothing changed (0 in the example above) or the other selected 630 // value (3 in the example above). 631 RecurrenceDescriptor::InstDesc 632 RecurrenceDescriptor::isSelectCmpPattern(Loop *Loop, PHINode *OrigPhi, 633 Instruction *I, InstDesc &Prev) { 634 // We must handle the select(cmp(),x,y) as a single instruction. Advance to 635 // the select. 636 CmpInst::Predicate Pred; 637 if (match(I, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) { 638 if (auto *Select = dyn_cast<SelectInst>(*I->user_begin())) 639 return InstDesc(Select, Prev.getRecKind()); 640 } 641 642 // Only match select with single use cmp condition. 643 if (!match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(), 644 m_Value()))) 645 return InstDesc(false, I); 646 647 SelectInst *SI = cast<SelectInst>(I); 648 Value *NonPhi = nullptr; 649 650 if (OrigPhi == dyn_cast<PHINode>(SI->getTrueValue())) 651 NonPhi = SI->getFalseValue(); 652 else if (OrigPhi == dyn_cast<PHINode>(SI->getFalseValue())) 653 NonPhi = SI->getTrueValue(); 654 else 655 return InstDesc(false, I); 656 657 // We are looking for selects of the form: 658 // select(cmp(), phi, loop_invariant) or 659 // select(cmp(), loop_invariant, phi) 660 if (!Loop->isLoopInvariant(NonPhi)) 661 return InstDesc(false, I); 662 663 return InstDesc(I, isa<ICmpInst>(I->getOperand(0)) ? RecurKind::SelectICmp 664 : RecurKind::SelectFCmp); 665 } 666 667 RecurrenceDescriptor::InstDesc 668 RecurrenceDescriptor::isMinMaxPattern(Instruction *I, RecurKind Kind, 669 const InstDesc &Prev) { 670 assert((isa<CmpInst>(I) || isa<SelectInst>(I) || isa<CallInst>(I)) && 671 "Expected a cmp or select or call instruction"); 672 if (!isMinMaxRecurrenceKind(Kind)) 673 return InstDesc(false, I); 674 675 // We must handle the select(cmp()) as a single instruction. Advance to the 676 // select. 677 CmpInst::Predicate Pred; 678 if (match(I, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) { 679 if (auto *Select = dyn_cast<SelectInst>(*I->user_begin())) 680 return InstDesc(Select, Prev.getRecKind()); 681 } 682 683 // Only match select with single use cmp condition, or a min/max intrinsic. 684 if (!isa<IntrinsicInst>(I) && 685 !match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(), 686 m_Value()))) 687 return InstDesc(false, I); 688 689 // Look for a min/max pattern. 690 if (match(I, m_UMin(m_Value(), m_Value()))) 691 return InstDesc(Kind == RecurKind::UMin, I); 692 if (match(I, m_UMax(m_Value(), m_Value()))) 693 return InstDesc(Kind == RecurKind::UMax, I); 694 if (match(I, m_SMax(m_Value(), m_Value()))) 695 return InstDesc(Kind == RecurKind::SMax, I); 696 if (match(I, m_SMin(m_Value(), m_Value()))) 697 return InstDesc(Kind == RecurKind::SMin, I); 698 if (match(I, m_OrdFMin(m_Value(), m_Value()))) 699 return InstDesc(Kind == RecurKind::FMin, I); 700 if (match(I, m_OrdFMax(m_Value(), m_Value()))) 701 return InstDesc(Kind == RecurKind::FMax, I); 702 if (match(I, m_UnordFMin(m_Value(), m_Value()))) 703 return InstDesc(Kind == RecurKind::FMin, I); 704 if (match(I, m_UnordFMax(m_Value(), m_Value()))) 705 return InstDesc(Kind == RecurKind::FMax, I); 706 if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) 707 return InstDesc(Kind == RecurKind::FMin, I); 708 if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) 709 return InstDesc(Kind == RecurKind::FMax, I); 710 711 return InstDesc(false, I); 712 } 713 714 /// Returns true if the select instruction has users in the compare-and-add 715 /// reduction pattern below. The select instruction argument is the last one 716 /// in the sequence. 717 /// 718 /// %sum.1 = phi ... 719 /// ... 720 /// %cmp = fcmp pred %0, %CFP 721 /// %add = fadd %0, %sum.1 722 /// %sum.2 = select %cmp, %add, %sum.1 723 RecurrenceDescriptor::InstDesc 724 RecurrenceDescriptor::isConditionalRdxPattern(RecurKind Kind, Instruction *I) { 725 SelectInst *SI = dyn_cast<SelectInst>(I); 726 if (!SI) 727 return InstDesc(false, I); 728 729 CmpInst *CI = dyn_cast<CmpInst>(SI->getCondition()); 730 // Only handle single use cases for now. 731 if (!CI || !CI->hasOneUse()) 732 return InstDesc(false, I); 733 734 Value *TrueVal = SI->getTrueValue(); 735 Value *FalseVal = SI->getFalseValue(); 736 // Handle only when either of operands of select instruction is a PHI 737 // node for now. 738 if ((isa<PHINode>(*TrueVal) && isa<PHINode>(*FalseVal)) || 739 (!isa<PHINode>(*TrueVal) && !isa<PHINode>(*FalseVal))) 740 return InstDesc(false, I); 741 742 Instruction *I1 = 743 isa<PHINode>(*TrueVal) ? dyn_cast<Instruction>(FalseVal) 744 : dyn_cast<Instruction>(TrueVal); 745 if (!I1 || !I1->isBinaryOp()) 746 return InstDesc(false, I); 747 748 Value *Op1, *Op2; 749 if ((m_FAdd(m_Value(Op1), m_Value(Op2)).match(I1) || 750 m_FSub(m_Value(Op1), m_Value(Op2)).match(I1)) && 751 I1->isFast()) 752 return InstDesc(Kind == RecurKind::FAdd, SI); 753 754 if (m_FMul(m_Value(Op1), m_Value(Op2)).match(I1) && (I1->isFast())) 755 return InstDesc(Kind == RecurKind::FMul, SI); 756 757 return InstDesc(false, I); 758 } 759 760 RecurrenceDescriptor::InstDesc 761 RecurrenceDescriptor::isRecurrenceInstr(Loop *L, PHINode *OrigPhi, 762 Instruction *I, RecurKind Kind, 763 InstDesc &Prev, FastMathFlags FuncFMF) { 764 assert(Prev.getRecKind() == RecurKind::None || Prev.getRecKind() == Kind); 765 switch (I->getOpcode()) { 766 default: 767 return InstDesc(false, I); 768 case Instruction::PHI: 769 return InstDesc(I, Prev.getRecKind(), Prev.getExactFPMathInst()); 770 case Instruction::Sub: 771 case Instruction::Add: 772 return InstDesc(Kind == RecurKind::Add, I); 773 case Instruction::Mul: 774 return InstDesc(Kind == RecurKind::Mul, I); 775 case Instruction::And: 776 return InstDesc(Kind == RecurKind::And, I); 777 case Instruction::Or: 778 return InstDesc(Kind == RecurKind::Or, I); 779 case Instruction::Xor: 780 return InstDesc(Kind == RecurKind::Xor, I); 781 case Instruction::FDiv: 782 case Instruction::FMul: 783 return InstDesc(Kind == RecurKind::FMul, I, 784 I->hasAllowReassoc() ? nullptr : I); 785 case Instruction::FSub: 786 case Instruction::FAdd: 787 return InstDesc(Kind == RecurKind::FAdd, I, 788 I->hasAllowReassoc() ? nullptr : I); 789 case Instruction::Select: 790 if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) 791 return isConditionalRdxPattern(Kind, I); 792 [[fallthrough]]; 793 case Instruction::FCmp: 794 case Instruction::ICmp: 795 case Instruction::Call: 796 if (isSelectCmpRecurrenceKind(Kind)) 797 return isSelectCmpPattern(L, OrigPhi, I, Prev); 798 if (isIntMinMaxRecurrenceKind(Kind) || 799 (((FuncFMF.noNaNs() && FuncFMF.noSignedZeros()) || 800 (isa<FPMathOperator>(I) && I->hasNoNaNs() && 801 I->hasNoSignedZeros())) && 802 isFPMinMaxRecurrenceKind(Kind))) 803 return isMinMaxPattern(I, Kind, Prev); 804 else if (isFMulAddIntrinsic(I)) 805 return InstDesc(Kind == RecurKind::FMulAdd, I, 806 I->hasAllowReassoc() ? nullptr : I); 807 return InstDesc(false, I); 808 } 809 } 810 811 bool RecurrenceDescriptor::hasMultipleUsesOf( 812 Instruction *I, SmallPtrSetImpl<Instruction *> &Insts, 813 unsigned MaxNumUses) { 814 unsigned NumUses = 0; 815 for (const Use &U : I->operands()) { 816 if (Insts.count(dyn_cast<Instruction>(U))) 817 ++NumUses; 818 if (NumUses > MaxNumUses) 819 return true; 820 } 821 822 return false; 823 } 824 825 bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop, 826 RecurrenceDescriptor &RedDes, 827 DemandedBits *DB, AssumptionCache *AC, 828 DominatorTree *DT, 829 ScalarEvolution *SE) { 830 BasicBlock *Header = TheLoop->getHeader(); 831 Function &F = *Header->getParent(); 832 FastMathFlags FMF; 833 FMF.setNoNaNs( 834 F.getFnAttribute("no-nans-fp-math").getValueAsBool()); 835 FMF.setNoSignedZeros( 836 F.getFnAttribute("no-signed-zeros-fp-math").getValueAsBool()); 837 838 if (AddReductionVar(Phi, RecurKind::Add, TheLoop, FMF, RedDes, DB, AC, DT, 839 SE)) { 840 LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n"); 841 return true; 842 } 843 if (AddReductionVar(Phi, RecurKind::Mul, TheLoop, FMF, RedDes, DB, AC, DT, 844 SE)) { 845 LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n"); 846 return true; 847 } 848 if (AddReductionVar(Phi, RecurKind::Or, TheLoop, FMF, RedDes, DB, AC, DT, 849 SE)) { 850 LLVM_DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n"); 851 return true; 852 } 853 if (AddReductionVar(Phi, RecurKind::And, TheLoop, FMF, RedDes, DB, AC, DT, 854 SE)) { 855 LLVM_DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n"); 856 return true; 857 } 858 if (AddReductionVar(Phi, RecurKind::Xor, TheLoop, FMF, RedDes, DB, AC, DT, 859 SE)) { 860 LLVM_DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n"); 861 return true; 862 } 863 if (AddReductionVar(Phi, RecurKind::SMax, TheLoop, FMF, RedDes, DB, AC, DT, 864 SE)) { 865 LLVM_DEBUG(dbgs() << "Found a SMAX reduction PHI." << *Phi << "\n"); 866 return true; 867 } 868 if (AddReductionVar(Phi, RecurKind::SMin, TheLoop, FMF, RedDes, DB, AC, DT, 869 SE)) { 870 LLVM_DEBUG(dbgs() << "Found a SMIN reduction PHI." << *Phi << "\n"); 871 return true; 872 } 873 if (AddReductionVar(Phi, RecurKind::UMax, TheLoop, FMF, RedDes, DB, AC, DT, 874 SE)) { 875 LLVM_DEBUG(dbgs() << "Found a UMAX reduction PHI." << *Phi << "\n"); 876 return true; 877 } 878 if (AddReductionVar(Phi, RecurKind::UMin, TheLoop, FMF, RedDes, DB, AC, DT, 879 SE)) { 880 LLVM_DEBUG(dbgs() << "Found a UMIN reduction PHI." << *Phi << "\n"); 881 return true; 882 } 883 if (AddReductionVar(Phi, RecurKind::SelectICmp, TheLoop, FMF, RedDes, DB, AC, 884 DT, SE)) { 885 LLVM_DEBUG(dbgs() << "Found an integer conditional select reduction PHI." 886 << *Phi << "\n"); 887 return true; 888 } 889 if (AddReductionVar(Phi, RecurKind::FMul, TheLoop, FMF, RedDes, DB, AC, DT, 890 SE)) { 891 LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n"); 892 return true; 893 } 894 if (AddReductionVar(Phi, RecurKind::FAdd, TheLoop, FMF, RedDes, DB, AC, DT, 895 SE)) { 896 LLVM_DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n"); 897 return true; 898 } 899 if (AddReductionVar(Phi, RecurKind::FMax, TheLoop, FMF, RedDes, DB, AC, DT, 900 SE)) { 901 LLVM_DEBUG(dbgs() << "Found a float MAX reduction PHI." << *Phi << "\n"); 902 return true; 903 } 904 if (AddReductionVar(Phi, RecurKind::FMin, TheLoop, FMF, RedDes, DB, AC, DT, 905 SE)) { 906 LLVM_DEBUG(dbgs() << "Found a float MIN reduction PHI." << *Phi << "\n"); 907 return true; 908 } 909 if (AddReductionVar(Phi, RecurKind::SelectFCmp, TheLoop, FMF, RedDes, DB, AC, 910 DT, SE)) { 911 LLVM_DEBUG(dbgs() << "Found a float conditional select reduction PHI." 912 << " PHI." << *Phi << "\n"); 913 return true; 914 } 915 if (AddReductionVar(Phi, RecurKind::FMulAdd, TheLoop, FMF, RedDes, DB, AC, DT, 916 SE)) { 917 LLVM_DEBUG(dbgs() << "Found an FMulAdd reduction PHI." << *Phi << "\n"); 918 return true; 919 } 920 // Not a reduction of known type. 921 return false; 922 } 923 924 bool RecurrenceDescriptor::isFixedOrderRecurrence( 925 PHINode *Phi, Loop *TheLoop, 926 MapVector<Instruction *, Instruction *> &SinkAfter, DominatorTree *DT) { 927 928 // Ensure the phi node is in the loop header and has two incoming values. 929 if (Phi->getParent() != TheLoop->getHeader() || 930 Phi->getNumIncomingValues() != 2) 931 return false; 932 933 // Ensure the loop has a preheader and a single latch block. The loop 934 // vectorizer will need the latch to set up the next iteration of the loop. 935 auto *Preheader = TheLoop->getLoopPreheader(); 936 auto *Latch = TheLoop->getLoopLatch(); 937 if (!Preheader || !Latch) 938 return false; 939 940 // Ensure the phi node's incoming blocks are the loop preheader and latch. 941 if (Phi->getBasicBlockIndex(Preheader) < 0 || 942 Phi->getBasicBlockIndex(Latch) < 0) 943 return false; 944 945 // Get the previous value. The previous value comes from the latch edge while 946 // the initial value comes from the preheader edge. 947 auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch)); 948 949 // If Previous is a phi in the header, go through incoming values from the 950 // latch until we find a non-phi value. Use this as the new Previous, all uses 951 // in the header will be dominated by the original phi, but need to be moved 952 // after the non-phi previous value. 953 SmallPtrSet<PHINode *, 4> SeenPhis; 954 while (auto *PrevPhi = dyn_cast_or_null<PHINode>(Previous)) { 955 if (PrevPhi->getParent() != Phi->getParent()) 956 return false; 957 if (!SeenPhis.insert(PrevPhi).second) 958 return false; 959 Previous = dyn_cast<Instruction>(PrevPhi->getIncomingValueForBlock(Latch)); 960 } 961 962 if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) || 963 SinkAfter.count(Previous)) // Cannot rely on dominance due to motion. 964 return false; 965 966 // Ensure every user of the phi node (recursively) is dominated by the 967 // previous value. The dominance requirement ensures the loop vectorizer will 968 // not need to vectorize the initial value prior to the first iteration of the 969 // loop. 970 // TODO: Consider extending this sinking to handle memory instructions. 971 972 // We optimistically assume we can sink all users after Previous. Keep a set 973 // of instructions to sink after Previous ordered by dominance in the common 974 // basic block. It will be applied to SinkAfter if all users can be sunk. 975 auto CompareByComesBefore = [](const Instruction *A, const Instruction *B) { 976 return A->comesBefore(B); 977 }; 978 std::set<Instruction *, decltype(CompareByComesBefore)> InstrsToSink( 979 CompareByComesBefore); 980 981 BasicBlock *PhiBB = Phi->getParent(); 982 SmallVector<Instruction *, 8> WorkList; 983 auto TryToPushSinkCandidate = [&](Instruction *SinkCandidate) { 984 // Already sunk SinkCandidate. 985 if (SinkCandidate->getParent() == PhiBB && 986 InstrsToSink.find(SinkCandidate) != InstrsToSink.end()) 987 return true; 988 989 // Cyclic dependence. 990 if (Previous == SinkCandidate) 991 return false; 992 993 if (DT->dominates(Previous, 994 SinkCandidate)) // We already are good w/o sinking. 995 return true; 996 997 if (SinkCandidate->getParent() != PhiBB || 998 SinkCandidate->mayHaveSideEffects() || 999 SinkCandidate->mayReadFromMemory() || SinkCandidate->isTerminator()) 1000 return false; 1001 1002 // Avoid sinking an instruction multiple times (if multiple operands are 1003 // fixed order recurrences) by sinking once - after the latest 'previous' 1004 // instruction. 1005 auto It = SinkAfter.find(SinkCandidate); 1006 if (It != SinkAfter.end()) { 1007 auto *OtherPrev = It->second; 1008 // Find the earliest entry in the 'sink-after' chain. The last entry in 1009 // the chain is the original 'Previous' for a recurrence handled earlier. 1010 auto EarlierIt = SinkAfter.find(OtherPrev); 1011 while (EarlierIt != SinkAfter.end()) { 1012 Instruction *EarlierInst = EarlierIt->second; 1013 EarlierIt = SinkAfter.find(EarlierInst); 1014 // Bail out if order has not been preserved. 1015 if (EarlierIt != SinkAfter.end() && 1016 !DT->dominates(EarlierInst, OtherPrev)) 1017 return false; 1018 OtherPrev = EarlierInst; 1019 } 1020 // Bail out if order has not been preserved. 1021 if (OtherPrev != It->second && !DT->dominates(It->second, OtherPrev)) 1022 return false; 1023 1024 // SinkCandidate is already being sunk after an instruction after 1025 // Previous. Nothing left to do. 1026 if (DT->dominates(Previous, OtherPrev) || Previous == OtherPrev) 1027 return true; 1028 1029 // If there are other instructions to be sunk after SinkCandidate, remove 1030 // and re-insert SinkCandidate can break those instructions. Bail out for 1031 // simplicity. 1032 if (any_of(SinkAfter, 1033 [SinkCandidate](const std::pair<Instruction *, Instruction *> &P) { 1034 return P.second == SinkCandidate; 1035 })) 1036 return false; 1037 1038 // Otherwise, Previous comes after OtherPrev and SinkCandidate needs to be 1039 // re-sunk to Previous, instead of sinking to OtherPrev. Remove 1040 // SinkCandidate from SinkAfter to ensure it's insert position is updated. 1041 SinkAfter.erase(SinkCandidate); 1042 } 1043 1044 // If we reach a PHI node that is not dominated by Previous, we reached a 1045 // header PHI. No need for sinking. 1046 if (isa<PHINode>(SinkCandidate)) 1047 return true; 1048 1049 // Sink User tentatively and check its users 1050 InstrsToSink.insert(SinkCandidate); 1051 WorkList.push_back(SinkCandidate); 1052 return true; 1053 }; 1054 1055 WorkList.push_back(Phi); 1056 // Try to recursively sink instructions and their users after Previous. 1057 while (!WorkList.empty()) { 1058 Instruction *Current = WorkList.pop_back_val(); 1059 for (User *User : Current->users()) { 1060 if (!TryToPushSinkCandidate(cast<Instruction>(User))) 1061 return false; 1062 } 1063 } 1064 1065 // We can sink all users of Phi. Update the mapping. 1066 for (Instruction *I : InstrsToSink) { 1067 SinkAfter[I] = Previous; 1068 Previous = I; 1069 } 1070 return true; 1071 } 1072 1073 /// This function returns the identity element (or neutral element) for 1074 /// the operation K. 1075 Value *RecurrenceDescriptor::getRecurrenceIdentity(RecurKind K, Type *Tp, 1076 FastMathFlags FMF) const { 1077 switch (K) { 1078 case RecurKind::Xor: 1079 case RecurKind::Add: 1080 case RecurKind::Or: 1081 // Adding, Xoring, Oring zero to a number does not change it. 1082 return ConstantInt::get(Tp, 0); 1083 case RecurKind::Mul: 1084 // Multiplying a number by 1 does not change it. 1085 return ConstantInt::get(Tp, 1); 1086 case RecurKind::And: 1087 // AND-ing a number with an all-1 value does not change it. 1088 return ConstantInt::get(Tp, -1, true); 1089 case RecurKind::FMul: 1090 // Multiplying a number by 1 does not change it. 1091 return ConstantFP::get(Tp, 1.0L); 1092 case RecurKind::FMulAdd: 1093 case RecurKind::FAdd: 1094 // Adding zero to a number does not change it. 1095 // FIXME: Ideally we should not need to check FMF for FAdd and should always 1096 // use -0.0. However, this will currently result in mixed vectors of 0.0/-0.0. 1097 // Instead, we should ensure that 1) the FMF from FAdd are propagated to the PHI 1098 // nodes where possible, and 2) PHIs with the nsz flag + -0.0 use 0.0. This would 1099 // mean we can then remove the check for noSignedZeros() below (see D98963). 1100 if (FMF.noSignedZeros()) 1101 return ConstantFP::get(Tp, 0.0L); 1102 return ConstantFP::get(Tp, -0.0L); 1103 case RecurKind::UMin: 1104 return ConstantInt::get(Tp, -1); 1105 case RecurKind::UMax: 1106 return ConstantInt::get(Tp, 0); 1107 case RecurKind::SMin: 1108 return ConstantInt::get(Tp, 1109 APInt::getSignedMaxValue(Tp->getIntegerBitWidth())); 1110 case RecurKind::SMax: 1111 return ConstantInt::get(Tp, 1112 APInt::getSignedMinValue(Tp->getIntegerBitWidth())); 1113 case RecurKind::FMin: 1114 assert((FMF.noNaNs() && FMF.noSignedZeros()) && 1115 "nnan, nsz is expected to be set for FP min reduction."); 1116 return ConstantFP::getInfinity(Tp, false /*Negative*/); 1117 case RecurKind::FMax: 1118 assert((FMF.noNaNs() && FMF.noSignedZeros()) && 1119 "nnan, nsz is expected to be set for FP max reduction."); 1120 return ConstantFP::getInfinity(Tp, true /*Negative*/); 1121 case RecurKind::SelectICmp: 1122 case RecurKind::SelectFCmp: 1123 return getRecurrenceStartValue(); 1124 break; 1125 default: 1126 llvm_unreachable("Unknown recurrence kind"); 1127 } 1128 } 1129 1130 unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) { 1131 switch (Kind) { 1132 case RecurKind::Add: 1133 return Instruction::Add; 1134 case RecurKind::Mul: 1135 return Instruction::Mul; 1136 case RecurKind::Or: 1137 return Instruction::Or; 1138 case RecurKind::And: 1139 return Instruction::And; 1140 case RecurKind::Xor: 1141 return Instruction::Xor; 1142 case RecurKind::FMul: 1143 return Instruction::FMul; 1144 case RecurKind::FMulAdd: 1145 case RecurKind::FAdd: 1146 return Instruction::FAdd; 1147 case RecurKind::SMax: 1148 case RecurKind::SMin: 1149 case RecurKind::UMax: 1150 case RecurKind::UMin: 1151 case RecurKind::SelectICmp: 1152 return Instruction::ICmp; 1153 case RecurKind::FMax: 1154 case RecurKind::FMin: 1155 case RecurKind::SelectFCmp: 1156 return Instruction::FCmp; 1157 default: 1158 llvm_unreachable("Unknown recurrence operation"); 1159 } 1160 } 1161 1162 SmallVector<Instruction *, 4> 1163 RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const { 1164 SmallVector<Instruction *, 4> ReductionOperations; 1165 unsigned RedOp = getOpcode(Kind); 1166 1167 // Search down from the Phi to the LoopExitInstr, looking for instructions 1168 // with a single user of the correct type for the reduction. 1169 1170 // Note that we check that the type of the operand is correct for each item in 1171 // the chain, including the last (the loop exit value). This can come up from 1172 // sub, which would otherwise be treated as an add reduction. MinMax also need 1173 // to check for a pair of icmp/select, for which we use getNextInstruction and 1174 // isCorrectOpcode functions to step the right number of instruction, and 1175 // check the icmp/select pair. 1176 // FIXME: We also do not attempt to look through Select's yet, which might 1177 // be part of the reduction chain, or attempt to looks through And's to find a 1178 // smaller bitwidth. Subs are also currently not allowed (which are usually 1179 // treated as part of a add reduction) as they are expected to generally be 1180 // more expensive than out-of-loop reductions, and need to be costed more 1181 // carefully. 1182 unsigned ExpectedUses = 1; 1183 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) 1184 ExpectedUses = 2; 1185 1186 auto getNextInstruction = [&](Instruction *Cur) -> Instruction * { 1187 for (auto *User : Cur->users()) { 1188 Instruction *UI = cast<Instruction>(User); 1189 if (isa<PHINode>(UI)) 1190 continue; 1191 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) { 1192 // We are expecting a icmp/select pair, which we go to the next select 1193 // instruction if we can. We already know that Cur has 2 uses. 1194 if (isa<SelectInst>(UI)) 1195 return UI; 1196 continue; 1197 } 1198 return UI; 1199 } 1200 return nullptr; 1201 }; 1202 auto isCorrectOpcode = [&](Instruction *Cur) { 1203 if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) { 1204 Value *LHS, *RHS; 1205 return SelectPatternResult::isMinOrMax( 1206 matchSelectPattern(Cur, LHS, RHS).Flavor); 1207 } 1208 // Recognize a call to the llvm.fmuladd intrinsic. 1209 if (isFMulAddIntrinsic(Cur)) 1210 return true; 1211 1212 return Cur->getOpcode() == RedOp; 1213 }; 1214 1215 // Attempt to look through Phis which are part of the reduction chain 1216 unsigned ExtraPhiUses = 0; 1217 Instruction *RdxInstr = LoopExitInstr; 1218 if (auto ExitPhi = dyn_cast<PHINode>(LoopExitInstr)) { 1219 if (ExitPhi->getNumIncomingValues() != 2) 1220 return {}; 1221 1222 Instruction *Inc0 = dyn_cast<Instruction>(ExitPhi->getIncomingValue(0)); 1223 Instruction *Inc1 = dyn_cast<Instruction>(ExitPhi->getIncomingValue(1)); 1224 1225 Instruction *Chain = nullptr; 1226 if (Inc0 == Phi) 1227 Chain = Inc1; 1228 else if (Inc1 == Phi) 1229 Chain = Inc0; 1230 else 1231 return {}; 1232 1233 RdxInstr = Chain; 1234 ExtraPhiUses = 1; 1235 } 1236 1237 // The loop exit instruction we check first (as a quick test) but add last. We 1238 // check the opcode is correct (and dont allow them to be Subs) and that they 1239 // have expected to have the expected number of uses. They will have one use 1240 // from the phi and one from a LCSSA value, no matter the type. 1241 if (!isCorrectOpcode(RdxInstr) || !LoopExitInstr->hasNUses(2)) 1242 return {}; 1243 1244 // Check that the Phi has one (or two for min/max) uses, plus an extra use 1245 // for conditional reductions. 1246 if (!Phi->hasNUses(ExpectedUses + ExtraPhiUses)) 1247 return {}; 1248 1249 Instruction *Cur = getNextInstruction(Phi); 1250 1251 // Each other instruction in the chain should have the expected number of uses 1252 // and be the correct opcode. 1253 while (Cur != RdxInstr) { 1254 if (!Cur || !isCorrectOpcode(Cur) || !Cur->hasNUses(ExpectedUses)) 1255 return {}; 1256 1257 ReductionOperations.push_back(Cur); 1258 Cur = getNextInstruction(Cur); 1259 } 1260 1261 ReductionOperations.push_back(Cur); 1262 return ReductionOperations; 1263 } 1264 1265 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K, 1266 const SCEV *Step, BinaryOperator *BOp, 1267 Type *ElementType, 1268 SmallVectorImpl<Instruction *> *Casts) 1269 : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp), 1270 ElementType(ElementType) { 1271 assert(IK != IK_NoInduction && "Not an induction"); 1272 1273 // Start value type should match the induction kind and the value 1274 // itself should not be null. 1275 assert(StartValue && "StartValue is null"); 1276 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) && 1277 "StartValue is not a pointer for pointer induction"); 1278 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) && 1279 "StartValue is not an integer for integer induction"); 1280 1281 // Check the Step Value. It should be non-zero integer value. 1282 assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) && 1283 "Step value is zero"); 1284 1285 assert((IK != IK_PtrInduction || getConstIntStepValue()) && 1286 "Step value should be constant for pointer induction"); 1287 assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) && 1288 "StepValue is not an integer"); 1289 1290 assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) && 1291 "StepValue is not FP for FpInduction"); 1292 assert((IK != IK_FpInduction || 1293 (InductionBinOp && 1294 (InductionBinOp->getOpcode() == Instruction::FAdd || 1295 InductionBinOp->getOpcode() == Instruction::FSub))) && 1296 "Binary opcode should be specified for FP induction"); 1297 1298 if (IK == IK_PtrInduction) 1299 assert(ElementType && "Pointer induction must have element type"); 1300 else 1301 assert(!ElementType && "Non-pointer induction cannot have element type"); 1302 1303 if (Casts) { 1304 for (auto &Inst : *Casts) { 1305 RedundantCasts.push_back(Inst); 1306 } 1307 } 1308 } 1309 1310 ConstantInt *InductionDescriptor::getConstIntStepValue() const { 1311 if (isa<SCEVConstant>(Step)) 1312 return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue()); 1313 return nullptr; 1314 } 1315 1316 bool InductionDescriptor::isFPInductionPHI(PHINode *Phi, const Loop *TheLoop, 1317 ScalarEvolution *SE, 1318 InductionDescriptor &D) { 1319 1320 // Here we only handle FP induction variables. 1321 assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type"); 1322 1323 if (TheLoop->getHeader() != Phi->getParent()) 1324 return false; 1325 1326 // The loop may have multiple entrances or multiple exits; we can analyze 1327 // this phi if it has a unique entry value and a unique backedge value. 1328 if (Phi->getNumIncomingValues() != 2) 1329 return false; 1330 Value *BEValue = nullptr, *StartValue = nullptr; 1331 if (TheLoop->contains(Phi->getIncomingBlock(0))) { 1332 BEValue = Phi->getIncomingValue(0); 1333 StartValue = Phi->getIncomingValue(1); 1334 } else { 1335 assert(TheLoop->contains(Phi->getIncomingBlock(1)) && 1336 "Unexpected Phi node in the loop"); 1337 BEValue = Phi->getIncomingValue(1); 1338 StartValue = Phi->getIncomingValue(0); 1339 } 1340 1341 BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue); 1342 if (!BOp) 1343 return false; 1344 1345 Value *Addend = nullptr; 1346 if (BOp->getOpcode() == Instruction::FAdd) { 1347 if (BOp->getOperand(0) == Phi) 1348 Addend = BOp->getOperand(1); 1349 else if (BOp->getOperand(1) == Phi) 1350 Addend = BOp->getOperand(0); 1351 } else if (BOp->getOpcode() == Instruction::FSub) 1352 if (BOp->getOperand(0) == Phi) 1353 Addend = BOp->getOperand(1); 1354 1355 if (!Addend) 1356 return false; 1357 1358 // The addend should be loop invariant 1359 if (auto *I = dyn_cast<Instruction>(Addend)) 1360 if (TheLoop->contains(I)) 1361 return false; 1362 1363 // FP Step has unknown SCEV 1364 const SCEV *Step = SE->getUnknown(Addend); 1365 D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp); 1366 return true; 1367 } 1368 1369 /// This function is called when we suspect that the update-chain of a phi node 1370 /// (whose symbolic SCEV expression sin \p PhiScev) contains redundant casts, 1371 /// that can be ignored. (This can happen when the PSCEV rewriter adds a runtime 1372 /// predicate P under which the SCEV expression for the phi can be the 1373 /// AddRecurrence \p AR; See createAddRecFromPHIWithCast). We want to find the 1374 /// cast instructions that are involved in the update-chain of this induction. 1375 /// A caller that adds the required runtime predicate can be free to drop these 1376 /// cast instructions, and compute the phi using \p AR (instead of some scev 1377 /// expression with casts). 1378 /// 1379 /// For example, without a predicate the scev expression can take the following 1380 /// form: 1381 /// (Ext ix (Trunc iy ( Start + i*Step ) to ix) to iy) 1382 /// 1383 /// It corresponds to the following IR sequence: 1384 /// %for.body: 1385 /// %x = phi i64 [ 0, %ph ], [ %add, %for.body ] 1386 /// %casted_phi = "ExtTrunc i64 %x" 1387 /// %add = add i64 %casted_phi, %step 1388 /// 1389 /// where %x is given in \p PN, 1390 /// PSE.getSCEV(%x) is equal to PSE.getSCEV(%casted_phi) under a predicate, 1391 /// and the IR sequence that "ExtTrunc i64 %x" represents can take one of 1392 /// several forms, for example, such as: 1393 /// ExtTrunc1: %casted_phi = and %x, 2^n-1 1394 /// or: 1395 /// ExtTrunc2: %t = shl %x, m 1396 /// %casted_phi = ashr %t, m 1397 /// 1398 /// If we are able to find such sequence, we return the instructions 1399 /// we found, namely %casted_phi and the instructions on its use-def chain up 1400 /// to the phi (not including the phi). 1401 static bool getCastsForInductionPHI(PredicatedScalarEvolution &PSE, 1402 const SCEVUnknown *PhiScev, 1403 const SCEVAddRecExpr *AR, 1404 SmallVectorImpl<Instruction *> &CastInsts) { 1405 1406 assert(CastInsts.empty() && "CastInsts is expected to be empty."); 1407 auto *PN = cast<PHINode>(PhiScev->getValue()); 1408 assert(PSE.getSCEV(PN) == AR && "Unexpected phi node SCEV expression"); 1409 const Loop *L = AR->getLoop(); 1410 1411 // Find any cast instructions that participate in the def-use chain of 1412 // PhiScev in the loop. 1413 // FORNOW/TODO: We currently expect the def-use chain to include only 1414 // two-operand instructions, where one of the operands is an invariant. 1415 // createAddRecFromPHIWithCasts() currently does not support anything more 1416 // involved than that, so we keep the search simple. This can be 1417 // extended/generalized as needed. 1418 1419 auto getDef = [&](const Value *Val) -> Value * { 1420 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val); 1421 if (!BinOp) 1422 return nullptr; 1423 Value *Op0 = BinOp->getOperand(0); 1424 Value *Op1 = BinOp->getOperand(1); 1425 Value *Def = nullptr; 1426 if (L->isLoopInvariant(Op0)) 1427 Def = Op1; 1428 else if (L->isLoopInvariant(Op1)) 1429 Def = Op0; 1430 return Def; 1431 }; 1432 1433 // Look for the instruction that defines the induction via the 1434 // loop backedge. 1435 BasicBlock *Latch = L->getLoopLatch(); 1436 if (!Latch) 1437 return false; 1438 Value *Val = PN->getIncomingValueForBlock(Latch); 1439 if (!Val) 1440 return false; 1441 1442 // Follow the def-use chain until the induction phi is reached. 1443 // If on the way we encounter a Value that has the same SCEV Expr as the 1444 // phi node, we can consider the instructions we visit from that point 1445 // as part of the cast-sequence that can be ignored. 1446 bool InCastSequence = false; 1447 auto *Inst = dyn_cast<Instruction>(Val); 1448 while (Val != PN) { 1449 // If we encountered a phi node other than PN, or if we left the loop, 1450 // we bail out. 1451 if (!Inst || !L->contains(Inst)) { 1452 return false; 1453 } 1454 auto *AddRec = dyn_cast<SCEVAddRecExpr>(PSE.getSCEV(Val)); 1455 if (AddRec && PSE.areAddRecsEqualWithPreds(AddRec, AR)) 1456 InCastSequence = true; 1457 if (InCastSequence) { 1458 // Only the last instruction in the cast sequence is expected to have 1459 // uses outside the induction def-use chain. 1460 if (!CastInsts.empty()) 1461 if (!Inst->hasOneUse()) 1462 return false; 1463 CastInsts.push_back(Inst); 1464 } 1465 Val = getDef(Val); 1466 if (!Val) 1467 return false; 1468 Inst = dyn_cast<Instruction>(Val); 1469 } 1470 1471 return InCastSequence; 1472 } 1473 1474 bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop, 1475 PredicatedScalarEvolution &PSE, 1476 InductionDescriptor &D, bool Assume) { 1477 Type *PhiTy = Phi->getType(); 1478 1479 // Handle integer and pointer inductions variables. 1480 // Now we handle also FP induction but not trying to make a 1481 // recurrent expression from the PHI node in-place. 1482 1483 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() && !PhiTy->isFloatTy() && 1484 !PhiTy->isDoubleTy() && !PhiTy->isHalfTy()) 1485 return false; 1486 1487 if (PhiTy->isFloatingPointTy()) 1488 return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D); 1489 1490 const SCEV *PhiScev = PSE.getSCEV(Phi); 1491 const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 1492 1493 // We need this expression to be an AddRecExpr. 1494 if (Assume && !AR) 1495 AR = PSE.getAsAddRec(Phi); 1496 1497 if (!AR) { 1498 LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 1499 return false; 1500 } 1501 1502 // Record any Cast instructions that participate in the induction update 1503 const auto *SymbolicPhi = dyn_cast<SCEVUnknown>(PhiScev); 1504 // If we started from an UnknownSCEV, and managed to build an addRecurrence 1505 // only after enabling Assume with PSCEV, this means we may have encountered 1506 // cast instructions that required adding a runtime check in order to 1507 // guarantee the correctness of the AddRecurrence respresentation of the 1508 // induction. 1509 if (PhiScev != AR && SymbolicPhi) { 1510 SmallVector<Instruction *, 2> Casts; 1511 if (getCastsForInductionPHI(PSE, SymbolicPhi, AR, Casts)) 1512 return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR, &Casts); 1513 } 1514 1515 return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR); 1516 } 1517 1518 bool InductionDescriptor::isInductionPHI( 1519 PHINode *Phi, const Loop *TheLoop, ScalarEvolution *SE, 1520 InductionDescriptor &D, const SCEV *Expr, 1521 SmallVectorImpl<Instruction *> *CastsToIgnore) { 1522 Type *PhiTy = Phi->getType(); 1523 // We only handle integer and pointer inductions variables. 1524 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 1525 return false; 1526 1527 // Check that the PHI is consecutive. 1528 const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi); 1529 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 1530 1531 if (!AR) { 1532 LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 1533 return false; 1534 } 1535 1536 if (AR->getLoop() != TheLoop) { 1537 // FIXME: We should treat this as a uniform. Unfortunately, we 1538 // don't currently know how to handled uniform PHIs. 1539 LLVM_DEBUG( 1540 dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n"); 1541 return false; 1542 } 1543 1544 Value *StartValue = 1545 Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader()); 1546 1547 BasicBlock *Latch = AR->getLoop()->getLoopLatch(); 1548 if (!Latch) 1549 return false; 1550 1551 const SCEV *Step = AR->getStepRecurrence(*SE); 1552 // Calculate the pointer stride and check if it is consecutive. 1553 // The stride may be a constant or a loop invariant integer value. 1554 const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step); 1555 if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop)) 1556 return false; 1557 1558 if (PhiTy->isIntegerTy()) { 1559 BinaryOperator *BOp = 1560 dyn_cast<BinaryOperator>(Phi->getIncomingValueForBlock(Latch)); 1561 D = InductionDescriptor(StartValue, IK_IntInduction, Step, BOp, 1562 /* ElementType */ nullptr, CastsToIgnore); 1563 return true; 1564 } 1565 1566 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 1567 // Pointer induction should be a constant. 1568 if (!ConstStep) 1569 return false; 1570 1571 // Always use i8 element type for opaque pointer inductions. 1572 PointerType *PtrTy = cast<PointerType>(PhiTy); 1573 Type *ElementType = PtrTy->isOpaque() 1574 ? Type::getInt8Ty(PtrTy->getContext()) 1575 : PtrTy->getNonOpaquePointerElementType(); 1576 if (!ElementType->isSized()) 1577 return false; 1578 1579 ConstantInt *CV = ConstStep->getValue(); 1580 const DataLayout &DL = Phi->getModule()->getDataLayout(); 1581 TypeSize TySize = DL.getTypeAllocSize(ElementType); 1582 // TODO: We could potentially support this for scalable vectors if we can 1583 // prove at compile time that the constant step is always a multiple of 1584 // the scalable type. 1585 if (TySize.isZero() || TySize.isScalable()) 1586 return false; 1587 1588 int64_t Size = static_cast<int64_t>(TySize.getFixedValue()); 1589 int64_t CVSize = CV->getSExtValue(); 1590 if (CVSize % Size) 1591 return false; 1592 auto *StepValue = 1593 SE->getConstant(CV->getType(), CVSize / Size, true /* signed */); 1594 D = InductionDescriptor(StartValue, IK_PtrInduction, StepValue, 1595 /* BinOp */ nullptr, ElementType); 1596 return true; 1597 } 1598