1 //===- LoopReroll.cpp - Loop rerolling pass -------------------------------===// 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 pass implements a simple loop reroller. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/APInt.h" 14 #include "llvm/ADT/BitVector.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/MapVector.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/AliasAnalysis.h" 23 #include "llvm/Analysis/AliasSetTracker.h" 24 #include "llvm/Analysis/LoopInfo.h" 25 #include "llvm/Analysis/LoopPass.h" 26 #include "llvm/Analysis/ScalarEvolution.h" 27 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 28 #include "llvm/Analysis/TargetLibraryInfo.h" 29 #include "llvm/Analysis/ValueTracking.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/Constants.h" 32 #include "llvm/IR/Dominators.h" 33 #include "llvm/IR/InstrTypes.h" 34 #include "llvm/IR/Instruction.h" 35 #include "llvm/IR/Instructions.h" 36 #include "llvm/IR/IntrinsicInst.h" 37 #include "llvm/IR/Module.h" 38 #include "llvm/IR/Type.h" 39 #include "llvm/IR/Use.h" 40 #include "llvm/IR/User.h" 41 #include "llvm/IR/Value.h" 42 #include "llvm/InitializePasses.h" 43 #include "llvm/Pass.h" 44 #include "llvm/Support/Casting.h" 45 #include "llvm/Support/CommandLine.h" 46 #include "llvm/Support/Debug.h" 47 #include "llvm/Support/raw_ostream.h" 48 #include "llvm/Transforms/Scalar.h" 49 #include "llvm/Transforms/Scalar/LoopReroll.h" 50 #include "llvm/Transforms/Utils.h" 51 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 52 #include "llvm/Transforms/Utils/Local.h" 53 #include "llvm/Transforms/Utils/LoopUtils.h" 54 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 55 #include <cassert> 56 #include <cstddef> 57 #include <cstdint> 58 #include <iterator> 59 #include <map> 60 #include <utility> 61 62 using namespace llvm; 63 64 #define DEBUG_TYPE "loop-reroll" 65 66 STATISTIC(NumRerolledLoops, "Number of rerolled loops"); 67 68 static cl::opt<unsigned> 69 NumToleratedFailedMatches("reroll-num-tolerated-failed-matches", cl::init(400), 70 cl::Hidden, 71 cl::desc("The maximum number of failures to tolerate" 72 " during fuzzy matching. (default: 400)")); 73 74 // This loop re-rolling transformation aims to transform loops like this: 75 // 76 // int foo(int a); 77 // void bar(int *x) { 78 // for (int i = 0; i < 500; i += 3) { 79 // foo(i); 80 // foo(i+1); 81 // foo(i+2); 82 // } 83 // } 84 // 85 // into a loop like this: 86 // 87 // void bar(int *x) { 88 // for (int i = 0; i < 500; ++i) 89 // foo(i); 90 // } 91 // 92 // It does this by looking for loops that, besides the latch code, are composed 93 // of isomorphic DAGs of instructions, with each DAG rooted at some increment 94 // to the induction variable, and where each DAG is isomorphic to the DAG 95 // rooted at the induction variable (excepting the sub-DAGs which root the 96 // other induction-variable increments). In other words, we're looking for loop 97 // bodies of the form: 98 // 99 // %iv = phi [ (preheader, ...), (body, %iv.next) ] 100 // f(%iv) 101 // %iv.1 = add %iv, 1 <-- a root increment 102 // f(%iv.1) 103 // %iv.2 = add %iv, 2 <-- a root increment 104 // f(%iv.2) 105 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment 106 // f(%iv.scale_m_1) 107 // ... 108 // %iv.next = add %iv, scale 109 // %cmp = icmp(%iv, ...) 110 // br %cmp, header, exit 111 // 112 // where each f(i) is a set of instructions that, collectively, are a function 113 // only of i (and other loop-invariant values). 114 // 115 // As a special case, we can also reroll loops like this: 116 // 117 // int foo(int); 118 // void bar(int *x) { 119 // for (int i = 0; i < 500; ++i) { 120 // x[3*i] = foo(0); 121 // x[3*i+1] = foo(0); 122 // x[3*i+2] = foo(0); 123 // } 124 // } 125 // 126 // into this: 127 // 128 // void bar(int *x) { 129 // for (int i = 0; i < 1500; ++i) 130 // x[i] = foo(0); 131 // } 132 // 133 // in which case, we're looking for inputs like this: 134 // 135 // %iv = phi [ (preheader, ...), (body, %iv.next) ] 136 // %scaled.iv = mul %iv, scale 137 // f(%scaled.iv) 138 // %scaled.iv.1 = add %scaled.iv, 1 139 // f(%scaled.iv.1) 140 // %scaled.iv.2 = add %scaled.iv, 2 141 // f(%scaled.iv.2) 142 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1 143 // f(%scaled.iv.scale_m_1) 144 // ... 145 // %iv.next = add %iv, 1 146 // %cmp = icmp(%iv, ...) 147 // br %cmp, header, exit 148 149 namespace { 150 151 enum IterationLimits { 152 /// The maximum number of iterations that we'll try and reroll. 153 IL_MaxRerollIterations = 32, 154 /// The bitvector index used by loop induction variables and other 155 /// instructions that belong to all iterations. 156 IL_All, 157 IL_End 158 }; 159 160 class LoopRerollLegacyPass : public LoopPass { 161 public: 162 static char ID; // Pass ID, replacement for typeid 163 164 LoopRerollLegacyPass() : LoopPass(ID) { 165 initializeLoopRerollLegacyPassPass(*PassRegistry::getPassRegistry()); 166 } 167 168 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 169 170 void getAnalysisUsage(AnalysisUsage &AU) const override { 171 AU.addRequired<TargetLibraryInfoWrapperPass>(); 172 getLoopAnalysisUsage(AU); 173 } 174 }; 175 176 class LoopReroll { 177 public: 178 LoopReroll(AliasAnalysis *AA, LoopInfo *LI, ScalarEvolution *SE, 179 TargetLibraryInfo *TLI, DominatorTree *DT, bool PreserveLCSSA) 180 : AA(AA), LI(LI), SE(SE), TLI(TLI), DT(DT), 181 PreserveLCSSA(PreserveLCSSA) {} 182 bool runOnLoop(Loop *L); 183 184 protected: 185 AliasAnalysis *AA; 186 LoopInfo *LI; 187 ScalarEvolution *SE; 188 TargetLibraryInfo *TLI; 189 DominatorTree *DT; 190 bool PreserveLCSSA; 191 192 using SmallInstructionVector = SmallVector<Instruction *, 16>; 193 using SmallInstructionSet = SmallPtrSet<Instruction *, 16>; 194 using TinyInstructionVector = SmallVector<Instruction *, 1>; 195 196 // Map between induction variable and its increment 197 DenseMap<Instruction *, int64_t> IVToIncMap; 198 199 // For loop with multiple induction variables, remember the ones used only to 200 // control the loop. 201 TinyInstructionVector LoopControlIVs; 202 203 // A chain of isomorphic instructions, identified by a single-use PHI 204 // representing a reduction. Only the last value may be used outside the 205 // loop. 206 struct SimpleLoopReduction { 207 SimpleLoopReduction(Instruction *P, Loop *L) : Instructions(1, P) { 208 assert(isa<PHINode>(P) && "First reduction instruction must be a PHI"); 209 add(L); 210 } 211 212 bool valid() const { 213 return Valid; 214 } 215 216 Instruction *getPHI() const { 217 assert(Valid && "Using invalid reduction"); 218 return Instructions.front(); 219 } 220 221 Instruction *getReducedValue() const { 222 assert(Valid && "Using invalid reduction"); 223 return Instructions.back(); 224 } 225 226 Instruction *get(size_t i) const { 227 assert(Valid && "Using invalid reduction"); 228 return Instructions[i+1]; 229 } 230 231 Instruction *operator [] (size_t i) const { return get(i); } 232 233 // The size, ignoring the initial PHI. 234 size_t size() const { 235 assert(Valid && "Using invalid reduction"); 236 return Instructions.size()-1; 237 } 238 239 using iterator = SmallInstructionVector::iterator; 240 using const_iterator = SmallInstructionVector::const_iterator; 241 242 iterator begin() { 243 assert(Valid && "Using invalid reduction"); 244 return std::next(Instructions.begin()); 245 } 246 247 const_iterator begin() const { 248 assert(Valid && "Using invalid reduction"); 249 return std::next(Instructions.begin()); 250 } 251 252 iterator end() { return Instructions.end(); } 253 const_iterator end() const { return Instructions.end(); } 254 255 protected: 256 bool Valid = false; 257 SmallInstructionVector Instructions; 258 259 void add(Loop *L); 260 }; 261 262 // The set of all reductions, and state tracking of possible reductions 263 // during loop instruction processing. 264 struct ReductionTracker { 265 using SmallReductionVector = SmallVector<SimpleLoopReduction, 16>; 266 267 // Add a new possible reduction. 268 void addSLR(SimpleLoopReduction &SLR) { PossibleReds.push_back(SLR); } 269 270 // Setup to track possible reductions corresponding to the provided 271 // rerolling scale. Only reductions with a number of non-PHI instructions 272 // that is divisible by the scale are considered. Three instructions sets 273 // are filled in: 274 // - A set of all possible instructions in eligible reductions. 275 // - A set of all PHIs in eligible reductions 276 // - A set of all reduced values (last instructions) in eligible 277 // reductions. 278 void restrictToScale(uint64_t Scale, 279 SmallInstructionSet &PossibleRedSet, 280 SmallInstructionSet &PossibleRedPHISet, 281 SmallInstructionSet &PossibleRedLastSet) { 282 PossibleRedIdx.clear(); 283 PossibleRedIter.clear(); 284 Reds.clear(); 285 286 for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i) 287 if (PossibleReds[i].size() % Scale == 0) { 288 PossibleRedLastSet.insert(PossibleReds[i].getReducedValue()); 289 PossibleRedPHISet.insert(PossibleReds[i].getPHI()); 290 291 PossibleRedSet.insert(PossibleReds[i].getPHI()); 292 PossibleRedIdx[PossibleReds[i].getPHI()] = i; 293 for (Instruction *J : PossibleReds[i]) { 294 PossibleRedSet.insert(J); 295 PossibleRedIdx[J] = i; 296 } 297 } 298 } 299 300 // The functions below are used while processing the loop instructions. 301 302 // Are the two instructions both from reductions, and furthermore, from 303 // the same reduction? 304 bool isPairInSame(Instruction *J1, Instruction *J2) { 305 DenseMap<Instruction *, int>::iterator J1I = PossibleRedIdx.find(J1); 306 if (J1I != PossibleRedIdx.end()) { 307 DenseMap<Instruction *, int>::iterator J2I = PossibleRedIdx.find(J2); 308 if (J2I != PossibleRedIdx.end() && J1I->second == J2I->second) 309 return true; 310 } 311 312 return false; 313 } 314 315 // The two provided instructions, the first from the base iteration, and 316 // the second from iteration i, form a matched pair. If these are part of 317 // a reduction, record that fact. 318 void recordPair(Instruction *J1, Instruction *J2, unsigned i) { 319 if (PossibleRedIdx.count(J1)) { 320 assert(PossibleRedIdx.count(J2) && 321 "Recording reduction vs. non-reduction instruction?"); 322 323 PossibleRedIter[J1] = 0; 324 PossibleRedIter[J2] = i; 325 326 int Idx = PossibleRedIdx[J1]; 327 assert(Idx == PossibleRedIdx[J2] && 328 "Recording pair from different reductions?"); 329 Reds.insert(Idx); 330 } 331 } 332 333 // The functions below can be called after we've finished processing all 334 // instructions in the loop, and we know which reductions were selected. 335 336 bool validateSelected(); 337 void replaceSelected(); 338 339 protected: 340 // The vector of all possible reductions (for any scale). 341 SmallReductionVector PossibleReds; 342 343 DenseMap<Instruction *, int> PossibleRedIdx; 344 DenseMap<Instruction *, int> PossibleRedIter; 345 DenseSet<int> Reds; 346 }; 347 348 // A DAGRootSet models an induction variable being used in a rerollable 349 // loop. For example, 350 // 351 // x[i*3+0] = y1 352 // x[i*3+1] = y2 353 // x[i*3+2] = y3 354 // 355 // Base instruction -> i*3 356 // +---+----+ 357 // / | \ 358 // ST[y1] +1 +2 <-- Roots 359 // | | 360 // ST[y2] ST[y3] 361 // 362 // There may be multiple DAGRoots, for example: 363 // 364 // x[i*2+0] = ... (1) 365 // x[i*2+1] = ... (1) 366 // x[i*2+4] = ... (2) 367 // x[i*2+5] = ... (2) 368 // x[(i+1234)*2+5678] = ... (3) 369 // x[(i+1234)*2+5679] = ... (3) 370 // 371 // The loop will be rerolled by adding a new loop induction variable, 372 // one for the Base instruction in each DAGRootSet. 373 // 374 struct DAGRootSet { 375 Instruction *BaseInst; 376 SmallInstructionVector Roots; 377 378 // The instructions between IV and BaseInst (but not including BaseInst). 379 SmallInstructionSet SubsumedInsts; 380 }; 381 382 // The set of all DAG roots, and state tracking of all roots 383 // for a particular induction variable. 384 struct DAGRootTracker { 385 DAGRootTracker(LoopReroll *Parent, Loop *L, Instruction *IV, 386 ScalarEvolution *SE, AliasAnalysis *AA, 387 TargetLibraryInfo *TLI, DominatorTree *DT, LoopInfo *LI, 388 bool PreserveLCSSA, 389 DenseMap<Instruction *, int64_t> &IncrMap, 390 TinyInstructionVector LoopCtrlIVs) 391 : Parent(Parent), L(L), SE(SE), AA(AA), TLI(TLI), DT(DT), LI(LI), 392 PreserveLCSSA(PreserveLCSSA), IV(IV), IVToIncMap(IncrMap), 393 LoopControlIVs(LoopCtrlIVs) {} 394 395 /// Stage 1: Find all the DAG roots for the induction variable. 396 bool findRoots(); 397 398 /// Stage 2: Validate if the found roots are valid. 399 bool validate(ReductionTracker &Reductions); 400 401 /// Stage 3: Assuming validate() returned true, perform the 402 /// replacement. 403 /// @param BackedgeTakenCount The backedge-taken count of L. 404 void replace(const SCEV *BackedgeTakenCount); 405 406 protected: 407 using UsesTy = MapVector<Instruction *, BitVector>; 408 409 void findRootsRecursive(Instruction *IVU, 410 SmallInstructionSet SubsumedInsts); 411 bool findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts); 412 bool collectPossibleRoots(Instruction *Base, 413 std::map<int64_t,Instruction*> &Roots); 414 bool validateRootSet(DAGRootSet &DRS); 415 416 bool collectUsedInstructions(SmallInstructionSet &PossibleRedSet); 417 void collectInLoopUserSet(const SmallInstructionVector &Roots, 418 const SmallInstructionSet &Exclude, 419 const SmallInstructionSet &Final, 420 DenseSet<Instruction *> &Users); 421 void collectInLoopUserSet(Instruction *Root, 422 const SmallInstructionSet &Exclude, 423 const SmallInstructionSet &Final, 424 DenseSet<Instruction *> &Users); 425 426 UsesTy::iterator nextInstr(int Val, UsesTy &In, 427 const SmallInstructionSet &Exclude, 428 UsesTy::iterator *StartI=nullptr); 429 bool isBaseInst(Instruction *I); 430 bool isRootInst(Instruction *I); 431 bool instrDependsOn(Instruction *I, 432 UsesTy::iterator Start, 433 UsesTy::iterator End); 434 void replaceIV(DAGRootSet &DRS, const SCEV *Start, const SCEV *IncrExpr); 435 436 LoopReroll *Parent; 437 438 // Members of Parent, replicated here for brevity. 439 Loop *L; 440 ScalarEvolution *SE; 441 AliasAnalysis *AA; 442 TargetLibraryInfo *TLI; 443 DominatorTree *DT; 444 LoopInfo *LI; 445 bool PreserveLCSSA; 446 447 // The loop induction variable. 448 Instruction *IV; 449 450 // Loop step amount. 451 int64_t Inc; 452 453 // Loop reroll count; if Inc == 1, this records the scaling applied 454 // to the indvar: a[i*2+0] = ...; a[i*2+1] = ... ; 455 // If Inc is not 1, Scale = Inc. 456 uint64_t Scale; 457 458 // The roots themselves. 459 SmallVector<DAGRootSet,16> RootSets; 460 461 // All increment instructions for IV. 462 SmallInstructionVector LoopIncs; 463 464 // Map of all instructions in the loop (in order) to the iterations 465 // they are used in (or specially, IL_All for instructions 466 // used in the loop increment mechanism). 467 UsesTy Uses; 468 469 // Map between induction variable and its increment 470 DenseMap<Instruction *, int64_t> &IVToIncMap; 471 472 TinyInstructionVector LoopControlIVs; 473 }; 474 475 // Check if it is a compare-like instruction whose user is a branch 476 bool isCompareUsedByBranch(Instruction *I) { 477 auto *TI = I->getParent()->getTerminator(); 478 if (!isa<BranchInst>(TI) || !isa<CmpInst>(I)) 479 return false; 480 return I->hasOneUse() && TI->getOperand(0) == I; 481 }; 482 483 bool isLoopControlIV(Loop *L, Instruction *IV); 484 void collectPossibleIVs(Loop *L, SmallInstructionVector &PossibleIVs); 485 void collectPossibleReductions(Loop *L, 486 ReductionTracker &Reductions); 487 bool reroll(Instruction *IV, Loop *L, BasicBlock *Header, 488 const SCEV *BackedgeTakenCount, ReductionTracker &Reductions); 489 }; 490 491 } // end anonymous namespace 492 493 char LoopRerollLegacyPass::ID = 0; 494 495 INITIALIZE_PASS_BEGIN(LoopRerollLegacyPass, "loop-reroll", "Reroll loops", 496 false, false) 497 INITIALIZE_PASS_DEPENDENCY(LoopPass) 498 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 499 INITIALIZE_PASS_END(LoopRerollLegacyPass, "loop-reroll", "Reroll loops", false, 500 false) 501 502 Pass *llvm::createLoopRerollPass() { return new LoopRerollLegacyPass; } 503 504 // Returns true if the provided instruction is used outside the given loop. 505 // This operates like Instruction::isUsedOutsideOfBlock, but considers PHIs in 506 // non-loop blocks to be outside the loop. 507 static bool hasUsesOutsideLoop(Instruction *I, Loop *L) { 508 for (User *U : I->users()) { 509 if (!L->contains(cast<Instruction>(U))) 510 return true; 511 } 512 return false; 513 } 514 515 // Check if an IV is only used to control the loop. There are two cases: 516 // 1. It only has one use which is loop increment, and the increment is only 517 // used by comparison and the PHI (could has sext with nsw in between), and the 518 // comparison is only used by branch. 519 // 2. It is used by loop increment and the comparison, the loop increment is 520 // only used by the PHI, and the comparison is used only by the branch. 521 bool LoopReroll::isLoopControlIV(Loop *L, Instruction *IV) { 522 unsigned IVUses = IV->getNumUses(); 523 if (IVUses != 2 && IVUses != 1) 524 return false; 525 526 for (auto *User : IV->users()) { 527 int32_t IncOrCmpUses = User->getNumUses(); 528 bool IsCompInst = isCompareUsedByBranch(cast<Instruction>(User)); 529 530 // User can only have one or two uses. 531 if (IncOrCmpUses != 2 && IncOrCmpUses != 1) 532 return false; 533 534 // Case 1 535 if (IVUses == 1) { 536 // The only user must be the loop increment. 537 // The loop increment must have two uses. 538 if (IsCompInst || IncOrCmpUses != 2) 539 return false; 540 } 541 542 // Case 2 543 if (IVUses == 2 && IncOrCmpUses != 1) 544 return false; 545 546 // The users of the IV must be a binary operation or a comparison 547 if (auto *BO = dyn_cast<BinaryOperator>(User)) { 548 if (BO->getOpcode() == Instruction::Add) { 549 // Loop Increment 550 // User of Loop Increment should be either PHI or CMP 551 for (auto *UU : User->users()) { 552 if (PHINode *PN = dyn_cast<PHINode>(UU)) { 553 if (PN != IV) 554 return false; 555 } 556 // Must be a CMP or an ext (of a value with nsw) then CMP 557 else { 558 auto *UUser = cast<Instruction>(UU); 559 // Skip SExt if we are extending an nsw value 560 // TODO: Allow ZExt too 561 if (BO->hasNoSignedWrap() && UUser->hasOneUse() && 562 isa<SExtInst>(UUser)) 563 UUser = cast<Instruction>(*(UUser->user_begin())); 564 if (!isCompareUsedByBranch(UUser)) 565 return false; 566 } 567 } 568 } else 569 return false; 570 // Compare : can only have one use, and must be branch 571 } else if (!IsCompInst) 572 return false; 573 } 574 return true; 575 } 576 577 // Collect the list of loop induction variables with respect to which it might 578 // be possible to reroll the loop. 579 void LoopReroll::collectPossibleIVs(Loop *L, 580 SmallInstructionVector &PossibleIVs) { 581 for (Instruction &IV : L->getHeader()->phis()) { 582 if (!IV.getType()->isIntegerTy() && !IV.getType()->isPointerTy()) 583 continue; 584 585 if (const SCEVAddRecExpr *PHISCEV = 586 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(&IV))) { 587 if (PHISCEV->getLoop() != L) 588 continue; 589 if (!PHISCEV->isAffine()) 590 continue; 591 const auto *IncSCEV = dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE)); 592 if (IncSCEV) { 593 IVToIncMap[&IV] = IncSCEV->getValue()->getSExtValue(); 594 LLVM_DEBUG(dbgs() << "LRR: Possible IV: " << IV << " = " << *PHISCEV 595 << "\n"); 596 597 if (isLoopControlIV(L, &IV)) { 598 LoopControlIVs.push_back(&IV); 599 LLVM_DEBUG(dbgs() << "LRR: Loop control only IV: " << IV 600 << " = " << *PHISCEV << "\n"); 601 } else 602 PossibleIVs.push_back(&IV); 603 } 604 } 605 } 606 } 607 608 // Add the remainder of the reduction-variable chain to the instruction vector 609 // (the initial PHINode has already been added). If successful, the object is 610 // marked as valid. 611 void LoopReroll::SimpleLoopReduction::add(Loop *L) { 612 assert(!Valid && "Cannot add to an already-valid chain"); 613 614 // The reduction variable must be a chain of single-use instructions 615 // (including the PHI), except for the last value (which is used by the PHI 616 // and also outside the loop). 617 Instruction *C = Instructions.front(); 618 if (C->user_empty()) 619 return; 620 621 do { 622 C = cast<Instruction>(*C->user_begin()); 623 if (C->hasOneUse()) { 624 if (!C->isBinaryOp()) 625 return; 626 627 if (!(isa<PHINode>(Instructions.back()) || 628 C->isSameOperationAs(Instructions.back()))) 629 return; 630 631 Instructions.push_back(C); 632 } 633 } while (C->hasOneUse()); 634 635 if (Instructions.size() < 2 || 636 !C->isSameOperationAs(Instructions.back()) || 637 C->use_empty()) 638 return; 639 640 // C is now the (potential) last instruction in the reduction chain. 641 for (User *U : C->users()) { 642 // The only in-loop user can be the initial PHI. 643 if (L->contains(cast<Instruction>(U))) 644 if (cast<Instruction>(U) != Instructions.front()) 645 return; 646 } 647 648 Instructions.push_back(C); 649 Valid = true; 650 } 651 652 // Collect the vector of possible reduction variables. 653 void LoopReroll::collectPossibleReductions(Loop *L, 654 ReductionTracker &Reductions) { 655 BasicBlock *Header = L->getHeader(); 656 for (BasicBlock::iterator I = Header->begin(), 657 IE = Header->getFirstInsertionPt(); I != IE; ++I) { 658 if (!isa<PHINode>(I)) 659 continue; 660 if (!I->getType()->isSingleValueType()) 661 continue; 662 663 SimpleLoopReduction SLR(&*I, L); 664 if (!SLR.valid()) 665 continue; 666 667 LLVM_DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with " 668 << SLR.size() << " chained instructions)\n"); 669 Reductions.addSLR(SLR); 670 } 671 } 672 673 // Collect the set of all users of the provided root instruction. This set of 674 // users contains not only the direct users of the root instruction, but also 675 // all users of those users, and so on. There are two exceptions: 676 // 677 // 1. Instructions in the set of excluded instructions are never added to the 678 // use set (even if they are users). This is used, for example, to exclude 679 // including root increments in the use set of the primary IV. 680 // 681 // 2. Instructions in the set of final instructions are added to the use set 682 // if they are users, but their users are not added. This is used, for 683 // example, to prevent a reduction update from forcing all later reduction 684 // updates into the use set. 685 void LoopReroll::DAGRootTracker::collectInLoopUserSet( 686 Instruction *Root, const SmallInstructionSet &Exclude, 687 const SmallInstructionSet &Final, 688 DenseSet<Instruction *> &Users) { 689 SmallInstructionVector Queue(1, Root); 690 while (!Queue.empty()) { 691 Instruction *I = Queue.pop_back_val(); 692 if (!Users.insert(I).second) 693 continue; 694 695 if (!Final.count(I)) 696 for (Use &U : I->uses()) { 697 Instruction *User = cast<Instruction>(U.getUser()); 698 if (PHINode *PN = dyn_cast<PHINode>(User)) { 699 // Ignore "wrap-around" uses to PHIs of this loop's header. 700 if (PN->getIncomingBlock(U) == L->getHeader()) 701 continue; 702 } 703 704 if (L->contains(User) && !Exclude.count(User)) { 705 Queue.push_back(User); 706 } 707 } 708 709 // We also want to collect single-user "feeder" values. 710 for (Use &U : I->operands()) { 711 if (Instruction *Op = dyn_cast<Instruction>(U)) 712 if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) && 713 !Final.count(Op)) 714 Queue.push_back(Op); 715 } 716 } 717 } 718 719 // Collect all of the users of all of the provided root instructions (combined 720 // into a single set). 721 void LoopReroll::DAGRootTracker::collectInLoopUserSet( 722 const SmallInstructionVector &Roots, 723 const SmallInstructionSet &Exclude, 724 const SmallInstructionSet &Final, 725 DenseSet<Instruction *> &Users) { 726 for (Instruction *Root : Roots) 727 collectInLoopUserSet(Root, Exclude, Final, Users); 728 } 729 730 static bool isUnorderedLoadStore(Instruction *I) { 731 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 732 return LI->isUnordered(); 733 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 734 return SI->isUnordered(); 735 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) 736 return !MI->isVolatile(); 737 return false; 738 } 739 740 /// Return true if IVU is a "simple" arithmetic operation. 741 /// This is used for narrowing the search space for DAGRoots; only arithmetic 742 /// and GEPs can be part of a DAGRoot. 743 static bool isSimpleArithmeticOp(User *IVU) { 744 if (Instruction *I = dyn_cast<Instruction>(IVU)) { 745 switch (I->getOpcode()) { 746 default: return false; 747 case Instruction::Add: 748 case Instruction::Sub: 749 case Instruction::Mul: 750 case Instruction::Shl: 751 case Instruction::AShr: 752 case Instruction::LShr: 753 case Instruction::GetElementPtr: 754 case Instruction::Trunc: 755 case Instruction::ZExt: 756 case Instruction::SExt: 757 return true; 758 } 759 } 760 return false; 761 } 762 763 static bool isLoopIncrement(User *U, Instruction *IV) { 764 BinaryOperator *BO = dyn_cast<BinaryOperator>(U); 765 766 if ((BO && BO->getOpcode() != Instruction::Add) || 767 (!BO && !isa<GetElementPtrInst>(U))) 768 return false; 769 770 for (auto *UU : U->users()) { 771 PHINode *PN = dyn_cast<PHINode>(UU); 772 if (PN && PN == IV) 773 return true; 774 } 775 return false; 776 } 777 778 bool LoopReroll::DAGRootTracker:: 779 collectPossibleRoots(Instruction *Base, std::map<int64_t,Instruction*> &Roots) { 780 SmallInstructionVector BaseUsers; 781 782 for (auto *I : Base->users()) { 783 ConstantInt *CI = nullptr; 784 785 if (isLoopIncrement(I, IV)) { 786 LoopIncs.push_back(cast<Instruction>(I)); 787 continue; 788 } 789 790 // The root nodes must be either GEPs, ORs or ADDs. 791 if (auto *BO = dyn_cast<BinaryOperator>(I)) { 792 if (BO->getOpcode() == Instruction::Add || 793 BO->getOpcode() == Instruction::Or) 794 CI = dyn_cast<ConstantInt>(BO->getOperand(1)); 795 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) { 796 Value *LastOperand = GEP->getOperand(GEP->getNumOperands()-1); 797 CI = dyn_cast<ConstantInt>(LastOperand); 798 } 799 800 if (!CI) { 801 if (Instruction *II = dyn_cast<Instruction>(I)) { 802 BaseUsers.push_back(II); 803 continue; 804 } else { 805 LLVM_DEBUG(dbgs() << "LRR: Aborting due to non-instruction: " << *I 806 << "\n"); 807 return false; 808 } 809 } 810 811 int64_t V = std::abs(CI->getValue().getSExtValue()); 812 if (Roots.find(V) != Roots.end()) 813 // No duplicates, please. 814 return false; 815 816 Roots[V] = cast<Instruction>(I); 817 } 818 819 // Make sure we have at least two roots. 820 if (Roots.empty() || (Roots.size() == 1 && BaseUsers.empty())) 821 return false; 822 823 // If we found non-loop-inc, non-root users of Base, assume they are 824 // for the zeroth root index. This is because "add %a, 0" gets optimized 825 // away. 826 if (BaseUsers.size()) { 827 if (Roots.find(0) != Roots.end()) { 828 LLVM_DEBUG(dbgs() << "LRR: Multiple roots found for base - aborting!\n"); 829 return false; 830 } 831 Roots[0] = Base; 832 } 833 834 // Calculate the number of users of the base, or lowest indexed, iteration. 835 unsigned NumBaseUses = BaseUsers.size(); 836 if (NumBaseUses == 0) 837 NumBaseUses = Roots.begin()->second->getNumUses(); 838 839 // Check that every node has the same number of users. 840 for (auto &KV : Roots) { 841 if (KV.first == 0) 842 continue; 843 if (!KV.second->hasNUses(NumBaseUses)) { 844 LLVM_DEBUG(dbgs() << "LRR: Aborting - Root and Base #users not the same: " 845 << "#Base=" << NumBaseUses 846 << ", #Root=" << KV.second->getNumUses() << "\n"); 847 return false; 848 } 849 } 850 851 return true; 852 } 853 854 void LoopReroll::DAGRootTracker:: 855 findRootsRecursive(Instruction *I, SmallInstructionSet SubsumedInsts) { 856 // Does the user look like it could be part of a root set? 857 // All its users must be simple arithmetic ops. 858 if (I->hasNUsesOrMore(IL_MaxRerollIterations + 1)) 859 return; 860 861 if (I != IV && findRootsBase(I, SubsumedInsts)) 862 return; 863 864 SubsumedInsts.insert(I); 865 866 for (User *V : I->users()) { 867 Instruction *I = cast<Instruction>(V); 868 if (is_contained(LoopIncs, I)) 869 continue; 870 871 if (!isSimpleArithmeticOp(I)) 872 continue; 873 874 // The recursive call makes a copy of SubsumedInsts. 875 findRootsRecursive(I, SubsumedInsts); 876 } 877 } 878 879 bool LoopReroll::DAGRootTracker::validateRootSet(DAGRootSet &DRS) { 880 if (DRS.Roots.empty()) 881 return false; 882 883 // If the value of the base instruction is used outside the loop, we cannot 884 // reroll the loop. Check for other root instructions is unnecessary because 885 // they don't match any base instructions if their values are used outside. 886 if (hasUsesOutsideLoop(DRS.BaseInst, L)) 887 return false; 888 889 // Consider a DAGRootSet with N-1 roots (so N different values including 890 // BaseInst). 891 // Define d = Roots[0] - BaseInst, which should be the same as 892 // Roots[I] - Roots[I-1] for all I in [1..N). 893 // Define D = BaseInst@J - BaseInst@J-1, where "@J" means the value at the 894 // loop iteration J. 895 // 896 // Now, For the loop iterations to be consecutive: 897 // D = d * N 898 const auto *ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst)); 899 if (!ADR) 900 return false; 901 902 // Check that the first root is evenly spaced. 903 unsigned N = DRS.Roots.size() + 1; 904 const SCEV *StepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), ADR); 905 if (isa<SCEVCouldNotCompute>(StepSCEV) || StepSCEV->getType()->isPointerTy()) 906 return false; 907 const SCEV *ScaleSCEV = SE->getConstant(StepSCEV->getType(), N); 908 if (ADR->getStepRecurrence(*SE) != SE->getMulExpr(StepSCEV, ScaleSCEV)) 909 return false; 910 911 // Check that the remainling roots are evenly spaced. 912 for (unsigned i = 1; i < N - 1; ++i) { 913 const SCEV *NewStepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[i]), 914 SE->getSCEV(DRS.Roots[i-1])); 915 if (NewStepSCEV != StepSCEV) 916 return false; 917 } 918 919 return true; 920 } 921 922 bool LoopReroll::DAGRootTracker:: 923 findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts) { 924 // The base of a RootSet must be an AddRec, so it can be erased. 925 const auto *IVU_ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IVU)); 926 if (!IVU_ADR || IVU_ADR->getLoop() != L) 927 return false; 928 929 std::map<int64_t, Instruction*> V; 930 if (!collectPossibleRoots(IVU, V)) 931 return false; 932 933 // If we didn't get a root for index zero, then IVU must be 934 // subsumed. 935 if (V.find(0) == V.end()) 936 SubsumedInsts.insert(IVU); 937 938 // Partition the vector into monotonically increasing indexes. 939 DAGRootSet DRS; 940 DRS.BaseInst = nullptr; 941 942 SmallVector<DAGRootSet, 16> PotentialRootSets; 943 944 for (auto &KV : V) { 945 if (!DRS.BaseInst) { 946 DRS.BaseInst = KV.second; 947 DRS.SubsumedInsts = SubsumedInsts; 948 } else if (DRS.Roots.empty()) { 949 DRS.Roots.push_back(KV.second); 950 } else if (V.find(KV.first - 1) != V.end()) { 951 DRS.Roots.push_back(KV.second); 952 } else { 953 // Linear sequence terminated. 954 if (!validateRootSet(DRS)) 955 return false; 956 957 // Construct a new DAGRootSet with the next sequence. 958 PotentialRootSets.push_back(DRS); 959 DRS.BaseInst = KV.second; 960 DRS.Roots.clear(); 961 } 962 } 963 964 if (!validateRootSet(DRS)) 965 return false; 966 967 PotentialRootSets.push_back(DRS); 968 969 RootSets.append(PotentialRootSets.begin(), PotentialRootSets.end()); 970 971 return true; 972 } 973 974 bool LoopReroll::DAGRootTracker::findRoots() { 975 Inc = IVToIncMap[IV]; 976 977 assert(RootSets.empty() && "Unclean state!"); 978 if (std::abs(Inc) == 1) { 979 for (auto *IVU : IV->users()) { 980 if (isLoopIncrement(IVU, IV)) 981 LoopIncs.push_back(cast<Instruction>(IVU)); 982 } 983 findRootsRecursive(IV, SmallInstructionSet()); 984 LoopIncs.push_back(IV); 985 } else { 986 if (!findRootsBase(IV, SmallInstructionSet())) 987 return false; 988 } 989 990 // Ensure all sets have the same size. 991 if (RootSets.empty()) { 992 LLVM_DEBUG(dbgs() << "LRR: Aborting because no root sets found!\n"); 993 return false; 994 } 995 for (auto &V : RootSets) { 996 if (V.Roots.empty() || V.Roots.size() != RootSets[0].Roots.size()) { 997 LLVM_DEBUG( 998 dbgs() 999 << "LRR: Aborting because not all root sets have the same size\n"); 1000 return false; 1001 } 1002 } 1003 1004 Scale = RootSets[0].Roots.size() + 1; 1005 1006 if (Scale > IL_MaxRerollIterations) { 1007 LLVM_DEBUG(dbgs() << "LRR: Aborting - too many iterations found. " 1008 << "#Found=" << Scale 1009 << ", #Max=" << IL_MaxRerollIterations << "\n"); 1010 return false; 1011 } 1012 1013 LLVM_DEBUG(dbgs() << "LRR: Successfully found roots: Scale=" << Scale 1014 << "\n"); 1015 1016 return true; 1017 } 1018 1019 bool LoopReroll::DAGRootTracker::collectUsedInstructions(SmallInstructionSet &PossibleRedSet) { 1020 // Populate the MapVector with all instructions in the block, in order first, 1021 // so we can iterate over the contents later in perfect order. 1022 for (auto &I : *L->getHeader()) { 1023 Uses[&I].resize(IL_End); 1024 } 1025 1026 SmallInstructionSet Exclude; 1027 for (auto &DRS : RootSets) { 1028 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end()); 1029 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end()); 1030 Exclude.insert(DRS.BaseInst); 1031 } 1032 Exclude.insert(LoopIncs.begin(), LoopIncs.end()); 1033 1034 for (auto &DRS : RootSets) { 1035 DenseSet<Instruction*> VBase; 1036 collectInLoopUserSet(DRS.BaseInst, Exclude, PossibleRedSet, VBase); 1037 for (auto *I : VBase) { 1038 Uses[I].set(0); 1039 } 1040 1041 unsigned Idx = 1; 1042 for (auto *Root : DRS.Roots) { 1043 DenseSet<Instruction*> V; 1044 collectInLoopUserSet(Root, Exclude, PossibleRedSet, V); 1045 1046 // While we're here, check the use sets are the same size. 1047 if (V.size() != VBase.size()) { 1048 LLVM_DEBUG(dbgs() << "LRR: Aborting - use sets are different sizes\n"); 1049 return false; 1050 } 1051 1052 for (auto *I : V) { 1053 Uses[I].set(Idx); 1054 } 1055 ++Idx; 1056 } 1057 1058 // Make sure our subsumed instructions are remembered too. 1059 for (auto *I : DRS.SubsumedInsts) { 1060 Uses[I].set(IL_All); 1061 } 1062 } 1063 1064 // Make sure the loop increments are also accounted for. 1065 1066 Exclude.clear(); 1067 for (auto &DRS : RootSets) { 1068 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end()); 1069 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end()); 1070 Exclude.insert(DRS.BaseInst); 1071 } 1072 1073 DenseSet<Instruction*> V; 1074 collectInLoopUserSet(LoopIncs, Exclude, PossibleRedSet, V); 1075 for (auto *I : V) { 1076 if (I->mayHaveSideEffects()) { 1077 LLVM_DEBUG(dbgs() << "LRR: Aborting - " 1078 << "An instruction which does not belong to any root " 1079 << "sets must not have side effects: " << *I); 1080 return false; 1081 } 1082 Uses[I].set(IL_All); 1083 } 1084 1085 return true; 1086 } 1087 1088 /// Get the next instruction in "In" that is a member of set Val. 1089 /// Start searching from StartI, and do not return anything in Exclude. 1090 /// If StartI is not given, start from In.begin(). 1091 LoopReroll::DAGRootTracker::UsesTy::iterator 1092 LoopReroll::DAGRootTracker::nextInstr(int Val, UsesTy &In, 1093 const SmallInstructionSet &Exclude, 1094 UsesTy::iterator *StartI) { 1095 UsesTy::iterator I = StartI ? *StartI : In.begin(); 1096 while (I != In.end() && (I->second.test(Val) == 0 || 1097 Exclude.contains(I->first))) 1098 ++I; 1099 return I; 1100 } 1101 1102 bool LoopReroll::DAGRootTracker::isBaseInst(Instruction *I) { 1103 for (auto &DRS : RootSets) { 1104 if (DRS.BaseInst == I) 1105 return true; 1106 } 1107 return false; 1108 } 1109 1110 bool LoopReroll::DAGRootTracker::isRootInst(Instruction *I) { 1111 for (auto &DRS : RootSets) { 1112 if (is_contained(DRS.Roots, I)) 1113 return true; 1114 } 1115 return false; 1116 } 1117 1118 /// Return true if instruction I depends on any instruction between 1119 /// Start and End. 1120 bool LoopReroll::DAGRootTracker::instrDependsOn(Instruction *I, 1121 UsesTy::iterator Start, 1122 UsesTy::iterator End) { 1123 for (auto *U : I->users()) { 1124 for (auto It = Start; It != End; ++It) 1125 if (U == It->first) 1126 return true; 1127 } 1128 return false; 1129 } 1130 1131 static bool isIgnorableInst(const Instruction *I) { 1132 if (isa<DbgInfoIntrinsic>(I)) 1133 return true; 1134 const IntrinsicInst* II = dyn_cast<IntrinsicInst>(I); 1135 if (!II) 1136 return false; 1137 switch (II->getIntrinsicID()) { 1138 default: 1139 return false; 1140 case Intrinsic::annotation: 1141 case Intrinsic::ptr_annotation: 1142 case Intrinsic::var_annotation: 1143 // TODO: the following intrinsics may also be allowed: 1144 // lifetime_start, lifetime_end, invariant_start, invariant_end 1145 return true; 1146 } 1147 return false; 1148 } 1149 1150 bool LoopReroll::DAGRootTracker::validate(ReductionTracker &Reductions) { 1151 // We now need to check for equivalence of the use graph of each root with 1152 // that of the primary induction variable (excluding the roots). Our goal 1153 // here is not to solve the full graph isomorphism problem, but rather to 1154 // catch common cases without a lot of work. As a result, we will assume 1155 // that the relative order of the instructions in each unrolled iteration 1156 // is the same (although we will not make an assumption about how the 1157 // different iterations are intermixed). Note that while the order must be 1158 // the same, the instructions may not be in the same basic block. 1159 1160 // An array of just the possible reductions for this scale factor. When we 1161 // collect the set of all users of some root instructions, these reduction 1162 // instructions are treated as 'final' (their uses are not considered). 1163 // This is important because we don't want the root use set to search down 1164 // the reduction chain. 1165 SmallInstructionSet PossibleRedSet; 1166 SmallInstructionSet PossibleRedLastSet; 1167 SmallInstructionSet PossibleRedPHISet; 1168 Reductions.restrictToScale(Scale, PossibleRedSet, 1169 PossibleRedPHISet, PossibleRedLastSet); 1170 1171 // Populate "Uses" with where each instruction is used. 1172 if (!collectUsedInstructions(PossibleRedSet)) 1173 return false; 1174 1175 // Make sure we mark the reduction PHIs as used in all iterations. 1176 for (auto *I : PossibleRedPHISet) { 1177 Uses[I].set(IL_All); 1178 } 1179 1180 // Make sure we mark loop-control-only PHIs as used in all iterations. See 1181 // comment above LoopReroll::isLoopControlIV for more information. 1182 BasicBlock *Header = L->getHeader(); 1183 for (Instruction *LoopControlIV : LoopControlIVs) { 1184 for (auto *U : LoopControlIV->users()) { 1185 Instruction *IVUser = dyn_cast<Instruction>(U); 1186 // IVUser could be loop increment or compare 1187 Uses[IVUser].set(IL_All); 1188 for (auto *UU : IVUser->users()) { 1189 Instruction *UUser = dyn_cast<Instruction>(UU); 1190 // UUser could be compare, PHI or branch 1191 Uses[UUser].set(IL_All); 1192 // Skip SExt 1193 if (isa<SExtInst>(UUser)) { 1194 UUser = dyn_cast<Instruction>(*(UUser->user_begin())); 1195 Uses[UUser].set(IL_All); 1196 } 1197 // Is UUser a compare instruction? 1198 if (UU->hasOneUse()) { 1199 Instruction *BI = dyn_cast<BranchInst>(*UUser->user_begin()); 1200 if (BI == cast<BranchInst>(Header->getTerminator())) 1201 Uses[BI].set(IL_All); 1202 } 1203 } 1204 } 1205 } 1206 1207 // Make sure all instructions in the loop are in one and only one 1208 // set. 1209 for (auto &KV : Uses) { 1210 if (KV.second.count() != 1 && !isIgnorableInst(KV.first)) { 1211 LLVM_DEBUG( 1212 dbgs() << "LRR: Aborting - instruction is not used in 1 iteration: " 1213 << *KV.first << " (#uses=" << KV.second.count() << ")\n"); 1214 return false; 1215 } 1216 } 1217 1218 LLVM_DEBUG(for (auto &KV 1219 : Uses) { 1220 dbgs() << "LRR: " << KV.second.find_first() << "\t" << *KV.first << "\n"; 1221 }); 1222 1223 BatchAAResults BatchAA(*AA); 1224 for (unsigned Iter = 1; Iter < Scale; ++Iter) { 1225 // In addition to regular aliasing information, we need to look for 1226 // instructions from later (future) iterations that have side effects 1227 // preventing us from reordering them past other instructions with side 1228 // effects. 1229 bool FutureSideEffects = false; 1230 AliasSetTracker AST(BatchAA); 1231 // The map between instructions in f(%iv.(i+1)) and f(%iv). 1232 DenseMap<Value *, Value *> BaseMap; 1233 1234 // Compare iteration Iter to the base. 1235 SmallInstructionSet Visited; 1236 auto BaseIt = nextInstr(0, Uses, Visited); 1237 auto RootIt = nextInstr(Iter, Uses, Visited); 1238 auto LastRootIt = Uses.begin(); 1239 1240 while (BaseIt != Uses.end() && RootIt != Uses.end()) { 1241 Instruction *BaseInst = BaseIt->first; 1242 Instruction *RootInst = RootIt->first; 1243 1244 // Skip over the IV or root instructions; only match their users. 1245 bool Continue = false; 1246 if (isBaseInst(BaseInst)) { 1247 Visited.insert(BaseInst); 1248 BaseIt = nextInstr(0, Uses, Visited); 1249 Continue = true; 1250 } 1251 if (isRootInst(RootInst)) { 1252 LastRootIt = RootIt; 1253 Visited.insert(RootInst); 1254 RootIt = nextInstr(Iter, Uses, Visited); 1255 Continue = true; 1256 } 1257 if (Continue) continue; 1258 1259 if (!BaseInst->isSameOperationAs(RootInst)) { 1260 // Last chance saloon. We don't try and solve the full isomorphism 1261 // problem, but try and at least catch the case where two instructions 1262 // *of different types* are round the wrong way. We won't be able to 1263 // efficiently tell, given two ADD instructions, which way around we 1264 // should match them, but given an ADD and a SUB, we can at least infer 1265 // which one is which. 1266 // 1267 // This should allow us to deal with a greater subset of the isomorphism 1268 // problem. It does however change a linear algorithm into a quadratic 1269 // one, so limit the number of probes we do. 1270 auto TryIt = RootIt; 1271 unsigned N = NumToleratedFailedMatches; 1272 while (TryIt != Uses.end() && 1273 !BaseInst->isSameOperationAs(TryIt->first) && 1274 N--) { 1275 ++TryIt; 1276 TryIt = nextInstr(Iter, Uses, Visited, &TryIt); 1277 } 1278 1279 if (TryIt == Uses.end() || TryIt == RootIt || 1280 instrDependsOn(TryIt->first, RootIt, TryIt)) { 1281 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " 1282 << *BaseInst << " vs. " << *RootInst << "\n"); 1283 return false; 1284 } 1285 1286 RootIt = TryIt; 1287 RootInst = TryIt->first; 1288 } 1289 1290 // All instructions between the last root and this root 1291 // may belong to some other iteration. If they belong to a 1292 // future iteration, then they're dangerous to alias with. 1293 // 1294 // Note that because we allow a limited amount of flexibility in the order 1295 // that we visit nodes, LastRootIt might be *before* RootIt, in which 1296 // case we've already checked this set of instructions so we shouldn't 1297 // do anything. 1298 for (; LastRootIt < RootIt; ++LastRootIt) { 1299 Instruction *I = LastRootIt->first; 1300 if (LastRootIt->second.find_first() < (int)Iter) 1301 continue; 1302 if (I->mayWriteToMemory()) 1303 AST.add(I); 1304 // Note: This is specifically guarded by a check on isa<PHINode>, 1305 // which while a valid (somewhat arbitrary) micro-optimization, is 1306 // needed because otherwise isSafeToSpeculativelyExecute returns 1307 // false on PHI nodes. 1308 if (!isa<PHINode>(I) && !isUnorderedLoadStore(I) && 1309 !isSafeToSpeculativelyExecute(I)) 1310 // Intervening instructions cause side effects. 1311 FutureSideEffects = true; 1312 } 1313 1314 // Make sure that this instruction, which is in the use set of this 1315 // root instruction, does not also belong to the base set or the set of 1316 // some other root instruction. 1317 if (RootIt->second.count() > 1) { 1318 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1319 << " vs. " << *RootInst << " (prev. case overlap)\n"); 1320 return false; 1321 } 1322 1323 // Make sure that we don't alias with any instruction in the alias set 1324 // tracker. If we do, then we depend on a future iteration, and we 1325 // can't reroll. 1326 if (RootInst->mayReadFromMemory()) { 1327 for (auto &K : AST) { 1328 if (isModOrRefSet(K.aliasesUnknownInst(RootInst, BatchAA))) { 1329 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " 1330 << *BaseInst << " vs. " << *RootInst 1331 << " (depends on future store)\n"); 1332 return false; 1333 } 1334 } 1335 } 1336 1337 // If we've past an instruction from a future iteration that may have 1338 // side effects, and this instruction might also, then we can't reorder 1339 // them, and this matching fails. As an exception, we allow the alias 1340 // set tracker to handle regular (unordered) load/store dependencies. 1341 if (FutureSideEffects && ((!isUnorderedLoadStore(BaseInst) && 1342 !isSafeToSpeculativelyExecute(BaseInst)) || 1343 (!isUnorderedLoadStore(RootInst) && 1344 !isSafeToSpeculativelyExecute(RootInst)))) { 1345 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1346 << " vs. " << *RootInst 1347 << " (side effects prevent reordering)\n"); 1348 return false; 1349 } 1350 1351 // For instructions that are part of a reduction, if the operation is 1352 // associative, then don't bother matching the operands (because we 1353 // already know that the instructions are isomorphic, and the order 1354 // within the iteration does not matter). For non-associative reductions, 1355 // we do need to match the operands, because we need to reject 1356 // out-of-order instructions within an iteration! 1357 // For example (assume floating-point addition), we need to reject this: 1358 // x += a[i]; x += b[i]; 1359 // x += a[i+1]; x += b[i+1]; 1360 // x += b[i+2]; x += a[i+2]; 1361 bool InReduction = Reductions.isPairInSame(BaseInst, RootInst); 1362 1363 if (!(InReduction && BaseInst->isAssociative())) { 1364 bool Swapped = false, SomeOpMatched = false; 1365 for (unsigned j = 0; j < BaseInst->getNumOperands(); ++j) { 1366 Value *Op2 = RootInst->getOperand(j); 1367 1368 // If this is part of a reduction (and the operation is not 1369 // associatve), then we match all operands, but not those that are 1370 // part of the reduction. 1371 if (InReduction) 1372 if (Instruction *Op2I = dyn_cast<Instruction>(Op2)) 1373 if (Reductions.isPairInSame(RootInst, Op2I)) 1374 continue; 1375 1376 DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2); 1377 if (BMI != BaseMap.end()) { 1378 Op2 = BMI->second; 1379 } else { 1380 for (auto &DRS : RootSets) { 1381 if (DRS.Roots[Iter-1] == (Instruction*) Op2) { 1382 Op2 = DRS.BaseInst; 1383 break; 1384 } 1385 } 1386 } 1387 1388 if (BaseInst->getOperand(Swapped ? unsigned(!j) : j) != Op2) { 1389 // If we've not already decided to swap the matched operands, and 1390 // we've not already matched our first operand (note that we could 1391 // have skipped matching the first operand because it is part of a 1392 // reduction above), and the instruction is commutative, then try 1393 // the swapped match. 1394 if (!Swapped && BaseInst->isCommutative() && !SomeOpMatched && 1395 BaseInst->getOperand(!j) == Op2) { 1396 Swapped = true; 1397 } else { 1398 LLVM_DEBUG(dbgs() 1399 << "LRR: iteration root match failed at " << *BaseInst 1400 << " vs. " << *RootInst << " (operand " << j << ")\n"); 1401 return false; 1402 } 1403 } 1404 1405 SomeOpMatched = true; 1406 } 1407 } 1408 1409 if ((!PossibleRedLastSet.count(BaseInst) && 1410 hasUsesOutsideLoop(BaseInst, L)) || 1411 (!PossibleRedLastSet.count(RootInst) && 1412 hasUsesOutsideLoop(RootInst, L))) { 1413 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1414 << " vs. " << *RootInst << " (uses outside loop)\n"); 1415 return false; 1416 } 1417 1418 Reductions.recordPair(BaseInst, RootInst, Iter); 1419 BaseMap.insert(std::make_pair(RootInst, BaseInst)); 1420 1421 LastRootIt = RootIt; 1422 Visited.insert(BaseInst); 1423 Visited.insert(RootInst); 1424 BaseIt = nextInstr(0, Uses, Visited); 1425 RootIt = nextInstr(Iter, Uses, Visited); 1426 } 1427 assert(BaseIt == Uses.end() && RootIt == Uses.end() && 1428 "Mismatched set sizes!"); 1429 } 1430 1431 LLVM_DEBUG(dbgs() << "LRR: Matched all iteration increments for " << *IV 1432 << "\n"); 1433 1434 return true; 1435 } 1436 1437 void LoopReroll::DAGRootTracker::replace(const SCEV *BackedgeTakenCount) { 1438 BasicBlock *Header = L->getHeader(); 1439 1440 // Compute the start and increment for each BaseInst before we start erasing 1441 // instructions. 1442 SmallVector<const SCEV *, 8> StartExprs; 1443 SmallVector<const SCEV *, 8> IncrExprs; 1444 for (auto &DRS : RootSets) { 1445 const SCEVAddRecExpr *IVSCEV = 1446 cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst)); 1447 StartExprs.push_back(IVSCEV->getStart()); 1448 IncrExprs.push_back(SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), IVSCEV)); 1449 } 1450 1451 // Remove instructions associated with non-base iterations. 1452 for (Instruction &Inst : llvm::make_early_inc_range(llvm::reverse(*Header))) { 1453 unsigned I = Uses[&Inst].find_first(); 1454 if (I > 0 && I < IL_All) { 1455 LLVM_DEBUG(dbgs() << "LRR: removing: " << Inst << "\n"); 1456 Inst.eraseFromParent(); 1457 } 1458 } 1459 1460 // Rewrite each BaseInst using SCEV. 1461 for (size_t i = 0, e = RootSets.size(); i != e; ++i) 1462 // Insert the new induction variable. 1463 replaceIV(RootSets[i], StartExprs[i], IncrExprs[i]); 1464 1465 { // Limit the lifetime of SCEVExpander. 1466 BranchInst *BI = cast<BranchInst>(Header->getTerminator()); 1467 const DataLayout &DL = Header->getModule()->getDataLayout(); 1468 SCEVExpander Expander(*SE, DL, "reroll"); 1469 auto Zero = SE->getZero(BackedgeTakenCount->getType()); 1470 auto One = SE->getOne(BackedgeTakenCount->getType()); 1471 auto NewIVSCEV = SE->getAddRecExpr(Zero, One, L, SCEV::FlagAnyWrap); 1472 Value *NewIV = 1473 Expander.expandCodeFor(NewIVSCEV, BackedgeTakenCount->getType(), 1474 Header->getFirstNonPHIOrDbg()); 1475 // FIXME: This arithmetic can overflow. 1476 auto TripCount = SE->getAddExpr(BackedgeTakenCount, One); 1477 auto ScaledTripCount = SE->getMulExpr( 1478 TripCount, SE->getConstant(BackedgeTakenCount->getType(), Scale)); 1479 auto ScaledBECount = SE->getMinusSCEV(ScaledTripCount, One); 1480 Value *TakenCount = 1481 Expander.expandCodeFor(ScaledBECount, BackedgeTakenCount->getType(), 1482 Header->getFirstNonPHIOrDbg()); 1483 Value *Cond = 1484 new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, TakenCount, "exitcond"); 1485 BI->setCondition(Cond); 1486 1487 if (BI->getSuccessor(1) != Header) 1488 BI->swapSuccessors(); 1489 } 1490 1491 SimplifyInstructionsInBlock(Header, TLI); 1492 DeleteDeadPHIs(Header, TLI); 1493 } 1494 1495 void LoopReroll::DAGRootTracker::replaceIV(DAGRootSet &DRS, 1496 const SCEV *Start, 1497 const SCEV *IncrExpr) { 1498 BasicBlock *Header = L->getHeader(); 1499 Instruction *Inst = DRS.BaseInst; 1500 1501 const SCEV *NewIVSCEV = 1502 SE->getAddRecExpr(Start, IncrExpr, L, SCEV::FlagAnyWrap); 1503 1504 { // Limit the lifetime of SCEVExpander. 1505 const DataLayout &DL = Header->getModule()->getDataLayout(); 1506 SCEVExpander Expander(*SE, DL, "reroll"); 1507 Value *NewIV = Expander.expandCodeFor(NewIVSCEV, Inst->getType(), 1508 Header->getFirstNonPHIOrDbg()); 1509 1510 for (auto &KV : Uses) 1511 if (KV.second.find_first() == 0) 1512 KV.first->replaceUsesOfWith(Inst, NewIV); 1513 } 1514 } 1515 1516 // Validate the selected reductions. All iterations must have an isomorphic 1517 // part of the reduction chain and, for non-associative reductions, the chain 1518 // entries must appear in order. 1519 bool LoopReroll::ReductionTracker::validateSelected() { 1520 // For a non-associative reduction, the chain entries must appear in order. 1521 for (int i : Reds) { 1522 int PrevIter = 0, BaseCount = 0, Count = 0; 1523 for (Instruction *J : PossibleReds[i]) { 1524 // Note that all instructions in the chain must have been found because 1525 // all instructions in the function must have been assigned to some 1526 // iteration. 1527 int Iter = PossibleRedIter[J]; 1528 if (Iter != PrevIter && Iter != PrevIter + 1 && 1529 !PossibleReds[i].getReducedValue()->isAssociative()) { 1530 LLVM_DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: " 1531 << J << "\n"); 1532 return false; 1533 } 1534 1535 if (Iter != PrevIter) { 1536 if (Count != BaseCount) { 1537 LLVM_DEBUG(dbgs() 1538 << "LRR: Iteration " << PrevIter << " reduction use count " 1539 << Count << " is not equal to the base use count " 1540 << BaseCount << "\n"); 1541 return false; 1542 } 1543 1544 Count = 0; 1545 } 1546 1547 ++Count; 1548 if (Iter == 0) 1549 ++BaseCount; 1550 1551 PrevIter = Iter; 1552 } 1553 } 1554 1555 return true; 1556 } 1557 1558 // For all selected reductions, remove all parts except those in the first 1559 // iteration (and the PHI). Replace outside uses of the reduced value with uses 1560 // of the first-iteration reduced value (in other words, reroll the selected 1561 // reductions). 1562 void LoopReroll::ReductionTracker::replaceSelected() { 1563 // Fixup reductions to refer to the last instruction associated with the 1564 // first iteration (not the last). 1565 for (int i : Reds) { 1566 int j = 0; 1567 for (int e = PossibleReds[i].size(); j != e; ++j) 1568 if (PossibleRedIter[PossibleReds[i][j]] != 0) { 1569 --j; 1570 break; 1571 } 1572 1573 // Replace users with the new end-of-chain value. 1574 SmallInstructionVector Users; 1575 for (User *U : PossibleReds[i].getReducedValue()->users()) { 1576 Users.push_back(cast<Instruction>(U)); 1577 } 1578 1579 for (Instruction *User : Users) 1580 User->replaceUsesOfWith(PossibleReds[i].getReducedValue(), 1581 PossibleReds[i][j]); 1582 } 1583 } 1584 1585 // Reroll the provided loop with respect to the provided induction variable. 1586 // Generally, we're looking for a loop like this: 1587 // 1588 // %iv = phi [ (preheader, ...), (body, %iv.next) ] 1589 // f(%iv) 1590 // %iv.1 = add %iv, 1 <-- a root increment 1591 // f(%iv.1) 1592 // %iv.2 = add %iv, 2 <-- a root increment 1593 // f(%iv.2) 1594 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment 1595 // f(%iv.scale_m_1) 1596 // ... 1597 // %iv.next = add %iv, scale 1598 // %cmp = icmp(%iv, ...) 1599 // br %cmp, header, exit 1600 // 1601 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of 1602 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can 1603 // be intermixed with eachother. The restriction imposed by this algorithm is 1604 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1), 1605 // etc. be the same. 1606 // 1607 // First, we collect the use set of %iv, excluding the other increment roots. 1608 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1) 1609 // times, having collected the use set of f(%iv.(i+1)), during which we: 1610 // - Ensure that the next unmatched instruction in f(%iv) is isomorphic to 1611 // the next unmatched instruction in f(%iv.(i+1)). 1612 // - Ensure that both matched instructions don't have any external users 1613 // (with the exception of last-in-chain reduction instructions). 1614 // - Track the (aliasing) write set, and other side effects, of all 1615 // instructions that belong to future iterations that come before the matched 1616 // instructions. If the matched instructions read from that write set, then 1617 // f(%iv) or f(%iv.(i+1)) has some dependency on instructions in 1618 // f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly, 1619 // if any of these future instructions had side effects (could not be 1620 // speculatively executed), and so do the matched instructions, when we 1621 // cannot reorder those side-effect-producing instructions, and rerolling 1622 // fails. 1623 // 1624 // Finally, we make sure that all loop instructions are either loop increment 1625 // roots, belong to simple latch code, parts of validated reductions, part of 1626 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions 1627 // have been validated), then we reroll the loop. 1628 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header, 1629 const SCEV *BackedgeTakenCount, 1630 ReductionTracker &Reductions) { 1631 DAGRootTracker DAGRoots(this, L, IV, SE, AA, TLI, DT, LI, PreserveLCSSA, 1632 IVToIncMap, LoopControlIVs); 1633 1634 if (!DAGRoots.findRoots()) 1635 return false; 1636 LLVM_DEBUG(dbgs() << "LRR: Found all root induction increments for: " << *IV 1637 << "\n"); 1638 1639 if (!DAGRoots.validate(Reductions)) 1640 return false; 1641 if (!Reductions.validateSelected()) 1642 return false; 1643 // At this point, we've validated the rerolling, and we're committed to 1644 // making changes! 1645 1646 Reductions.replaceSelected(); 1647 DAGRoots.replace(BackedgeTakenCount); 1648 1649 ++NumRerolledLoops; 1650 return true; 1651 } 1652 1653 bool LoopReroll::runOnLoop(Loop *L) { 1654 BasicBlock *Header = L->getHeader(); 1655 LLVM_DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() << "] Loop %" 1656 << Header->getName() << " (" << L->getNumBlocks() 1657 << " block(s))\n"); 1658 1659 // For now, we'll handle only single BB loops. 1660 if (L->getNumBlocks() > 1) 1661 return false; 1662 1663 if (!SE->hasLoopInvariantBackedgeTakenCount(L)) 1664 return false; 1665 1666 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1667 LLVM_DEBUG(dbgs() << "\n Before Reroll:\n" << *(L->getHeader()) << "\n"); 1668 LLVM_DEBUG(dbgs() << "LRR: backedge-taken count = " << *BackedgeTakenCount 1669 << "\n"); 1670 1671 // First, we need to find the induction variable with respect to which we can 1672 // reroll (there may be several possible options). 1673 SmallInstructionVector PossibleIVs; 1674 IVToIncMap.clear(); 1675 LoopControlIVs.clear(); 1676 collectPossibleIVs(L, PossibleIVs); 1677 1678 if (PossibleIVs.empty()) { 1679 LLVM_DEBUG(dbgs() << "LRR: No possible IVs found\n"); 1680 return false; 1681 } 1682 1683 ReductionTracker Reductions; 1684 collectPossibleReductions(L, Reductions); 1685 bool Changed = false; 1686 1687 // For each possible IV, collect the associated possible set of 'root' nodes 1688 // (i+1, i+2, etc.). 1689 for (Instruction *PossibleIV : PossibleIVs) 1690 if (reroll(PossibleIV, L, Header, BackedgeTakenCount, Reductions)) { 1691 Changed = true; 1692 break; 1693 } 1694 LLVM_DEBUG(dbgs() << "\n After Reroll:\n" << *(L->getHeader()) << "\n"); 1695 1696 // Trip count of L has changed so SE must be re-evaluated. 1697 if (Changed) 1698 SE->forgetLoop(L); 1699 1700 return Changed; 1701 } 1702 1703 bool LoopRerollLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { 1704 if (skipLoop(L)) 1705 return false; 1706 1707 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 1708 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1709 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1710 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( 1711 *L->getHeader()->getParent()); 1712 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1713 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 1714 1715 return LoopReroll(AA, LI, SE, TLI, DT, PreserveLCSSA).runOnLoop(L); 1716 } 1717 1718 PreservedAnalyses LoopRerollPass::run(Loop &L, LoopAnalysisManager &AM, 1719 LoopStandardAnalysisResults &AR, 1720 LPMUpdater &U) { 1721 return LoopReroll(&AR.AA, &AR.LI, &AR.SE, &AR.TLI, &AR.DT, true).runOnLoop(&L) 1722 ? getLoopPassPreservedAnalyses() 1723 : PreservedAnalyses::all(); 1724 } 1725