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>; 194 195 // Map between induction variable and its increment 196 DenseMap<Instruction *, int64_t> IVToIncMap; 197 198 // For loop with multiple induction variable, remember the one used only to 199 // control the loop. 200 Instruction *LoopControlIV; 201 202 // A chain of isomorphic instructions, identified by a single-use PHI 203 // representing a reduction. Only the last value may be used outside the 204 // loop. 205 struct SimpleLoopReduction { 206 SimpleLoopReduction(Instruction *P, Loop *L) : Instructions(1, P) { 207 assert(isa<PHINode>(P) && "First reduction instruction must be a PHI"); 208 add(L); 209 } 210 211 bool valid() const { 212 return Valid; 213 } 214 215 Instruction *getPHI() const { 216 assert(Valid && "Using invalid reduction"); 217 return Instructions.front(); 218 } 219 220 Instruction *getReducedValue() const { 221 assert(Valid && "Using invalid reduction"); 222 return Instructions.back(); 223 } 224 225 Instruction *get(size_t i) const { 226 assert(Valid && "Using invalid reduction"); 227 return Instructions[i+1]; 228 } 229 230 Instruction *operator [] (size_t i) const { return get(i); } 231 232 // The size, ignoring the initial PHI. 233 size_t size() const { 234 assert(Valid && "Using invalid reduction"); 235 return Instructions.size()-1; 236 } 237 238 using iterator = SmallInstructionVector::iterator; 239 using const_iterator = SmallInstructionVector::const_iterator; 240 241 iterator begin() { 242 assert(Valid && "Using invalid reduction"); 243 return std::next(Instructions.begin()); 244 } 245 246 const_iterator begin() const { 247 assert(Valid && "Using invalid reduction"); 248 return std::next(Instructions.begin()); 249 } 250 251 iterator end() { return Instructions.end(); } 252 const_iterator end() const { return Instructions.end(); } 253 254 protected: 255 bool Valid = false; 256 SmallInstructionVector Instructions; 257 258 void add(Loop *L); 259 }; 260 261 // The set of all reductions, and state tracking of possible reductions 262 // during loop instruction processing. 263 struct ReductionTracker { 264 using SmallReductionVector = SmallVector<SimpleLoopReduction, 16>; 265 266 // Add a new possible reduction. 267 void addSLR(SimpleLoopReduction &SLR) { PossibleReds.push_back(SLR); } 268 269 // Setup to track possible reductions corresponding to the provided 270 // rerolling scale. Only reductions with a number of non-PHI instructions 271 // that is divisible by the scale are considered. Three instructions sets 272 // are filled in: 273 // - A set of all possible instructions in eligible reductions. 274 // - A set of all PHIs in eligible reductions 275 // - A set of all reduced values (last instructions) in eligible 276 // reductions. 277 void restrictToScale(uint64_t Scale, 278 SmallInstructionSet &PossibleRedSet, 279 SmallInstructionSet &PossibleRedPHISet, 280 SmallInstructionSet &PossibleRedLastSet) { 281 PossibleRedIdx.clear(); 282 PossibleRedIter.clear(); 283 Reds.clear(); 284 285 for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i) 286 if (PossibleReds[i].size() % Scale == 0) { 287 PossibleRedLastSet.insert(PossibleReds[i].getReducedValue()); 288 PossibleRedPHISet.insert(PossibleReds[i].getPHI()); 289 290 PossibleRedSet.insert(PossibleReds[i].getPHI()); 291 PossibleRedIdx[PossibleReds[i].getPHI()] = i; 292 for (Instruction *J : PossibleReds[i]) { 293 PossibleRedSet.insert(J); 294 PossibleRedIdx[J] = i; 295 } 296 } 297 } 298 299 // The functions below are used while processing the loop instructions. 300 301 // Are the two instructions both from reductions, and furthermore, from 302 // the same reduction? 303 bool isPairInSame(Instruction *J1, Instruction *J2) { 304 DenseMap<Instruction *, int>::iterator J1I = PossibleRedIdx.find(J1); 305 if (J1I != PossibleRedIdx.end()) { 306 DenseMap<Instruction *, int>::iterator J2I = PossibleRedIdx.find(J2); 307 if (J2I != PossibleRedIdx.end() && J1I->second == J2I->second) 308 return true; 309 } 310 311 return false; 312 } 313 314 // The two provided instructions, the first from the base iteration, and 315 // the second from iteration i, form a matched pair. If these are part of 316 // a reduction, record that fact. 317 void recordPair(Instruction *J1, Instruction *J2, unsigned i) { 318 if (PossibleRedIdx.count(J1)) { 319 assert(PossibleRedIdx.count(J2) && 320 "Recording reduction vs. non-reduction instruction?"); 321 322 PossibleRedIter[J1] = 0; 323 PossibleRedIter[J2] = i; 324 325 int Idx = PossibleRedIdx[J1]; 326 assert(Idx == PossibleRedIdx[J2] && 327 "Recording pair from different reductions?"); 328 Reds.insert(Idx); 329 } 330 } 331 332 // The functions below can be called after we've finished processing all 333 // instructions in the loop, and we know which reductions were selected. 334 335 bool validateSelected(); 336 void replaceSelected(); 337 338 protected: 339 // The vector of all possible reductions (for any scale). 340 SmallReductionVector PossibleReds; 341 342 DenseMap<Instruction *, int> PossibleRedIdx; 343 DenseMap<Instruction *, int> PossibleRedIter; 344 DenseSet<int> Reds; 345 }; 346 347 // A DAGRootSet models an induction variable being used in a rerollable 348 // loop. For example, 349 // 350 // x[i*3+0] = y1 351 // x[i*3+1] = y2 352 // x[i*3+2] = y3 353 // 354 // Base instruction -> i*3 355 // +---+----+ 356 // / | \ 357 // ST[y1] +1 +2 <-- Roots 358 // | | 359 // ST[y2] ST[y3] 360 // 361 // There may be multiple DAGRoots, for example: 362 // 363 // x[i*2+0] = ... (1) 364 // x[i*2+1] = ... (1) 365 // x[i*2+4] = ... (2) 366 // x[i*2+5] = ... (2) 367 // x[(i+1234)*2+5678] = ... (3) 368 // x[(i+1234)*2+5679] = ... (3) 369 // 370 // The loop will be rerolled by adding a new loop induction variable, 371 // one for the Base instruction in each DAGRootSet. 372 // 373 struct DAGRootSet { 374 Instruction *BaseInst; 375 SmallInstructionVector Roots; 376 377 // The instructions between IV and BaseInst (but not including BaseInst). 378 SmallInstructionSet SubsumedInsts; 379 }; 380 381 // The set of all DAG roots, and state tracking of all roots 382 // for a particular induction variable. 383 struct DAGRootTracker { 384 DAGRootTracker(LoopReroll *Parent, Loop *L, Instruction *IV, 385 ScalarEvolution *SE, AliasAnalysis *AA, 386 TargetLibraryInfo *TLI, DominatorTree *DT, LoopInfo *LI, 387 bool PreserveLCSSA, 388 DenseMap<Instruction *, int64_t> &IncrMap, 389 Instruction *LoopCtrlIV) 390 : Parent(Parent), L(L), SE(SE), AA(AA), TLI(TLI), DT(DT), LI(LI), 391 PreserveLCSSA(PreserveLCSSA), IV(IV), IVToIncMap(IncrMap), 392 LoopControlIV(LoopCtrlIV) {} 393 394 /// Stage 1: Find all the DAG roots for the induction variable. 395 bool findRoots(); 396 397 /// Stage 2: Validate if the found roots are valid. 398 bool validate(ReductionTracker &Reductions); 399 400 /// Stage 3: Assuming validate() returned true, perform the 401 /// replacement. 402 /// @param BackedgeTakenCount The backedge-taken count of L. 403 void replace(const SCEV *BackedgeTakenCount); 404 405 protected: 406 using UsesTy = MapVector<Instruction *, BitVector>; 407 408 void findRootsRecursive(Instruction *IVU, 409 SmallInstructionSet SubsumedInsts); 410 bool findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts); 411 bool collectPossibleRoots(Instruction *Base, 412 std::map<int64_t,Instruction*> &Roots); 413 bool validateRootSet(DAGRootSet &DRS); 414 415 bool collectUsedInstructions(SmallInstructionSet &PossibleRedSet); 416 void collectInLoopUserSet(const SmallInstructionVector &Roots, 417 const SmallInstructionSet &Exclude, 418 const SmallInstructionSet &Final, 419 DenseSet<Instruction *> &Users); 420 void collectInLoopUserSet(Instruction *Root, 421 const SmallInstructionSet &Exclude, 422 const SmallInstructionSet &Final, 423 DenseSet<Instruction *> &Users); 424 425 UsesTy::iterator nextInstr(int Val, UsesTy &In, 426 const SmallInstructionSet &Exclude, 427 UsesTy::iterator *StartI=nullptr); 428 bool isBaseInst(Instruction *I); 429 bool isRootInst(Instruction *I); 430 bool instrDependsOn(Instruction *I, 431 UsesTy::iterator Start, 432 UsesTy::iterator End); 433 void replaceIV(DAGRootSet &DRS, const SCEV *Start, const SCEV *IncrExpr); 434 435 LoopReroll *Parent; 436 437 // Members of Parent, replicated here for brevity. 438 Loop *L; 439 ScalarEvolution *SE; 440 AliasAnalysis *AA; 441 TargetLibraryInfo *TLI; 442 DominatorTree *DT; 443 LoopInfo *LI; 444 bool PreserveLCSSA; 445 446 // The loop induction variable. 447 Instruction *IV; 448 449 // Loop step amount. 450 int64_t Inc; 451 452 // Loop reroll count; if Inc == 1, this records the scaling applied 453 // to the indvar: a[i*2+0] = ...; a[i*2+1] = ... ; 454 // If Inc is not 1, Scale = Inc. 455 uint64_t Scale; 456 457 // The roots themselves. 458 SmallVector<DAGRootSet,16> RootSets; 459 460 // All increment instructions for IV. 461 SmallInstructionVector LoopIncs; 462 463 // Map of all instructions in the loop (in order) to the iterations 464 // they are used in (or specially, IL_All for instructions 465 // used in the loop increment mechanism). 466 UsesTy Uses; 467 468 // Map between induction variable and its increment 469 DenseMap<Instruction *, int64_t> &IVToIncMap; 470 471 Instruction *LoopControlIV; 472 }; 473 474 // Check if it is a compare-like instruction whose user is a branch 475 bool isCompareUsedByBranch(Instruction *I) { 476 auto *TI = I->getParent()->getTerminator(); 477 if (!isa<BranchInst>(TI) || !isa<CmpInst>(I)) 478 return false; 479 return I->hasOneUse() && TI->getOperand(0) == I; 480 }; 481 482 bool isLoopControlIV(Loop *L, Instruction *IV); 483 void collectPossibleIVs(Loop *L, SmallInstructionVector &PossibleIVs); 484 void collectPossibleReductions(Loop *L, 485 ReductionTracker &Reductions); 486 bool reroll(Instruction *IV, Loop *L, BasicBlock *Header, 487 const SCEV *BackedgeTakenCount, ReductionTracker &Reductions); 488 }; 489 490 } // end anonymous namespace 491 492 char LoopRerollLegacyPass::ID = 0; 493 494 INITIALIZE_PASS_BEGIN(LoopRerollLegacyPass, "loop-reroll", "Reroll loops", 495 false, false) 496 INITIALIZE_PASS_DEPENDENCY(LoopPass) 497 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 498 INITIALIZE_PASS_END(LoopRerollLegacyPass, "loop-reroll", "Reroll loops", false, 499 false) 500 501 Pass *llvm::createLoopRerollPass() { return new LoopRerollLegacyPass; } 502 503 // Returns true if the provided instruction is used outside the given loop. 504 // This operates like Instruction::isUsedOutsideOfBlock, but considers PHIs in 505 // non-loop blocks to be outside the loop. 506 static bool hasUsesOutsideLoop(Instruction *I, Loop *L) { 507 for (User *U : I->users()) { 508 if (!L->contains(cast<Instruction>(U))) 509 return true; 510 } 511 return false; 512 } 513 514 // Check if an IV is only used to control the loop. There are two cases: 515 // 1. It only has one use which is loop increment, and the increment is only 516 // used by comparison and the PHI (could has sext with nsw in between), and the 517 // comparison is only used by branch. 518 // 2. It is used by loop increment and the comparison, the loop increment is 519 // only used by the PHI, and the comparison is used only by the branch. 520 bool LoopReroll::isLoopControlIV(Loop *L, Instruction *IV) { 521 unsigned IVUses = IV->getNumUses(); 522 if (IVUses != 2 && IVUses != 1) 523 return false; 524 525 for (auto *User : IV->users()) { 526 int32_t IncOrCmpUses = User->getNumUses(); 527 bool IsCompInst = isCompareUsedByBranch(cast<Instruction>(User)); 528 529 // User can only have one or two uses. 530 if (IncOrCmpUses != 2 && IncOrCmpUses != 1) 531 return false; 532 533 // Case 1 534 if (IVUses == 1) { 535 // The only user must be the loop increment. 536 // The loop increment must have two uses. 537 if (IsCompInst || IncOrCmpUses != 2) 538 return false; 539 } 540 541 // Case 2 542 if (IVUses == 2 && IncOrCmpUses != 1) 543 return false; 544 545 // The users of the IV must be a binary operation or a comparison 546 if (auto *BO = dyn_cast<BinaryOperator>(User)) { 547 if (BO->getOpcode() == Instruction::Add) { 548 // Loop Increment 549 // User of Loop Increment should be either PHI or CMP 550 for (auto *UU : User->users()) { 551 if (PHINode *PN = dyn_cast<PHINode>(UU)) { 552 if (PN != IV) 553 return false; 554 } 555 // Must be a CMP or an ext (of a value with nsw) then CMP 556 else { 557 auto *UUser = cast<Instruction>(UU); 558 // Skip SExt if we are extending an nsw value 559 // TODO: Allow ZExt too 560 if (BO->hasNoSignedWrap() && UUser->hasOneUse() && 561 isa<SExtInst>(UUser)) 562 UUser = cast<Instruction>(*(UUser->user_begin())); 563 if (!isCompareUsedByBranch(UUser)) 564 return false; 565 } 566 } 567 } else 568 return false; 569 // Compare : can only have one use, and must be branch 570 } else if (!IsCompInst) 571 return false; 572 } 573 return true; 574 } 575 576 // Collect the list of loop induction variables with respect to which it might 577 // be possible to reroll the loop. 578 void LoopReroll::collectPossibleIVs(Loop *L, 579 SmallInstructionVector &PossibleIVs) { 580 BasicBlock *Header = L->getHeader(); 581 for (BasicBlock::iterator I = Header->begin(), 582 IE = Header->getFirstInsertionPt(); I != IE; ++I) { 583 if (!isa<PHINode>(I)) 584 continue; 585 if (!I->getType()->isIntegerTy() && !I->getType()->isPointerTy()) 586 continue; 587 588 if (const SCEVAddRecExpr *PHISCEV = 589 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(&*I))) { 590 if (PHISCEV->getLoop() != L) 591 continue; 592 if (!PHISCEV->isAffine()) 593 continue; 594 auto IncSCEV = dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE)); 595 if (IncSCEV) { 596 IVToIncMap[&*I] = IncSCEV->getValue()->getSExtValue(); 597 LLVM_DEBUG(dbgs() << "LRR: Possible IV: " << *I << " = " << *PHISCEV 598 << "\n"); 599 600 if (isLoopControlIV(L, &*I)) { 601 assert(!LoopControlIV && "Found two loop control only IV"); 602 LoopControlIV = &(*I); 603 LLVM_DEBUG(dbgs() << "LRR: Possible loop control only IV: " << *I 604 << " = " << *PHISCEV << "\n"); 605 } else 606 PossibleIVs.push_back(&*I); 607 } 608 } 609 } 610 } 611 612 // Add the remainder of the reduction-variable chain to the instruction vector 613 // (the initial PHINode has already been added). If successful, the object is 614 // marked as valid. 615 void LoopReroll::SimpleLoopReduction::add(Loop *L) { 616 assert(!Valid && "Cannot add to an already-valid chain"); 617 618 // The reduction variable must be a chain of single-use instructions 619 // (including the PHI), except for the last value (which is used by the PHI 620 // and also outside the loop). 621 Instruction *C = Instructions.front(); 622 if (C->user_empty()) 623 return; 624 625 do { 626 C = cast<Instruction>(*C->user_begin()); 627 if (C->hasOneUse()) { 628 if (!C->isBinaryOp()) 629 return; 630 631 if (!(isa<PHINode>(Instructions.back()) || 632 C->isSameOperationAs(Instructions.back()))) 633 return; 634 635 Instructions.push_back(C); 636 } 637 } while (C->hasOneUse()); 638 639 if (Instructions.size() < 2 || 640 !C->isSameOperationAs(Instructions.back()) || 641 C->use_empty()) 642 return; 643 644 // C is now the (potential) last instruction in the reduction chain. 645 for (User *U : C->users()) { 646 // The only in-loop user can be the initial PHI. 647 if (L->contains(cast<Instruction>(U))) 648 if (cast<Instruction>(U) != Instructions.front()) 649 return; 650 } 651 652 Instructions.push_back(C); 653 Valid = true; 654 } 655 656 // Collect the vector of possible reduction variables. 657 void LoopReroll::collectPossibleReductions(Loop *L, 658 ReductionTracker &Reductions) { 659 BasicBlock *Header = L->getHeader(); 660 for (BasicBlock::iterator I = Header->begin(), 661 IE = Header->getFirstInsertionPt(); I != IE; ++I) { 662 if (!isa<PHINode>(I)) 663 continue; 664 if (!I->getType()->isSingleValueType()) 665 continue; 666 667 SimpleLoopReduction SLR(&*I, L); 668 if (!SLR.valid()) 669 continue; 670 671 LLVM_DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with " 672 << SLR.size() << " chained instructions)\n"); 673 Reductions.addSLR(SLR); 674 } 675 } 676 677 // Collect the set of all users of the provided root instruction. This set of 678 // users contains not only the direct users of the root instruction, but also 679 // all users of those users, and so on. There are two exceptions: 680 // 681 // 1. Instructions in the set of excluded instructions are never added to the 682 // use set (even if they are users). This is used, for example, to exclude 683 // including root increments in the use set of the primary IV. 684 // 685 // 2. Instructions in the set of final instructions are added to the use set 686 // if they are users, but their users are not added. This is used, for 687 // example, to prevent a reduction update from forcing all later reduction 688 // updates into the use set. 689 void LoopReroll::DAGRootTracker::collectInLoopUserSet( 690 Instruction *Root, const SmallInstructionSet &Exclude, 691 const SmallInstructionSet &Final, 692 DenseSet<Instruction *> &Users) { 693 SmallInstructionVector Queue(1, Root); 694 while (!Queue.empty()) { 695 Instruction *I = Queue.pop_back_val(); 696 if (!Users.insert(I).second) 697 continue; 698 699 if (!Final.count(I)) 700 for (Use &U : I->uses()) { 701 Instruction *User = cast<Instruction>(U.getUser()); 702 if (PHINode *PN = dyn_cast<PHINode>(User)) { 703 // Ignore "wrap-around" uses to PHIs of this loop's header. 704 if (PN->getIncomingBlock(U) == L->getHeader()) 705 continue; 706 } 707 708 if (L->contains(User) && !Exclude.count(User)) { 709 Queue.push_back(User); 710 } 711 } 712 713 // We also want to collect single-user "feeder" values. 714 for (Use &U : I->operands()) { 715 if (Instruction *Op = dyn_cast<Instruction>(U)) 716 if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) && 717 !Final.count(Op)) 718 Queue.push_back(Op); 719 } 720 } 721 } 722 723 // Collect all of the users of all of the provided root instructions (combined 724 // into a single set). 725 void LoopReroll::DAGRootTracker::collectInLoopUserSet( 726 const SmallInstructionVector &Roots, 727 const SmallInstructionSet &Exclude, 728 const SmallInstructionSet &Final, 729 DenseSet<Instruction *> &Users) { 730 for (Instruction *Root : Roots) 731 collectInLoopUserSet(Root, Exclude, Final, Users); 732 } 733 734 static bool isUnorderedLoadStore(Instruction *I) { 735 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 736 return LI->isUnordered(); 737 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 738 return SI->isUnordered(); 739 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) 740 return !MI->isVolatile(); 741 return false; 742 } 743 744 /// Return true if IVU is a "simple" arithmetic operation. 745 /// This is used for narrowing the search space for DAGRoots; only arithmetic 746 /// and GEPs can be part of a DAGRoot. 747 static bool isSimpleArithmeticOp(User *IVU) { 748 if (Instruction *I = dyn_cast<Instruction>(IVU)) { 749 switch (I->getOpcode()) { 750 default: return false; 751 case Instruction::Add: 752 case Instruction::Sub: 753 case Instruction::Mul: 754 case Instruction::Shl: 755 case Instruction::AShr: 756 case Instruction::LShr: 757 case Instruction::GetElementPtr: 758 case Instruction::Trunc: 759 case Instruction::ZExt: 760 case Instruction::SExt: 761 return true; 762 } 763 } 764 return false; 765 } 766 767 static bool isLoopIncrement(User *U, Instruction *IV) { 768 BinaryOperator *BO = dyn_cast<BinaryOperator>(U); 769 770 if ((BO && BO->getOpcode() != Instruction::Add) || 771 (!BO && !isa<GetElementPtrInst>(U))) 772 return false; 773 774 for (auto *UU : U->users()) { 775 PHINode *PN = dyn_cast<PHINode>(UU); 776 if (PN && PN == IV) 777 return true; 778 } 779 return false; 780 } 781 782 bool LoopReroll::DAGRootTracker:: 783 collectPossibleRoots(Instruction *Base, std::map<int64_t,Instruction*> &Roots) { 784 SmallInstructionVector BaseUsers; 785 786 for (auto *I : Base->users()) { 787 ConstantInt *CI = nullptr; 788 789 if (isLoopIncrement(I, IV)) { 790 LoopIncs.push_back(cast<Instruction>(I)); 791 continue; 792 } 793 794 // The root nodes must be either GEPs, ORs or ADDs. 795 if (auto *BO = dyn_cast<BinaryOperator>(I)) { 796 if (BO->getOpcode() == Instruction::Add || 797 BO->getOpcode() == Instruction::Or) 798 CI = dyn_cast<ConstantInt>(BO->getOperand(1)); 799 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) { 800 Value *LastOperand = GEP->getOperand(GEP->getNumOperands()-1); 801 CI = dyn_cast<ConstantInt>(LastOperand); 802 } 803 804 if (!CI) { 805 if (Instruction *II = dyn_cast<Instruction>(I)) { 806 BaseUsers.push_back(II); 807 continue; 808 } else { 809 LLVM_DEBUG(dbgs() << "LRR: Aborting due to non-instruction: " << *I 810 << "\n"); 811 return false; 812 } 813 } 814 815 int64_t V = std::abs(CI->getValue().getSExtValue()); 816 if (Roots.find(V) != Roots.end()) 817 // No duplicates, please. 818 return false; 819 820 Roots[V] = cast<Instruction>(I); 821 } 822 823 // Make sure we have at least two roots. 824 if (Roots.empty() || (Roots.size() == 1 && BaseUsers.empty())) 825 return false; 826 827 // If we found non-loop-inc, non-root users of Base, assume they are 828 // for the zeroth root index. This is because "add %a, 0" gets optimized 829 // away. 830 if (BaseUsers.size()) { 831 if (Roots.find(0) != Roots.end()) { 832 LLVM_DEBUG(dbgs() << "LRR: Multiple roots found for base - aborting!\n"); 833 return false; 834 } 835 Roots[0] = Base; 836 } 837 838 // Calculate the number of users of the base, or lowest indexed, iteration. 839 unsigned NumBaseUses = BaseUsers.size(); 840 if (NumBaseUses == 0) 841 NumBaseUses = Roots.begin()->second->getNumUses(); 842 843 // Check that every node has the same number of users. 844 for (auto &KV : Roots) { 845 if (KV.first == 0) 846 continue; 847 if (!KV.second->hasNUses(NumBaseUses)) { 848 LLVM_DEBUG(dbgs() << "LRR: Aborting - Root and Base #users not the same: " 849 << "#Base=" << NumBaseUses 850 << ", #Root=" << KV.second->getNumUses() << "\n"); 851 return false; 852 } 853 } 854 855 return true; 856 } 857 858 void LoopReroll::DAGRootTracker:: 859 findRootsRecursive(Instruction *I, SmallInstructionSet SubsumedInsts) { 860 // Does the user look like it could be part of a root set? 861 // All its users must be simple arithmetic ops. 862 if (I->hasNUsesOrMore(IL_MaxRerollIterations + 1)) 863 return; 864 865 if (I != IV && findRootsBase(I, SubsumedInsts)) 866 return; 867 868 SubsumedInsts.insert(I); 869 870 for (User *V : I->users()) { 871 Instruction *I = cast<Instruction>(V); 872 if (is_contained(LoopIncs, I)) 873 continue; 874 875 if (!isSimpleArithmeticOp(I)) 876 continue; 877 878 // The recursive call makes a copy of SubsumedInsts. 879 findRootsRecursive(I, SubsumedInsts); 880 } 881 } 882 883 bool LoopReroll::DAGRootTracker::validateRootSet(DAGRootSet &DRS) { 884 if (DRS.Roots.empty()) 885 return false; 886 887 // If the value of the base instruction is used outside the loop, we cannot 888 // reroll the loop. Check for other root instructions is unnecessary because 889 // they don't match any base instructions if their values are used outside. 890 if (hasUsesOutsideLoop(DRS.BaseInst, L)) 891 return false; 892 893 // Consider a DAGRootSet with N-1 roots (so N different values including 894 // BaseInst). 895 // Define d = Roots[0] - BaseInst, which should be the same as 896 // Roots[I] - Roots[I-1] for all I in [1..N). 897 // Define D = BaseInst@J - BaseInst@J-1, where "@J" means the value at the 898 // loop iteration J. 899 // 900 // Now, For the loop iterations to be consecutive: 901 // D = d * N 902 const auto *ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst)); 903 if (!ADR) 904 return false; 905 906 // Check that the first root is evenly spaced. 907 unsigned N = DRS.Roots.size() + 1; 908 const SCEV *StepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), ADR); 909 if (isa<SCEVCouldNotCompute>(StepSCEV) || StepSCEV->getType()->isPointerTy()) 910 return false; 911 const SCEV *ScaleSCEV = SE->getConstant(StepSCEV->getType(), N); 912 if (ADR->getStepRecurrence(*SE) != SE->getMulExpr(StepSCEV, ScaleSCEV)) 913 return false; 914 915 // Check that the remainling roots are evenly spaced. 916 for (unsigned i = 1; i < N - 1; ++i) { 917 const SCEV *NewStepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[i]), 918 SE->getSCEV(DRS.Roots[i-1])); 919 if (NewStepSCEV != StepSCEV) 920 return false; 921 } 922 923 return true; 924 } 925 926 bool LoopReroll::DAGRootTracker:: 927 findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts) { 928 // The base of a RootSet must be an AddRec, so it can be erased. 929 const auto *IVU_ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IVU)); 930 if (!IVU_ADR || IVU_ADR->getLoop() != L) 931 return false; 932 933 std::map<int64_t, Instruction*> V; 934 if (!collectPossibleRoots(IVU, V)) 935 return false; 936 937 // If we didn't get a root for index zero, then IVU must be 938 // subsumed. 939 if (V.find(0) == V.end()) 940 SubsumedInsts.insert(IVU); 941 942 // Partition the vector into monotonically increasing indexes. 943 DAGRootSet DRS; 944 DRS.BaseInst = nullptr; 945 946 SmallVector<DAGRootSet, 16> PotentialRootSets; 947 948 for (auto &KV : V) { 949 if (!DRS.BaseInst) { 950 DRS.BaseInst = KV.second; 951 DRS.SubsumedInsts = SubsumedInsts; 952 } else if (DRS.Roots.empty()) { 953 DRS.Roots.push_back(KV.second); 954 } else if (V.find(KV.first - 1) != V.end()) { 955 DRS.Roots.push_back(KV.second); 956 } else { 957 // Linear sequence terminated. 958 if (!validateRootSet(DRS)) 959 return false; 960 961 // Construct a new DAGRootSet with the next sequence. 962 PotentialRootSets.push_back(DRS); 963 DRS.BaseInst = KV.second; 964 DRS.Roots.clear(); 965 } 966 } 967 968 if (!validateRootSet(DRS)) 969 return false; 970 971 PotentialRootSets.push_back(DRS); 972 973 RootSets.append(PotentialRootSets.begin(), PotentialRootSets.end()); 974 975 return true; 976 } 977 978 bool LoopReroll::DAGRootTracker::findRoots() { 979 Inc = IVToIncMap[IV]; 980 981 assert(RootSets.empty() && "Unclean state!"); 982 if (std::abs(Inc) == 1) { 983 for (auto *IVU : IV->users()) { 984 if (isLoopIncrement(IVU, IV)) 985 LoopIncs.push_back(cast<Instruction>(IVU)); 986 } 987 findRootsRecursive(IV, SmallInstructionSet()); 988 LoopIncs.push_back(IV); 989 } else { 990 if (!findRootsBase(IV, SmallInstructionSet())) 991 return false; 992 } 993 994 // Ensure all sets have the same size. 995 if (RootSets.empty()) { 996 LLVM_DEBUG(dbgs() << "LRR: Aborting because no root sets found!\n"); 997 return false; 998 } 999 for (auto &V : RootSets) { 1000 if (V.Roots.empty() || V.Roots.size() != RootSets[0].Roots.size()) { 1001 LLVM_DEBUG( 1002 dbgs() 1003 << "LRR: Aborting because not all root sets have the same size\n"); 1004 return false; 1005 } 1006 } 1007 1008 Scale = RootSets[0].Roots.size() + 1; 1009 1010 if (Scale > IL_MaxRerollIterations) { 1011 LLVM_DEBUG(dbgs() << "LRR: Aborting - too many iterations found. " 1012 << "#Found=" << Scale 1013 << ", #Max=" << IL_MaxRerollIterations << "\n"); 1014 return false; 1015 } 1016 1017 LLVM_DEBUG(dbgs() << "LRR: Successfully found roots: Scale=" << Scale 1018 << "\n"); 1019 1020 return true; 1021 } 1022 1023 bool LoopReroll::DAGRootTracker::collectUsedInstructions(SmallInstructionSet &PossibleRedSet) { 1024 // Populate the MapVector with all instructions in the block, in order first, 1025 // so we can iterate over the contents later in perfect order. 1026 for (auto &I : *L->getHeader()) { 1027 Uses[&I].resize(IL_End); 1028 } 1029 1030 SmallInstructionSet Exclude; 1031 for (auto &DRS : RootSets) { 1032 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end()); 1033 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end()); 1034 Exclude.insert(DRS.BaseInst); 1035 } 1036 Exclude.insert(LoopIncs.begin(), LoopIncs.end()); 1037 1038 for (auto &DRS : RootSets) { 1039 DenseSet<Instruction*> VBase; 1040 collectInLoopUserSet(DRS.BaseInst, Exclude, PossibleRedSet, VBase); 1041 for (auto *I : VBase) { 1042 Uses[I].set(0); 1043 } 1044 1045 unsigned Idx = 1; 1046 for (auto *Root : DRS.Roots) { 1047 DenseSet<Instruction*> V; 1048 collectInLoopUserSet(Root, Exclude, PossibleRedSet, V); 1049 1050 // While we're here, check the use sets are the same size. 1051 if (V.size() != VBase.size()) { 1052 LLVM_DEBUG(dbgs() << "LRR: Aborting - use sets are different sizes\n"); 1053 return false; 1054 } 1055 1056 for (auto *I : V) { 1057 Uses[I].set(Idx); 1058 } 1059 ++Idx; 1060 } 1061 1062 // Make sure our subsumed instructions are remembered too. 1063 for (auto *I : DRS.SubsumedInsts) { 1064 Uses[I].set(IL_All); 1065 } 1066 } 1067 1068 // Make sure the loop increments are also accounted for. 1069 1070 Exclude.clear(); 1071 for (auto &DRS : RootSets) { 1072 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end()); 1073 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end()); 1074 Exclude.insert(DRS.BaseInst); 1075 } 1076 1077 DenseSet<Instruction*> V; 1078 collectInLoopUserSet(LoopIncs, Exclude, PossibleRedSet, V); 1079 for (auto *I : V) { 1080 if (I->mayHaveSideEffects()) { 1081 LLVM_DEBUG(dbgs() << "LRR: Aborting - " 1082 << "An instruction which does not belong to any root " 1083 << "sets must not have side effects: " << *I); 1084 return false; 1085 } 1086 Uses[I].set(IL_All); 1087 } 1088 1089 return true; 1090 } 1091 1092 /// Get the next instruction in "In" that is a member of set Val. 1093 /// Start searching from StartI, and do not return anything in Exclude. 1094 /// If StartI is not given, start from In.begin(). 1095 LoopReroll::DAGRootTracker::UsesTy::iterator 1096 LoopReroll::DAGRootTracker::nextInstr(int Val, UsesTy &In, 1097 const SmallInstructionSet &Exclude, 1098 UsesTy::iterator *StartI) { 1099 UsesTy::iterator I = StartI ? *StartI : In.begin(); 1100 while (I != In.end() && (I->second.test(Val) == 0 || 1101 Exclude.contains(I->first))) 1102 ++I; 1103 return I; 1104 } 1105 1106 bool LoopReroll::DAGRootTracker::isBaseInst(Instruction *I) { 1107 for (auto &DRS : RootSets) { 1108 if (DRS.BaseInst == I) 1109 return true; 1110 } 1111 return false; 1112 } 1113 1114 bool LoopReroll::DAGRootTracker::isRootInst(Instruction *I) { 1115 for (auto &DRS : RootSets) { 1116 if (is_contained(DRS.Roots, I)) 1117 return true; 1118 } 1119 return false; 1120 } 1121 1122 /// Return true if instruction I depends on any instruction between 1123 /// Start and End. 1124 bool LoopReroll::DAGRootTracker::instrDependsOn(Instruction *I, 1125 UsesTy::iterator Start, 1126 UsesTy::iterator End) { 1127 for (auto *U : I->users()) { 1128 for (auto It = Start; It != End; ++It) 1129 if (U == It->first) 1130 return true; 1131 } 1132 return false; 1133 } 1134 1135 static bool isIgnorableInst(const Instruction *I) { 1136 if (isa<DbgInfoIntrinsic>(I)) 1137 return true; 1138 const IntrinsicInst* II = dyn_cast<IntrinsicInst>(I); 1139 if (!II) 1140 return false; 1141 switch (II->getIntrinsicID()) { 1142 default: 1143 return false; 1144 case Intrinsic::annotation: 1145 case Intrinsic::ptr_annotation: 1146 case Intrinsic::var_annotation: 1147 // TODO: the following intrinsics may also be allowed: 1148 // lifetime_start, lifetime_end, invariant_start, invariant_end 1149 return true; 1150 } 1151 return false; 1152 } 1153 1154 bool LoopReroll::DAGRootTracker::validate(ReductionTracker &Reductions) { 1155 // We now need to check for equivalence of the use graph of each root with 1156 // that of the primary induction variable (excluding the roots). Our goal 1157 // here is not to solve the full graph isomorphism problem, but rather to 1158 // catch common cases without a lot of work. As a result, we will assume 1159 // that the relative order of the instructions in each unrolled iteration 1160 // is the same (although we will not make an assumption about how the 1161 // different iterations are intermixed). Note that while the order must be 1162 // the same, the instructions may not be in the same basic block. 1163 1164 // An array of just the possible reductions for this scale factor. When we 1165 // collect the set of all users of some root instructions, these reduction 1166 // instructions are treated as 'final' (their uses are not considered). 1167 // This is important because we don't want the root use set to search down 1168 // the reduction chain. 1169 SmallInstructionSet PossibleRedSet; 1170 SmallInstructionSet PossibleRedLastSet; 1171 SmallInstructionSet PossibleRedPHISet; 1172 Reductions.restrictToScale(Scale, PossibleRedSet, 1173 PossibleRedPHISet, PossibleRedLastSet); 1174 1175 // Populate "Uses" with where each instruction is used. 1176 if (!collectUsedInstructions(PossibleRedSet)) 1177 return false; 1178 1179 // Make sure we mark the reduction PHIs as used in all iterations. 1180 for (auto *I : PossibleRedPHISet) { 1181 Uses[I].set(IL_All); 1182 } 1183 1184 // Make sure we mark loop-control-only PHIs as used in all iterations. See 1185 // comment above LoopReroll::isLoopControlIV for more information. 1186 BasicBlock *Header = L->getHeader(); 1187 if (LoopControlIV && LoopControlIV != IV) { 1188 for (auto *U : LoopControlIV->users()) { 1189 Instruction *IVUser = dyn_cast<Instruction>(U); 1190 // IVUser could be loop increment or compare 1191 Uses[IVUser].set(IL_All); 1192 for (auto *UU : IVUser->users()) { 1193 Instruction *UUser = dyn_cast<Instruction>(UU); 1194 // UUser could be compare, PHI or branch 1195 Uses[UUser].set(IL_All); 1196 // Skip SExt 1197 if (isa<SExtInst>(UUser)) { 1198 UUser = dyn_cast<Instruction>(*(UUser->user_begin())); 1199 Uses[UUser].set(IL_All); 1200 } 1201 // Is UUser a compare instruction? 1202 if (UU->hasOneUse()) { 1203 Instruction *BI = dyn_cast<BranchInst>(*UUser->user_begin()); 1204 if (BI == cast<BranchInst>(Header->getTerminator())) 1205 Uses[BI].set(IL_All); 1206 } 1207 } 1208 } 1209 } 1210 1211 // Make sure all instructions in the loop are in one and only one 1212 // set. 1213 for (auto &KV : Uses) { 1214 if (KV.second.count() != 1 && !isIgnorableInst(KV.first)) { 1215 LLVM_DEBUG( 1216 dbgs() << "LRR: Aborting - instruction is not used in 1 iteration: " 1217 << *KV.first << " (#uses=" << KV.second.count() << ")\n"); 1218 return false; 1219 } 1220 } 1221 1222 LLVM_DEBUG(for (auto &KV 1223 : Uses) { 1224 dbgs() << "LRR: " << KV.second.find_first() << "\t" << *KV.first << "\n"; 1225 }); 1226 1227 for (unsigned Iter = 1; Iter < Scale; ++Iter) { 1228 // In addition to regular aliasing information, we need to look for 1229 // instructions from later (future) iterations that have side effects 1230 // preventing us from reordering them past other instructions with side 1231 // effects. 1232 bool FutureSideEffects = false; 1233 AliasSetTracker AST(*AA); 1234 // The map between instructions in f(%iv.(i+1)) and f(%iv). 1235 DenseMap<Value *, Value *> BaseMap; 1236 1237 // Compare iteration Iter to the base. 1238 SmallInstructionSet Visited; 1239 auto BaseIt = nextInstr(0, Uses, Visited); 1240 auto RootIt = nextInstr(Iter, Uses, Visited); 1241 auto LastRootIt = Uses.begin(); 1242 1243 while (BaseIt != Uses.end() && RootIt != Uses.end()) { 1244 Instruction *BaseInst = BaseIt->first; 1245 Instruction *RootInst = RootIt->first; 1246 1247 // Skip over the IV or root instructions; only match their users. 1248 bool Continue = false; 1249 if (isBaseInst(BaseInst)) { 1250 Visited.insert(BaseInst); 1251 BaseIt = nextInstr(0, Uses, Visited); 1252 Continue = true; 1253 } 1254 if (isRootInst(RootInst)) { 1255 LastRootIt = RootIt; 1256 Visited.insert(RootInst); 1257 RootIt = nextInstr(Iter, Uses, Visited); 1258 Continue = true; 1259 } 1260 if (Continue) continue; 1261 1262 if (!BaseInst->isSameOperationAs(RootInst)) { 1263 // Last chance saloon. We don't try and solve the full isomorphism 1264 // problem, but try and at least catch the case where two instructions 1265 // *of different types* are round the wrong way. We won't be able to 1266 // efficiently tell, given two ADD instructions, which way around we 1267 // should match them, but given an ADD and a SUB, we can at least infer 1268 // which one is which. 1269 // 1270 // This should allow us to deal with a greater subset of the isomorphism 1271 // problem. It does however change a linear algorithm into a quadratic 1272 // one, so limit the number of probes we do. 1273 auto TryIt = RootIt; 1274 unsigned N = NumToleratedFailedMatches; 1275 while (TryIt != Uses.end() && 1276 !BaseInst->isSameOperationAs(TryIt->first) && 1277 N--) { 1278 ++TryIt; 1279 TryIt = nextInstr(Iter, Uses, Visited, &TryIt); 1280 } 1281 1282 if (TryIt == Uses.end() || TryIt == RootIt || 1283 instrDependsOn(TryIt->first, RootIt, TryIt)) { 1284 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " 1285 << *BaseInst << " vs. " << *RootInst << "\n"); 1286 return false; 1287 } 1288 1289 RootIt = TryIt; 1290 RootInst = TryIt->first; 1291 } 1292 1293 // All instructions between the last root and this root 1294 // may belong to some other iteration. If they belong to a 1295 // future iteration, then they're dangerous to alias with. 1296 // 1297 // Note that because we allow a limited amount of flexibility in the order 1298 // that we visit nodes, LastRootIt might be *before* RootIt, in which 1299 // case we've already checked this set of instructions so we shouldn't 1300 // do anything. 1301 for (; LastRootIt < RootIt; ++LastRootIt) { 1302 Instruction *I = LastRootIt->first; 1303 if (LastRootIt->second.find_first() < (int)Iter) 1304 continue; 1305 if (I->mayWriteToMemory()) 1306 AST.add(I); 1307 // Note: This is specifically guarded by a check on isa<PHINode>, 1308 // which while a valid (somewhat arbitrary) micro-optimization, is 1309 // needed because otherwise isSafeToSpeculativelyExecute returns 1310 // false on PHI nodes. 1311 if (!isa<PHINode>(I) && !isUnorderedLoadStore(I) && 1312 !isSafeToSpeculativelyExecute(I)) 1313 // Intervening instructions cause side effects. 1314 FutureSideEffects = true; 1315 } 1316 1317 // Make sure that this instruction, which is in the use set of this 1318 // root instruction, does not also belong to the base set or the set of 1319 // some other root instruction. 1320 if (RootIt->second.count() > 1) { 1321 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1322 << " vs. " << *RootInst << " (prev. case overlap)\n"); 1323 return false; 1324 } 1325 1326 // Make sure that we don't alias with any instruction in the alias set 1327 // tracker. If we do, then we depend on a future iteration, and we 1328 // can't reroll. 1329 if (RootInst->mayReadFromMemory()) 1330 for (auto &K : AST) { 1331 if (K.aliasesUnknownInst(RootInst, *AA)) { 1332 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " 1333 << *BaseInst << " vs. " << *RootInst 1334 << " (depends on future store)\n"); 1335 return false; 1336 } 1337 } 1338 1339 // If we've past an instruction from a future iteration that may have 1340 // side effects, and this instruction might also, then we can't reorder 1341 // them, and this matching fails. As an exception, we allow the alias 1342 // set tracker to handle regular (unordered) load/store dependencies. 1343 if (FutureSideEffects && ((!isUnorderedLoadStore(BaseInst) && 1344 !isSafeToSpeculativelyExecute(BaseInst)) || 1345 (!isUnorderedLoadStore(RootInst) && 1346 !isSafeToSpeculativelyExecute(RootInst)))) { 1347 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1348 << " vs. " << *RootInst 1349 << " (side effects prevent reordering)\n"); 1350 return false; 1351 } 1352 1353 // For instructions that are part of a reduction, if the operation is 1354 // associative, then don't bother matching the operands (because we 1355 // already know that the instructions are isomorphic, and the order 1356 // within the iteration does not matter). For non-associative reductions, 1357 // we do need to match the operands, because we need to reject 1358 // out-of-order instructions within an iteration! 1359 // For example (assume floating-point addition), we need to reject this: 1360 // x += a[i]; x += b[i]; 1361 // x += a[i+1]; x += b[i+1]; 1362 // x += b[i+2]; x += a[i+2]; 1363 bool InReduction = Reductions.isPairInSame(BaseInst, RootInst); 1364 1365 if (!(InReduction && BaseInst->isAssociative())) { 1366 bool Swapped = false, SomeOpMatched = false; 1367 for (unsigned j = 0; j < BaseInst->getNumOperands(); ++j) { 1368 Value *Op2 = RootInst->getOperand(j); 1369 1370 // If this is part of a reduction (and the operation is not 1371 // associatve), then we match all operands, but not those that are 1372 // part of the reduction. 1373 if (InReduction) 1374 if (Instruction *Op2I = dyn_cast<Instruction>(Op2)) 1375 if (Reductions.isPairInSame(RootInst, Op2I)) 1376 continue; 1377 1378 DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2); 1379 if (BMI != BaseMap.end()) { 1380 Op2 = BMI->second; 1381 } else { 1382 for (auto &DRS : RootSets) { 1383 if (DRS.Roots[Iter-1] == (Instruction*) Op2) { 1384 Op2 = DRS.BaseInst; 1385 break; 1386 } 1387 } 1388 } 1389 1390 if (BaseInst->getOperand(Swapped ? unsigned(!j) : j) != Op2) { 1391 // If we've not already decided to swap the matched operands, and 1392 // we've not already matched our first operand (note that we could 1393 // have skipped matching the first operand because it is part of a 1394 // reduction above), and the instruction is commutative, then try 1395 // the swapped match. 1396 if (!Swapped && BaseInst->isCommutative() && !SomeOpMatched && 1397 BaseInst->getOperand(!j) == Op2) { 1398 Swapped = true; 1399 } else { 1400 LLVM_DEBUG(dbgs() 1401 << "LRR: iteration root match failed at " << *BaseInst 1402 << " vs. " << *RootInst << " (operand " << j << ")\n"); 1403 return false; 1404 } 1405 } 1406 1407 SomeOpMatched = true; 1408 } 1409 } 1410 1411 if ((!PossibleRedLastSet.count(BaseInst) && 1412 hasUsesOutsideLoop(BaseInst, L)) || 1413 (!PossibleRedLastSet.count(RootInst) && 1414 hasUsesOutsideLoop(RootInst, L))) { 1415 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst 1416 << " vs. " << *RootInst << " (uses outside loop)\n"); 1417 return false; 1418 } 1419 1420 Reductions.recordPair(BaseInst, RootInst, Iter); 1421 BaseMap.insert(std::make_pair(RootInst, BaseInst)); 1422 1423 LastRootIt = RootIt; 1424 Visited.insert(BaseInst); 1425 Visited.insert(RootInst); 1426 BaseIt = nextInstr(0, Uses, Visited); 1427 RootIt = nextInstr(Iter, Uses, Visited); 1428 } 1429 assert(BaseIt == Uses.end() && RootIt == Uses.end() && 1430 "Mismatched set sizes!"); 1431 } 1432 1433 LLVM_DEBUG(dbgs() << "LRR: Matched all iteration increments for " << *IV 1434 << "\n"); 1435 1436 return true; 1437 } 1438 1439 void LoopReroll::DAGRootTracker::replace(const SCEV *BackedgeTakenCount) { 1440 BasicBlock *Header = L->getHeader(); 1441 1442 // Compute the start and increment for each BaseInst before we start erasing 1443 // instructions. 1444 SmallVector<const SCEV *, 8> StartExprs; 1445 SmallVector<const SCEV *, 8> IncrExprs; 1446 for (auto &DRS : RootSets) { 1447 const SCEVAddRecExpr *IVSCEV = 1448 cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst)); 1449 StartExprs.push_back(IVSCEV->getStart()); 1450 IncrExprs.push_back(SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), IVSCEV)); 1451 } 1452 1453 // Remove instructions associated with non-base iterations. 1454 for (Instruction &Inst : llvm::make_early_inc_range(llvm::reverse(*Header))) { 1455 unsigned I = Uses[&Inst].find_first(); 1456 if (I > 0 && I < IL_All) { 1457 LLVM_DEBUG(dbgs() << "LRR: removing: " << Inst << "\n"); 1458 Inst.eraseFromParent(); 1459 } 1460 } 1461 1462 // Rewrite each BaseInst using SCEV. 1463 for (size_t i = 0, e = RootSets.size(); i != e; ++i) 1464 // Insert the new induction variable. 1465 replaceIV(RootSets[i], StartExprs[i], IncrExprs[i]); 1466 1467 { // Limit the lifetime of SCEVExpander. 1468 BranchInst *BI = cast<BranchInst>(Header->getTerminator()); 1469 const DataLayout &DL = Header->getModule()->getDataLayout(); 1470 SCEVExpander Expander(*SE, DL, "reroll"); 1471 auto Zero = SE->getZero(BackedgeTakenCount->getType()); 1472 auto One = SE->getOne(BackedgeTakenCount->getType()); 1473 auto NewIVSCEV = SE->getAddRecExpr(Zero, One, L, SCEV::FlagAnyWrap); 1474 Value *NewIV = 1475 Expander.expandCodeFor(NewIVSCEV, BackedgeTakenCount->getType(), 1476 Header->getFirstNonPHIOrDbg()); 1477 // FIXME: This arithmetic can overflow. 1478 auto TripCount = SE->getAddExpr(BackedgeTakenCount, One); 1479 auto ScaledTripCount = SE->getMulExpr( 1480 TripCount, SE->getConstant(BackedgeTakenCount->getType(), Scale)); 1481 auto ScaledBECount = SE->getMinusSCEV(ScaledTripCount, One); 1482 Value *TakenCount = 1483 Expander.expandCodeFor(ScaledBECount, BackedgeTakenCount->getType(), 1484 Header->getFirstNonPHIOrDbg()); 1485 Value *Cond = 1486 new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, TakenCount, "exitcond"); 1487 BI->setCondition(Cond); 1488 1489 if (BI->getSuccessor(1) != Header) 1490 BI->swapSuccessors(); 1491 } 1492 1493 SimplifyInstructionsInBlock(Header, TLI); 1494 DeleteDeadPHIs(Header, TLI); 1495 } 1496 1497 void LoopReroll::DAGRootTracker::replaceIV(DAGRootSet &DRS, 1498 const SCEV *Start, 1499 const SCEV *IncrExpr) { 1500 BasicBlock *Header = L->getHeader(); 1501 Instruction *Inst = DRS.BaseInst; 1502 1503 const SCEV *NewIVSCEV = 1504 SE->getAddRecExpr(Start, IncrExpr, L, SCEV::FlagAnyWrap); 1505 1506 { // Limit the lifetime of SCEVExpander. 1507 const DataLayout &DL = Header->getModule()->getDataLayout(); 1508 SCEVExpander Expander(*SE, DL, "reroll"); 1509 Value *NewIV = Expander.expandCodeFor(NewIVSCEV, Inst->getType(), 1510 Header->getFirstNonPHIOrDbg()); 1511 1512 for (auto &KV : Uses) 1513 if (KV.second.find_first() == 0) 1514 KV.first->replaceUsesOfWith(Inst, NewIV); 1515 } 1516 } 1517 1518 // Validate the selected reductions. All iterations must have an isomorphic 1519 // part of the reduction chain and, for non-associative reductions, the chain 1520 // entries must appear in order. 1521 bool LoopReroll::ReductionTracker::validateSelected() { 1522 // For a non-associative reduction, the chain entries must appear in order. 1523 for (int i : Reds) { 1524 int PrevIter = 0, BaseCount = 0, Count = 0; 1525 for (Instruction *J : PossibleReds[i]) { 1526 // Note that all instructions in the chain must have been found because 1527 // all instructions in the function must have been assigned to some 1528 // iteration. 1529 int Iter = PossibleRedIter[J]; 1530 if (Iter != PrevIter && Iter != PrevIter + 1 && 1531 !PossibleReds[i].getReducedValue()->isAssociative()) { 1532 LLVM_DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: " 1533 << J << "\n"); 1534 return false; 1535 } 1536 1537 if (Iter != PrevIter) { 1538 if (Count != BaseCount) { 1539 LLVM_DEBUG(dbgs() 1540 << "LRR: Iteration " << PrevIter << " reduction use count " 1541 << Count << " is not equal to the base use count " 1542 << BaseCount << "\n"); 1543 return false; 1544 } 1545 1546 Count = 0; 1547 } 1548 1549 ++Count; 1550 if (Iter == 0) 1551 ++BaseCount; 1552 1553 PrevIter = Iter; 1554 } 1555 } 1556 1557 return true; 1558 } 1559 1560 // For all selected reductions, remove all parts except those in the first 1561 // iteration (and the PHI). Replace outside uses of the reduced value with uses 1562 // of the first-iteration reduced value (in other words, reroll the selected 1563 // reductions). 1564 void LoopReroll::ReductionTracker::replaceSelected() { 1565 // Fixup reductions to refer to the last instruction associated with the 1566 // first iteration (not the last). 1567 for (int i : Reds) { 1568 int j = 0; 1569 for (int e = PossibleReds[i].size(); j != e; ++j) 1570 if (PossibleRedIter[PossibleReds[i][j]] != 0) { 1571 --j; 1572 break; 1573 } 1574 1575 // Replace users with the new end-of-chain value. 1576 SmallInstructionVector Users; 1577 for (User *U : PossibleReds[i].getReducedValue()->users()) { 1578 Users.push_back(cast<Instruction>(U)); 1579 } 1580 1581 for (Instruction *User : Users) 1582 User->replaceUsesOfWith(PossibleReds[i].getReducedValue(), 1583 PossibleReds[i][j]); 1584 } 1585 } 1586 1587 // Reroll the provided loop with respect to the provided induction variable. 1588 // Generally, we're looking for a loop like this: 1589 // 1590 // %iv = phi [ (preheader, ...), (body, %iv.next) ] 1591 // f(%iv) 1592 // %iv.1 = add %iv, 1 <-- a root increment 1593 // f(%iv.1) 1594 // %iv.2 = add %iv, 2 <-- a root increment 1595 // f(%iv.2) 1596 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment 1597 // f(%iv.scale_m_1) 1598 // ... 1599 // %iv.next = add %iv, scale 1600 // %cmp = icmp(%iv, ...) 1601 // br %cmp, header, exit 1602 // 1603 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of 1604 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can 1605 // be intermixed with eachother. The restriction imposed by this algorithm is 1606 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1), 1607 // etc. be the same. 1608 // 1609 // First, we collect the use set of %iv, excluding the other increment roots. 1610 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1) 1611 // times, having collected the use set of f(%iv.(i+1)), during which we: 1612 // - Ensure that the next unmatched instruction in f(%iv) is isomorphic to 1613 // the next unmatched instruction in f(%iv.(i+1)). 1614 // - Ensure that both matched instructions don't have any external users 1615 // (with the exception of last-in-chain reduction instructions). 1616 // - Track the (aliasing) write set, and other side effects, of all 1617 // instructions that belong to future iterations that come before the matched 1618 // instructions. If the matched instructions read from that write set, then 1619 // f(%iv) or f(%iv.(i+1)) has some dependency on instructions in 1620 // f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly, 1621 // if any of these future instructions had side effects (could not be 1622 // speculatively executed), and so do the matched instructions, when we 1623 // cannot reorder those side-effect-producing instructions, and rerolling 1624 // fails. 1625 // 1626 // Finally, we make sure that all loop instructions are either loop increment 1627 // roots, belong to simple latch code, parts of validated reductions, part of 1628 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions 1629 // have been validated), then we reroll the loop. 1630 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header, 1631 const SCEV *BackedgeTakenCount, 1632 ReductionTracker &Reductions) { 1633 DAGRootTracker DAGRoots(this, L, IV, SE, AA, TLI, DT, LI, PreserveLCSSA, 1634 IVToIncMap, LoopControlIV); 1635 1636 if (!DAGRoots.findRoots()) 1637 return false; 1638 LLVM_DEBUG(dbgs() << "LRR: Found all root induction increments for: " << *IV 1639 << "\n"); 1640 1641 if (!DAGRoots.validate(Reductions)) 1642 return false; 1643 if (!Reductions.validateSelected()) 1644 return false; 1645 // At this point, we've validated the rerolling, and we're committed to 1646 // making changes! 1647 1648 Reductions.replaceSelected(); 1649 DAGRoots.replace(BackedgeTakenCount); 1650 1651 ++NumRerolledLoops; 1652 return true; 1653 } 1654 1655 bool LoopReroll::runOnLoop(Loop *L) { 1656 BasicBlock *Header = L->getHeader(); 1657 LLVM_DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() << "] Loop %" 1658 << Header->getName() << " (" << L->getNumBlocks() 1659 << " block(s))\n"); 1660 1661 // For now, we'll handle only single BB loops. 1662 if (L->getNumBlocks() > 1) 1663 return false; 1664 1665 if (!SE->hasLoopInvariantBackedgeTakenCount(L)) 1666 return false; 1667 1668 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1669 LLVM_DEBUG(dbgs() << "\n Before Reroll:\n" << *(L->getHeader()) << "\n"); 1670 LLVM_DEBUG(dbgs() << "LRR: backedge-taken count = " << *BackedgeTakenCount 1671 << "\n"); 1672 1673 // First, we need to find the induction variable with respect to which we can 1674 // reroll (there may be several possible options). 1675 SmallInstructionVector PossibleIVs; 1676 IVToIncMap.clear(); 1677 LoopControlIV = nullptr; 1678 collectPossibleIVs(L, PossibleIVs); 1679 1680 if (PossibleIVs.empty()) { 1681 LLVM_DEBUG(dbgs() << "LRR: No possible IVs found\n"); 1682 return false; 1683 } 1684 1685 ReductionTracker Reductions; 1686 collectPossibleReductions(L, Reductions); 1687 bool Changed = false; 1688 1689 // For each possible IV, collect the associated possible set of 'root' nodes 1690 // (i+1, i+2, etc.). 1691 for (Instruction *PossibleIV : PossibleIVs) 1692 if (reroll(PossibleIV, L, Header, BackedgeTakenCount, Reductions)) { 1693 Changed = true; 1694 break; 1695 } 1696 LLVM_DEBUG(dbgs() << "\n After Reroll:\n" << *(L->getHeader()) << "\n"); 1697 1698 // Trip count of L has changed so SE must be re-evaluated. 1699 if (Changed) 1700 SE->forgetLoop(L); 1701 1702 return Changed; 1703 } 1704 1705 bool LoopRerollLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { 1706 if (skipLoop(L)) 1707 return false; 1708 1709 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 1710 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1711 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1712 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( 1713 *L->getHeader()->getParent()); 1714 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1715 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 1716 1717 return LoopReroll(AA, LI, SE, TLI, DT, PreserveLCSSA).runOnLoop(L); 1718 } 1719 1720 PreservedAnalyses LoopRerollPass::run(Loop &L, LoopAnalysisManager &AM, 1721 LoopStandardAnalysisResults &AR, 1722 LPMUpdater &U) { 1723 return LoopReroll(&AR.AA, &AR.LI, &AR.SE, &AR.TLI, &AR.DT, true).runOnLoop(&L) 1724 ? getLoopPassPreservedAnalyses() 1725 : PreservedAnalyses::all(); 1726 } 1727