1 //===- LoopInterchange.cpp - Loop interchange 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 handles loop interchange transform. 10 // This pass interchanges loops to provide a more cache-friendly memory access 11 // patterns. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Scalar/LoopInterchange.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/Analysis/DependenceAnalysis.h" 21 #include "llvm/Analysis/LoopCacheAnalysis.h" 22 #include "llvm/Analysis/LoopInfo.h" 23 #include "llvm/Analysis/LoopNestAnalysis.h" 24 #include "llvm/Analysis/LoopPass.h" 25 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 26 #include "llvm/Analysis/ScalarEvolution.h" 27 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 28 #include "llvm/IR/BasicBlock.h" 29 #include "llvm/IR/Constants.h" 30 #include "llvm/IR/DiagnosticInfo.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/Function.h" 33 #include "llvm/IR/IRBuilder.h" 34 #include "llvm/IR/InstrTypes.h" 35 #include "llvm/IR/Instruction.h" 36 #include "llvm/IR/Instructions.h" 37 #include "llvm/IR/User.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/InitializePasses.h" 40 #include "llvm/Pass.h" 41 #include "llvm/Support/Casting.h" 42 #include "llvm/Support/CommandLine.h" 43 #include "llvm/Support/Debug.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/raw_ostream.h" 46 #include "llvm/Transforms/Scalar.h" 47 #include "llvm/Transforms/Scalar/LoopPassManager.h" 48 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 49 #include "llvm/Transforms/Utils/LoopUtils.h" 50 #include <cassert> 51 #include <utility> 52 #include <vector> 53 54 using namespace llvm; 55 56 #define DEBUG_TYPE "loop-interchange" 57 58 STATISTIC(LoopsInterchanged, "Number of loops interchanged"); 59 60 static cl::opt<int> LoopInterchangeCostThreshold( 61 "loop-interchange-threshold", cl::init(0), cl::Hidden, 62 cl::desc("Interchange if you gain more than this number")); 63 64 namespace { 65 66 using LoopVector = SmallVector<Loop *, 8>; 67 68 // TODO: Check if we can use a sparse matrix here. 69 using CharMatrix = std::vector<std::vector<char>>; 70 71 } // end anonymous namespace 72 73 // Maximum number of dependencies that can be handled in the dependency matrix. 74 static const unsigned MaxMemInstrCount = 100; 75 76 // Maximum loop depth supported. 77 static const unsigned MaxLoopNestDepth = 10; 78 79 #ifdef DUMP_DEP_MATRICIES 80 static void printDepMatrix(CharMatrix &DepMatrix) { 81 for (auto &Row : DepMatrix) { 82 for (auto D : Row) 83 LLVM_DEBUG(dbgs() << D << " "); 84 LLVM_DEBUG(dbgs() << "\n"); 85 } 86 } 87 #endif 88 89 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level, 90 Loop *L, DependenceInfo *DI, 91 ScalarEvolution *SE) { 92 using ValueVector = SmallVector<Value *, 16>; 93 94 ValueVector MemInstr; 95 96 // For each block. 97 for (BasicBlock *BB : L->blocks()) { 98 // Scan the BB and collect legal loads and stores. 99 for (Instruction &I : *BB) { 100 if (!isa<Instruction>(I)) 101 return false; 102 if (auto *Ld = dyn_cast<LoadInst>(&I)) { 103 if (!Ld->isSimple()) 104 return false; 105 MemInstr.push_back(&I); 106 } else if (auto *St = dyn_cast<StoreInst>(&I)) { 107 if (!St->isSimple()) 108 return false; 109 MemInstr.push_back(&I); 110 } 111 } 112 } 113 114 LLVM_DEBUG(dbgs() << "Found " << MemInstr.size() 115 << " Loads and Stores to analyze\n"); 116 117 ValueVector::iterator I, IE, J, JE; 118 119 for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) { 120 for (J = I, JE = MemInstr.end(); J != JE; ++J) { 121 std::vector<char> Dep; 122 Instruction *Src = cast<Instruction>(*I); 123 Instruction *Dst = cast<Instruction>(*J); 124 // Ignore Input dependencies. 125 if (isa<LoadInst>(Src) && isa<LoadInst>(Dst)) 126 continue; 127 // Track Output, Flow, and Anti dependencies. 128 if (auto D = DI->depends(Src, Dst, true)) { 129 assert(D->isOrdered() && "Expected an output, flow or anti dep."); 130 // If the direction vector is negative, normalize it to 131 // make it non-negative. 132 if (D->normalize(SE)) 133 LLVM_DEBUG(dbgs() << "Negative dependence vector normalized.\n"); 134 LLVM_DEBUG(StringRef DepType = 135 D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; 136 dbgs() << "Found " << DepType 137 << " dependency between Src and Dst\n" 138 << " Src:" << *Src << "\n Dst:" << *Dst << '\n'); 139 unsigned Levels = D->getLevels(); 140 char Direction; 141 for (unsigned II = 1; II <= Levels; ++II) { 142 if (D->isScalar(II)) { 143 Direction = 'S'; 144 Dep.push_back(Direction); 145 } else { 146 unsigned Dir = D->getDirection(II); 147 if (Dir == Dependence::DVEntry::LT || 148 Dir == Dependence::DVEntry::LE) 149 Direction = '<'; 150 else if (Dir == Dependence::DVEntry::GT || 151 Dir == Dependence::DVEntry::GE) 152 Direction = '>'; 153 else if (Dir == Dependence::DVEntry::EQ) 154 Direction = '='; 155 else 156 Direction = '*'; 157 Dep.push_back(Direction); 158 } 159 } 160 while (Dep.size() != Level) { 161 Dep.push_back('I'); 162 } 163 164 DepMatrix.push_back(Dep); 165 if (DepMatrix.size() > MaxMemInstrCount) { 166 LLVM_DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount 167 << " dependencies inside loop\n"); 168 return false; 169 } 170 } 171 } 172 } 173 174 return true; 175 } 176 177 // A loop is moved from index 'from' to an index 'to'. Update the Dependence 178 // matrix by exchanging the two columns. 179 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx, 180 unsigned ToIndx) { 181 for (unsigned I = 0, E = DepMatrix.size(); I < E; ++I) 182 std::swap(DepMatrix[I][ToIndx], DepMatrix[I][FromIndx]); 183 } 184 185 // After interchanging, check if the direction vector is valid. 186 // [Theorem] A permutation of the loops in a perfect nest is legal if and only 187 // if the direction matrix, after the same permutation is applied to its 188 // columns, has no ">" direction as the leftmost non-"=" direction in any row. 189 static bool isLexicographicallyPositive(std::vector<char> &DV) { 190 for (unsigned Level = 0; Level < DV.size(); ++Level) { 191 unsigned char Direction = DV[Level]; 192 if (Direction == '<') 193 return true; 194 if (Direction == '>' || Direction == '*') 195 return false; 196 } 197 return true; 198 } 199 200 // Checks if it is legal to interchange 2 loops. 201 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix, 202 unsigned InnerLoopId, 203 unsigned OuterLoopId) { 204 unsigned NumRows = DepMatrix.size(); 205 std::vector<char> Cur; 206 // For each row check if it is valid to interchange. 207 for (unsigned Row = 0; Row < NumRows; ++Row) { 208 // Create temporary DepVector check its lexicographical order 209 // before and after swapping OuterLoop vs InnerLoop 210 Cur = DepMatrix[Row]; 211 if (!isLexicographicallyPositive(Cur)) 212 return false; 213 std::swap(Cur[InnerLoopId], Cur[OuterLoopId]); 214 if (!isLexicographicallyPositive(Cur)) 215 return false; 216 } 217 return true; 218 } 219 220 static void populateWorklist(Loop &L, LoopVector &LoopList) { 221 LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: " 222 << L.getHeader()->getParent()->getName() << " Loop: %" 223 << L.getHeader()->getName() << '\n'); 224 assert(LoopList.empty() && "LoopList should initially be empty!"); 225 Loop *CurrentLoop = &L; 226 const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops(); 227 while (!Vec->empty()) { 228 // The current loop has multiple subloops in it hence it is not tightly 229 // nested. 230 // Discard all loops above it added into Worklist. 231 if (Vec->size() != 1) { 232 LoopList = {}; 233 return; 234 } 235 236 LoopList.push_back(CurrentLoop); 237 CurrentLoop = Vec->front(); 238 Vec = &CurrentLoop->getSubLoops(); 239 } 240 LoopList.push_back(CurrentLoop); 241 } 242 243 namespace { 244 245 /// LoopInterchangeLegality checks if it is legal to interchange the loop. 246 class LoopInterchangeLegality { 247 public: 248 LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE, 249 OptimizationRemarkEmitter *ORE) 250 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {} 251 252 /// Check if the loops can be interchanged. 253 bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId, 254 CharMatrix &DepMatrix); 255 256 /// Discover induction PHIs in the header of \p L. Induction 257 /// PHIs are added to \p Inductions. 258 bool findInductions(Loop *L, SmallVectorImpl<PHINode *> &Inductions); 259 260 /// Check if the loop structure is understood. We do not handle triangular 261 /// loops for now. 262 bool isLoopStructureUnderstood(); 263 264 bool currentLimitations(); 265 266 const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const { 267 return OuterInnerReductions; 268 } 269 270 const SmallVectorImpl<PHINode *> &getInnerLoopInductions() const { 271 return InnerLoopInductions; 272 } 273 274 private: 275 bool tightlyNested(Loop *Outer, Loop *Inner); 276 bool containsUnsafeInstructions(BasicBlock *BB); 277 278 /// Discover induction and reduction PHIs in the header of \p L. Induction 279 /// PHIs are added to \p Inductions, reductions are added to 280 /// OuterInnerReductions. When the outer loop is passed, the inner loop needs 281 /// to be passed as \p InnerLoop. 282 bool findInductionAndReductions(Loop *L, 283 SmallVector<PHINode *, 8> &Inductions, 284 Loop *InnerLoop); 285 286 Loop *OuterLoop; 287 Loop *InnerLoop; 288 289 ScalarEvolution *SE; 290 291 /// Interface to emit optimization remarks. 292 OptimizationRemarkEmitter *ORE; 293 294 /// Set of reduction PHIs taking part of a reduction across the inner and 295 /// outer loop. 296 SmallPtrSet<PHINode *, 4> OuterInnerReductions; 297 298 /// Set of inner loop induction PHIs 299 SmallVector<PHINode *, 8> InnerLoopInductions; 300 }; 301 302 /// LoopInterchangeProfitability checks if it is profitable to interchange the 303 /// loop. 304 class LoopInterchangeProfitability { 305 public: 306 LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE, 307 OptimizationRemarkEmitter *ORE) 308 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {} 309 310 /// Check if the loop interchange is profitable. 311 bool isProfitable(const Loop *InnerLoop, const Loop *OuterLoop, 312 unsigned InnerLoopId, unsigned OuterLoopId, 313 CharMatrix &DepMatrix, 314 const DenseMap<const Loop *, unsigned> &CostMap, 315 std::unique_ptr<CacheCost> &CC); 316 317 private: 318 int getInstrOrderCost(); 319 std::optional<bool> isProfitablePerLoopCacheAnalysis( 320 const DenseMap<const Loop *, unsigned> &CostMap, 321 std::unique_ptr<CacheCost> &CC); 322 std::optional<bool> isProfitablePerInstrOrderCost(); 323 std::optional<bool> isProfitableForVectorization(unsigned InnerLoopId, 324 unsigned OuterLoopId, 325 CharMatrix &DepMatrix); 326 Loop *OuterLoop; 327 Loop *InnerLoop; 328 329 /// Scev analysis. 330 ScalarEvolution *SE; 331 332 /// Interface to emit optimization remarks. 333 OptimizationRemarkEmitter *ORE; 334 }; 335 336 /// LoopInterchangeTransform interchanges the loop. 337 class LoopInterchangeTransform { 338 public: 339 LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE, 340 LoopInfo *LI, DominatorTree *DT, 341 const LoopInterchangeLegality &LIL) 342 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), LIL(LIL) {} 343 344 /// Interchange OuterLoop and InnerLoop. 345 bool transform(); 346 void restructureLoops(Loop *NewInner, Loop *NewOuter, 347 BasicBlock *OrigInnerPreHeader, 348 BasicBlock *OrigOuterPreHeader); 349 void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop); 350 351 private: 352 bool adjustLoopLinks(); 353 bool adjustLoopBranches(); 354 355 Loop *OuterLoop; 356 Loop *InnerLoop; 357 358 /// Scev analysis. 359 ScalarEvolution *SE; 360 361 LoopInfo *LI; 362 DominatorTree *DT; 363 364 const LoopInterchangeLegality &LIL; 365 }; 366 367 struct LoopInterchange { 368 ScalarEvolution *SE = nullptr; 369 LoopInfo *LI = nullptr; 370 DependenceInfo *DI = nullptr; 371 DominatorTree *DT = nullptr; 372 std::unique_ptr<CacheCost> CC = nullptr; 373 374 /// Interface to emit optimization remarks. 375 OptimizationRemarkEmitter *ORE; 376 377 LoopInterchange(ScalarEvolution *SE, LoopInfo *LI, DependenceInfo *DI, 378 DominatorTree *DT, std::unique_ptr<CacheCost> &CC, 379 OptimizationRemarkEmitter *ORE) 380 : SE(SE), LI(LI), DI(DI), DT(DT), CC(std::move(CC)), ORE(ORE) {} 381 382 bool run(Loop *L) { 383 if (L->getParentLoop()) 384 return false; 385 SmallVector<Loop *, 8> LoopList; 386 populateWorklist(*L, LoopList); 387 return processLoopList(LoopList); 388 } 389 390 bool run(LoopNest &LN) { 391 SmallVector<Loop *, 8> LoopList(LN.getLoops().begin(), LN.getLoops().end()); 392 for (unsigned I = 1; I < LoopList.size(); ++I) 393 if (LoopList[I]->getParentLoop() != LoopList[I - 1]) 394 return false; 395 return processLoopList(LoopList); 396 } 397 398 bool isComputableLoopNest(ArrayRef<Loop *> LoopList) { 399 for (Loop *L : LoopList) { 400 const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L); 401 if (isa<SCEVCouldNotCompute>(ExitCountOuter)) { 402 LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n"); 403 return false; 404 } 405 if (L->getNumBackEdges() != 1) { 406 LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n"); 407 return false; 408 } 409 if (!L->getExitingBlock()) { 410 LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n"); 411 return false; 412 } 413 } 414 return true; 415 } 416 417 unsigned selectLoopForInterchange(ArrayRef<Loop *> LoopList) { 418 // TODO: Add a better heuristic to select the loop to be interchanged based 419 // on the dependence matrix. Currently we select the innermost loop. 420 return LoopList.size() - 1; 421 } 422 423 bool processLoopList(SmallVectorImpl<Loop *> &LoopList) { 424 bool Changed = false; 425 unsigned LoopNestDepth = LoopList.size(); 426 if (LoopNestDepth < 2) { 427 LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n"); 428 return false; 429 } 430 if (LoopNestDepth > MaxLoopNestDepth) { 431 LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than " 432 << MaxLoopNestDepth << "\n"); 433 return false; 434 } 435 if (!isComputableLoopNest(LoopList)) { 436 LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n"); 437 return false; 438 } 439 440 LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth 441 << "\n"); 442 443 CharMatrix DependencyMatrix; 444 Loop *OuterMostLoop = *(LoopList.begin()); 445 if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth, 446 OuterMostLoop, DI, SE)) { 447 LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n"); 448 return false; 449 } 450 #ifdef DUMP_DEP_MATRICIES 451 LLVM_DEBUG(dbgs() << "Dependence before interchange\n"); 452 printDepMatrix(DependencyMatrix); 453 #endif 454 455 // Get the Outermost loop exit. 456 BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock(); 457 if (!LoopNestExit) { 458 LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block"); 459 return false; 460 } 461 462 unsigned SelecLoopId = selectLoopForInterchange(LoopList); 463 // Obtain the loop vector returned from loop cache analysis beforehand, 464 // and put each <Loop, index> pair into a map for constant time query 465 // later. Indices in loop vector reprsent the optimal order of the 466 // corresponding loop, e.g., given a loopnest with depth N, index 0 467 // indicates the loop should be placed as the outermost loop and index N 468 // indicates the loop should be placed as the innermost loop. 469 // 470 // For the old pass manager CacheCost would be null. 471 DenseMap<const Loop *, unsigned> CostMap; 472 if (CC != nullptr) { 473 const auto &LoopCosts = CC->getLoopCosts(); 474 for (unsigned i = 0; i < LoopCosts.size(); i++) { 475 CostMap[LoopCosts[i].first] = i; 476 } 477 } 478 // We try to achieve the globally optimal memory access for the loopnest, 479 // and do interchange based on a bubble-sort fasion. We start from 480 // the innermost loop, move it outwards to the best possible position 481 // and repeat this process. 482 for (unsigned j = SelecLoopId; j > 0; j--) { 483 bool ChangedPerIter = false; 484 for (unsigned i = SelecLoopId; i > SelecLoopId - j; i--) { 485 bool Interchanged = processLoop(LoopList[i], LoopList[i - 1], i, i - 1, 486 DependencyMatrix, CostMap); 487 if (!Interchanged) 488 continue; 489 // Loops interchanged, update LoopList accordingly. 490 std::swap(LoopList[i - 1], LoopList[i]); 491 // Update the DependencyMatrix 492 interChangeDependencies(DependencyMatrix, i, i - 1); 493 #ifdef DUMP_DEP_MATRICIES 494 LLVM_DEBUG(dbgs() << "Dependence after interchange\n"); 495 printDepMatrix(DependencyMatrix); 496 #endif 497 ChangedPerIter |= Interchanged; 498 Changed |= Interchanged; 499 } 500 // Early abort if there was no interchange during an entire round of 501 // moving loops outwards. 502 if (!ChangedPerIter) 503 break; 504 } 505 return Changed; 506 } 507 508 bool processLoop(Loop *InnerLoop, Loop *OuterLoop, unsigned InnerLoopId, 509 unsigned OuterLoopId, 510 std::vector<std::vector<char>> &DependencyMatrix, 511 const DenseMap<const Loop *, unsigned> &CostMap) { 512 LLVM_DEBUG(dbgs() << "Processing InnerLoopId = " << InnerLoopId 513 << " and OuterLoopId = " << OuterLoopId << "\n"); 514 LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE); 515 if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) { 516 LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n"); 517 return false; 518 } 519 LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n"); 520 LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE); 521 if (!LIP.isProfitable(InnerLoop, OuterLoop, InnerLoopId, OuterLoopId, 522 DependencyMatrix, CostMap, CC)) { 523 LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n"); 524 return false; 525 } 526 527 ORE->emit([&]() { 528 return OptimizationRemark(DEBUG_TYPE, "Interchanged", 529 InnerLoop->getStartLoc(), 530 InnerLoop->getHeader()) 531 << "Loop interchanged with enclosing loop."; 532 }); 533 534 LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LIL); 535 LIT.transform(); 536 LLVM_DEBUG(dbgs() << "Loops interchanged.\n"); 537 LoopsInterchanged++; 538 539 llvm::formLCSSARecursively(*OuterLoop, *DT, LI, SE); 540 return true; 541 } 542 }; 543 544 } // end anonymous namespace 545 546 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) { 547 return any_of(*BB, [](const Instruction &I) { 548 return I.mayHaveSideEffects() || I.mayReadFromMemory(); 549 }); 550 } 551 552 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) { 553 BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); 554 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); 555 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); 556 557 LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n"); 558 559 // A perfectly nested loop will not have any branch in between the outer and 560 // inner block i.e. outer header will branch to either inner preheader and 561 // outerloop latch. 562 BranchInst *OuterLoopHeaderBI = 563 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator()); 564 if (!OuterLoopHeaderBI) 565 return false; 566 567 for (BasicBlock *Succ : successors(OuterLoopHeaderBI)) 568 if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() && 569 Succ != OuterLoopLatch) 570 return false; 571 572 LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n"); 573 // We do not have any basic block in between now make sure the outer header 574 // and outer loop latch doesn't contain any unsafe instructions. 575 if (containsUnsafeInstructions(OuterLoopHeader) || 576 containsUnsafeInstructions(OuterLoopLatch)) 577 return false; 578 579 // Also make sure the inner loop preheader does not contain any unsafe 580 // instructions. Note that all instructions in the preheader will be moved to 581 // the outer loop header when interchanging. 582 if (InnerLoopPreHeader != OuterLoopHeader && 583 containsUnsafeInstructions(InnerLoopPreHeader)) 584 return false; 585 586 BasicBlock *InnerLoopExit = InnerLoop->getExitBlock(); 587 // Ensure the inner loop exit block flows to the outer loop latch possibly 588 // through empty blocks. 589 const BasicBlock &SuccInner = 590 LoopNest::skipEmptyBlockUntil(InnerLoopExit, OuterLoopLatch); 591 if (&SuccInner != OuterLoopLatch) { 592 LLVM_DEBUG(dbgs() << "Inner loop exit block " << *InnerLoopExit 593 << " does not lead to the outer loop latch.\n";); 594 return false; 595 } 596 // The inner loop exit block does flow to the outer loop latch and not some 597 // other BBs, now make sure it contains safe instructions, since it will be 598 // moved into the (new) inner loop after interchange. 599 if (containsUnsafeInstructions(InnerLoopExit)) 600 return false; 601 602 LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n"); 603 // We have a perfect loop nest. 604 return true; 605 } 606 607 bool LoopInterchangeLegality::isLoopStructureUnderstood() { 608 BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader(); 609 for (PHINode *InnerInduction : InnerLoopInductions) { 610 unsigned Num = InnerInduction->getNumOperands(); 611 for (unsigned i = 0; i < Num; ++i) { 612 Value *Val = InnerInduction->getOperand(i); 613 if (isa<Constant>(Val)) 614 continue; 615 Instruction *I = dyn_cast<Instruction>(Val); 616 if (!I) 617 return false; 618 // TODO: Handle triangular loops. 619 // e.g. for(int i=0;i<N;i++) 620 // for(int j=i;j<N;j++) 621 unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i); 622 if (InnerInduction->getIncomingBlock(IncomBlockIndx) == 623 InnerLoopPreheader && 624 !OuterLoop->isLoopInvariant(I)) { 625 return false; 626 } 627 } 628 } 629 630 // TODO: Handle triangular loops of another form. 631 // e.g. for(int i=0;i<N;i++) 632 // for(int j=0;j<i;j++) 633 // or, 634 // for(int i=0;i<N;i++) 635 // for(int j=0;j*i<N;j++) 636 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 637 BranchInst *InnerLoopLatchBI = 638 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator()); 639 if (!InnerLoopLatchBI->isConditional()) 640 return false; 641 if (CmpInst *InnerLoopCmp = 642 dyn_cast<CmpInst>(InnerLoopLatchBI->getCondition())) { 643 Value *Op0 = InnerLoopCmp->getOperand(0); 644 Value *Op1 = InnerLoopCmp->getOperand(1); 645 646 // LHS and RHS of the inner loop exit condition, e.g., 647 // in "for(int j=0;j<i;j++)", LHS is j and RHS is i. 648 Value *Left = nullptr; 649 Value *Right = nullptr; 650 651 // Check if V only involves inner loop induction variable. 652 // Return true if V is InnerInduction, or a cast from 653 // InnerInduction, or a binary operator that involves 654 // InnerInduction and a constant. 655 std::function<bool(Value *)> IsPathToInnerIndVar; 656 IsPathToInnerIndVar = [this, &IsPathToInnerIndVar](const Value *V) -> bool { 657 if (llvm::is_contained(InnerLoopInductions, V)) 658 return true; 659 if (isa<Constant>(V)) 660 return true; 661 const Instruction *I = dyn_cast<Instruction>(V); 662 if (!I) 663 return false; 664 if (isa<CastInst>(I)) 665 return IsPathToInnerIndVar(I->getOperand(0)); 666 if (isa<BinaryOperator>(I)) 667 return IsPathToInnerIndVar(I->getOperand(0)) && 668 IsPathToInnerIndVar(I->getOperand(1)); 669 return false; 670 }; 671 672 // In case of multiple inner loop indvars, it is okay if LHS and RHS 673 // are both inner indvar related variables. 674 if (IsPathToInnerIndVar(Op0) && IsPathToInnerIndVar(Op1)) 675 return true; 676 677 // Otherwise we check if the cmp instruction compares an inner indvar 678 // related variable (Left) with a outer loop invariant (Right). 679 if (IsPathToInnerIndVar(Op0) && !isa<Constant>(Op0)) { 680 Left = Op0; 681 Right = Op1; 682 } else if (IsPathToInnerIndVar(Op1) && !isa<Constant>(Op1)) { 683 Left = Op1; 684 Right = Op0; 685 } 686 687 if (Left == nullptr) 688 return false; 689 690 const SCEV *S = SE->getSCEV(Right); 691 if (!SE->isLoopInvariant(S, OuterLoop)) 692 return false; 693 } 694 695 return true; 696 } 697 698 // If SV is a LCSSA PHI node with a single incoming value, return the incoming 699 // value. 700 static Value *followLCSSA(Value *SV) { 701 PHINode *PHI = dyn_cast<PHINode>(SV); 702 if (!PHI) 703 return SV; 704 705 if (PHI->getNumIncomingValues() != 1) 706 return SV; 707 return followLCSSA(PHI->getIncomingValue(0)); 708 } 709 710 // Check V's users to see if it is involved in a reduction in L. 711 static PHINode *findInnerReductionPhi(Loop *L, Value *V) { 712 // Reduction variables cannot be constants. 713 if (isa<Constant>(V)) 714 return nullptr; 715 716 for (Value *User : V->users()) { 717 if (PHINode *PHI = dyn_cast<PHINode>(User)) { 718 if (PHI->getNumIncomingValues() == 1) 719 continue; 720 RecurrenceDescriptor RD; 721 if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD)) { 722 // Detect floating point reduction only when it can be reordered. 723 if (RD.getExactFPMathInst() != nullptr) 724 return nullptr; 725 return PHI; 726 } 727 return nullptr; 728 } 729 } 730 731 return nullptr; 732 } 733 734 bool LoopInterchangeLegality::findInductionAndReductions( 735 Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) { 736 if (!L->getLoopLatch() || !L->getLoopPredecessor()) 737 return false; 738 for (PHINode &PHI : L->getHeader()->phis()) { 739 RecurrenceDescriptor RD; 740 InductionDescriptor ID; 741 if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID)) 742 Inductions.push_back(&PHI); 743 else { 744 // PHIs in inner loops need to be part of a reduction in the outer loop, 745 // discovered when checking the PHIs of the outer loop earlier. 746 if (!InnerLoop) { 747 if (!OuterInnerReductions.count(&PHI)) { 748 LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions " 749 "across the outer loop.\n"); 750 return false; 751 } 752 } else { 753 assert(PHI.getNumIncomingValues() == 2 && 754 "Phis in loop header should have exactly 2 incoming values"); 755 // Check if we have a PHI node in the outer loop that has a reduction 756 // result from the inner loop as an incoming value. 757 Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch())); 758 PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V); 759 if (!InnerRedPhi || 760 !llvm::is_contained(InnerRedPhi->incoming_values(), &PHI)) { 761 LLVM_DEBUG( 762 dbgs() 763 << "Failed to recognize PHI as an induction or reduction.\n"); 764 return false; 765 } 766 OuterInnerReductions.insert(&PHI); 767 OuterInnerReductions.insert(InnerRedPhi); 768 } 769 } 770 } 771 return true; 772 } 773 774 // This function indicates the current limitations in the transform as a result 775 // of which we do not proceed. 776 bool LoopInterchangeLegality::currentLimitations() { 777 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 778 779 // transform currently expects the loop latches to also be the exiting 780 // blocks. 781 if (InnerLoop->getExitingBlock() != InnerLoopLatch || 782 OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() || 783 !isa<BranchInst>(InnerLoopLatch->getTerminator()) || 784 !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) { 785 LLVM_DEBUG( 786 dbgs() << "Loops where the latch is not the exiting block are not" 787 << " supported currently.\n"); 788 ORE->emit([&]() { 789 return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch", 790 OuterLoop->getStartLoc(), 791 OuterLoop->getHeader()) 792 << "Loops where the latch is not the exiting block cannot be" 793 " interchange currently."; 794 }); 795 return true; 796 } 797 798 SmallVector<PHINode *, 8> Inductions; 799 if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) { 800 LLVM_DEBUG( 801 dbgs() << "Only outer loops with induction or reduction PHI nodes " 802 << "are supported currently.\n"); 803 ORE->emit([&]() { 804 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter", 805 OuterLoop->getStartLoc(), 806 OuterLoop->getHeader()) 807 << "Only outer loops with induction or reduction PHI nodes can be" 808 " interchanged currently."; 809 }); 810 return true; 811 } 812 813 Inductions.clear(); 814 // For multi-level loop nests, make sure that all phi nodes for inner loops 815 // at all levels can be recognized as a induction or reduction phi. Bail out 816 // if a phi node at a certain nesting level cannot be properly recognized. 817 Loop *CurLevelLoop = OuterLoop; 818 while (!CurLevelLoop->getSubLoops().empty()) { 819 // We already made sure that the loop nest is tightly nested. 820 CurLevelLoop = CurLevelLoop->getSubLoops().front(); 821 if (!findInductionAndReductions(CurLevelLoop, Inductions, nullptr)) { 822 LLVM_DEBUG( 823 dbgs() << "Only inner loops with induction or reduction PHI nodes " 824 << "are supported currently.\n"); 825 ORE->emit([&]() { 826 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner", 827 CurLevelLoop->getStartLoc(), 828 CurLevelLoop->getHeader()) 829 << "Only inner loops with induction or reduction PHI nodes can be" 830 " interchange currently."; 831 }); 832 return true; 833 } 834 } 835 836 // TODO: Triangular loops are not handled for now. 837 if (!isLoopStructureUnderstood()) { 838 LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n"); 839 ORE->emit([&]() { 840 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner", 841 InnerLoop->getStartLoc(), 842 InnerLoop->getHeader()) 843 << "Inner loop structure not understood currently."; 844 }); 845 return true; 846 } 847 848 return false; 849 } 850 851 bool LoopInterchangeLegality::findInductions( 852 Loop *L, SmallVectorImpl<PHINode *> &Inductions) { 853 for (PHINode &PHI : L->getHeader()->phis()) { 854 InductionDescriptor ID; 855 if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID)) 856 Inductions.push_back(&PHI); 857 } 858 return !Inductions.empty(); 859 } 860 861 // We currently only support LCSSA PHI nodes in the inner loop exit, if their 862 // users are either reduction PHIs or PHIs outside the outer loop (which means 863 // the we are only interested in the final value after the loop). 864 static bool 865 areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL, 866 SmallPtrSetImpl<PHINode *> &Reductions) { 867 BasicBlock *InnerExit = OuterL->getUniqueExitBlock(); 868 for (PHINode &PHI : InnerExit->phis()) { 869 // Reduction lcssa phi will have only 1 incoming block that from loop latch. 870 if (PHI.getNumIncomingValues() > 1) 871 return false; 872 if (any_of(PHI.users(), [&Reductions, OuterL](User *U) { 873 PHINode *PN = dyn_cast<PHINode>(U); 874 return !PN || 875 (!Reductions.count(PN) && OuterL->contains(PN->getParent())); 876 })) { 877 return false; 878 } 879 } 880 return true; 881 } 882 883 // We currently support LCSSA PHI nodes in the outer loop exit, if their 884 // incoming values do not come from the outer loop latch or if the 885 // outer loop latch has a single predecessor. In that case, the value will 886 // be available if both the inner and outer loop conditions are true, which 887 // will still be true after interchanging. If we have multiple predecessor, 888 // that may not be the case, e.g. because the outer loop latch may be executed 889 // if the inner loop is not executed. 890 static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) { 891 BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock(); 892 for (PHINode &PHI : LoopNestExit->phis()) { 893 for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) { 894 Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i)); 895 if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch()) 896 continue; 897 898 // The incoming value is defined in the outer loop latch. Currently we 899 // only support that in case the outer loop latch has a single predecessor. 900 // This guarantees that the outer loop latch is executed if and only if 901 // the inner loop is executed (because tightlyNested() guarantees that the 902 // outer loop header only branches to the inner loop or the outer loop 903 // latch). 904 // FIXME: We could weaken this logic and allow multiple predecessors, 905 // if the values are produced outside the loop latch. We would need 906 // additional logic to update the PHI nodes in the exit block as 907 // well. 908 if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr) 909 return false; 910 } 911 } 912 return true; 913 } 914 915 // In case of multi-level nested loops, it may occur that lcssa phis exist in 916 // the latch of InnerLoop, i.e., when defs of the incoming values are further 917 // inside the loopnest. Sometimes those incoming values are not available 918 // after interchange, since the original inner latch will become the new outer 919 // latch which may have predecessor paths that do not include those incoming 920 // values. 921 // TODO: Handle transformation of lcssa phis in the InnerLoop latch in case of 922 // multi-level loop nests. 923 static bool areInnerLoopLatchPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) { 924 if (InnerLoop->getSubLoops().empty()) 925 return true; 926 // If the original outer latch has only one predecessor, then values defined 927 // further inside the looploop, e.g., in the innermost loop, will be available 928 // at the new outer latch after interchange. 929 if (OuterLoop->getLoopLatch()->getUniquePredecessor() != nullptr) 930 return true; 931 932 // The outer latch has more than one predecessors, i.e., the inner 933 // exit and the inner header. 934 // PHI nodes in the inner latch are lcssa phis where the incoming values 935 // are defined further inside the loopnest. Check if those phis are used 936 // in the original inner latch. If that is the case then bail out since 937 // those incoming values may not be available at the new outer latch. 938 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 939 for (PHINode &PHI : InnerLoopLatch->phis()) { 940 for (auto *U : PHI.users()) { 941 Instruction *UI = cast<Instruction>(U); 942 if (InnerLoopLatch == UI->getParent()) 943 return false; 944 } 945 } 946 return true; 947 } 948 949 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId, 950 unsigned OuterLoopId, 951 CharMatrix &DepMatrix) { 952 if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) { 953 LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId 954 << " and OuterLoopId = " << OuterLoopId 955 << " due to dependence\n"); 956 ORE->emit([&]() { 957 return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence", 958 InnerLoop->getStartLoc(), 959 InnerLoop->getHeader()) 960 << "Cannot interchange loops due to dependences."; 961 }); 962 return false; 963 } 964 // Check if outer and inner loop contain legal instructions only. 965 for (auto *BB : OuterLoop->blocks()) 966 for (Instruction &I : BB->instructionsWithoutDebug()) 967 if (CallInst *CI = dyn_cast<CallInst>(&I)) { 968 // readnone functions do not prevent interchanging. 969 if (CI->onlyWritesMemory()) 970 continue; 971 LLVM_DEBUG( 972 dbgs() << "Loops with call instructions cannot be interchanged " 973 << "safely."); 974 ORE->emit([&]() { 975 return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst", 976 CI->getDebugLoc(), 977 CI->getParent()) 978 << "Cannot interchange loops due to call instruction."; 979 }); 980 981 return false; 982 } 983 984 if (!findInductions(InnerLoop, InnerLoopInductions)) { 985 LLVM_DEBUG(dbgs() << "Cound not find inner loop induction variables.\n"); 986 return false; 987 } 988 989 if (!areInnerLoopLatchPHIsSupported(OuterLoop, InnerLoop)) { 990 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop latch.\n"); 991 ORE->emit([&]() { 992 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedInnerLatchPHI", 993 InnerLoop->getStartLoc(), 994 InnerLoop->getHeader()) 995 << "Cannot interchange loops because unsupported PHI nodes found " 996 "in inner loop latch."; 997 }); 998 return false; 999 } 1000 1001 // TODO: The loops could not be interchanged due to current limitations in the 1002 // transform module. 1003 if (currentLimitations()) { 1004 LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n"); 1005 return false; 1006 } 1007 1008 // Check if the loops are tightly nested. 1009 if (!tightlyNested(OuterLoop, InnerLoop)) { 1010 LLVM_DEBUG(dbgs() << "Loops not tightly nested\n"); 1011 ORE->emit([&]() { 1012 return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested", 1013 InnerLoop->getStartLoc(), 1014 InnerLoop->getHeader()) 1015 << "Cannot interchange loops because they are not tightly " 1016 "nested."; 1017 }); 1018 return false; 1019 } 1020 1021 if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop, 1022 OuterInnerReductions)) { 1023 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n"); 1024 ORE->emit([&]() { 1025 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI", 1026 InnerLoop->getStartLoc(), 1027 InnerLoop->getHeader()) 1028 << "Found unsupported PHI node in loop exit."; 1029 }); 1030 return false; 1031 } 1032 1033 if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) { 1034 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n"); 1035 ORE->emit([&]() { 1036 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI", 1037 OuterLoop->getStartLoc(), 1038 OuterLoop->getHeader()) 1039 << "Found unsupported PHI node in loop exit."; 1040 }); 1041 return false; 1042 } 1043 1044 return true; 1045 } 1046 1047 int LoopInterchangeProfitability::getInstrOrderCost() { 1048 unsigned GoodOrder, BadOrder; 1049 BadOrder = GoodOrder = 0; 1050 for (BasicBlock *BB : InnerLoop->blocks()) { 1051 for (Instruction &Ins : *BB) { 1052 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) { 1053 unsigned NumOp = GEP->getNumOperands(); 1054 bool FoundInnerInduction = false; 1055 bool FoundOuterInduction = false; 1056 for (unsigned i = 0; i < NumOp; ++i) { 1057 // Skip operands that are not SCEV-able. 1058 if (!SE->isSCEVable(GEP->getOperand(i)->getType())) 1059 continue; 1060 1061 const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i)); 1062 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal); 1063 if (!AR) 1064 continue; 1065 1066 // If we find the inner induction after an outer induction e.g. 1067 // for(int i=0;i<N;i++) 1068 // for(int j=0;j<N;j++) 1069 // A[i][j] = A[i-1][j-1]+k; 1070 // then it is a good order. 1071 if (AR->getLoop() == InnerLoop) { 1072 // We found an InnerLoop induction after OuterLoop induction. It is 1073 // a good order. 1074 FoundInnerInduction = true; 1075 if (FoundOuterInduction) { 1076 GoodOrder++; 1077 break; 1078 } 1079 } 1080 // If we find the outer induction after an inner induction e.g. 1081 // for(int i=0;i<N;i++) 1082 // for(int j=0;j<N;j++) 1083 // A[j][i] = A[j-1][i-1]+k; 1084 // then it is a bad order. 1085 if (AR->getLoop() == OuterLoop) { 1086 // We found an OuterLoop induction after InnerLoop induction. It is 1087 // a bad order. 1088 FoundOuterInduction = true; 1089 if (FoundInnerInduction) { 1090 BadOrder++; 1091 break; 1092 } 1093 } 1094 } 1095 } 1096 } 1097 } 1098 return GoodOrder - BadOrder; 1099 } 1100 1101 std::optional<bool> 1102 LoopInterchangeProfitability::isProfitablePerLoopCacheAnalysis( 1103 const DenseMap<const Loop *, unsigned> &CostMap, 1104 std::unique_ptr<CacheCost> &CC) { 1105 // This is the new cost model returned from loop cache analysis. 1106 // A smaller index means the loop should be placed an outer loop, and vice 1107 // versa. 1108 if (CostMap.find(InnerLoop) != CostMap.end() && 1109 CostMap.find(OuterLoop) != CostMap.end()) { 1110 unsigned InnerIndex = 0, OuterIndex = 0; 1111 InnerIndex = CostMap.find(InnerLoop)->second; 1112 OuterIndex = CostMap.find(OuterLoop)->second; 1113 LLVM_DEBUG(dbgs() << "InnerIndex = " << InnerIndex 1114 << ", OuterIndex = " << OuterIndex << "\n"); 1115 if (InnerIndex < OuterIndex) 1116 return std::optional<bool>(true); 1117 assert(InnerIndex != OuterIndex && "CostMap should assign unique " 1118 "numbers to each loop"); 1119 if (CC->getLoopCost(*OuterLoop) == CC->getLoopCost(*InnerLoop)) 1120 return std::nullopt; 1121 return std::optional<bool>(false); 1122 } 1123 return std::nullopt; 1124 } 1125 1126 std::optional<bool> 1127 LoopInterchangeProfitability::isProfitablePerInstrOrderCost() { 1128 // Legacy cost model: this is rough cost estimation algorithm. It counts the 1129 // good and bad order of induction variables in the instruction and allows 1130 // reordering if number of bad orders is more than good. 1131 int Cost = getInstrOrderCost(); 1132 LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n"); 1133 if (Cost < 0 && Cost < LoopInterchangeCostThreshold) 1134 return std::optional<bool>(true); 1135 1136 return std::nullopt; 1137 } 1138 1139 std::optional<bool> LoopInterchangeProfitability::isProfitableForVectorization( 1140 unsigned InnerLoopId, unsigned OuterLoopId, CharMatrix &DepMatrix) { 1141 for (auto &Row : DepMatrix) { 1142 // If the inner loop is loop independent or doesn't carry any dependency 1143 // it is not profitable to move this to outer position, since we are 1144 // likely able to do inner loop vectorization already. 1145 if (Row[InnerLoopId] == 'I' || Row[InnerLoopId] == '=') 1146 return std::optional<bool>(false); 1147 1148 // If the outer loop is not loop independent it is not profitable to move 1149 // this to inner position, since doing so would not enable inner loop 1150 // parallelism. 1151 if (Row[OuterLoopId] != 'I' && Row[OuterLoopId] != '=') 1152 return std::optional<bool>(false); 1153 } 1154 // If inner loop has dependence and outer loop is loop independent then it 1155 // is/ profitable to interchange to enable inner loop parallelism. 1156 // If there are no dependences, interchanging will not improve anything. 1157 return std::optional<bool>(!DepMatrix.empty()); 1158 } 1159 1160 bool LoopInterchangeProfitability::isProfitable( 1161 const Loop *InnerLoop, const Loop *OuterLoop, unsigned InnerLoopId, 1162 unsigned OuterLoopId, CharMatrix &DepMatrix, 1163 const DenseMap<const Loop *, unsigned> &CostMap, 1164 std::unique_ptr<CacheCost> &CC) { 1165 // isProfitable() is structured to avoid endless loop interchange. 1166 // If loop cache analysis could decide the profitability then, 1167 // profitability check will stop and return the analysis result. 1168 // If cache analysis failed to analyze the loopnest (e.g., 1169 // due to delinearization issues) then only check whether it is 1170 // profitable for InstrOrderCost. Likewise, if InstrOrderCost failed to 1171 // analysis the profitability then only, isProfitableForVectorization 1172 // will decide. 1173 std::optional<bool> shouldInterchange = 1174 isProfitablePerLoopCacheAnalysis(CostMap, CC); 1175 if (!shouldInterchange.has_value()) { 1176 shouldInterchange = isProfitablePerInstrOrderCost(); 1177 if (!shouldInterchange.has_value()) 1178 shouldInterchange = 1179 isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix); 1180 } 1181 if (!shouldInterchange.has_value()) { 1182 ORE->emit([&]() { 1183 return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable", 1184 InnerLoop->getStartLoc(), 1185 InnerLoop->getHeader()) 1186 << "Insufficient information to calculate the cost of loop for " 1187 "interchange."; 1188 }); 1189 return false; 1190 } else if (!shouldInterchange.value()) { 1191 ORE->emit([&]() { 1192 return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable", 1193 InnerLoop->getStartLoc(), 1194 InnerLoop->getHeader()) 1195 << "Interchanging loops is not considered to improve cache " 1196 "locality nor vectorization."; 1197 }); 1198 return false; 1199 } 1200 return true; 1201 } 1202 1203 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop, 1204 Loop *InnerLoop) { 1205 for (Loop *L : *OuterLoop) 1206 if (L == InnerLoop) { 1207 OuterLoop->removeChildLoop(L); 1208 return; 1209 } 1210 llvm_unreachable("Couldn't find loop"); 1211 } 1212 1213 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the 1214 /// new inner and outer loop after interchanging: NewInner is the original 1215 /// outer loop and NewOuter is the original inner loop. 1216 /// 1217 /// Before interchanging, we have the following structure 1218 /// Outer preheader 1219 // Outer header 1220 // Inner preheader 1221 // Inner header 1222 // Inner body 1223 // Inner latch 1224 // outer bbs 1225 // Outer latch 1226 // 1227 // After interchanging: 1228 // Inner preheader 1229 // Inner header 1230 // Outer preheader 1231 // Outer header 1232 // Inner body 1233 // outer bbs 1234 // Outer latch 1235 // Inner latch 1236 void LoopInterchangeTransform::restructureLoops( 1237 Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader, 1238 BasicBlock *OrigOuterPreHeader) { 1239 Loop *OuterLoopParent = OuterLoop->getParentLoop(); 1240 // The original inner loop preheader moves from the new inner loop to 1241 // the parent loop, if there is one. 1242 NewInner->removeBlockFromLoop(OrigInnerPreHeader); 1243 LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent); 1244 1245 // Switch the loop levels. 1246 if (OuterLoopParent) { 1247 // Remove the loop from its parent loop. 1248 removeChildLoop(OuterLoopParent, NewInner); 1249 removeChildLoop(NewInner, NewOuter); 1250 OuterLoopParent->addChildLoop(NewOuter); 1251 } else { 1252 removeChildLoop(NewInner, NewOuter); 1253 LI->changeTopLevelLoop(NewInner, NewOuter); 1254 } 1255 while (!NewOuter->isInnermost()) 1256 NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin())); 1257 NewOuter->addChildLoop(NewInner); 1258 1259 // BBs from the original inner loop. 1260 SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks()); 1261 1262 // Add BBs from the original outer loop to the original inner loop (excluding 1263 // BBs already in inner loop) 1264 for (BasicBlock *BB : NewInner->blocks()) 1265 if (LI->getLoopFor(BB) == NewInner) 1266 NewOuter->addBlockEntry(BB); 1267 1268 // Now remove inner loop header and latch from the new inner loop and move 1269 // other BBs (the loop body) to the new inner loop. 1270 BasicBlock *OuterHeader = NewOuter->getHeader(); 1271 BasicBlock *OuterLatch = NewOuter->getLoopLatch(); 1272 for (BasicBlock *BB : OrigInnerBBs) { 1273 // Nothing will change for BBs in child loops. 1274 if (LI->getLoopFor(BB) != NewOuter) 1275 continue; 1276 // Remove the new outer loop header and latch from the new inner loop. 1277 if (BB == OuterHeader || BB == OuterLatch) 1278 NewInner->removeBlockFromLoop(BB); 1279 else 1280 LI->changeLoopFor(BB, NewInner); 1281 } 1282 1283 // The preheader of the original outer loop becomes part of the new 1284 // outer loop. 1285 NewOuter->addBlockEntry(OrigOuterPreHeader); 1286 LI->changeLoopFor(OrigOuterPreHeader, NewOuter); 1287 1288 // Tell SE that we move the loops around. 1289 SE->forgetLoop(NewOuter); 1290 } 1291 1292 bool LoopInterchangeTransform::transform() { 1293 bool Transformed = false; 1294 1295 if (InnerLoop->getSubLoops().empty()) { 1296 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); 1297 LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n"); 1298 auto &InductionPHIs = LIL.getInnerLoopInductions(); 1299 if (InductionPHIs.empty()) { 1300 LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n"); 1301 return false; 1302 } 1303 1304 SmallVector<Instruction *, 8> InnerIndexVarList; 1305 for (PHINode *CurInductionPHI : InductionPHIs) { 1306 if (CurInductionPHI->getIncomingBlock(0) == InnerLoopPreHeader) 1307 InnerIndexVarList.push_back( 1308 dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(1))); 1309 else 1310 InnerIndexVarList.push_back( 1311 dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(0))); 1312 } 1313 1314 // Create a new latch block for the inner loop. We split at the 1315 // current latch's terminator and then move the condition and all 1316 // operands that are not either loop-invariant or the induction PHI into the 1317 // new latch block. 1318 BasicBlock *NewLatch = 1319 SplitBlock(InnerLoop->getLoopLatch(), 1320 InnerLoop->getLoopLatch()->getTerminator(), DT, LI); 1321 1322 SmallSetVector<Instruction *, 4> WorkList; 1323 unsigned i = 0; 1324 auto MoveInstructions = [&i, &WorkList, this, &InductionPHIs, NewLatch]() { 1325 for (; i < WorkList.size(); i++) { 1326 // Duplicate instruction and move it the new latch. Update uses that 1327 // have been moved. 1328 Instruction *NewI = WorkList[i]->clone(); 1329 NewI->insertBefore(NewLatch->getFirstNonPHI()); 1330 assert(!NewI->mayHaveSideEffects() && 1331 "Moving instructions with side-effects may change behavior of " 1332 "the loop nest!"); 1333 for (Use &U : llvm::make_early_inc_range(WorkList[i]->uses())) { 1334 Instruction *UserI = cast<Instruction>(U.getUser()); 1335 if (!InnerLoop->contains(UserI->getParent()) || 1336 UserI->getParent() == NewLatch || 1337 llvm::is_contained(InductionPHIs, UserI)) 1338 U.set(NewI); 1339 } 1340 // Add operands of moved instruction to the worklist, except if they are 1341 // outside the inner loop or are the induction PHI. 1342 for (Value *Op : WorkList[i]->operands()) { 1343 Instruction *OpI = dyn_cast<Instruction>(Op); 1344 if (!OpI || 1345 this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop || 1346 llvm::is_contained(InductionPHIs, OpI)) 1347 continue; 1348 WorkList.insert(OpI); 1349 } 1350 } 1351 }; 1352 1353 // FIXME: Should we interchange when we have a constant condition? 1354 Instruction *CondI = dyn_cast<Instruction>( 1355 cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator()) 1356 ->getCondition()); 1357 if (CondI) 1358 WorkList.insert(CondI); 1359 MoveInstructions(); 1360 for (Instruction *InnerIndexVar : InnerIndexVarList) 1361 WorkList.insert(cast<Instruction>(InnerIndexVar)); 1362 MoveInstructions(); 1363 } 1364 1365 // Ensure the inner loop phi nodes have a separate basic block. 1366 BasicBlock *InnerLoopHeader = InnerLoop->getHeader(); 1367 if (InnerLoopHeader->getFirstNonPHI() != InnerLoopHeader->getTerminator()) { 1368 SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI); 1369 LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n"); 1370 } 1371 1372 // Instructions in the original inner loop preheader may depend on values 1373 // defined in the outer loop header. Move them there, because the original 1374 // inner loop preheader will become the entry into the interchanged loop nest. 1375 // Currently we move all instructions and rely on LICM to move invariant 1376 // instructions outside the loop nest. 1377 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); 1378 BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); 1379 if (InnerLoopPreHeader != OuterLoopHeader) { 1380 SmallPtrSet<Instruction *, 4> NeedsMoving; 1381 for (Instruction &I : 1382 make_early_inc_range(make_range(InnerLoopPreHeader->begin(), 1383 std::prev(InnerLoopPreHeader->end())))) 1384 I.moveBefore(OuterLoopHeader->getTerminator()); 1385 } 1386 1387 Transformed |= adjustLoopLinks(); 1388 if (!Transformed) { 1389 LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n"); 1390 return false; 1391 } 1392 1393 return true; 1394 } 1395 1396 /// \brief Move all instructions except the terminator from FromBB right before 1397 /// InsertBefore 1398 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) { 1399 BasicBlock *ToBB = InsertBefore->getParent(); 1400 1401 ToBB->splice(InsertBefore->getIterator(), FromBB, FromBB->begin(), 1402 FromBB->getTerminator()->getIterator()); 1403 } 1404 1405 /// Swap instructions between \p BB1 and \p BB2 but keep terminators intact. 1406 static void swapBBContents(BasicBlock *BB1, BasicBlock *BB2) { 1407 // Save all non-terminator instructions of BB1 into TempInstrs and unlink them 1408 // from BB1 afterwards. 1409 auto Iter = map_range(*BB1, [](Instruction &I) { return &I; }); 1410 SmallVector<Instruction *, 4> TempInstrs(Iter.begin(), std::prev(Iter.end())); 1411 for (Instruction *I : TempInstrs) 1412 I->removeFromParent(); 1413 1414 // Move instructions from BB2 to BB1. 1415 moveBBContents(BB2, BB1->getTerminator()); 1416 1417 // Move instructions from TempInstrs to BB2. 1418 for (Instruction *I : TempInstrs) 1419 I->insertBefore(BB2->getTerminator()); 1420 } 1421 1422 // Update BI to jump to NewBB instead of OldBB. Records updates to the 1423 // dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that 1424 // \p OldBB is exactly once in BI's successor list. 1425 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB, 1426 BasicBlock *NewBB, 1427 std::vector<DominatorTree::UpdateType> &DTUpdates, 1428 bool MustUpdateOnce = true) { 1429 assert((!MustUpdateOnce || 1430 llvm::count_if(successors(BI), 1431 [OldBB](BasicBlock *BB) { 1432 return BB == OldBB; 1433 }) == 1) && "BI must jump to OldBB exactly once."); 1434 bool Changed = false; 1435 for (Use &Op : BI->operands()) 1436 if (Op == OldBB) { 1437 Op.set(NewBB); 1438 Changed = true; 1439 } 1440 1441 if (Changed) { 1442 DTUpdates.push_back( 1443 {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB}); 1444 DTUpdates.push_back( 1445 {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB}); 1446 } 1447 assert(Changed && "Expected a successor to be updated"); 1448 } 1449 1450 // Move Lcssa PHIs to the right place. 1451 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader, 1452 BasicBlock *InnerLatch, BasicBlock *OuterHeader, 1453 BasicBlock *OuterLatch, BasicBlock *OuterExit, 1454 Loop *InnerLoop, LoopInfo *LI) { 1455 1456 // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are 1457 // defined either in the header or latch. Those blocks will become header and 1458 // latch of the new outer loop, and the only possible users can PHI nodes 1459 // in the exit block of the loop nest or the outer loop header (reduction 1460 // PHIs, in that case, the incoming value must be defined in the inner loop 1461 // header). We can just substitute the user with the incoming value and remove 1462 // the PHI. 1463 for (PHINode &P : make_early_inc_range(InnerExit->phis())) { 1464 assert(P.getNumIncomingValues() == 1 && 1465 "Only loops with a single exit are supported!"); 1466 1467 // Incoming values are guaranteed be instructions currently. 1468 auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch)); 1469 // In case of multi-level nested loops, follow LCSSA to find the incoming 1470 // value defined from the innermost loop. 1471 auto IncIInnerMost = cast<Instruction>(followLCSSA(IncI)); 1472 // Skip phis with incoming values from the inner loop body, excluding the 1473 // header and latch. 1474 if (IncIInnerMost->getParent() != InnerLatch && 1475 IncIInnerMost->getParent() != InnerHeader) 1476 continue; 1477 1478 assert(all_of(P.users(), 1479 [OuterHeader, OuterExit, IncI, InnerHeader](User *U) { 1480 return (cast<PHINode>(U)->getParent() == OuterHeader && 1481 IncI->getParent() == InnerHeader) || 1482 cast<PHINode>(U)->getParent() == OuterExit; 1483 }) && 1484 "Can only replace phis iff the uses are in the loop nest exit or " 1485 "the incoming value is defined in the inner header (it will " 1486 "dominate all loop blocks after interchanging)"); 1487 P.replaceAllUsesWith(IncI); 1488 P.eraseFromParent(); 1489 } 1490 1491 SmallVector<PHINode *, 8> LcssaInnerExit; 1492 for (PHINode &P : InnerExit->phis()) 1493 LcssaInnerExit.push_back(&P); 1494 1495 SmallVector<PHINode *, 8> LcssaInnerLatch; 1496 for (PHINode &P : InnerLatch->phis()) 1497 LcssaInnerLatch.push_back(&P); 1498 1499 // Lcssa PHIs for values used outside the inner loop are in InnerExit. 1500 // If a PHI node has users outside of InnerExit, it has a use outside the 1501 // interchanged loop and we have to preserve it. We move these to 1502 // InnerLatch, which will become the new exit block for the innermost 1503 // loop after interchanging. 1504 for (PHINode *P : LcssaInnerExit) 1505 P->moveBefore(InnerLatch->getFirstNonPHI()); 1506 1507 // If the inner loop latch contains LCSSA PHIs, those come from a child loop 1508 // and we have to move them to the new inner latch. 1509 for (PHINode *P : LcssaInnerLatch) 1510 P->moveBefore(InnerExit->getFirstNonPHI()); 1511 1512 // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have 1513 // incoming values defined in the outer loop, we have to add a new PHI 1514 // in the inner loop latch, which became the exit block of the outer loop, 1515 // after interchanging. 1516 if (OuterExit) { 1517 for (PHINode &P : OuterExit->phis()) { 1518 if (P.getNumIncomingValues() != 1) 1519 continue; 1520 // Skip Phis with incoming values defined in the inner loop. Those should 1521 // already have been updated. 1522 auto I = dyn_cast<Instruction>(P.getIncomingValue(0)); 1523 if (!I || LI->getLoopFor(I->getParent()) == InnerLoop) 1524 continue; 1525 1526 PHINode *NewPhi = dyn_cast<PHINode>(P.clone()); 1527 NewPhi->setIncomingValue(0, P.getIncomingValue(0)); 1528 NewPhi->setIncomingBlock(0, OuterLatch); 1529 // We might have incoming edges from other BBs, i.e., the original outer 1530 // header. 1531 for (auto *Pred : predecessors(InnerLatch)) { 1532 if (Pred == OuterLatch) 1533 continue; 1534 NewPhi->addIncoming(P.getIncomingValue(0), Pred); 1535 } 1536 NewPhi->insertBefore(InnerLatch->getFirstNonPHI()); 1537 P.setIncomingValue(0, NewPhi); 1538 } 1539 } 1540 1541 // Now adjust the incoming blocks for the LCSSA PHIs. 1542 // For PHIs moved from Inner's exit block, we need to replace Inner's latch 1543 // with the new latch. 1544 InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch); 1545 } 1546 1547 bool LoopInterchangeTransform::adjustLoopBranches() { 1548 LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n"); 1549 std::vector<DominatorTree::UpdateType> DTUpdates; 1550 1551 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader(); 1552 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); 1553 1554 assert(OuterLoopPreHeader != OuterLoop->getHeader() && 1555 InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader && 1556 InnerLoopPreHeader && "Guaranteed by loop-simplify form"); 1557 // Ensure that both preheaders do not contain PHI nodes and have single 1558 // predecessors. This allows us to move them easily. We use 1559 // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing 1560 // preheaders do not satisfy those conditions. 1561 if (isa<PHINode>(OuterLoopPreHeader->begin()) || 1562 !OuterLoopPreHeader->getUniquePredecessor()) 1563 OuterLoopPreHeader = 1564 InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true); 1565 if (InnerLoopPreHeader == OuterLoop->getHeader()) 1566 InnerLoopPreHeader = 1567 InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true); 1568 1569 // Adjust the loop preheader 1570 BasicBlock *InnerLoopHeader = InnerLoop->getHeader(); 1571 BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); 1572 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 1573 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); 1574 BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor(); 1575 BasicBlock *InnerLoopLatchPredecessor = 1576 InnerLoopLatch->getUniquePredecessor(); 1577 BasicBlock *InnerLoopLatchSuccessor; 1578 BasicBlock *OuterLoopLatchSuccessor; 1579 1580 BranchInst *OuterLoopLatchBI = 1581 dyn_cast<BranchInst>(OuterLoopLatch->getTerminator()); 1582 BranchInst *InnerLoopLatchBI = 1583 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator()); 1584 BranchInst *OuterLoopHeaderBI = 1585 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator()); 1586 BranchInst *InnerLoopHeaderBI = 1587 dyn_cast<BranchInst>(InnerLoopHeader->getTerminator()); 1588 1589 if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor || 1590 !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI || 1591 !InnerLoopHeaderBI) 1592 return false; 1593 1594 BranchInst *InnerLoopLatchPredecessorBI = 1595 dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator()); 1596 BranchInst *OuterLoopPredecessorBI = 1597 dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator()); 1598 1599 if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI) 1600 return false; 1601 BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor(); 1602 if (!InnerLoopHeaderSuccessor) 1603 return false; 1604 1605 // Adjust Loop Preheader and headers. 1606 // The branches in the outer loop predecessor and the outer loop header can 1607 // be unconditional branches or conditional branches with duplicates. Consider 1608 // this when updating the successors. 1609 updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader, 1610 InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false); 1611 // The outer loop header might or might not branch to the outer latch. 1612 // We are guaranteed to branch to the inner loop preheader. 1613 if (llvm::is_contained(OuterLoopHeaderBI->successors(), OuterLoopLatch)) { 1614 // In this case the outerLoopHeader should branch to the InnerLoopLatch. 1615 updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, InnerLoopLatch, 1616 DTUpdates, 1617 /*MustUpdateOnce=*/false); 1618 } 1619 updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader, 1620 InnerLoopHeaderSuccessor, DTUpdates, 1621 /*MustUpdateOnce=*/false); 1622 1623 // Adjust reduction PHI's now that the incoming block has changed. 1624 InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader, 1625 OuterLoopHeader); 1626 1627 updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor, 1628 OuterLoopPreHeader, DTUpdates); 1629 1630 // -------------Adjust loop latches----------- 1631 if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader) 1632 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1); 1633 else 1634 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0); 1635 1636 updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch, 1637 InnerLoopLatchSuccessor, DTUpdates); 1638 1639 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader) 1640 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1); 1641 else 1642 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0); 1643 1644 updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor, 1645 OuterLoopLatchSuccessor, DTUpdates); 1646 updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch, 1647 DTUpdates); 1648 1649 DT->applyUpdates(DTUpdates); 1650 restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader, 1651 OuterLoopPreHeader); 1652 1653 moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch, 1654 OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(), 1655 InnerLoop, LI); 1656 // For PHIs in the exit block of the outer loop, outer's latch has been 1657 // replaced by Inners'. 1658 OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch); 1659 1660 auto &OuterInnerReductions = LIL.getOuterInnerReductions(); 1661 // Now update the reduction PHIs in the inner and outer loop headers. 1662 SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs; 1663 for (PHINode &PHI : InnerLoopHeader->phis()) 1664 if (OuterInnerReductions.contains(&PHI)) 1665 InnerLoopPHIs.push_back(&PHI); 1666 1667 for (PHINode &PHI : OuterLoopHeader->phis()) 1668 if (OuterInnerReductions.contains(&PHI)) 1669 OuterLoopPHIs.push_back(&PHI); 1670 1671 // Now move the remaining reduction PHIs from outer to inner loop header and 1672 // vice versa. The PHI nodes must be part of a reduction across the inner and 1673 // outer loop and all the remains to do is and updating the incoming blocks. 1674 for (PHINode *PHI : OuterLoopPHIs) { 1675 LLVM_DEBUG(dbgs() << "Outer loop reduction PHIs:\n"; PHI->dump();); 1676 PHI->moveBefore(InnerLoopHeader->getFirstNonPHI()); 1677 assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node"); 1678 } 1679 for (PHINode *PHI : InnerLoopPHIs) { 1680 LLVM_DEBUG(dbgs() << "Inner loop reduction PHIs:\n"; PHI->dump();); 1681 PHI->moveBefore(OuterLoopHeader->getFirstNonPHI()); 1682 assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node"); 1683 } 1684 1685 // Update the incoming blocks for moved PHI nodes. 1686 OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader); 1687 OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch); 1688 InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader); 1689 InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch); 1690 1691 // Values defined in the outer loop header could be used in the inner loop 1692 // latch. In that case, we need to create LCSSA phis for them, because after 1693 // interchanging they will be defined in the new inner loop and used in the 1694 // new outer loop. 1695 IRBuilder<> Builder(OuterLoopHeader->getContext()); 1696 SmallVector<Instruction *, 4> MayNeedLCSSAPhis; 1697 for (Instruction &I : 1698 make_range(OuterLoopHeader->begin(), std::prev(OuterLoopHeader->end()))) 1699 MayNeedLCSSAPhis.push_back(&I); 1700 formLCSSAForInstructions(MayNeedLCSSAPhis, *DT, *LI, SE, Builder); 1701 1702 return true; 1703 } 1704 1705 bool LoopInterchangeTransform::adjustLoopLinks() { 1706 // Adjust all branches in the inner and outer loop. 1707 bool Changed = adjustLoopBranches(); 1708 if (Changed) { 1709 // We have interchanged the preheaders so we need to interchange the data in 1710 // the preheaders as well. This is because the content of the inner 1711 // preheader was previously executed inside the outer loop. 1712 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader(); 1713 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); 1714 swapBBContents(OuterLoopPreHeader, InnerLoopPreHeader); 1715 } 1716 return Changed; 1717 } 1718 1719 namespace { 1720 /// Main LoopInterchange Pass. 1721 struct LoopInterchangeLegacyPass : public LoopPass { 1722 static char ID; 1723 1724 LoopInterchangeLegacyPass() : LoopPass(ID) { 1725 initializeLoopInterchangeLegacyPassPass(*PassRegistry::getPassRegistry()); 1726 } 1727 1728 void getAnalysisUsage(AnalysisUsage &AU) const override { 1729 AU.addRequired<DependenceAnalysisWrapperPass>(); 1730 AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); 1731 1732 getLoopAnalysisUsage(AU); 1733 } 1734 1735 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 1736 if (skipLoop(L)) 1737 return false; 1738 1739 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1740 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1741 auto *DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI(); 1742 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1743 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); 1744 std::unique_ptr<CacheCost> CC = nullptr; 1745 return LoopInterchange(SE, LI, DI, DT, CC, ORE).run(L); 1746 } 1747 }; 1748 } // namespace 1749 1750 char LoopInterchangeLegacyPass::ID = 0; 1751 1752 INITIALIZE_PASS_BEGIN(LoopInterchangeLegacyPass, "loop-interchange", 1753 "Interchanges loops for cache reuse", false, false) 1754 INITIALIZE_PASS_DEPENDENCY(LoopPass) 1755 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass) 1756 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) 1757 1758 INITIALIZE_PASS_END(LoopInterchangeLegacyPass, "loop-interchange", 1759 "Interchanges loops for cache reuse", false, false) 1760 1761 Pass *llvm::createLoopInterchangePass() { 1762 return new LoopInterchangeLegacyPass(); 1763 } 1764 1765 PreservedAnalyses LoopInterchangePass::run(LoopNest &LN, 1766 LoopAnalysisManager &AM, 1767 LoopStandardAnalysisResults &AR, 1768 LPMUpdater &U) { 1769 Function &F = *LN.getParent(); 1770 1771 DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI); 1772 std::unique_ptr<CacheCost> CC = 1773 CacheCost::getCacheCost(LN.getOutermostLoop(), AR, DI); 1774 OptimizationRemarkEmitter ORE(&F); 1775 if (!LoopInterchange(&AR.SE, &AR.LI, &DI, &AR.DT, CC, &ORE).run(LN)) 1776 return PreservedAnalyses::all(); 1777 U.markLoopNestChanged(true); 1778 return getLoopPassPreservedAnalyses(); 1779 } 1780