| File: | lib/Transforms/Scalar/LoopInterchange.cpp |
| Warning: | line 1002, column 5 Value stored to 'OuterLoopPreHeader' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LoopInterchange.cpp - Loop interchange pass-------------------------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This Pass handles loop interchange transform. |
| 11 | // This pass interchanges loops to provide a more cache-friendly memory access |
| 12 | // patterns. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #include "llvm/ADT/STLExtras.h" |
| 17 | #include "llvm/ADT/SmallVector.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/Analysis/AliasAnalysis.h" |
| 20 | #include "llvm/Analysis/DependenceAnalysis.h" |
| 21 | #include "llvm/Analysis/LoopInfo.h" |
| 22 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 23 | #include "llvm/Analysis/ScalarEvolution.h" |
| 24 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 25 | #include "llvm/IR/BasicBlock.h" |
| 26 | #include "llvm/IR/Constants.h" |
| 27 | #include "llvm/IR/DiagnosticInfo.h" |
| 28 | #include "llvm/IR/Dominators.h" |
| 29 | #include "llvm/IR/Function.h" |
| 30 | #include "llvm/IR/InstrTypes.h" |
| 31 | #include "llvm/IR/Instruction.h" |
| 32 | #include "llvm/IR/Instructions.h" |
| 33 | #include "llvm/IR/Type.h" |
| 34 | #include "llvm/IR/User.h" |
| 35 | #include "llvm/IR/Value.h" |
| 36 | #include "llvm/Pass.h" |
| 37 | #include "llvm/Support/Casting.h" |
| 38 | #include "llvm/Support/CommandLine.h" |
| 39 | #include "llvm/Support/Debug.h" |
| 40 | #include "llvm/Support/ErrorHandling.h" |
| 41 | #include "llvm/Support/raw_ostream.h" |
| 42 | #include "llvm/Transforms/Scalar.h" |
| 43 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 44 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 45 | #include <cassert> |
| 46 | #include <utility> |
| 47 | #include <vector> |
| 48 | |
| 49 | using namespace llvm; |
| 50 | |
| 51 | #define DEBUG_TYPE"loop-interchange" "loop-interchange" |
| 52 | |
| 53 | static cl::opt<int> LoopInterchangeCostThreshold( |
| 54 | "loop-interchange-threshold", cl::init(0), cl::Hidden, |
| 55 | cl::desc("Interchange if you gain more than this number")); |
| 56 | |
| 57 | namespace { |
| 58 | |
| 59 | using LoopVector = SmallVector<Loop *, 8>; |
| 60 | |
| 61 | // TODO: Check if we can use a sparse matrix here. |
| 62 | using CharMatrix = std::vector<std::vector<char>>; |
| 63 | |
| 64 | } // end anonymous namespace |
| 65 | |
| 66 | // Maximum number of dependencies that can be handled in the dependency matrix. |
| 67 | static const unsigned MaxMemInstrCount = 100; |
| 68 | |
| 69 | // Maximum loop depth supported. |
| 70 | static const unsigned MaxLoopNestDepth = 10; |
| 71 | |
| 72 | #ifdef DUMP_DEP_MATRICIES |
| 73 | static void printDepMatrix(CharMatrix &DepMatrix) { |
| 74 | for (auto &Row : DepMatrix) { |
| 75 | for (auto D : Row) |
| 76 | DEBUG(dbgs() << D << " ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << D << " "; } } while (false); |
| 77 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "\n"; } } while (false ); |
| 78 | } |
| 79 | } |
| 80 | #endif |
| 81 | |
| 82 | static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level, |
| 83 | Loop *L, DependenceInfo *DI) { |
| 84 | using ValueVector = SmallVector<Value *, 16>; |
| 85 | |
| 86 | ValueVector MemInstr; |
| 87 | |
| 88 | // For each block. |
| 89 | for (BasicBlock *BB : L->blocks()) { |
| 90 | // Scan the BB and collect legal loads and stores. |
| 91 | for (Instruction &I : *BB) { |
| 92 | if (!isa<Instruction>(I)) |
| 93 | return false; |
| 94 | if (auto *Ld = dyn_cast<LoadInst>(&I)) { |
| 95 | if (!Ld->isSimple()) |
| 96 | return false; |
| 97 | MemInstr.push_back(&I); |
| 98 | } else if (auto *St = dyn_cast<StoreInst>(&I)) { |
| 99 | if (!St->isSimple()) |
| 100 | return false; |
| 101 | MemInstr.push_back(&I); |
| 102 | } |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | DEBUG(dbgs() << "Found " << MemInstr.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Found " << MemInstr .size() << " Loads and Stores to analyze\n"; } } while ( false) |
| 107 | << " Loads and Stores to analyze\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Found " << MemInstr .size() << " Loads and Stores to analyze\n"; } } while ( false); |
| 108 | |
| 109 | ValueVector::iterator I, IE, J, JE; |
| 110 | |
| 111 | for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) { |
| 112 | for (J = I, JE = MemInstr.end(); J != JE; ++J) { |
| 113 | std::vector<char> Dep; |
| 114 | Instruction *Src = cast<Instruction>(*I); |
| 115 | Instruction *Dst = cast<Instruction>(*J); |
| 116 | if (Src == Dst) |
| 117 | continue; |
| 118 | // Ignore Input dependencies. |
| 119 | if (isa<LoadInst>(Src) && isa<LoadInst>(Dst)) |
| 120 | continue; |
| 121 | // Track Output, Flow, and Anti dependencies. |
| 122 | if (auto D = DI->depends(Src, Dst, true)) { |
| 123 | assert(D->isOrdered() && "Expected an output, flow or anti dep.")(static_cast <bool> (D->isOrdered() && "Expected an output, flow or anti dep." ) ? void (0) : __assert_fail ("D->isOrdered() && \"Expected an output, flow or anti dep.\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 123, __extension__ __PRETTY_FUNCTION__)); |
| 124 | DEBUG(StringRef DepType =do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; dbgs() << "Found " << DepType << " dependency between Src and Dst\n" << " Src:" << *Src << "\n Dst:" << * Dst << '\n'; } } while (false) |
| 125 | D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; dbgs() << "Found " << DepType << " dependency between Src and Dst\n" << " Src:" << *Src << "\n Dst:" << * Dst << '\n'; } } while (false) |
| 126 | dbgs() << "Found " << DepTypedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; dbgs() << "Found " << DepType << " dependency between Src and Dst\n" << " Src:" << *Src << "\n Dst:" << * Dst << '\n'; } } while (false) |
| 127 | << " dependency between Src and Dst\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; dbgs() << "Found " << DepType << " dependency between Src and Dst\n" << " Src:" << *Src << "\n Dst:" << * Dst << '\n'; } } while (false) |
| 128 | << " Src:" << *Src << "\n Dst:" << *Dst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output"; dbgs() << "Found " << DepType << " dependency between Src and Dst\n" << " Src:" << *Src << "\n Dst:" << * Dst << '\n'; } } while (false); |
| 129 | unsigned Levels = D->getLevels(); |
| 130 | char Direction; |
| 131 | for (unsigned II = 1; II <= Levels; ++II) { |
| 132 | const SCEV *Distance = D->getDistance(II); |
| 133 | const SCEVConstant *SCEVConst = |
| 134 | dyn_cast_or_null<SCEVConstant>(Distance); |
| 135 | if (SCEVConst) { |
| 136 | const ConstantInt *CI = SCEVConst->getValue(); |
| 137 | if (CI->isNegative()) |
| 138 | Direction = '<'; |
| 139 | else if (CI->isZero()) |
| 140 | Direction = '='; |
| 141 | else |
| 142 | Direction = '>'; |
| 143 | Dep.push_back(Direction); |
| 144 | } else if (D->isScalar(II)) { |
| 145 | Direction = 'S'; |
| 146 | Dep.push_back(Direction); |
| 147 | } else { |
| 148 | unsigned Dir = D->getDirection(II); |
| 149 | if (Dir == Dependence::DVEntry::LT || |
| 150 | Dir == Dependence::DVEntry::LE) |
| 151 | Direction = '<'; |
| 152 | else if (Dir == Dependence::DVEntry::GT || |
| 153 | Dir == Dependence::DVEntry::GE) |
| 154 | Direction = '>'; |
| 155 | else if (Dir == Dependence::DVEntry::EQ) |
| 156 | Direction = '='; |
| 157 | else |
| 158 | Direction = '*'; |
| 159 | Dep.push_back(Direction); |
| 160 | } |
| 161 | } |
| 162 | while (Dep.size() != Level) { |
| 163 | Dep.push_back('I'); |
| 164 | } |
| 165 | |
| 166 | DepMatrix.push_back(Dep); |
| 167 | if (DepMatrix.size() > MaxMemInstrCount) { |
| 168 | DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCountdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Cannot handle more than " << MaxMemInstrCount << " dependencies inside loop\n" ; } } while (false) |
| 169 | << " dependencies inside loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Cannot handle more than " << MaxMemInstrCount << " dependencies inside loop\n" ; } } while (false); |
| 170 | return false; |
| 171 | } |
| 172 | } |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | return true; |
| 177 | } |
| 178 | |
| 179 | // A loop is moved from index 'from' to an index 'to'. Update the Dependence |
| 180 | // matrix by exchanging the two columns. |
| 181 | static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx, |
| 182 | unsigned ToIndx) { |
| 183 | unsigned numRows = DepMatrix.size(); |
| 184 | for (unsigned i = 0; i < numRows; ++i) { |
| 185 | char TmpVal = DepMatrix[i][ToIndx]; |
| 186 | DepMatrix[i][ToIndx] = DepMatrix[i][FromIndx]; |
| 187 | DepMatrix[i][FromIndx] = TmpVal; |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | // Checks if outermost non '=','S'or'I' dependence in the dependence matrix is |
| 192 | // '>' |
| 193 | static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row, |
| 194 | unsigned Column) { |
| 195 | for (unsigned i = 0; i <= Column; ++i) { |
| 196 | if (DepMatrix[Row][i] == '<') |
| 197 | return false; |
| 198 | if (DepMatrix[Row][i] == '>') |
| 199 | return true; |
| 200 | } |
| 201 | // All dependencies were '=','S' or 'I' |
| 202 | return false; |
| 203 | } |
| 204 | |
| 205 | // Checks if no dependence exist in the dependency matrix in Row before Column. |
| 206 | static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row, |
| 207 | unsigned Column) { |
| 208 | for (unsigned i = 0; i < Column; ++i) { |
| 209 | if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' && |
| 210 | DepMatrix[Row][i] != 'I') |
| 211 | return false; |
| 212 | } |
| 213 | return true; |
| 214 | } |
| 215 | |
| 216 | static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row, |
| 217 | unsigned OuterLoopId, char InnerDep, |
| 218 | char OuterDep) { |
| 219 | if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId)) |
| 220 | return false; |
| 221 | |
| 222 | if (InnerDep == OuterDep) |
| 223 | return true; |
| 224 | |
| 225 | // It is legal to interchange if and only if after interchange no row has a |
| 226 | // '>' direction as the leftmost non-'='. |
| 227 | |
| 228 | if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I') |
| 229 | return true; |
| 230 | |
| 231 | if (InnerDep == '<') |
| 232 | return true; |
| 233 | |
| 234 | if (InnerDep == '>') { |
| 235 | // If OuterLoopId represents outermost loop then interchanging will make the |
| 236 | // 1st dependency as '>' |
| 237 | if (OuterLoopId == 0) |
| 238 | return false; |
| 239 | |
| 240 | // If all dependencies before OuterloopId are '=','S'or 'I'. Then |
| 241 | // interchanging will result in this row having an outermost non '=' |
| 242 | // dependency of '>' |
| 243 | if (!containsNoDependence(DepMatrix, Row, OuterLoopId)) |
| 244 | return true; |
| 245 | } |
| 246 | |
| 247 | return false; |
| 248 | } |
| 249 | |
| 250 | // Checks if it is legal to interchange 2 loops. |
| 251 | // [Theorem] A permutation of the loops in a perfect nest is legal if and only |
| 252 | // if the direction matrix, after the same permutation is applied to its |
| 253 | // columns, has no ">" direction as the leftmost non-"=" direction in any row. |
| 254 | static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix, |
| 255 | unsigned InnerLoopId, |
| 256 | unsigned OuterLoopId) { |
| 257 | unsigned NumRows = DepMatrix.size(); |
| 258 | // For each row check if it is valid to interchange. |
| 259 | for (unsigned Row = 0; Row < NumRows; ++Row) { |
| 260 | char InnerDep = DepMatrix[Row][InnerLoopId]; |
| 261 | char OuterDep = DepMatrix[Row][OuterLoopId]; |
| 262 | if (InnerDep == '*' || OuterDep == '*') |
| 263 | return false; |
| 264 | if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep)) |
| 265 | return false; |
| 266 | } |
| 267 | return true; |
| 268 | } |
| 269 | |
| 270 | static void populateWorklist(Loop &L, SmallVector<LoopVector, 8> &V) { |
| 271 | DEBUG(dbgs() << "Calling populateWorklist on Func: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: " << L.getHeader()->getParent()->getName() << " Loop: %" << L.getHeader()->getName() << '\n' ; } } while (false) |
| 272 | << L.getHeader()->getParent()->getName() << " Loop: %"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: " << L.getHeader()->getParent()->getName() << " Loop: %" << L.getHeader()->getName() << '\n' ; } } while (false) |
| 273 | << L.getHeader()->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: " << L.getHeader()->getParent()->getName() << " Loop: %" << L.getHeader()->getName() << '\n' ; } } while (false); |
| 274 | LoopVector LoopList; |
| 275 | Loop *CurrentLoop = &L; |
| 276 | const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops(); |
| 277 | while (!Vec->empty()) { |
| 278 | // The current loop has multiple subloops in it hence it is not tightly |
| 279 | // nested. |
| 280 | // Discard all loops above it added into Worklist. |
| 281 | if (Vec->size() != 1) { |
| 282 | LoopList.clear(); |
| 283 | return; |
| 284 | } |
| 285 | LoopList.push_back(CurrentLoop); |
| 286 | CurrentLoop = Vec->front(); |
| 287 | Vec = &CurrentLoop->getSubLoops(); |
| 288 | } |
| 289 | LoopList.push_back(CurrentLoop); |
| 290 | V.push_back(std::move(LoopList)); |
| 291 | } |
| 292 | |
| 293 | static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) { |
| 294 | PHINode *InnerIndexVar = L->getCanonicalInductionVariable(); |
| 295 | if (InnerIndexVar) |
| 296 | return InnerIndexVar; |
| 297 | if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr) |
| 298 | return nullptr; |
| 299 | for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { |
| 300 | PHINode *PhiVar = cast<PHINode>(I); |
| 301 | Type *PhiTy = PhiVar->getType(); |
| 302 | if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() && |
| 303 | !PhiTy->isPointerTy()) |
| 304 | return nullptr; |
| 305 | const SCEVAddRecExpr *AddRec = |
| 306 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar)); |
| 307 | if (!AddRec || !AddRec->isAffine()) |
| 308 | continue; |
| 309 | const SCEV *Step = AddRec->getStepRecurrence(*SE); |
| 310 | if (!isa<SCEVConstant>(Step)) |
| 311 | continue; |
| 312 | // Found the induction variable. |
| 313 | // FIXME: Handle loops with more than one induction variable. Note that, |
| 314 | // currently, legality makes sure we have only one induction variable. |
| 315 | return PhiVar; |
| 316 | } |
| 317 | return nullptr; |
| 318 | } |
| 319 | |
| 320 | namespace { |
| 321 | |
| 322 | /// LoopInterchangeLegality checks if it is legal to interchange the loop. |
| 323 | class LoopInterchangeLegality { |
| 324 | public: |
| 325 | LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE, |
| 326 | LoopInfo *LI, DominatorTree *DT, bool PreserveLCSSA, |
| 327 | OptimizationRemarkEmitter *ORE) |
| 328 | : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), |
| 329 | PreserveLCSSA(PreserveLCSSA), ORE(ORE) {} |
| 330 | |
| 331 | /// Check if the loops can be interchanged. |
| 332 | bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId, |
| 333 | CharMatrix &DepMatrix); |
| 334 | |
| 335 | /// Check if the loop structure is understood. We do not handle triangular |
| 336 | /// loops for now. |
| 337 | bool isLoopStructureUnderstood(PHINode *InnerInductionVar); |
| 338 | |
| 339 | bool currentLimitations(); |
| 340 | |
| 341 | bool hasInnerLoopReduction() { return InnerLoopHasReduction; } |
| 342 | |
| 343 | private: |
| 344 | bool tightlyNested(Loop *Outer, Loop *Inner); |
| 345 | bool containsUnsafeInstructionsInHeader(BasicBlock *BB); |
| 346 | bool areAllUsesReductions(Instruction *Ins, Loop *L); |
| 347 | bool containsUnsafeInstructionsInLatch(BasicBlock *BB); |
| 348 | bool findInductionAndReductions(Loop *L, |
| 349 | SmallVector<PHINode *, 8> &Inductions, |
| 350 | SmallVector<PHINode *, 8> &Reductions); |
| 351 | |
| 352 | Loop *OuterLoop; |
| 353 | Loop *InnerLoop; |
| 354 | |
| 355 | ScalarEvolution *SE; |
| 356 | LoopInfo *LI; |
| 357 | DominatorTree *DT; |
| 358 | bool PreserveLCSSA; |
| 359 | |
| 360 | /// Interface to emit optimization remarks. |
| 361 | OptimizationRemarkEmitter *ORE; |
| 362 | |
| 363 | bool InnerLoopHasReduction = false; |
| 364 | }; |
| 365 | |
| 366 | /// LoopInterchangeProfitability checks if it is profitable to interchange the |
| 367 | /// loop. |
| 368 | class LoopInterchangeProfitability { |
| 369 | public: |
| 370 | LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE, |
| 371 | OptimizationRemarkEmitter *ORE) |
| 372 | : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {} |
| 373 | |
| 374 | /// Check if the loop interchange is profitable. |
| 375 | bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId, |
| 376 | CharMatrix &DepMatrix); |
| 377 | |
| 378 | private: |
| 379 | int getInstrOrderCost(); |
| 380 | |
| 381 | Loop *OuterLoop; |
| 382 | Loop *InnerLoop; |
| 383 | |
| 384 | /// Scev analysis. |
| 385 | ScalarEvolution *SE; |
| 386 | |
| 387 | /// Interface to emit optimization remarks. |
| 388 | OptimizationRemarkEmitter *ORE; |
| 389 | }; |
| 390 | |
| 391 | /// LoopInterchangeTransform interchanges the loop. |
| 392 | class LoopInterchangeTransform { |
| 393 | public: |
| 394 | LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE, |
| 395 | LoopInfo *LI, DominatorTree *DT, |
| 396 | BasicBlock *LoopNestExit, |
| 397 | bool InnerLoopContainsReductions) |
| 398 | : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), |
| 399 | LoopExit(LoopNestExit), |
| 400 | InnerLoopHasReduction(InnerLoopContainsReductions) {} |
| 401 | |
| 402 | /// Interchange OuterLoop and InnerLoop. |
| 403 | bool transform(); |
| 404 | void restructureLoops(Loop *InnerLoop, Loop *OuterLoop); |
| 405 | void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop); |
| 406 | |
| 407 | private: |
| 408 | void splitInnerLoopLatch(Instruction *); |
| 409 | void splitInnerLoopHeader(); |
| 410 | bool adjustLoopLinks(); |
| 411 | void adjustLoopPreheaders(); |
| 412 | bool adjustLoopBranches(); |
| 413 | void updateIncomingBlock(BasicBlock *CurrBlock, BasicBlock *OldPred, |
| 414 | BasicBlock *NewPred); |
| 415 | |
| 416 | Loop *OuterLoop; |
| 417 | Loop *InnerLoop; |
| 418 | |
| 419 | /// Scev analysis. |
| 420 | ScalarEvolution *SE; |
| 421 | |
| 422 | LoopInfo *LI; |
| 423 | DominatorTree *DT; |
| 424 | BasicBlock *LoopExit; |
| 425 | bool InnerLoopHasReduction; |
| 426 | }; |
| 427 | |
| 428 | // Main LoopInterchange Pass. |
| 429 | struct LoopInterchange : public FunctionPass { |
| 430 | static char ID; |
| 431 | ScalarEvolution *SE = nullptr; |
| 432 | LoopInfo *LI = nullptr; |
| 433 | DependenceInfo *DI = nullptr; |
| 434 | DominatorTree *DT = nullptr; |
| 435 | bool PreserveLCSSA; |
| 436 | |
| 437 | /// Interface to emit optimization remarks. |
| 438 | OptimizationRemarkEmitter *ORE; |
| 439 | |
| 440 | LoopInterchange() : FunctionPass(ID) { |
| 441 | initializeLoopInterchangePass(*PassRegistry::getPassRegistry()); |
| 442 | } |
| 443 | |
| 444 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 445 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
| 446 | AU.addRequired<AAResultsWrapperPass>(); |
| 447 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 448 | AU.addRequired<LoopInfoWrapperPass>(); |
| 449 | AU.addRequired<DependenceAnalysisWrapperPass>(); |
| 450 | AU.addRequiredID(LoopSimplifyID); |
| 451 | AU.addRequiredID(LCSSAID); |
| 452 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
| 453 | |
| 454 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 455 | } |
| 456 | |
| 457 | bool runOnFunction(Function &F) override { |
| 458 | if (skipFunction(F)) |
| 459 | return false; |
| 460 | |
| 461 | SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
| 462 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 463 | DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI(); |
| 464 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 465 | ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); |
| 466 | PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); |
| 467 | |
| 468 | // Build up a worklist of loop pairs to analyze. |
| 469 | SmallVector<LoopVector, 8> Worklist; |
| 470 | |
| 471 | for (Loop *L : *LI) |
| 472 | populateWorklist(*L, Worklist); |
| 473 | |
| 474 | DEBUG(dbgs() << "Worklist size = " << Worklist.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Worklist size = " << Worklist.size() << "\n"; } } while (false); |
| 475 | bool Changed = true; |
| 476 | while (!Worklist.empty()) { |
| 477 | LoopVector LoopList = Worklist.pop_back_val(); |
| 478 | Changed = processLoopList(LoopList, F); |
| 479 | } |
| 480 | return Changed; |
| 481 | } |
| 482 | |
| 483 | bool isComputableLoopNest(LoopVector LoopList) { |
| 484 | for (Loop *L : LoopList) { |
| 485 | const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L); |
| 486 | if (ExitCountOuter == SE->getCouldNotCompute()) { |
| 487 | DEBUG(dbgs() << "Couldn't compute backedge count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Couldn't compute backedge count\n" ; } } while (false); |
| 488 | return false; |
| 489 | } |
| 490 | if (L->getNumBackEdges() != 1) { |
| 491 | DEBUG(dbgs() << "NumBackEdges is not equal to 1\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "NumBackEdges is not equal to 1\n" ; } } while (false); |
| 492 | return false; |
| 493 | } |
| 494 | if (!L->getExitingBlock()) { |
| 495 | DEBUG(dbgs() << "Loop doesn't have unique exit block\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loop doesn't have unique exit block\n" ; } } while (false); |
| 496 | return false; |
| 497 | } |
| 498 | } |
| 499 | return true; |
| 500 | } |
| 501 | |
| 502 | unsigned selectLoopForInterchange(const LoopVector &LoopList) { |
| 503 | // TODO: Add a better heuristic to select the loop to be interchanged based |
| 504 | // on the dependence matrix. Currently we select the innermost loop. |
| 505 | return LoopList.size() - 1; |
| 506 | } |
| 507 | |
| 508 | bool processLoopList(LoopVector LoopList, Function &F) { |
| 509 | bool Changed = false; |
| 510 | unsigned LoopNestDepth = LoopList.size(); |
| 511 | if (LoopNestDepth < 2) { |
| 512 | DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loop doesn't contain minimum nesting level.\n" ; } } while (false); |
| 513 | return false; |
| 514 | } |
| 515 | if (LoopNestDepth > MaxLoopNestDepth) { |
| 516 | DEBUG(dbgs() << "Cannot handle loops of depth greater than "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Cannot handle loops of depth greater than " << MaxLoopNestDepth << "\n"; } } while (false) |
| 517 | << MaxLoopNestDepth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Cannot handle loops of depth greater than " << MaxLoopNestDepth << "\n"; } } while (false); |
| 518 | return false; |
| 519 | } |
| 520 | if (!isComputableLoopNest(LoopList)) { |
| 521 | DEBUG(dbgs() << "Not valid loop candidate for interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Not valid loop candidate for interchange\n" ; } } while (false); |
| 522 | return false; |
| 523 | } |
| 524 | |
| 525 | DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Processing LoopList of size = " << LoopNestDepth << "\n"; } } while (false); |
| 526 | |
| 527 | CharMatrix DependencyMatrix; |
| 528 | Loop *OuterMostLoop = *(LoopList.begin()); |
| 529 | if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth, |
| 530 | OuterMostLoop, DI)) { |
| 531 | DEBUG(dbgs() << "Populating dependency matrix failed\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Populating dependency matrix failed\n" ; } } while (false); |
| 532 | return false; |
| 533 | } |
| 534 | #ifdef DUMP_DEP_MATRICIES |
| 535 | DEBUG(dbgs() << "Dependence before interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Dependence before interchange\n" ; } } while (false); |
| 536 | printDepMatrix(DependencyMatrix); |
| 537 | #endif |
| 538 | |
| 539 | BasicBlock *OuterMostLoopLatch = OuterMostLoop->getLoopLatch(); |
| 540 | BranchInst *OuterMostLoopLatchBI = |
| 541 | dyn_cast<BranchInst>(OuterMostLoopLatch->getTerminator()); |
| 542 | if (!OuterMostLoopLatchBI || OuterMostLoopLatchBI->getNumSuccessors() != 2) |
| 543 | return false; |
| 544 | |
| 545 | // Since we currently do not handle LCSSA PHI's any failure in loop |
| 546 | // condition will now branch to LoopNestExit. |
| 547 | // TODO: This should be removed once we handle LCSSA PHI nodes. |
| 548 | |
| 549 | // Get the Outermost loop exit. |
| 550 | BasicBlock *LoopNestExit; |
| 551 | if (OuterMostLoopLatchBI->getSuccessor(0) == OuterMostLoop->getHeader()) |
| 552 | LoopNestExit = OuterMostLoopLatchBI->getSuccessor(1); |
| 553 | else |
| 554 | LoopNestExit = OuterMostLoopLatchBI->getSuccessor(0); |
| 555 | |
| 556 | if (isa<PHINode>(LoopNestExit->begin())) { |
| 557 | DEBUG(dbgs() << "PHI Nodes in loop nest exit is not handled for now "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "PHI Nodes in loop nest exit is not handled for now " "since on failure all loops branch to loop nest exit.\n"; } } while (false) |
| 558 | "since on failure all loops branch to loop nest exit.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "PHI Nodes in loop nest exit is not handled for now " "since on failure all loops branch to loop nest exit.\n"; } } while (false); |
| 559 | return false; |
| 560 | } |
| 561 | |
| 562 | unsigned SelecLoopId = selectLoopForInterchange(LoopList); |
| 563 | // Move the selected loop outwards to the best possible position. |
| 564 | for (unsigned i = SelecLoopId; i > 0; i--) { |
| 565 | bool Interchanged = |
| 566 | processLoop(LoopList, i, i - 1, LoopNestExit, DependencyMatrix); |
| 567 | if (!Interchanged) |
| 568 | return Changed; |
| 569 | // Loops interchanged reflect the same in LoopList |
| 570 | std::swap(LoopList[i - 1], LoopList[i]); |
| 571 | |
| 572 | // Update the DependencyMatrix |
| 573 | interChangeDependencies(DependencyMatrix, i, i - 1); |
| 574 | #ifdef DUMP_DEP_MATRICIES |
| 575 | DEBUG(dbgs() << "Dependence after interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Dependence after interchange\n" ; } } while (false); |
| 576 | printDepMatrix(DependencyMatrix); |
| 577 | #endif |
| 578 | Changed |= Interchanged; |
| 579 | } |
| 580 | return Changed; |
| 581 | } |
| 582 | |
| 583 | bool processLoop(LoopVector LoopList, unsigned InnerLoopId, |
| 584 | unsigned OuterLoopId, BasicBlock *LoopNestExit, |
| 585 | std::vector<std::vector<char>> &DependencyMatrix) { |
| 586 | DEBUG(dbgs() << "Processing Inner Loop Id = " << InnerLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Processing Inner Loop Id = " << InnerLoopId << " and OuterLoopId = " << OuterLoopId << "\n"; } } while (false) |
| 587 | << " and OuterLoopId = " << OuterLoopId << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Processing Inner Loop Id = " << InnerLoopId << " and OuterLoopId = " << OuterLoopId << "\n"; } } while (false); |
| 588 | Loop *InnerLoop = LoopList[InnerLoopId]; |
| 589 | Loop *OuterLoop = LoopList[OuterLoopId]; |
| 590 | |
| 591 | LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, LI, DT, |
| 592 | PreserveLCSSA, ORE); |
| 593 | if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) { |
| 594 | DEBUG(dbgs() << "Not interchanging Loops. Cannot prove legality\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Not interchanging Loops. Cannot prove legality\n" ; } } while (false); |
| 595 | return false; |
| 596 | } |
| 597 | DEBUG(dbgs() << "Loops are legal to interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops are legal to interchange\n" ; } } while (false); |
| 598 | LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE); |
| 599 | if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) { |
| 600 | DEBUG(dbgs() << "Interchanging loops not profitable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Interchanging loops not profitable\n" ; } } while (false); |
| 601 | return false; |
| 602 | } |
| 603 | |
| 604 | ORE->emit([&]() { |
| 605 | return OptimizationRemark(DEBUG_TYPE"loop-interchange", "Interchanged", |
| 606 | InnerLoop->getStartLoc(), |
| 607 | InnerLoop->getHeader()) |
| 608 | << "Loop interchanged with enclosing loop."; |
| 609 | }); |
| 610 | |
| 611 | LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, |
| 612 | LoopNestExit, LIL.hasInnerLoopReduction()); |
| 613 | LIT.transform(); |
| 614 | DEBUG(dbgs() << "Loops interchanged\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops interchanged\n" ; } } while (false); |
| 615 | return true; |
| 616 | } |
| 617 | }; |
| 618 | |
| 619 | } // end anonymous namespace |
| 620 | |
| 621 | bool LoopInterchangeLegality::areAllUsesReductions(Instruction *Ins, Loop *L) { |
| 622 | return llvm::none_of(Ins->users(), [=](User *U) -> bool { |
| 623 | auto *UserIns = dyn_cast<PHINode>(U); |
| 624 | RecurrenceDescriptor RD; |
| 625 | return !UserIns || !RecurrenceDescriptor::isReductionPHI(UserIns, L, RD); |
| 626 | }); |
| 627 | } |
| 628 | |
| 629 | bool LoopInterchangeLegality::containsUnsafeInstructionsInHeader( |
| 630 | BasicBlock *BB) { |
| 631 | for (auto I = BB->begin(), E = BB->end(); I != E; ++I) { |
| 632 | // Load corresponding to reduction PHI's are safe while concluding if |
| 633 | // tightly nested. |
| 634 | if (LoadInst *L = dyn_cast<LoadInst>(I)) { |
| 635 | if (!areAllUsesReductions(L, InnerLoop)) |
| 636 | return true; |
| 637 | } else if (I->mayHaveSideEffects() || I->mayReadFromMemory()) |
| 638 | return true; |
| 639 | } |
| 640 | return false; |
| 641 | } |
| 642 | |
| 643 | bool LoopInterchangeLegality::containsUnsafeInstructionsInLatch( |
| 644 | BasicBlock *BB) { |
| 645 | for (auto I = BB->begin(), E = BB->end(); I != E; ++I) { |
| 646 | // Stores corresponding to reductions are safe while concluding if tightly |
| 647 | // nested. |
| 648 | if (StoreInst *L = dyn_cast<StoreInst>(I)) { |
| 649 | if (!isa<PHINode>(L->getOperand(0))) |
| 650 | return true; |
| 651 | } else if (I->mayHaveSideEffects() || I->mayReadFromMemory()) |
| 652 | return true; |
| 653 | } |
| 654 | return false; |
| 655 | } |
| 656 | |
| 657 | bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) { |
| 658 | BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); |
| 659 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 660 | BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); |
| 661 | |
| 662 | DEBUG(dbgs() << "Checking if loops are tightly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Checking if loops are tightly nested\n" ; } } while (false); |
| 663 | |
| 664 | // A perfectly nested loop will not have any branch in between the outer and |
| 665 | // inner block i.e. outer header will branch to either inner preheader and |
| 666 | // outerloop latch. |
| 667 | BranchInst *OuterLoopHeaderBI = |
| 668 | dyn_cast<BranchInst>(OuterLoopHeader->getTerminator()); |
| 669 | if (!OuterLoopHeaderBI) |
| 670 | return false; |
| 671 | |
| 672 | for (BasicBlock *Succ : OuterLoopHeaderBI->successors()) |
| 673 | if (Succ != InnerLoopPreHeader && Succ != OuterLoopLatch) |
| 674 | return false; |
| 675 | |
| 676 | DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Checking instructions in Loop header and Loop latch\n" ; } } while (false); |
| 677 | // We do not have any basic block in between now make sure the outer header |
| 678 | // and outer loop latch doesn't contain any unsafe instructions. |
| 679 | if (containsUnsafeInstructionsInHeader(OuterLoopHeader) || |
| 680 | containsUnsafeInstructionsInLatch(OuterLoopLatch)) |
| 681 | return false; |
| 682 | |
| 683 | DEBUG(dbgs() << "Loops are perfectly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops are perfectly nested\n" ; } } while (false); |
| 684 | // We have a perfect loop nest. |
| 685 | return true; |
| 686 | } |
| 687 | |
| 688 | bool LoopInterchangeLegality::isLoopStructureUnderstood( |
| 689 | PHINode *InnerInduction) { |
| 690 | unsigned Num = InnerInduction->getNumOperands(); |
| 691 | BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader(); |
| 692 | for (unsigned i = 0; i < Num; ++i) { |
| 693 | Value *Val = InnerInduction->getOperand(i); |
| 694 | if (isa<Constant>(Val)) |
| 695 | continue; |
| 696 | Instruction *I = dyn_cast<Instruction>(Val); |
| 697 | if (!I) |
| 698 | return false; |
| 699 | // TODO: Handle triangular loops. |
| 700 | // e.g. for(int i=0;i<N;i++) |
| 701 | // for(int j=i;j<N;j++) |
| 702 | unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i); |
| 703 | if (InnerInduction->getIncomingBlock(IncomBlockIndx) == |
| 704 | InnerLoopPreheader && |
| 705 | !OuterLoop->isLoopInvariant(I)) { |
| 706 | return false; |
| 707 | } |
| 708 | } |
| 709 | return true; |
| 710 | } |
| 711 | |
| 712 | bool LoopInterchangeLegality::findInductionAndReductions( |
| 713 | Loop *L, SmallVector<PHINode *, 8> &Inductions, |
| 714 | SmallVector<PHINode *, 8> &Reductions) { |
| 715 | if (!L->getLoopLatch() || !L->getLoopPredecessor()) |
| 716 | return false; |
| 717 | for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { |
| 718 | RecurrenceDescriptor RD; |
| 719 | InductionDescriptor ID; |
| 720 | PHINode *PHI = cast<PHINode>(I); |
| 721 | if (InductionDescriptor::isInductionPHI(PHI, L, SE, ID)) |
| 722 | Inductions.push_back(PHI); |
| 723 | else if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD)) |
| 724 | Reductions.push_back(PHI); |
| 725 | else { |
| 726 | DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed to recognize PHI as an induction or reduction.\n" ; } } while (false) |
| 727 | dbgs() << "Failed to recognize PHI as an induction or reduction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed to recognize PHI as an induction or reduction.\n" ; } } while (false); |
| 728 | return false; |
| 729 | } |
| 730 | } |
| 731 | return true; |
| 732 | } |
| 733 | |
| 734 | static bool containsSafePHI(BasicBlock *Block, bool isOuterLoopExitBlock) { |
| 735 | for (auto I = Block->begin(); isa<PHINode>(I); ++I) { |
| 736 | PHINode *PHI = cast<PHINode>(I); |
| 737 | // Reduction lcssa phi will have only 1 incoming block that from loop latch. |
| 738 | if (PHI->getNumIncomingValues() > 1) |
| 739 | return false; |
| 740 | Instruction *Ins = dyn_cast<Instruction>(PHI->getIncomingValue(0)); |
| 741 | if (!Ins) |
| 742 | return false; |
| 743 | // Incoming value for lcssa phi's in outer loop exit can only be inner loop |
| 744 | // exits lcssa phi else it would not be tightly nested. |
| 745 | if (!isa<PHINode>(Ins) && isOuterLoopExitBlock) |
| 746 | return false; |
| 747 | } |
| 748 | return true; |
| 749 | } |
| 750 | |
| 751 | static BasicBlock *getLoopLatchExitBlock(BasicBlock *LatchBlock, |
| 752 | BasicBlock *LoopHeader) { |
| 753 | if (BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator())) { |
| 754 | assert(BI->getNumSuccessors() == 2 &&(static_cast <bool> (BI->getNumSuccessors() == 2 && "Branch leaving loop latch must have 2 successors") ? void ( 0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Branch leaving loop latch must have 2 successors\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 755, __extension__ __PRETTY_FUNCTION__)) |
| 755 | "Branch leaving loop latch must have 2 successors")(static_cast <bool> (BI->getNumSuccessors() == 2 && "Branch leaving loop latch must have 2 successors") ? void ( 0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Branch leaving loop latch must have 2 successors\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 755, __extension__ __PRETTY_FUNCTION__)); |
| 756 | for (BasicBlock *Succ : BI->successors()) { |
| 757 | if (Succ == LoopHeader) |
| 758 | continue; |
| 759 | return Succ; |
| 760 | } |
| 761 | } |
| 762 | return nullptr; |
| 763 | } |
| 764 | |
| 765 | // This function indicates the current limitations in the transform as a result |
| 766 | // of which we do not proceed. |
| 767 | bool LoopInterchangeLegality::currentLimitations() { |
| 768 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 769 | BasicBlock *InnerLoopHeader = InnerLoop->getHeader(); |
| 770 | BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); |
| 771 | BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); |
| 772 | BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); |
| 773 | |
| 774 | PHINode *InnerInductionVar; |
| 775 | SmallVector<PHINode *, 8> Inductions; |
| 776 | SmallVector<PHINode *, 8> Reductions; |
| 777 | if (!findInductionAndReductions(InnerLoop, Inductions, Reductions)) { |
| 778 | DEBUG(dbgs() << "Only inner loops with induction or reduction PHI nodes "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Only inner loops with induction or reduction PHI nodes " << "are supported currently.\n"; } } while (false) |
| 779 | << "are supported currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Only inner loops with induction or reduction PHI nodes " << "are supported currently.\n"; } } while (false); |
| 780 | ORE->emit([&]() { |
| 781 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "UnsupportedPHIInner", |
| 782 | InnerLoop->getStartLoc(), |
| 783 | InnerLoop->getHeader()) |
| 784 | << "Only inner loops with induction or reduction PHI nodes can be" |
| 785 | " interchange currently."; |
| 786 | }); |
| 787 | return true; |
| 788 | } |
| 789 | |
| 790 | // TODO: Currently we handle only loops with 1 induction variable. |
| 791 | if (Inductions.size() != 1) { |
| 792 | DEBUG(dbgs() << "We currently only support loops with 1 induction variable."do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "We currently only support loops with 1 induction variable." << "Failed to interchange due to current limitation\n" ; } } while (false) |
| 793 | << "Failed to interchange due to current limitation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "We currently only support loops with 1 induction variable." << "Failed to interchange due to current limitation\n" ; } } while (false); |
| 794 | ORE->emit([&]() { |
| 795 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "MultiInductionInner", |
| 796 | InnerLoop->getStartLoc(), |
| 797 | InnerLoop->getHeader()) |
| 798 | << "Only inner loops with 1 induction variable can be " |
| 799 | "interchanged currently."; |
| 800 | }); |
| 801 | return true; |
| 802 | } |
| 803 | if (Reductions.size() > 0) |
| 804 | InnerLoopHasReduction = true; |
| 805 | |
| 806 | InnerInductionVar = Inductions.pop_back_val(); |
| 807 | Reductions.clear(); |
| 808 | if (!findInductionAndReductions(OuterLoop, Inductions, Reductions)) { |
| 809 | DEBUG(dbgs() << "Only outer loops with induction or reduction PHI nodes "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Only outer loops with induction or reduction PHI nodes " << "are supported currently.\n"; } } while (false) |
| 810 | << "are supported currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Only outer loops with induction or reduction PHI nodes " << "are supported currently.\n"; } } while (false); |
| 811 | ORE->emit([&]() { |
| 812 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "UnsupportedPHIOuter", |
| 813 | OuterLoop->getStartLoc(), |
| 814 | OuterLoop->getHeader()) |
| 815 | << "Only outer loops with induction or reduction PHI nodes can be" |
| 816 | " interchanged currently."; |
| 817 | }); |
| 818 | return true; |
| 819 | } |
| 820 | |
| 821 | // Outer loop cannot have reduction because then loops will not be tightly |
| 822 | // nested. |
| 823 | if (!Reductions.empty()) { |
| 824 | DEBUG(dbgs() << "Outer loops with reductions are not supported "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Outer loops with reductions are not supported " << "currently.\n"; } } while (false) |
| 825 | << "currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Outer loops with reductions are not supported " << "currently.\n"; } } while (false); |
| 826 | ORE->emit([&]() { |
| 827 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "ReductionsOuter", |
| 828 | OuterLoop->getStartLoc(), |
| 829 | OuterLoop->getHeader()) |
| 830 | << "Outer loops with reductions cannot be interchangeed " |
| 831 | "currently."; |
| 832 | }); |
| 833 | return true; |
| 834 | } |
| 835 | // TODO: Currently we handle only loops with 1 induction variable. |
| 836 | if (Inductions.size() != 1) { |
| 837 | DEBUG(dbgs() << "Loops with more than 1 induction variables are not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops with more than 1 induction variables are not " << "supported currently.\n"; } } while (false) |
| 838 | << "supported currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops with more than 1 induction variables are not " << "supported currently.\n"; } } while (false); |
| 839 | ORE->emit([&]() { |
| 840 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "MultiIndutionOuter", |
| 841 | OuterLoop->getStartLoc(), |
| 842 | OuterLoop->getHeader()) |
| 843 | << "Only outer loops with 1 induction variable can be " |
| 844 | "interchanged currently."; |
| 845 | }); |
| 846 | return true; |
| 847 | } |
| 848 | |
| 849 | // TODO: Triangular loops are not handled for now. |
| 850 | if (!isLoopStructureUnderstood(InnerInductionVar)) { |
| 851 | DEBUG(dbgs() << "Loop structure not understood by pass\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loop structure not understood by pass\n" ; } } while (false); |
| 852 | ORE->emit([&]() { |
| 853 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "UnsupportedStructureInner", |
| 854 | InnerLoop->getStartLoc(), |
| 855 | InnerLoop->getHeader()) |
| 856 | << "Inner loop structure not understood currently."; |
| 857 | }); |
| 858 | return true; |
| 859 | } |
| 860 | |
| 861 | // TODO: We only handle LCSSA PHI's corresponding to reduction for now. |
| 862 | BasicBlock *LoopExitBlock = |
| 863 | getLoopLatchExitBlock(OuterLoopLatch, OuterLoopHeader); |
| 864 | if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, true)) { |
| 865 | DEBUG(dbgs() << "Can only handle LCSSA PHIs in outer loops currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Can only handle LCSSA PHIs in outer loops currently.\n" ; } } while (false); |
| 866 | ORE->emit([&]() { |
| 867 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "NoLCSSAPHIOuter", |
| 868 | OuterLoop->getStartLoc(), |
| 869 | OuterLoop->getHeader()) |
| 870 | << "Only outer loops with LCSSA PHIs can be interchange " |
| 871 | "currently."; |
| 872 | }); |
| 873 | return true; |
| 874 | } |
| 875 | |
| 876 | LoopExitBlock = getLoopLatchExitBlock(InnerLoopLatch, InnerLoopHeader); |
| 877 | if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, false)) { |
| 878 | DEBUG(dbgs() << "Can only handle LCSSA PHIs in inner loops currently.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Can only handle LCSSA PHIs in inner loops currently.\n" ; } } while (false); |
| 879 | ORE->emit([&]() { |
| 880 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "NoLCSSAPHIOuterInner", |
| 881 | InnerLoop->getStartLoc(), |
| 882 | InnerLoop->getHeader()) |
| 883 | << "Only inner loops with LCSSA PHIs can be interchange " |
| 884 | "currently."; |
| 885 | }); |
| 886 | return true; |
| 887 | } |
| 888 | |
| 889 | // TODO: Current limitation: Since we split the inner loop latch at the point |
| 890 | // were induction variable is incremented (induction.next); We cannot have |
| 891 | // more than 1 user of induction.next since it would result in broken code |
| 892 | // after split. |
| 893 | // e.g. |
| 894 | // for(i=0;i<N;i++) { |
| 895 | // for(j = 0;j<M;j++) { |
| 896 | // A[j+1][i+2] = A[j][i]+k; |
| 897 | // } |
| 898 | // } |
| 899 | Instruction *InnerIndexVarInc = nullptr; |
| 900 | if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader) |
| 901 | InnerIndexVarInc = |
| 902 | dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1)); |
| 903 | else |
| 904 | InnerIndexVarInc = |
| 905 | dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0)); |
| 906 | |
| 907 | if (!InnerIndexVarInc) { |
| 908 | DEBUG(dbgs() << "Did not find an instruction to increment the induction "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Did not find an instruction to increment the induction " << "variable.\n"; } } while (false) |
| 909 | << "variable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Did not find an instruction to increment the induction " << "variable.\n"; } } while (false); |
| 910 | ORE->emit([&]() { |
| 911 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "NoIncrementInInner", |
| 912 | InnerLoop->getStartLoc(), |
| 913 | InnerLoop->getHeader()) |
| 914 | << "The inner loop does not increment the induction variable."; |
| 915 | }); |
| 916 | return true; |
| 917 | } |
| 918 | |
| 919 | // Since we split the inner loop latch on this induction variable. Make sure |
| 920 | // we do not have any instruction between the induction variable and branch |
| 921 | // instruction. |
| 922 | |
| 923 | bool FoundInduction = false; |
| 924 | for (const Instruction &I : llvm::reverse(*InnerLoopLatch)) { |
| 925 | if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I) || |
| 926 | isa<ZExtInst>(I)) |
| 927 | continue; |
| 928 | |
| 929 | // We found an instruction. If this is not induction variable then it is not |
| 930 | // safe to split this loop latch. |
| 931 | if (!I.isIdenticalTo(InnerIndexVarInc)) { |
| 932 | DEBUG(dbgs() << "Found unsupported instructions between induction "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Found unsupported instructions between induction " << "variable increment and branch.\n"; } } while (false ) |
| 933 | << "variable increment and branch.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Found unsupported instructions between induction " << "variable increment and branch.\n"; } } while (false ); |
| 934 | ORE->emit([&]() { |
| 935 | return OptimizationRemarkMissed( |
| 936 | DEBUG_TYPE"loop-interchange", "UnsupportedInsBetweenInduction", |
| 937 | InnerLoop->getStartLoc(), InnerLoop->getHeader()) |
| 938 | << "Found unsupported instruction between induction variable " |
| 939 | "increment and branch."; |
| 940 | }); |
| 941 | return true; |
| 942 | } |
| 943 | |
| 944 | FoundInduction = true; |
| 945 | break; |
| 946 | } |
| 947 | // The loop latch ended and we didn't find the induction variable return as |
| 948 | // current limitation. |
| 949 | if (!FoundInduction) { |
| 950 | DEBUG(dbgs() << "Did not find the induction variable.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Did not find the induction variable.\n" ; } } while (false); |
| 951 | ORE->emit([&]() { |
| 952 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "NoIndutionVariable", |
| 953 | InnerLoop->getStartLoc(), |
| 954 | InnerLoop->getHeader()) |
| 955 | << "Did not find the induction variable."; |
| 956 | }); |
| 957 | return true; |
| 958 | } |
| 959 | return false; |
| 960 | } |
| 961 | |
| 962 | bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId, |
| 963 | unsigned OuterLoopId, |
| 964 | CharMatrix &DepMatrix) { |
| 965 | if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) { |
| 966 | DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId << " and OuterLoopId = " << OuterLoopId << " due to dependence\n"; } } while (false ) |
| 967 | << " and OuterLoopId = " << OuterLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId << " and OuterLoopId = " << OuterLoopId << " due to dependence\n"; } } while (false ) |
| 968 | << " due to dependence\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId << " and OuterLoopId = " << OuterLoopId << " due to dependence\n"; } } while (false ); |
| 969 | ORE->emit([&]() { |
| 970 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "Dependence", |
| 971 | InnerLoop->getStartLoc(), |
| 972 | InnerLoop->getHeader()) |
| 973 | << "Cannot interchange loops due to dependences."; |
| 974 | }); |
| 975 | return false; |
| 976 | } |
| 977 | |
| 978 | // Check if outer and inner loop contain legal instructions only. |
| 979 | for (auto *BB : OuterLoop->blocks()) |
| 980 | for (Instruction &I : *BB) |
| 981 | if (CallInst *CI = dyn_cast<CallInst>(&I)) { |
| 982 | // readnone functions do not prevent interchanging. |
| 983 | if (CI->doesNotReadMemory()) |
| 984 | continue; |
| 985 | DEBUG(dbgs() << "Loops with call instructions cannot be interchanged "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops with call instructions cannot be interchanged " << "safely."; } } while (false) |
| 986 | << "safely.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops with call instructions cannot be interchanged " << "safely."; } } while (false); |
| 987 | return false; |
| 988 | } |
| 989 | |
| 990 | // Create unique Preheaders if we already do not have one. |
| 991 | BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader(); |
| 992 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 993 | |
| 994 | // Create a unique outer preheader - |
| 995 | // 1) If OuterLoop preheader is not present. |
| 996 | // 2) If OuterLoop Preheader is same as OuterLoop Header |
| 997 | // 3) If OuterLoop Preheader is same as Header of the previous loop. |
| 998 | // 4) If OuterLoop Preheader is Entry node. |
| 999 | if (!OuterLoopPreHeader || OuterLoopPreHeader == OuterLoop->getHeader() || |
| 1000 | isa<PHINode>(OuterLoopPreHeader->begin()) || |
| 1001 | !OuterLoopPreHeader->getUniquePredecessor()) { |
| 1002 | OuterLoopPreHeader = |
Value stored to 'OuterLoopPreHeader' is never read | |
| 1003 | InsertPreheaderForLoop(OuterLoop, DT, LI, PreserveLCSSA); |
| 1004 | } |
| 1005 | |
| 1006 | if (!InnerLoopPreHeader || InnerLoopPreHeader == InnerLoop->getHeader() || |
| 1007 | InnerLoopPreHeader == OuterLoop->getHeader()) { |
| 1008 | InnerLoopPreHeader = |
| 1009 | InsertPreheaderForLoop(InnerLoop, DT, LI, PreserveLCSSA); |
| 1010 | } |
| 1011 | |
| 1012 | // TODO: The loops could not be interchanged due to current limitations in the |
| 1013 | // transform module. |
| 1014 | if (currentLimitations()) { |
| 1015 | DEBUG(dbgs() << "Not legal because of current transform limitation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Not legal because of current transform limitation\n" ; } } while (false); |
| 1016 | return false; |
| 1017 | } |
| 1018 | |
| 1019 | // Check if the loops are tightly nested. |
| 1020 | if (!tightlyNested(OuterLoop, InnerLoop)) { |
| 1021 | DEBUG(dbgs() << "Loops not tightly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Loops not tightly nested\n" ; } } while (false); |
| 1022 | ORE->emit([&]() { |
| 1023 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "NotTightlyNested", |
| 1024 | InnerLoop->getStartLoc(), |
| 1025 | InnerLoop->getHeader()) |
| 1026 | << "Cannot interchange loops because they are not tightly " |
| 1027 | "nested."; |
| 1028 | }); |
| 1029 | return false; |
| 1030 | } |
| 1031 | |
| 1032 | return true; |
| 1033 | } |
| 1034 | |
| 1035 | int LoopInterchangeProfitability::getInstrOrderCost() { |
| 1036 | unsigned GoodOrder, BadOrder; |
| 1037 | BadOrder = GoodOrder = 0; |
| 1038 | for (BasicBlock *BB : InnerLoop->blocks()) { |
| 1039 | for (Instruction &Ins : *BB) { |
| 1040 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) { |
| 1041 | unsigned NumOp = GEP->getNumOperands(); |
| 1042 | bool FoundInnerInduction = false; |
| 1043 | bool FoundOuterInduction = false; |
| 1044 | for (unsigned i = 0; i < NumOp; ++i) { |
| 1045 | const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i)); |
| 1046 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal); |
| 1047 | if (!AR) |
| 1048 | continue; |
| 1049 | |
| 1050 | // If we find the inner induction after an outer induction e.g. |
| 1051 | // for(int i=0;i<N;i++) |
| 1052 | // for(int j=0;j<N;j++) |
| 1053 | // A[i][j] = A[i-1][j-1]+k; |
| 1054 | // then it is a good order. |
| 1055 | if (AR->getLoop() == InnerLoop) { |
| 1056 | // We found an InnerLoop induction after OuterLoop induction. It is |
| 1057 | // a good order. |
| 1058 | FoundInnerInduction = true; |
| 1059 | if (FoundOuterInduction) { |
| 1060 | GoodOrder++; |
| 1061 | break; |
| 1062 | } |
| 1063 | } |
| 1064 | // If we find the outer induction after an inner induction e.g. |
| 1065 | // for(int i=0;i<N;i++) |
| 1066 | // for(int j=0;j<N;j++) |
| 1067 | // A[j][i] = A[j-1][i-1]+k; |
| 1068 | // then it is a bad order. |
| 1069 | if (AR->getLoop() == OuterLoop) { |
| 1070 | // We found an OuterLoop induction after InnerLoop induction. It is |
| 1071 | // a bad order. |
| 1072 | FoundOuterInduction = true; |
| 1073 | if (FoundInnerInduction) { |
| 1074 | BadOrder++; |
| 1075 | break; |
| 1076 | } |
| 1077 | } |
| 1078 | } |
| 1079 | } |
| 1080 | } |
| 1081 | } |
| 1082 | return GoodOrder - BadOrder; |
| 1083 | } |
| 1084 | |
| 1085 | static bool isProfitableForVectorization(unsigned InnerLoopId, |
| 1086 | unsigned OuterLoopId, |
| 1087 | CharMatrix &DepMatrix) { |
| 1088 | // TODO: Improve this heuristic to catch more cases. |
| 1089 | // If the inner loop is loop independent or doesn't carry any dependency it is |
| 1090 | // profitable to move this to outer position. |
| 1091 | for (auto &Row : DepMatrix) { |
| 1092 | if (Row[InnerLoopId] != 'S' && Row[InnerLoopId] != 'I') |
| 1093 | return false; |
| 1094 | // TODO: We need to improve this heuristic. |
| 1095 | if (Row[OuterLoopId] != '=') |
| 1096 | return false; |
| 1097 | } |
| 1098 | // If outer loop has dependence and inner loop is loop independent then it is |
| 1099 | // profitable to interchange to enable parallelism. |
| 1100 | return true; |
| 1101 | } |
| 1102 | |
| 1103 | bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId, |
| 1104 | unsigned OuterLoopId, |
| 1105 | CharMatrix &DepMatrix) { |
| 1106 | // TODO: Add better profitability checks. |
| 1107 | // e.g |
| 1108 | // 1) Construct dependency matrix and move the one with no loop carried dep |
| 1109 | // inside to enable vectorization. |
| 1110 | |
| 1111 | // This is rough cost estimation algorithm. It counts the good and bad order |
| 1112 | // of induction variables in the instruction and allows reordering if number |
| 1113 | // of bad orders is more than good. |
| 1114 | int Cost = getInstrOrderCost(); |
| 1115 | DEBUG(dbgs() << "Cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Cost = " << Cost << "\n"; } } while (false); |
| 1116 | if (Cost < -LoopInterchangeCostThreshold) |
| 1117 | return true; |
| 1118 | |
| 1119 | // It is not profitable as per current cache profitability model. But check if |
| 1120 | // we can move this loop outside to improve parallelism. |
| 1121 | if (isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix)) |
| 1122 | return true; |
| 1123 | |
| 1124 | ORE->emit([&]() { |
| 1125 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-interchange", "InterchangeNotProfitable", |
| 1126 | InnerLoop->getStartLoc(), |
| 1127 | InnerLoop->getHeader()) |
| 1128 | << "Interchanging loops is too costly (cost=" |
| 1129 | << ore::NV("Cost", Cost) << ", threshold=" |
| 1130 | << ore::NV("Threshold", LoopInterchangeCostThreshold) |
| 1131 | << ") and it does not improve parallelism."; |
| 1132 | }); |
| 1133 | return false; |
| 1134 | } |
| 1135 | |
| 1136 | void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop, |
| 1137 | Loop *InnerLoop) { |
| 1138 | for (Loop::iterator I = OuterLoop->begin(), E = OuterLoop->end(); I != E; |
| 1139 | ++I) { |
| 1140 | if (*I == InnerLoop) { |
| 1141 | OuterLoop->removeChildLoop(I); |
| 1142 | return; |
| 1143 | } |
| 1144 | } |
| 1145 | llvm_unreachable("Couldn't find loop")::llvm::llvm_unreachable_internal("Couldn't find loop", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 1145); |
| 1146 | } |
| 1147 | |
| 1148 | void LoopInterchangeTransform::restructureLoops(Loop *InnerLoop, |
| 1149 | Loop *OuterLoop) { |
| 1150 | Loop *OuterLoopParent = OuterLoop->getParentLoop(); |
| 1151 | if (OuterLoopParent) { |
| 1152 | // Remove the loop from its parent loop. |
| 1153 | removeChildLoop(OuterLoopParent, OuterLoop); |
| 1154 | removeChildLoop(OuterLoop, InnerLoop); |
| 1155 | OuterLoopParent->addChildLoop(InnerLoop); |
| 1156 | } else { |
| 1157 | removeChildLoop(OuterLoop, InnerLoop); |
| 1158 | LI->changeTopLevelLoop(OuterLoop, InnerLoop); |
| 1159 | } |
| 1160 | |
| 1161 | while (!InnerLoop->empty()) |
| 1162 | OuterLoop->addChildLoop(InnerLoop->removeChildLoop(InnerLoop->begin())); |
| 1163 | |
| 1164 | InnerLoop->addChildLoop(OuterLoop); |
| 1165 | } |
| 1166 | |
| 1167 | bool LoopInterchangeTransform::transform() { |
| 1168 | bool Transformed = false; |
| 1169 | Instruction *InnerIndexVar; |
| 1170 | |
| 1171 | if (InnerLoop->getSubLoops().empty()) { |
| 1172 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 1173 | DEBUG(dbgs() << "Calling Split Inner Loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Calling Split Inner Loop\n" ; } } while (false); |
| 1174 | PHINode *InductionPHI = getInductionVariable(InnerLoop, SE); |
| 1175 | if (!InductionPHI) { |
| 1176 | DEBUG(dbgs() << "Failed to find the point to split loop latch \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Failed to find the point to split loop latch \n" ; } } while (false); |
| 1177 | return false; |
| 1178 | } |
| 1179 | |
| 1180 | if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader) |
| 1181 | InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1)); |
| 1182 | else |
| 1183 | InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0)); |
| 1184 | |
| 1185 | // Ensure that InductionPHI is the first Phi node as required by |
| 1186 | // splitInnerLoopHeader |
| 1187 | if (&InductionPHI->getParent()->front() != InductionPHI) |
| 1188 | InductionPHI->moveBefore(&InductionPHI->getParent()->front()); |
| 1189 | |
| 1190 | // Split at the place were the induction variable is |
| 1191 | // incremented/decremented. |
| 1192 | // TODO: This splitting logic may not work always. Fix this. |
| 1193 | splitInnerLoopLatch(InnerIndexVar); |
| 1194 | DEBUG(dbgs() << "splitInnerLoopLatch done\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "splitInnerLoopLatch done\n" ; } } while (false); |
| 1195 | |
| 1196 | // Splits the inner loops phi nodes out into a separate basic block. |
| 1197 | splitInnerLoopHeader(); |
| 1198 | DEBUG(dbgs() << "splitInnerLoopHeader done\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "splitInnerLoopHeader done\n" ; } } while (false); |
| 1199 | } |
| 1200 | |
| 1201 | Transformed |= adjustLoopLinks(); |
| 1202 | if (!Transformed) { |
| 1203 | DEBUG(dbgs() << "adjustLoopLinks failed\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "adjustLoopLinks failed\n" ; } } while (false); |
| 1204 | return false; |
| 1205 | } |
| 1206 | |
| 1207 | restructureLoops(InnerLoop, OuterLoop); |
| 1208 | return true; |
| 1209 | } |
| 1210 | |
| 1211 | void LoopInterchangeTransform::splitInnerLoopLatch(Instruction *Inc) { |
| 1212 | BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); |
| 1213 | BasicBlock *InnerLoopLatchPred = InnerLoopLatch; |
| 1214 | InnerLoopLatch = SplitBlock(InnerLoopLatchPred, Inc, DT, LI); |
| 1215 | } |
| 1216 | |
| 1217 | void LoopInterchangeTransform::splitInnerLoopHeader() { |
| 1218 | // Split the inner loop header out. Here make sure that the reduction PHI's |
| 1219 | // stay in the innerloop body. |
| 1220 | BasicBlock *InnerLoopHeader = InnerLoop->getHeader(); |
| 1221 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 1222 | SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI); |
| 1223 | if (InnerLoopHasReduction) { |
| 1224 | // Adjust Reduction PHI's in the block. The induction PHI must be the first |
| 1225 | // PHI in InnerLoopHeader for this to work. |
| 1226 | SmallVector<PHINode *, 8> PHIVec; |
| 1227 | for (auto I = std::next(InnerLoopHeader->begin()); isa<PHINode>(I); ++I) { |
| 1228 | PHINode *PHI = dyn_cast<PHINode>(I); |
| 1229 | Value *V = PHI->getIncomingValueForBlock(InnerLoopPreHeader); |
| 1230 | PHI->replaceAllUsesWith(V); |
| 1231 | PHIVec.push_back((PHI)); |
| 1232 | } |
| 1233 | for (PHINode *P : PHIVec) { |
| 1234 | P->eraseFromParent(); |
| 1235 | } |
| 1236 | } |
| 1237 | |
| 1238 | DEBUG(dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & " "InnerLoopHeader\n"; } } while (false) |
| 1239 | "InnerLoopHeader\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & " "InnerLoopHeader\n"; } } while (false); |
| 1240 | } |
| 1241 | |
| 1242 | /// \brief Move all instructions except the terminator from FromBB right before |
| 1243 | /// InsertBefore |
| 1244 | static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) { |
| 1245 | auto &ToList = InsertBefore->getParent()->getInstList(); |
| 1246 | auto &FromList = FromBB->getInstList(); |
| 1247 | |
| 1248 | ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(), |
| 1249 | FromBB->getTerminator()->getIterator()); |
| 1250 | } |
| 1251 | |
| 1252 | void LoopInterchangeTransform::updateIncomingBlock(BasicBlock *CurrBlock, |
| 1253 | BasicBlock *OldPred, |
| 1254 | BasicBlock *NewPred) { |
| 1255 | for (auto I = CurrBlock->begin(); isa<PHINode>(I); ++I) { |
| 1256 | PHINode *PHI = cast<PHINode>(I); |
| 1257 | unsigned Num = PHI->getNumIncomingValues(); |
| 1258 | for (unsigned i = 0; i < Num; ++i) { |
| 1259 | if (PHI->getIncomingBlock(i) == OldPred) |
| 1260 | PHI->setIncomingBlock(i, NewPred); |
| 1261 | } |
| 1262 | } |
| 1263 | } |
| 1264 | |
| 1265 | /// \brief Update BI to jump to NewBB instead of OldBB. Records updates to |
| 1266 | /// the dominator tree in DTUpdates, if DT should be preserved. |
| 1267 | static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB, |
| 1268 | BasicBlock *NewBB, |
| 1269 | std::vector<DominatorTree::UpdateType> &DTUpdates) { |
| 1270 | assert(llvm::count_if(BI->successors(),(static_cast <bool> (llvm::count_if(BI->successors() , [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && "BI must jump to OldBB at most once.") ? void (0) : __assert_fail ("llvm::count_if(BI->successors(), [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && \"BI must jump to OldBB at most once.\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 1272, __extension__ __PRETTY_FUNCTION__)) |
| 1271 | [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 &&(static_cast <bool> (llvm::count_if(BI->successors() , [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && "BI must jump to OldBB at most once.") ? void (0) : __assert_fail ("llvm::count_if(BI->successors(), [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && \"BI must jump to OldBB at most once.\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 1272, __extension__ __PRETTY_FUNCTION__)) |
| 1272 | "BI must jump to OldBB at most once.")(static_cast <bool> (llvm::count_if(BI->successors() , [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && "BI must jump to OldBB at most once.") ? void (0) : __assert_fail ("llvm::count_if(BI->successors(), [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 && \"BI must jump to OldBB at most once.\"" , "/build/llvm-toolchain-snapshot-7~svn326551/lib/Transforms/Scalar/LoopInterchange.cpp" , 1272, __extension__ __PRETTY_FUNCTION__)); |
| 1273 | for (unsigned i = 0, e = BI->getNumSuccessors(); i < e; ++i) { |
| 1274 | if (BI->getSuccessor(i) == OldBB) { |
| 1275 | BI->setSuccessor(i, NewBB); |
| 1276 | |
| 1277 | DTUpdates.push_back( |
| 1278 | {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB}); |
| 1279 | DTUpdates.push_back( |
| 1280 | {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB}); |
| 1281 | break; |
| 1282 | } |
| 1283 | } |
| 1284 | } |
| 1285 | |
| 1286 | bool LoopInterchangeTransform::adjustLoopBranches() { |
| 1287 | DEBUG(dbgs() << "adjustLoopBranches called\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-interchange")) { dbgs() << "adjustLoopBranches called\n" ; } } while (false); |
| 1288 | std::vector<DominatorTree::UpdateType> DTUpdates; |
| 1289 | |
| 1290 | // Adjust the loop preheader |
| 1291 | BasicBlock *InnerLoopHeader = InnerLoop->getHeader(); |
| 1292 | BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); |
| 1293 | BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); |
| 1294 | BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); |
| 1295 | BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader(); |
| 1296 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 1297 | BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor(); |
| 1298 | BasicBlock *InnerLoopLatchPredecessor = |
| 1299 | InnerLoopLatch->getUniquePredecessor(); |
| 1300 | BasicBlock *InnerLoopLatchSuccessor; |
| 1301 | BasicBlock *OuterLoopLatchSuccessor; |
| 1302 | |
| 1303 | BranchInst *OuterLoopLatchBI = |
| 1304 | dyn_cast<BranchInst>(OuterLoopLatch->getTerminator()); |
| 1305 | BranchInst *InnerLoopLatchBI = |
| 1306 | dyn_cast<BranchInst>(InnerLoopLatch->getTerminator()); |
| 1307 | BranchInst *OuterLoopHeaderBI = |
| 1308 | dyn_cast<BranchInst>(OuterLoopHeader->getTerminator()); |
| 1309 | BranchInst *InnerLoopHeaderBI = |
| 1310 | dyn_cast<BranchInst>(InnerLoopHeader->getTerminator()); |
| 1311 | |
| 1312 | if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor || |
| 1313 | !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI || |
| 1314 | !InnerLoopHeaderBI) |
| 1315 | return false; |
| 1316 | |
| 1317 | BranchInst *InnerLoopLatchPredecessorBI = |
| 1318 | dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator()); |
| 1319 | BranchInst *OuterLoopPredecessorBI = |
| 1320 | dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator()); |
| 1321 | |
| 1322 | if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI) |
| 1323 | return false; |
| 1324 | BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor(); |
| 1325 | if (!InnerLoopHeaderSuccessor) |
| 1326 | return false; |
| 1327 | |
| 1328 | // Adjust Loop Preheader and headers |
| 1329 | updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader, |
| 1330 | InnerLoopPreHeader, DTUpdates); |
| 1331 | updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, LoopExit, DTUpdates); |
| 1332 | updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader, |
| 1333 | InnerLoopHeaderSuccessor, DTUpdates); |
| 1334 | |
| 1335 | // Adjust reduction PHI's now that the incoming block has changed. |
| 1336 | updateIncomingBlock(InnerLoopHeaderSuccessor, InnerLoopHeader, |
| 1337 | OuterLoopHeader); |
| 1338 | |
| 1339 | updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor, |
| 1340 | OuterLoopPreHeader, DTUpdates); |
| 1341 | |
| 1342 | // -------------Adjust loop latches----------- |
| 1343 | if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader) |
| 1344 | InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1); |
| 1345 | else |
| 1346 | InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0); |
| 1347 | |
| 1348 | updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch, |
| 1349 | InnerLoopLatchSuccessor, DTUpdates); |
| 1350 | |
| 1351 | // Adjust PHI nodes in InnerLoopLatchSuccessor. Update all uses of PHI with |
| 1352 | // the value and remove this PHI node from inner loop. |
| 1353 | SmallVector<PHINode *, 8> LcssaVec; |
| 1354 | for (auto I = InnerLoopLatchSuccessor->begin(); isa<PHINode>(I); ++I) { |
| 1355 | PHINode *LcssaPhi = cast<PHINode>(I); |
| 1356 | LcssaVec.push_back(LcssaPhi); |
| 1357 | } |
| 1358 | for (PHINode *P : LcssaVec) { |
| 1359 | Value *Incoming = P->getIncomingValueForBlock(InnerLoopLatch); |
| 1360 | P->replaceAllUsesWith(Incoming); |
| 1361 | P->eraseFromParent(); |
| 1362 | } |
| 1363 | |
| 1364 | if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader) |
| 1365 | OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1); |
| 1366 | else |
| 1367 | OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0); |
| 1368 | |
| 1369 | updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor, |
| 1370 | OuterLoopLatchSuccessor, DTUpdates); |
| 1371 | updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch, |
| 1372 | DTUpdates); |
| 1373 | |
| 1374 | updateIncomingBlock(OuterLoopLatchSuccessor, OuterLoopLatch, InnerLoopLatch); |
| 1375 | |
| 1376 | DT->applyUpdates(DTUpdates); |
| 1377 | return true; |
| 1378 | } |
| 1379 | |
| 1380 | void LoopInterchangeTransform::adjustLoopPreheaders() { |
| 1381 | // We have interchanged the preheaders so we need to interchange the data in |
| 1382 | // the preheader as well. |
| 1383 | // This is because the content of inner preheader was previously executed |
| 1384 | // inside the outer loop. |
| 1385 | BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader(); |
| 1386 | BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader(); |
| 1387 | BasicBlock *OuterLoopHeader = OuterLoop->getHeader(); |
| 1388 | BranchInst *InnerTermBI = |
| 1389 | cast<BranchInst>(InnerLoopPreHeader->getTerminator()); |
| 1390 | |
| 1391 | // These instructions should now be executed inside the loop. |
| 1392 | // Move instruction into a new block after outer header. |
| 1393 | moveBBContents(InnerLoopPreHeader, OuterLoopHeader->getTerminator()); |
| 1394 | // These instructions were not executed previously in the loop so move them to |
| 1395 | // the older inner loop preheader. |
| 1396 | moveBBContents(OuterLoopPreHeader, InnerTermBI); |
| 1397 | } |
| 1398 | |
| 1399 | bool LoopInterchangeTransform::adjustLoopLinks() { |
| 1400 | // Adjust all branches in the inner and outer loop. |
| 1401 | bool Changed = adjustLoopBranches(); |
| 1402 | if (Changed) |
| 1403 | adjustLoopPreheaders(); |
| 1404 | return Changed; |
| 1405 | } |
| 1406 | |
| 1407 | char LoopInterchange::ID = 0; |
| 1408 | |
| 1409 | INITIALIZE_PASS_BEGIN(LoopInterchange, "loop-interchange",static void *initializeLoopInterchangePassOnce(PassRegistry & Registry) { |
| 1410 | "Interchanges loops for cache reuse", false, false)static void *initializeLoopInterchangePassOnce(PassRegistry & Registry) { |
| 1411 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
| 1412 | INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)initializeDependenceAnalysisWrapperPassPass(Registry); |
| 1413 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
| 1414 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); |
| 1415 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); |
| 1416 | INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)initializeLCSSAWrapperPassPass(Registry); |
| 1417 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); |
| 1418 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); |
| 1419 | |
| 1420 | INITIALIZE_PASS_END(LoopInterchange, "loop-interchange",PassInfo *PI = new PassInfo( "Interchanges loops for cache reuse" , "loop-interchange", &LoopInterchange::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopInterchange>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopInterchangePassFlag; void llvm::initializeLoopInterchangePass (PassRegistry &Registry) { llvm::call_once(InitializeLoopInterchangePassFlag , initializeLoopInterchangePassOnce, std::ref(Registry)); } |
| 1421 | "Interchanges loops for cache reuse", false, false)PassInfo *PI = new PassInfo( "Interchanges loops for cache reuse" , "loop-interchange", &LoopInterchange::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopInterchange>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopInterchangePassFlag; void llvm::initializeLoopInterchangePass (PassRegistry &Registry) { llvm::call_once(InitializeLoopInterchangePassFlag , initializeLoopInterchangePassOnce, std::ref(Registry)); } |
| 1422 | |
| 1423 | Pass *llvm::createLoopInterchangePass() { return new LoopInterchange(); } |