Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 1 | //===- LoopDistribute.cpp - Loop Distribution 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 file implements the Loop Distribution Pass. Its main focus is to |
| 11 | // distribute loops that cannot be vectorized due to dependence cycles. It |
| 12 | // tries to isolate the offending dependences into a new loop allowing |
| 13 | // vectorization of the remaining parts. |
| 14 | // |
| 15 | // For dependence analysis, the pass uses the LoopVectorizer's |
| 16 | // LoopAccessAnalysis. Because this analysis presumes no change in the order of |
| 17 | // memory operations, special care is taken to preserve the lexical order of |
| 18 | // these operations. |
| 19 | // |
| 20 | // Similarly to the Vectorizer, the pass also supports loop versioning to |
| 21 | // run-time disambiguate potentially overlapping arrays. |
| 22 | // |
| 23 | //===----------------------------------------------------------------------===// |
| 24 | |
| 25 | #include "llvm/ADT/DepthFirstIterator.h" |
| 26 | #include "llvm/ADT/EquivalenceClasses.h" |
| 27 | #include "llvm/ADT/STLExtras.h" |
| 28 | #include "llvm/ADT/Statistic.h" |
| 29 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
| 30 | #include "llvm/Analysis/LoopInfo.h" |
| 31 | #include "llvm/IR/Dominators.h" |
| 32 | #include "llvm/Pass.h" |
| 33 | #include "llvm/Support/CommandLine.h" |
| 34 | #include "llvm/Support/Debug.h" |
| 35 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 36 | #include "llvm/Transforms/Utils/Cloning.h" |
Adam Nemet | 215746b | 2015-07-10 18:55:13 +0000 | [diff] [blame] | 37 | #include "llvm/Transforms/Utils/LoopVersioning.h" |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 38 | #include <list> |
| 39 | |
| 40 | #define LDIST_NAME "loop-distribute" |
| 41 | #define DEBUG_TYPE LDIST_NAME |
| 42 | |
| 43 | using namespace llvm; |
| 44 | |
| 45 | static cl::opt<bool> |
| 46 | LDistVerify("loop-distribute-verify", cl::Hidden, |
| 47 | cl::desc("Turn on DominatorTree and LoopInfo verification " |
| 48 | "after Loop Distribution"), |
| 49 | cl::init(false)); |
| 50 | |
| 51 | static cl::opt<bool> DistributeNonIfConvertible( |
| 52 | "loop-distribute-non-if-convertible", cl::Hidden, |
| 53 | cl::desc("Whether to distribute into a loop that may not be " |
| 54 | "if-convertible by the loop vectorizer"), |
| 55 | cl::init(false)); |
| 56 | |
| 57 | STATISTIC(NumLoopsDistributed, "Number of loops distributed"); |
| 58 | |
Adam Nemet | 2f85b73 | 2015-05-14 12:33:32 +0000 | [diff] [blame] | 59 | namespace { |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 60 | /// \brief Maintains the set of instructions of the loop for a partition before |
| 61 | /// cloning. After cloning, it hosts the new loop. |
| 62 | class InstPartition { |
| 63 | typedef SmallPtrSet<Instruction *, 8> InstructionSet; |
| 64 | |
| 65 | public: |
| 66 | InstPartition(Instruction *I, Loop *L, bool DepCycle = false) |
| 67 | : DepCycle(DepCycle), OrigLoop(L), ClonedLoop(nullptr) { |
| 68 | Set.insert(I); |
| 69 | } |
| 70 | |
| 71 | /// \brief Returns whether this partition contains a dependence cycle. |
| 72 | bool hasDepCycle() const { return DepCycle; } |
| 73 | |
| 74 | /// \brief Adds an instruction to this partition. |
| 75 | void add(Instruction *I) { Set.insert(I); } |
| 76 | |
| 77 | /// \brief Collection accessors. |
| 78 | InstructionSet::iterator begin() { return Set.begin(); } |
| 79 | InstructionSet::iterator end() { return Set.end(); } |
| 80 | InstructionSet::const_iterator begin() const { return Set.begin(); } |
| 81 | InstructionSet::const_iterator end() const { return Set.end(); } |
| 82 | bool empty() const { return Set.empty(); } |
| 83 | |
| 84 | /// \brief Moves this partition into \p Other. This partition becomes empty |
| 85 | /// after this. |
| 86 | void moveTo(InstPartition &Other) { |
| 87 | Other.Set.insert(Set.begin(), Set.end()); |
| 88 | Set.clear(); |
| 89 | Other.DepCycle |= DepCycle; |
| 90 | } |
| 91 | |
| 92 | /// \brief Populates the partition with a transitive closure of all the |
| 93 | /// instructions that the seeded instructions dependent on. |
| 94 | void populateUsedSet() { |
| 95 | // FIXME: We currently don't use control-dependence but simply include all |
| 96 | // blocks (possibly empty at the end) and let simplifycfg mostly clean this |
| 97 | // up. |
| 98 | for (auto *B : OrigLoop->getBlocks()) |
| 99 | Set.insert(B->getTerminator()); |
| 100 | |
| 101 | // Follow the use-def chains to form a transitive closure of all the |
| 102 | // instructions that the originally seeded instructions depend on. |
| 103 | SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end()); |
| 104 | while (!Worklist.empty()) { |
| 105 | Instruction *I = Worklist.pop_back_val(); |
| 106 | // Insert instructions from the loop that we depend on. |
| 107 | for (Value *V : I->operand_values()) { |
| 108 | auto *I = dyn_cast<Instruction>(V); |
| 109 | if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second) |
| 110 | Worklist.push_back(I); |
| 111 | } |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | /// \brief Clones the original loop. |
| 116 | /// |
| 117 | /// Updates LoopInfo and DominatorTree using the information that block \p |
| 118 | /// LoopDomBB dominates the loop. |
| 119 | Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB, |
| 120 | unsigned Index, LoopInfo *LI, |
| 121 | DominatorTree *DT) { |
| 122 | ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop, |
| 123 | VMap, Twine(".ldist") + Twine(Index), |
| 124 | LI, DT, ClonedLoopBlocks); |
| 125 | return ClonedLoop; |
| 126 | } |
| 127 | |
| 128 | /// \brief The cloned loop. If this partition is mapped to the original loop, |
| 129 | /// this is null. |
| 130 | const Loop *getClonedLoop() const { return ClonedLoop; } |
| 131 | |
| 132 | /// \brief Returns the loop where this partition ends up after distribution. |
| 133 | /// If this partition is mapped to the original loop then use the block from |
| 134 | /// the loop. |
| 135 | const Loop *getDistributedLoop() const { |
| 136 | return ClonedLoop ? ClonedLoop : OrigLoop; |
| 137 | } |
| 138 | |
| 139 | /// \brief The VMap that is populated by cloning and then used in |
| 140 | /// remapinstruction to remap the cloned instructions. |
| 141 | ValueToValueMapTy &getVMap() { return VMap; } |
| 142 | |
| 143 | /// \brief Remaps the cloned instructions using VMap. |
Adam Nemet | 1a68918 | 2015-07-10 18:55:09 +0000 | [diff] [blame] | 144 | void remapInstructions() { |
| 145 | remapInstructionsInBlocks(ClonedLoopBlocks, VMap); |
| 146 | } |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 147 | |
| 148 | /// \brief Based on the set of instructions selected for this partition, |
| 149 | /// removes the unnecessary ones. |
| 150 | void removeUnusedInsts() { |
| 151 | SmallVector<Instruction *, 8> Unused; |
| 152 | |
| 153 | for (auto *Block : OrigLoop->getBlocks()) |
| 154 | for (auto &Inst : *Block) |
| 155 | if (!Set.count(&Inst)) { |
| 156 | Instruction *NewInst = &Inst; |
| 157 | if (!VMap.empty()) |
| 158 | NewInst = cast<Instruction>(VMap[NewInst]); |
| 159 | |
| 160 | assert(!isa<BranchInst>(NewInst) && |
| 161 | "Branches are marked used early on"); |
| 162 | Unused.push_back(NewInst); |
| 163 | } |
| 164 | |
| 165 | // Delete the instructions backwards, as it has a reduced likelihood of |
| 166 | // having to update as many def-use and use-def chains. |
| 167 | for (auto I = Unused.rbegin(), E = Unused.rend(); I != E; ++I) { |
| 168 | auto *Inst = *I; |
| 169 | |
| 170 | if (!Inst->use_empty()) |
| 171 | Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); |
| 172 | Inst->eraseFromParent(); |
| 173 | } |
| 174 | } |
| 175 | |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 176 | void print() const { |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 177 | if (DepCycle) |
| 178 | dbgs() << " (cycle)\n"; |
| 179 | for (auto *I : Set) |
| 180 | // Prefix with the block name. |
| 181 | dbgs() << " " << I->getParent()->getName() << ":" << *I << "\n"; |
| 182 | } |
| 183 | |
| 184 | void printBlocks() const { |
| 185 | for (auto *BB : getDistributedLoop()->getBlocks()) |
| 186 | dbgs() << *BB; |
| 187 | } |
| 188 | |
| 189 | private: |
| 190 | /// \brief Instructions from OrigLoop selected for this partition. |
| 191 | InstructionSet Set; |
| 192 | |
| 193 | /// \brief Whether this partition contains a dependence cycle. |
| 194 | bool DepCycle; |
| 195 | |
| 196 | /// \brief The original loop. |
| 197 | Loop *OrigLoop; |
| 198 | |
| 199 | /// \brief The cloned loop. If this partition is mapped to the original loop, |
| 200 | /// this is null. |
| 201 | Loop *ClonedLoop; |
| 202 | |
| 203 | /// \brief The blocks of ClonedLoop including the preheader. If this |
| 204 | /// partition is mapped to the original loop, this is empty. |
| 205 | SmallVector<BasicBlock *, 8> ClonedLoopBlocks; |
| 206 | |
| 207 | /// \brief These gets populated once the set of instructions have been |
| 208 | /// finalized. If this partition is mapped to the original loop, these are not |
| 209 | /// set. |
| 210 | ValueToValueMapTy VMap; |
| 211 | }; |
| 212 | |
| 213 | /// \brief Holds the set of Partitions. It populates them, merges them and then |
| 214 | /// clones the loops. |
| 215 | class InstPartitionContainer { |
| 216 | typedef DenseMap<Instruction *, int> InstToPartitionIdT; |
| 217 | |
| 218 | public: |
| 219 | InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT) |
| 220 | : L(L), LI(LI), DT(DT) {} |
| 221 | |
| 222 | /// \brief Returns the number of partitions. |
| 223 | unsigned getSize() const { return PartitionContainer.size(); } |
| 224 | |
| 225 | /// \brief Adds \p Inst into the current partition if that is marked to |
| 226 | /// contain cycles. Otherwise start a new partition for it. |
| 227 | void addToCyclicPartition(Instruction *Inst) { |
| 228 | // If the current partition is non-cyclic. Start a new one. |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 229 | if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle()) |
| 230 | PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 231 | else |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 232 | PartitionContainer.back().add(Inst); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 233 | } |
| 234 | |
| 235 | /// \brief Adds \p Inst into a partition that is not marked to contain |
| 236 | /// dependence cycles. |
| 237 | /// |
| 238 | // Initially we isolate memory instructions into as many partitions as |
| 239 | // possible, then later we may merge them back together. |
| 240 | void addToNewNonCyclicPartition(Instruction *Inst) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 241 | PartitionContainer.emplace_back(Inst, L); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 242 | } |
| 243 | |
| 244 | /// \brief Merges adjacent non-cyclic partitions. |
| 245 | /// |
| 246 | /// The idea is that we currently only want to isolate the non-vectorizable |
| 247 | /// partition. We could later allow more distribution among these partition |
| 248 | /// too. |
| 249 | void mergeAdjacentNonCyclic() { |
| 250 | mergeAdjacentPartitionsIf( |
| 251 | [](const InstPartition *P) { return !P->hasDepCycle(); }); |
| 252 | } |
| 253 | |
| 254 | /// \brief If a partition contains only conditional stores, we won't vectorize |
| 255 | /// it. Try to merge it with a previous cyclic partition. |
| 256 | void mergeNonIfConvertible() { |
| 257 | mergeAdjacentPartitionsIf([&](const InstPartition *Partition) { |
| 258 | if (Partition->hasDepCycle()) |
| 259 | return true; |
| 260 | |
| 261 | // Now, check if all stores are conditional in this partition. |
| 262 | bool seenStore = false; |
| 263 | |
| 264 | for (auto *Inst : *Partition) |
| 265 | if (isa<StoreInst>(Inst)) { |
| 266 | seenStore = true; |
| 267 | if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT)) |
| 268 | return false; |
| 269 | } |
| 270 | return seenStore; |
| 271 | }); |
| 272 | } |
| 273 | |
| 274 | /// \brief Merges the partitions according to various heuristics. |
| 275 | void mergeBeforePopulating() { |
| 276 | mergeAdjacentNonCyclic(); |
| 277 | if (!DistributeNonIfConvertible) |
| 278 | mergeNonIfConvertible(); |
| 279 | } |
| 280 | |
| 281 | /// \brief Merges partitions in order to ensure that no loads are duplicated. |
| 282 | /// |
| 283 | /// We can't duplicate loads because that could potentially reorder them. |
| 284 | /// LoopAccessAnalysis provides dependency information with the context that |
| 285 | /// the order of memory operation is preserved. |
| 286 | /// |
| 287 | /// Return if any partitions were merged. |
| 288 | bool mergeToAvoidDuplicatedLoads() { |
| 289 | typedef DenseMap<Instruction *, InstPartition *> LoadToPartitionT; |
| 290 | typedef EquivalenceClasses<InstPartition *> ToBeMergedT; |
| 291 | |
| 292 | LoadToPartitionT LoadToPartition; |
| 293 | ToBeMergedT ToBeMerged; |
| 294 | |
| 295 | // Step through the partitions and create equivalence between partitions |
| 296 | // that contain the same load. Also put partitions in between them in the |
| 297 | // same equivalence class to avoid reordering of memory operations. |
| 298 | for (PartitionContainerT::iterator I = PartitionContainer.begin(), |
| 299 | E = PartitionContainer.end(); |
| 300 | I != E; ++I) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 301 | auto *PartI = &*I; |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 302 | |
| 303 | // If a load occurs in two partitions PartI and PartJ, merge all |
| 304 | // partitions (PartI, PartJ] into PartI. |
| 305 | for (Instruction *Inst : *PartI) |
| 306 | if (isa<LoadInst>(Inst)) { |
| 307 | bool NewElt; |
| 308 | LoadToPartitionT::iterator LoadToPart; |
| 309 | |
| 310 | std::tie(LoadToPart, NewElt) = |
| 311 | LoadToPartition.insert(std::make_pair(Inst, PartI)); |
| 312 | if (!NewElt) { |
| 313 | DEBUG(dbgs() << "Merging partitions due to this load in multiple " |
| 314 | << "partitions: " << PartI << ", " |
| 315 | << LoadToPart->second << "\n" << *Inst << "\n"); |
| 316 | |
| 317 | auto PartJ = I; |
| 318 | do { |
| 319 | --PartJ; |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 320 | ToBeMerged.unionSets(PartI, &*PartJ); |
| 321 | } while (&*PartJ != LoadToPart->second); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 322 | } |
| 323 | } |
| 324 | } |
| 325 | if (ToBeMerged.empty()) |
| 326 | return false; |
| 327 | |
| 328 | // Merge the member of an equivalence class into its class leader. This |
| 329 | // makes the members empty. |
| 330 | for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end(); |
| 331 | I != E; ++I) { |
| 332 | if (!I->isLeader()) |
| 333 | continue; |
| 334 | |
| 335 | auto PartI = I->getData(); |
| 336 | for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)), |
| 337 | ToBeMerged.member_end())) { |
| 338 | PartJ->moveTo(*PartI); |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | // Remove the empty partitions. |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 343 | PartitionContainer.remove_if( |
| 344 | [](const InstPartition &P) { return P.empty(); }); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 345 | |
| 346 | return true; |
| 347 | } |
| 348 | |
| 349 | /// \brief Sets up the mapping between instructions to partitions. If the |
| 350 | /// instruction is duplicated across multiple partitions, set the entry to -1. |
| 351 | void setupPartitionIdOnInstructions() { |
| 352 | int PartitionID = 0; |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 353 | for (const auto &Partition : PartitionContainer) { |
| 354 | for (Instruction *Inst : Partition) { |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 355 | bool NewElt; |
| 356 | InstToPartitionIdT::iterator Iter; |
| 357 | |
| 358 | std::tie(Iter, NewElt) = |
| 359 | InstToPartitionId.insert(std::make_pair(Inst, PartitionID)); |
| 360 | if (!NewElt) |
| 361 | Iter->second = -1; |
| 362 | } |
| 363 | ++PartitionID; |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | /// \brief Populates the partition with everything that the seeding |
| 368 | /// instructions require. |
| 369 | void populateUsedSet() { |
| 370 | for (auto &P : PartitionContainer) |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 371 | P.populateUsedSet(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 372 | } |
| 373 | |
| 374 | /// \brief This performs the main chunk of the work of cloning the loops for |
| 375 | /// the partitions. |
| 376 | void cloneLoops(Pass *P) { |
| 377 | BasicBlock *OrigPH = L->getLoopPreheader(); |
| 378 | // At this point the predecessor of the preheader is either the memcheck |
| 379 | // block or the top part of the original preheader. |
| 380 | BasicBlock *Pred = OrigPH->getSinglePredecessor(); |
| 381 | assert(Pred && "Preheader does not have a single predecessor"); |
| 382 | BasicBlock *ExitBlock = L->getExitBlock(); |
| 383 | assert(ExitBlock && "No single exit block"); |
| 384 | Loop *NewLoop; |
| 385 | |
| 386 | assert(!PartitionContainer.empty() && "at least two partitions expected"); |
| 387 | // We're cloning the preheader along with the loop so we already made sure |
| 388 | // it was empty. |
| 389 | assert(&*OrigPH->begin() == OrigPH->getTerminator() && |
| 390 | "preheader not empty"); |
| 391 | |
| 392 | // Create a loop for each partition except the last. Clone the original |
| 393 | // loop before PH along with adding a preheader for the cloned loop. Then |
| 394 | // update PH to point to the newly added preheader. |
| 395 | BasicBlock *TopPH = OrigPH; |
| 396 | unsigned Index = getSize() - 1; |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 397 | for (auto I = std::next(PartitionContainer.rbegin()), |
| 398 | E = PartitionContainer.rend(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 399 | I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 400 | auto *Part = &*I; |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 401 | |
| 402 | NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT); |
| 403 | |
| 404 | Part->getVMap()[ExitBlock] = TopPH; |
| 405 | Part->remapInstructions(); |
| 406 | } |
| 407 | Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH); |
| 408 | |
| 409 | // Now go in forward order and update the immediate dominator for the |
| 410 | // preheaders with the exiting block of the previous loop. Dominance |
| 411 | // within the loop is updated in cloneLoopWithPreheader. |
| 412 | for (auto Curr = PartitionContainer.cbegin(), |
| 413 | Next = std::next(PartitionContainer.cbegin()), |
| 414 | E = PartitionContainer.cend(); |
| 415 | Next != E; ++Curr, ++Next) |
| 416 | DT->changeImmediateDominator( |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 417 | Next->getDistributedLoop()->getLoopPreheader(), |
| 418 | Curr->getDistributedLoop()->getExitingBlock()); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 419 | } |
| 420 | |
| 421 | /// \brief Removes the dead instructions from the cloned loops. |
| 422 | void removeUnusedInsts() { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 423 | for (auto &Partition : PartitionContainer) |
| 424 | Partition.removeUnusedInsts(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 425 | } |
| 426 | |
| 427 | /// \brief For each memory pointer, it computes the partitionId the pointer is |
| 428 | /// used in. |
| 429 | /// |
| 430 | /// This returns an array of int where the I-th entry corresponds to I-th |
| 431 | /// entry in LAI.getRuntimePointerCheck(). If the pointer is used in multiple |
| 432 | /// partitions its entry is set to -1. |
| 433 | SmallVector<int, 8> |
| 434 | computePartitionSetForPointers(const LoopAccessInfo &LAI) { |
Adam Nemet | 7cdebac | 2015-07-14 22:32:44 +0000 | [diff] [blame] | 435 | const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 436 | |
| 437 | unsigned N = RtPtrCheck->Pointers.size(); |
| 438 | SmallVector<int, 8> PtrToPartitions(N); |
| 439 | for (unsigned I = 0; I < N; ++I) { |
Adam Nemet | 9f7dedc | 2015-07-14 22:32:50 +0000 | [diff] [blame] | 440 | Value *Ptr = RtPtrCheck->Pointers[I].PointerValue; |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 441 | auto Instructions = |
Adam Nemet | 9f7dedc | 2015-07-14 22:32:50 +0000 | [diff] [blame] | 442 | LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 443 | |
| 444 | int &Partition = PtrToPartitions[I]; |
| 445 | // First set it to uninitialized. |
| 446 | Partition = -2; |
| 447 | for (Instruction *Inst : Instructions) { |
| 448 | // Note that this could be -1 if Inst is duplicated across multiple |
| 449 | // partitions. |
| 450 | int ThisPartition = this->InstToPartitionId[Inst]; |
| 451 | if (Partition == -2) |
| 452 | Partition = ThisPartition; |
| 453 | // -1 means belonging to multiple partitions. |
| 454 | else if (Partition == -1) |
| 455 | break; |
| 456 | else if (Partition != (int)ThisPartition) |
| 457 | Partition = -1; |
| 458 | } |
| 459 | assert(Partition != -2 && "Pointer not belonging to any partition"); |
| 460 | } |
| 461 | |
| 462 | return PtrToPartitions; |
| 463 | } |
| 464 | |
| 465 | void print(raw_ostream &OS) const { |
| 466 | unsigned Index = 0; |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 467 | for (const auto &P : PartitionContainer) { |
| 468 | OS << "Partition " << Index++ << " (" << &P << "):\n"; |
| 469 | P.print(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 470 | } |
| 471 | } |
| 472 | |
| 473 | void dump() const { print(dbgs()); } |
| 474 | |
| 475 | #ifndef NDEBUG |
| 476 | friend raw_ostream &operator<<(raw_ostream &OS, |
| 477 | const InstPartitionContainer &Partitions) { |
| 478 | Partitions.print(OS); |
| 479 | return OS; |
| 480 | } |
| 481 | #endif |
| 482 | |
| 483 | void printBlocks() const { |
| 484 | unsigned Index = 0; |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 485 | for (const auto &P : PartitionContainer) { |
| 486 | dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n"; |
| 487 | P.printBlocks(); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 488 | } |
| 489 | } |
| 490 | |
| 491 | private: |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 492 | typedef std::list<InstPartition> PartitionContainerT; |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 493 | |
| 494 | /// \brief List of partitions. |
| 495 | PartitionContainerT PartitionContainer; |
| 496 | |
| 497 | /// \brief Mapping from Instruction to partition Id. If the instruction |
| 498 | /// belongs to multiple partitions the entry contains -1. |
| 499 | InstToPartitionIdT InstToPartitionId; |
| 500 | |
| 501 | Loop *L; |
| 502 | LoopInfo *LI; |
| 503 | DominatorTree *DT; |
| 504 | |
| 505 | /// \brief The control structure to merge adjacent partitions if both satisfy |
| 506 | /// the \p Predicate. |
| 507 | template <class UnaryPredicate> |
| 508 | void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) { |
| 509 | InstPartition *PrevMatch = nullptr; |
| 510 | for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 511 | auto DoesMatch = Predicate(&*I); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 512 | if (PrevMatch == nullptr && DoesMatch) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 513 | PrevMatch = &*I; |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 514 | ++I; |
| 515 | } else if (PrevMatch != nullptr && DoesMatch) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 516 | I->moveTo(*PrevMatch); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 517 | I = PartitionContainer.erase(I); |
| 518 | } else { |
| 519 | PrevMatch = nullptr; |
| 520 | ++I; |
| 521 | } |
| 522 | } |
| 523 | } |
| 524 | }; |
| 525 | |
| 526 | /// \brief For each memory instruction, this class maintains difference of the |
| 527 | /// number of unsafe dependences that start out from this instruction minus |
| 528 | /// those that end here. |
| 529 | /// |
| 530 | /// By traversing the memory instructions in program order and accumulating this |
| 531 | /// number, we know whether any unsafe dependence crosses over a program point. |
| 532 | class MemoryInstructionDependences { |
| 533 | typedef MemoryDepChecker::Dependence Dependence; |
| 534 | |
| 535 | public: |
| 536 | struct Entry { |
| 537 | Instruction *Inst; |
| 538 | unsigned NumUnsafeDependencesStartOrEnd; |
| 539 | |
| 540 | Entry(Instruction *Inst) : Inst(Inst), NumUnsafeDependencesStartOrEnd(0) {} |
| 541 | }; |
| 542 | |
| 543 | typedef SmallVector<Entry, 8> AccessesType; |
| 544 | |
| 545 | AccessesType::const_iterator begin() const { return Accesses.begin(); } |
| 546 | AccessesType::const_iterator end() const { return Accesses.end(); } |
| 547 | |
| 548 | MemoryInstructionDependences( |
| 549 | const SmallVectorImpl<Instruction *> &Instructions, |
| 550 | const SmallVectorImpl<Dependence> &InterestingDependences) { |
Benjamin Kramer | e6987bf | 2015-05-21 18:32:07 +0000 | [diff] [blame] | 551 | Accesses.append(Instructions.begin(), Instructions.end()); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 552 | |
| 553 | DEBUG(dbgs() << "Backward dependences:\n"); |
| 554 | for (auto &Dep : InterestingDependences) |
| 555 | if (Dep.isPossiblyBackward()) { |
| 556 | // Note that the designations source and destination follow the program |
| 557 | // order, i.e. source is always first. (The direction is given by the |
| 558 | // DepType.) |
| 559 | ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd; |
| 560 | --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd; |
| 561 | |
| 562 | DEBUG(Dep.print(dbgs(), 2, Instructions)); |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | private: |
| 567 | AccessesType Accesses; |
| 568 | }; |
| 569 | |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 570 | /// \brief Returns the instructions that use values defined in the loop. |
| 571 | static SmallVector<Instruction *, 8> findDefsUsedOutsideOfLoop(Loop *L) { |
| 572 | SmallVector<Instruction *, 8> UsedOutside; |
| 573 | |
| 574 | for (auto *Block : L->getBlocks()) |
| 575 | // FIXME: I believe that this could use copy_if if the Inst reference could |
| 576 | // be adapted into a pointer. |
| 577 | for (auto &Inst : *Block) { |
| 578 | auto Users = Inst.users(); |
| 579 | if (std::any_of(Users.begin(), Users.end(), [&](User *U) { |
| 580 | auto *Use = cast<Instruction>(U); |
| 581 | return !L->contains(Use->getParent()); |
| 582 | })) |
| 583 | UsedOutside.push_back(&Inst); |
| 584 | } |
| 585 | |
| 586 | return UsedOutside; |
| 587 | } |
| 588 | |
| 589 | /// \brief The pass class. |
| 590 | class LoopDistribute : public FunctionPass { |
| 591 | public: |
| 592 | LoopDistribute() : FunctionPass(ID) { |
| 593 | initializeLoopDistributePass(*PassRegistry::getPassRegistry()); |
| 594 | } |
| 595 | |
| 596 | bool runOnFunction(Function &F) override { |
| 597 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 598 | LAA = &getAnalysis<LoopAccessAnalysis>(); |
| 599 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 600 | |
| 601 | // Build up a worklist of inner-loops to vectorize. This is necessary as the |
| 602 | // act of distributing a loop creates new loops and can invalidate iterators |
| 603 | // across the loops. |
| 604 | SmallVector<Loop *, 8> Worklist; |
| 605 | |
| 606 | for (Loop *TopLevelLoop : *LI) |
| 607 | for (Loop *L : depth_first(TopLevelLoop)) |
| 608 | // We only handle inner-most loops. |
| 609 | if (L->empty()) |
| 610 | Worklist.push_back(L); |
| 611 | |
| 612 | // Now walk the identified inner loops. |
| 613 | bool Changed = false; |
| 614 | for (Loop *L : Worklist) |
| 615 | Changed |= processLoop(L); |
| 616 | |
| 617 | // Process each loop nest in the function. |
| 618 | return Changed; |
| 619 | } |
| 620 | |
| 621 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 622 | AU.addRequired<LoopInfoWrapperPass>(); |
| 623 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 624 | AU.addRequired<LoopAccessAnalysis>(); |
| 625 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 626 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 627 | } |
| 628 | |
| 629 | static char ID; |
| 630 | |
| 631 | private: |
| 632 | /// \brief Try to distribute an inner-most loop. |
| 633 | bool processLoop(Loop *L) { |
| 634 | assert(L->empty() && "Only process inner loops."); |
| 635 | |
| 636 | DEBUG(dbgs() << "\nLDist: In \"" << L->getHeader()->getParent()->getName() |
| 637 | << "\" checking " << *L << "\n"); |
| 638 | |
| 639 | BasicBlock *PH = L->getLoopPreheader(); |
| 640 | if (!PH) { |
| 641 | DEBUG(dbgs() << "Skipping; no preheader"); |
| 642 | return false; |
| 643 | } |
| 644 | if (!L->getExitBlock()) { |
| 645 | DEBUG(dbgs() << "Skipping; multiple exit blocks"); |
| 646 | return false; |
| 647 | } |
| 648 | // LAA will check that we only have a single exiting block. |
| 649 | |
| 650 | const LoopAccessInfo &LAI = LAA->getInfo(L, ValueToValueMap()); |
| 651 | |
| 652 | // Currently, we only distribute to isolate the part of the loop with |
| 653 | // dependence cycles to enable partial vectorization. |
| 654 | if (LAI.canVectorizeMemory()) { |
| 655 | DEBUG(dbgs() << "Skipping; memory operations are safe for vectorization"); |
| 656 | return false; |
| 657 | } |
| 658 | auto *InterestingDependences = |
| 659 | LAI.getDepChecker().getInterestingDependences(); |
| 660 | if (!InterestingDependences || InterestingDependences->empty()) { |
| 661 | DEBUG(dbgs() << "Skipping; No unsafe dependences to isolate"); |
| 662 | return false; |
| 663 | } |
| 664 | |
| 665 | InstPartitionContainer Partitions(L, LI, DT); |
| 666 | |
| 667 | // First, go through each memory operation and assign them to consecutive |
| 668 | // partitions (the order of partitions follows program order). Put those |
| 669 | // with unsafe dependences into "cyclic" partition otherwise put each store |
| 670 | // in its own "non-cyclic" partition (we'll merge these later). |
| 671 | // |
| 672 | // Note that a memory operation (e.g. Load2 below) at a program point that |
| 673 | // has an unsafe dependence (Store3->Load1) spanning over it must be |
| 674 | // included in the same cyclic partition as the dependent operations. This |
| 675 | // is to preserve the original program order after distribution. E.g.: |
| 676 | // |
| 677 | // NumUnsafeDependencesStartOrEnd NumUnsafeDependencesActive |
| 678 | // Load1 -. 1 0->1 |
| 679 | // Load2 | /Unsafe/ 0 1 |
| 680 | // Store3 -' -1 1->0 |
| 681 | // Load4 0 0 |
| 682 | // |
| 683 | // NumUnsafeDependencesActive > 0 indicates this situation and in this case |
| 684 | // we just keep assigning to the same cyclic partition until |
| 685 | // NumUnsafeDependencesActive reaches 0. |
| 686 | const MemoryDepChecker &DepChecker = LAI.getDepChecker(); |
| 687 | MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(), |
| 688 | *InterestingDependences); |
| 689 | |
| 690 | int NumUnsafeDependencesActive = 0; |
| 691 | for (auto &InstDep : MID) { |
| 692 | Instruction *I = InstDep.Inst; |
| 693 | // We update NumUnsafeDependencesActive post-instruction, catch the |
| 694 | // start of a dependence directly via NumUnsafeDependencesStartOrEnd. |
| 695 | if (NumUnsafeDependencesActive || |
| 696 | InstDep.NumUnsafeDependencesStartOrEnd > 0) |
| 697 | Partitions.addToCyclicPartition(I); |
| 698 | else |
| 699 | Partitions.addToNewNonCyclicPartition(I); |
| 700 | NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd; |
| 701 | assert(NumUnsafeDependencesActive >= 0 && |
| 702 | "Negative number of dependences active"); |
| 703 | } |
| 704 | |
| 705 | // Add partitions for values used outside. These partitions can be out of |
| 706 | // order from the original program order. This is OK because if the |
| 707 | // partition uses a load we will merge this partition with the original |
| 708 | // partition of the load that we set up in the previous loop (see |
| 709 | // mergeToAvoidDuplicatedLoads). |
| 710 | auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L); |
| 711 | for (auto *Inst : DefsUsedOutside) |
| 712 | Partitions.addToNewNonCyclicPartition(Inst); |
| 713 | |
| 714 | DEBUG(dbgs() << "Seeded partitions:\n" << Partitions); |
| 715 | if (Partitions.getSize() < 2) |
| 716 | return false; |
| 717 | |
| 718 | // Run the merge heuristics: Merge non-cyclic adjacent partitions since we |
| 719 | // should be able to vectorize these together. |
| 720 | Partitions.mergeBeforePopulating(); |
| 721 | DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions); |
| 722 | if (Partitions.getSize() < 2) |
| 723 | return false; |
| 724 | |
| 725 | // Now, populate the partitions with non-memory operations. |
| 726 | Partitions.populateUsedSet(); |
| 727 | DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions); |
| 728 | |
| 729 | // In order to preserve original lexical order for loads, keep them in the |
| 730 | // partition that we set up in the MemoryInstructionDependences loop. |
| 731 | if (Partitions.mergeToAvoidDuplicatedLoads()) { |
| 732 | DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n" |
| 733 | << Partitions); |
| 734 | if (Partitions.getSize() < 2) |
| 735 | return false; |
| 736 | } |
| 737 | |
| 738 | DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n"); |
| 739 | // We're done forming the partitions set up the reverse mapping from |
| 740 | // instructions to partitions. |
| 741 | Partitions.setupPartitionIdOnInstructions(); |
| 742 | |
| 743 | // To keep things simple have an empty preheader before we version or clone |
| 744 | // the loop. (Also split if this has no predecessor, i.e. entry, because we |
| 745 | // rely on PH having a predecessor.) |
| 746 | if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator()) |
| 747 | SplitBlock(PH, PH->getTerminator(), DT, LI); |
| 748 | |
| 749 | // If we need run-time checks to disambiguate pointers are run-time, version |
| 750 | // the loop now. |
Adam Nemet | 772a150 | 2015-06-19 19:32:41 +0000 | [diff] [blame] | 751 | auto PtrToPartition = Partitions.computePartitionSetForPointers(LAI); |
Adam Nemet | 7632500 | 2015-06-19 19:32:48 +0000 | [diff] [blame] | 752 | LoopVersioning LVer(LAI, L, LI, DT, &PtrToPartition); |
| 753 | if (LVer.needsRuntimeChecks()) { |
Adam Nemet | 772a150 | 2015-06-19 19:32:41 +0000 | [diff] [blame] | 754 | DEBUG(dbgs() << "\nPointers:\n"); |
Adam Nemet | 7cdebac | 2015-07-14 22:32:44 +0000 | [diff] [blame] | 755 | DEBUG(LAI.getRuntimePointerChecking()->print(dbgs(), 0, &PtrToPartition)); |
Adam Nemet | 7632500 | 2015-06-19 19:32:48 +0000 | [diff] [blame] | 756 | LVer.versionLoop(this); |
| 757 | LVer.addPHINodes(DefsUsedOutside); |
Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 758 | } |
| 759 | |
| 760 | // Create identical copies of the original loop for each partition and hook |
| 761 | // them up sequentially. |
| 762 | Partitions.cloneLoops(this); |
| 763 | |
| 764 | // Now, we remove the instruction from each loop that don't belong to that |
| 765 | // partition. |
| 766 | Partitions.removeUnusedInsts(); |
| 767 | DEBUG(dbgs() << "\nAfter removing unused Instrs:\n"); |
| 768 | DEBUG(Partitions.printBlocks()); |
| 769 | |
| 770 | if (LDistVerify) { |
| 771 | LI->verify(); |
| 772 | DT->verifyDomTree(); |
| 773 | } |
| 774 | |
| 775 | ++NumLoopsDistributed; |
| 776 | return true; |
| 777 | } |
| 778 | |
| 779 | // Analyses used. |
| 780 | LoopInfo *LI; |
| 781 | LoopAccessAnalysis *LAA; |
| 782 | DominatorTree *DT; |
| 783 | }; |
| 784 | } // anonymous namespace |
| 785 | |
| 786 | char LoopDistribute::ID; |
| 787 | static const char ldist_name[] = "Loop Distribition"; |
| 788 | |
| 789 | INITIALIZE_PASS_BEGIN(LoopDistribute, LDIST_NAME, ldist_name, false, false) |
| 790 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
| 791 | INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis) |
| 792 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 793 | INITIALIZE_PASS_END(LoopDistribute, LDIST_NAME, ldist_name, false, false) |
| 794 | |
| 795 | namespace llvm { |
| 796 | FunctionPass *createLoopDistributePass() { return new LoopDistribute(); } |
| 797 | } |