| File: | llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp |
| Warning: | line 513, column 15 Although the value stored to 'PatternValue' is used in the enclosing expression, the value is never actually read from 'PatternValue' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This pass implements an idiom recognizer that transforms simple loops into a |
| 10 | // non-loop form. In cases that this kicks in, it can be a significant |
| 11 | // performance win. |
| 12 | // |
| 13 | // If compiling for code size we avoid idiom recognition if the resulting |
| 14 | // code could be larger than the code for the original loop. One way this could |
| 15 | // happen is if the loop is not removable after idiom recognition due to the |
| 16 | // presence of non-idiom instructions. The initial implementation of the |
| 17 | // heuristics applies to idioms in multi-block loops. |
| 18 | // |
| 19 | //===----------------------------------------------------------------------===// |
| 20 | // |
| 21 | // TODO List: |
| 22 | // |
| 23 | // Future loop memory idioms to recognize: |
| 24 | // memcmp, memmove, strlen, etc. |
| 25 | // Future floating point idioms to recognize in -ffast-math mode: |
| 26 | // fpowi |
| 27 | // Future integer operation idioms to recognize: |
| 28 | // ctpop |
| 29 | // |
| 30 | // Beware that isel's default lowering for ctpop is highly inefficient for |
| 31 | // i64 and larger types when i64 is legal and the value has few bits set. It |
| 32 | // would be good to enhance isel to emit a loop for ctpop in this case. |
| 33 | // |
| 34 | // This could recognize common matrix multiplies and dot product idioms and |
| 35 | // replace them with calls to BLAS (if linked in??). |
| 36 | // |
| 37 | //===----------------------------------------------------------------------===// |
| 38 | |
| 39 | #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h" |
| 40 | #include "llvm/ADT/APInt.h" |
| 41 | #include "llvm/ADT/ArrayRef.h" |
| 42 | #include "llvm/ADT/DenseMap.h" |
| 43 | #include "llvm/ADT/MapVector.h" |
| 44 | #include "llvm/ADT/SetVector.h" |
| 45 | #include "llvm/ADT/SmallPtrSet.h" |
| 46 | #include "llvm/ADT/SmallVector.h" |
| 47 | #include "llvm/ADT/Statistic.h" |
| 48 | #include "llvm/ADT/StringRef.h" |
| 49 | #include "llvm/Analysis/AliasAnalysis.h" |
| 50 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
| 51 | #include "llvm/Analysis/LoopInfo.h" |
| 52 | #include "llvm/Analysis/LoopPass.h" |
| 53 | #include "llvm/Analysis/MemoryLocation.h" |
| 54 | #include "llvm/Analysis/MemorySSA.h" |
| 55 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 56 | #include "llvm/Analysis/MustExecute.h" |
| 57 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 58 | #include "llvm/Analysis/ScalarEvolution.h" |
| 59 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 60 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 61 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 62 | #include "llvm/Analysis/ValueTracking.h" |
| 63 | #include "llvm/IR/Attributes.h" |
| 64 | #include "llvm/IR/BasicBlock.h" |
| 65 | #include "llvm/IR/Constant.h" |
| 66 | #include "llvm/IR/Constants.h" |
| 67 | #include "llvm/IR/DataLayout.h" |
| 68 | #include "llvm/IR/DebugLoc.h" |
| 69 | #include "llvm/IR/DerivedTypes.h" |
| 70 | #include "llvm/IR/Dominators.h" |
| 71 | #include "llvm/IR/GlobalValue.h" |
| 72 | #include "llvm/IR/GlobalVariable.h" |
| 73 | #include "llvm/IR/IRBuilder.h" |
| 74 | #include "llvm/IR/InstrTypes.h" |
| 75 | #include "llvm/IR/Instruction.h" |
| 76 | #include "llvm/IR/Instructions.h" |
| 77 | #include "llvm/IR/IntrinsicInst.h" |
| 78 | #include "llvm/IR/Intrinsics.h" |
| 79 | #include "llvm/IR/LLVMContext.h" |
| 80 | #include "llvm/IR/Module.h" |
| 81 | #include "llvm/IR/PassManager.h" |
| 82 | #include "llvm/IR/Type.h" |
| 83 | #include "llvm/IR/User.h" |
| 84 | #include "llvm/IR/Value.h" |
| 85 | #include "llvm/IR/ValueHandle.h" |
| 86 | #include "llvm/InitializePasses.h" |
| 87 | #include "llvm/Pass.h" |
| 88 | #include "llvm/Support/Casting.h" |
| 89 | #include "llvm/Support/CommandLine.h" |
| 90 | #include "llvm/Support/Debug.h" |
| 91 | #include "llvm/Support/raw_ostream.h" |
| 92 | #include "llvm/Transforms/Scalar.h" |
| 93 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
| 94 | #include "llvm/Transforms/Utils/Local.h" |
| 95 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 96 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
| 97 | #include <algorithm> |
| 98 | #include <cassert> |
| 99 | #include <cstdint> |
| 100 | #include <utility> |
| 101 | #include <vector> |
| 102 | |
| 103 | using namespace llvm; |
| 104 | |
| 105 | #define DEBUG_TYPE"loop-idiom" "loop-idiom" |
| 106 | |
| 107 | STATISTIC(NumMemSet, "Number of memset's formed from loop stores")static llvm::Statistic NumMemSet = {"loop-idiom", "NumMemSet" , "Number of memset's formed from loop stores"}; |
| 108 | STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores")static llvm::Statistic NumMemCpy = {"loop-idiom", "NumMemCpy" , "Number of memcpy's formed from loop load+stores"}; |
| 109 | |
| 110 | bool DisableLIRP::All; |
| 111 | static cl::opt<bool, true> |
| 112 | DisableLIRPAll("disable-" DEBUG_TYPE"loop-idiom" "-all", |
| 113 | cl::desc("Options to disable Loop Idiom Recognize Pass."), |
| 114 | cl::location(DisableLIRP::All), cl::init(false), |
| 115 | cl::ReallyHidden); |
| 116 | |
| 117 | bool DisableLIRP::Memset; |
| 118 | static cl::opt<bool, true> |
| 119 | DisableLIRPMemset("disable-" DEBUG_TYPE"loop-idiom" "-memset", |
| 120 | cl::desc("Proceed with loop idiom recognize pass, but do " |
| 121 | "not convert loop(s) to memset."), |
| 122 | cl::location(DisableLIRP::Memset), cl::init(false), |
| 123 | cl::ReallyHidden); |
| 124 | |
| 125 | bool DisableLIRP::Memcpy; |
| 126 | static cl::opt<bool, true> |
| 127 | DisableLIRPMemcpy("disable-" DEBUG_TYPE"loop-idiom" "-memcpy", |
| 128 | cl::desc("Proceed with loop idiom recognize pass, but do " |
| 129 | "not convert loop(s) to memcpy."), |
| 130 | cl::location(DisableLIRP::Memcpy), cl::init(false), |
| 131 | cl::ReallyHidden); |
| 132 | |
| 133 | static cl::opt<bool> UseLIRCodeSizeHeurs( |
| 134 | "use-lir-code-size-heurs", |
| 135 | cl::desc("Use loop idiom recognition code size heuristics when compiling" |
| 136 | "with -Os/-Oz"), |
| 137 | cl::init(true), cl::Hidden); |
| 138 | |
| 139 | namespace { |
| 140 | |
| 141 | class LoopIdiomRecognize { |
| 142 | Loop *CurLoop = nullptr; |
| 143 | AliasAnalysis *AA; |
| 144 | DominatorTree *DT; |
| 145 | LoopInfo *LI; |
| 146 | ScalarEvolution *SE; |
| 147 | TargetLibraryInfo *TLI; |
| 148 | const TargetTransformInfo *TTI; |
| 149 | const DataLayout *DL; |
| 150 | OptimizationRemarkEmitter &ORE; |
| 151 | bool ApplyCodeSizeHeuristics; |
| 152 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
| 153 | |
| 154 | public: |
| 155 | explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
| 156 | LoopInfo *LI, ScalarEvolution *SE, |
| 157 | TargetLibraryInfo *TLI, |
| 158 | const TargetTransformInfo *TTI, MemorySSA *MSSA, |
| 159 | const DataLayout *DL, |
| 160 | OptimizationRemarkEmitter &ORE) |
| 161 | : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { |
| 162 | if (MSSA) |
| 163 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); |
| 164 | } |
| 165 | |
| 166 | bool runOnLoop(Loop *L); |
| 167 | |
| 168 | private: |
| 169 | using StoreList = SmallVector<StoreInst *, 8>; |
| 170 | using StoreListMap = MapVector<Value *, StoreList>; |
| 171 | |
| 172 | StoreListMap StoreRefsForMemset; |
| 173 | StoreListMap StoreRefsForMemsetPattern; |
| 174 | StoreList StoreRefsForMemcpy; |
| 175 | bool HasMemset; |
| 176 | bool HasMemsetPattern; |
| 177 | bool HasMemcpy; |
| 178 | |
| 179 | /// Return code for isLegalStore() |
| 180 | enum LegalStoreKind { |
| 181 | None = 0, |
| 182 | Memset, |
| 183 | MemsetPattern, |
| 184 | Memcpy, |
| 185 | UnorderedAtomicMemcpy, |
| 186 | DontUse // Dummy retval never to be used. Allows catching errors in retval |
| 187 | // handling. |
| 188 | }; |
| 189 | |
| 190 | /// \name Countable Loop Idiom Handling |
| 191 | /// @{ |
| 192 | |
| 193 | bool runOnCountableLoop(); |
| 194 | bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, |
| 195 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
| 196 | |
| 197 | void collectStores(BasicBlock *BB); |
| 198 | LegalStoreKind isLegalStore(StoreInst *SI); |
| 199 | enum class ForMemset { No, Yes }; |
| 200 | bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount, |
| 201 | ForMemset For); |
| 202 | bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); |
| 203 | |
| 204 | bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize, |
| 205 | MaybeAlign StoreAlignment, Value *StoredVal, |
| 206 | Instruction *TheStore, |
| 207 | SmallPtrSetImpl<Instruction *> &Stores, |
| 208 | const SCEVAddRecExpr *Ev, const SCEV *BECount, |
| 209 | bool NegStride, bool IsLoopMemset = false); |
| 210 | bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); |
| 211 | bool avoidLIRForMultiBlockLoop(bool IsMemset = false, |
| 212 | bool IsLoopMemset = false); |
| 213 | |
| 214 | /// @} |
| 215 | /// \name Noncountable Loop Idiom Handling |
| 216 | /// @{ |
| 217 | |
| 218 | bool runOnNoncountableLoop(); |
| 219 | |
| 220 | bool recognizePopcount(); |
| 221 | void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, |
| 222 | PHINode *CntPhi, Value *Var); |
| 223 | bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz |
| 224 | void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, |
| 225 | Instruction *CntInst, PHINode *CntPhi, |
| 226 | Value *Var, Instruction *DefX, |
| 227 | const DebugLoc &DL, bool ZeroCheck, |
| 228 | bool IsCntPhiUsedOutsideLoop); |
| 229 | |
| 230 | /// @} |
| 231 | }; |
| 232 | |
| 233 | class LoopIdiomRecognizeLegacyPass : public LoopPass { |
| 234 | public: |
| 235 | static char ID; |
| 236 | |
| 237 | explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) { |
| 238 | initializeLoopIdiomRecognizeLegacyPassPass( |
| 239 | *PassRegistry::getPassRegistry()); |
| 240 | } |
| 241 | |
| 242 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { |
| 243 | if (DisableLIRP::All) |
| 244 | return false; |
| 245 | |
| 246 | if (skipLoop(L)) |
| 247 | return false; |
| 248 | |
| 249 | AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 250 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 251 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 252 | ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
| 253 | TargetLibraryInfo *TLI = |
| 254 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
| 255 | *L->getHeader()->getParent()); |
| 256 | const TargetTransformInfo *TTI = |
| 257 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( |
| 258 | *L->getHeader()->getParent()); |
| 259 | const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout(); |
| 260 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
| 261 | MemorySSA *MSSA = nullptr; |
| 262 | if (MSSAAnalysis) |
| 263 | MSSA = &MSSAAnalysis->getMSSA(); |
| 264 | |
| 265 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis |
| 266 | // pass. Function analyses need to be preserved across loop transformations |
| 267 | // but ORE cannot be preserved (see comment before the pass definition). |
| 268 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); |
| 269 | |
| 270 | LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE); |
| 271 | return LIR.runOnLoop(L); |
| 272 | } |
| 273 | |
| 274 | /// This transformation requires natural loop information & requires that |
| 275 | /// loop preheaders be inserted into the CFG. |
| 276 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 277 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 278 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 279 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 280 | getLoopAnalysisUsage(AU); |
| 281 | } |
| 282 | }; |
| 283 | |
| 284 | } // end anonymous namespace |
| 285 | |
| 286 | char LoopIdiomRecognizeLegacyPass::ID = 0; |
| 287 | |
| 288 | PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, |
| 289 | LoopStandardAnalysisResults &AR, |
| 290 | LPMUpdater &) { |
| 291 | if (DisableLIRP::All) |
| 292 | return PreservedAnalyses::all(); |
| 293 | |
| 294 | const auto *DL = &L.getHeader()->getModule()->getDataLayout(); |
| 295 | |
| 296 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
| 297 | // pass. Function analyses need to be preserved across loop transformations |
| 298 | // but ORE cannot be preserved (see comment before the pass definition). |
| 299 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
| 300 | |
| 301 | LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, |
| 302 | AR.MSSA, DL, ORE); |
| 303 | if (!LIR.runOnLoop(&L)) |
| 304 | return PreservedAnalyses::all(); |
| 305 | |
| 306 | auto PA = getLoopPassPreservedAnalyses(); |
| 307 | if (AR.MSSA) |
| 308 | PA.preserve<MemorySSAAnalysis>(); |
| 309 | return PA; |
| 310 | } |
| 311 | |
| 312 | INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
| 313 | "Recognize loop idioms", false, false)static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
| 314 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); |
| 315 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
| 316 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); |
| 317 | INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
| 318 | "Recognize loop idioms", false, false)PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
| 319 | |
| 320 | Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); } |
| 321 | |
| 322 | static void deleteDeadInstruction(Instruction *I) { |
| 323 | I->replaceAllUsesWith(UndefValue::get(I->getType())); |
| 324 | I->eraseFromParent(); |
| 325 | } |
| 326 | |
| 327 | //===----------------------------------------------------------------------===// |
| 328 | // |
| 329 | // Implementation of LoopIdiomRecognize |
| 330 | // |
| 331 | //===----------------------------------------------------------------------===// |
| 332 | |
| 333 | bool LoopIdiomRecognize::runOnLoop(Loop *L) { |
| 334 | CurLoop = L; |
| 335 | // If the loop could not be converted to canonical form, it must have an |
| 336 | // indirectbr in it, just give up. |
| 337 | if (!L->getLoopPreheader()) |
| 338 | return false; |
| 339 | |
| 340 | // Disable loop idiom recognition if the function's name is a common idiom. |
| 341 | StringRef Name = L->getHeader()->getParent()->getName(); |
| 342 | if (Name == "memset" || Name == "memcpy") |
| 343 | return false; |
| 344 | |
| 345 | // Determine if code size heuristics need to be applied. |
| 346 | ApplyCodeSizeHeuristics = |
| 347 | L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; |
| 348 | |
| 349 | HasMemset = TLI->has(LibFunc_memset); |
| 350 | HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); |
| 351 | HasMemcpy = TLI->has(LibFunc_memcpy); |
| 352 | |
| 353 | if (HasMemset || HasMemsetPattern || HasMemcpy) |
| 354 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
| 355 | return runOnCountableLoop(); |
| 356 | |
| 357 | return runOnNoncountableLoop(); |
| 358 | } |
| 359 | |
| 360 | bool LoopIdiomRecognize::runOnCountableLoop() { |
| 361 | const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); |
| 362 | assert(!isa<SCEVCouldNotCompute>(BECount) &&((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)) |
| 363 | "runOnCountableLoop() called on a loop without a predictable"((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)) |
| 364 | "backedge-taken count")((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)); |
| 365 | |
| 366 | // If this loop executes exactly one time, then it should be peeled, not |
| 367 | // optimized by this pass. |
| 368 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
| 369 | if (BECst->getAPInt() == 0) |
| 370 | return false; |
| 371 | |
| 372 | SmallVector<BasicBlock *, 8> ExitBlocks; |
| 373 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
| 374 | |
| 375 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 376 | << CurLoop->getHeader()->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 377 | << "] Countable Loop %" << CurLoop->getHeader()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 378 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false); |
| 379 | |
| 380 | // The following transforms hoist stores/memsets into the loop pre-header. |
| 381 | // Give up if the loop has instructions that may throw. |
| 382 | SimpleLoopSafetyInfo SafetyInfo; |
| 383 | SafetyInfo.computeLoopSafetyInfo(CurLoop); |
| 384 | if (SafetyInfo.anyBlockMayThrow()) |
| 385 | return false; |
| 386 | |
| 387 | bool MadeChange = false; |
| 388 | |
| 389 | // Scan all the blocks in the loop that are not in subloops. |
| 390 | for (auto *BB : CurLoop->getBlocks()) { |
| 391 | // Ignore blocks in subloops. |
| 392 | if (LI->getLoopFor(BB) != CurLoop) |
| 393 | continue; |
| 394 | |
| 395 | MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); |
| 396 | } |
| 397 | return MadeChange; |
| 398 | } |
| 399 | |
| 400 | static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { |
| 401 | const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1)); |
| 402 | return ConstStride->getAPInt(); |
| 403 | } |
| 404 | |
| 405 | /// getMemSetPatternValue - If a strided store of the specified value is safe to |
| 406 | /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should |
| 407 | /// be passed in. Otherwise, return null. |
| 408 | /// |
| 409 | /// Note that we don't ever attempt to use memset_pattern8 or 4, because these |
| 410 | /// just replicate their input array and then pass on to memset_pattern16. |
| 411 | static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { |
| 412 | // FIXME: This could check for UndefValue because it can be merged into any |
| 413 | // other valid pattern. |
| 414 | |
| 415 | // If the value isn't a constant, we can't promote it to being in a constant |
| 416 | // array. We could theoretically do a store to an alloca or something, but |
| 417 | // that doesn't seem worthwhile. |
| 418 | Constant *C = dyn_cast<Constant>(V); |
| 419 | if (!C) |
| 420 | return nullptr; |
| 421 | |
| 422 | // Only handle simple values that are a power of two bytes in size. |
| 423 | uint64_t Size = DL->getTypeSizeInBits(V->getType()); |
| 424 | if (Size == 0 || (Size & 7) || (Size & (Size - 1))) |
| 425 | return nullptr; |
| 426 | |
| 427 | // Don't care enough about darwin/ppc to implement this. |
| 428 | if (DL->isBigEndian()) |
| 429 | return nullptr; |
| 430 | |
| 431 | // Convert to size in bytes. |
| 432 | Size /= 8; |
| 433 | |
| 434 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
| 435 | // if the top and bottom are the same (e.g. for vectors and large integers). |
| 436 | if (Size > 16) |
| 437 | return nullptr; |
| 438 | |
| 439 | // If the constant is exactly 16 bytes, just use it. |
| 440 | if (Size == 16) |
| 441 | return C; |
| 442 | |
| 443 | // Otherwise, we'll use an array of the constants. |
| 444 | unsigned ArraySize = 16 / Size; |
| 445 | ArrayType *AT = ArrayType::get(V->getType(), ArraySize); |
| 446 | return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C)); |
| 447 | } |
| 448 | |
| 449 | LoopIdiomRecognize::LegalStoreKind |
| 450 | LoopIdiomRecognize::isLegalStore(StoreInst *SI) { |
| 451 | // Don't touch volatile stores. |
| 452 | if (SI->isVolatile()) |
| 453 | return LegalStoreKind::None; |
| 454 | // We only want simple or unordered-atomic stores. |
| 455 | if (!SI->isUnordered()) |
| 456 | return LegalStoreKind::None; |
| 457 | |
| 458 | // Don't convert stores of non-integral pointer types to memsets (which stores |
| 459 | // integers). |
| 460 | if (DL->isNonIntegralPointerType(SI->getValueOperand()->getType())) |
| 461 | return LegalStoreKind::None; |
| 462 | |
| 463 | // Avoid merging nontemporal stores. |
| 464 | if (SI->getMetadata(LLVMContext::MD_nontemporal)) |
| 465 | return LegalStoreKind::None; |
| 466 | |
| 467 | Value *StoredVal = SI->getValueOperand(); |
| 468 | Value *StorePtr = SI->getPointerOperand(); |
| 469 | |
| 470 | // Reject stores that are so large that they overflow an unsigned. |
| 471 | // When storing out scalable vectors we bail out for now, since the code |
| 472 | // below currently only works for constant strides. |
| 473 | TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); |
| 474 | if (SizeInBits.isScalable() || (SizeInBits.getFixedSize() & 7) || |
| 475 | (SizeInBits.getFixedSize() >> 32) != 0) |
| 476 | return LegalStoreKind::None; |
| 477 | |
| 478 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 479 | // loop, which indicates a strided store. If we have something else, it's a |
| 480 | // random store we can't handle. |
| 481 | const SCEVAddRecExpr *StoreEv = |
| 482 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 483 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
| 484 | return LegalStoreKind::None; |
| 485 | |
| 486 | // Check to see if we have a constant stride. |
| 487 | if (!isa<SCEVConstant>(StoreEv->getOperand(1))) |
| 488 | return LegalStoreKind::None; |
| 489 | |
| 490 | // See if the store can be turned into a memset. |
| 491 | |
| 492 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
| 493 | // turned into a memset of i8 -1, assuming that all the consecutive bytes |
| 494 | // are stored. A store of i32 0x01020304 can never be turned into a memset, |
| 495 | // but it can be turned into memset_pattern if the target supports it. |
| 496 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
| 497 | Constant *PatternValue = nullptr; |
| 498 | |
| 499 | // Note: memset and memset_pattern on unordered-atomic is yet not supported |
| 500 | bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); |
| 501 | |
| 502 | // If we're allowed to form a memset, and the stored value would be |
| 503 | // acceptable for memset, use it. |
| 504 | if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && |
| 505 | // Verify that the stored value is loop invariant. If not, we can't |
| 506 | // promote the memset. |
| 507 | CurLoop->isLoopInvariant(SplatValue)) { |
| 508 | // It looks like we can use SplatValue. |
| 509 | return LegalStoreKind::Memset; |
| 510 | } else if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && |
| 511 | // Don't create memset_pattern16s with address spaces. |
| 512 | StorePtr->getType()->getPointerAddressSpace() == 0 && |
| 513 | (PatternValue = getMemSetPatternValue(StoredVal, DL))) { |
Although the value stored to 'PatternValue' is used in the enclosing expression, the value is never actually read from 'PatternValue' | |
| 514 | // It looks like we can use PatternValue! |
| 515 | return LegalStoreKind::MemsetPattern; |
| 516 | } |
| 517 | |
| 518 | // Otherwise, see if the store can be turned into a memcpy. |
| 519 | if (HasMemcpy && !DisableLIRP::Memcpy) { |
| 520 | // Check to see if the stride matches the size of the store. If so, then we |
| 521 | // know that every byte is touched in the loop. |
| 522 | APInt Stride = getStoreStride(StoreEv); |
| 523 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
| 524 | if (StoreSize != Stride && StoreSize != -Stride) |
| 525 | return LegalStoreKind::None; |
| 526 | |
| 527 | // The store must be feeding a non-volatile load. |
| 528 | LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); |
| 529 | |
| 530 | // Only allow non-volatile loads |
| 531 | if (!LI || LI->isVolatile()) |
| 532 | return LegalStoreKind::None; |
| 533 | // Only allow simple or unordered-atomic loads |
| 534 | if (!LI->isUnordered()) |
| 535 | return LegalStoreKind::None; |
| 536 | |
| 537 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 538 | // loop, which indicates a strided load. If we have something else, it's a |
| 539 | // random load we can't handle. |
| 540 | const SCEVAddRecExpr *LoadEv = |
| 541 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
| 542 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
| 543 | return LegalStoreKind::None; |
| 544 | |
| 545 | // The store and load must share the same stride. |
| 546 | if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) |
| 547 | return LegalStoreKind::None; |
| 548 | |
| 549 | // Success. This store can be converted into a memcpy. |
| 550 | UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); |
| 551 | return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy |
| 552 | : LegalStoreKind::Memcpy; |
| 553 | } |
| 554 | // This store can't be transformed into a memset/memcpy. |
| 555 | return LegalStoreKind::None; |
| 556 | } |
| 557 | |
| 558 | void LoopIdiomRecognize::collectStores(BasicBlock *BB) { |
| 559 | StoreRefsForMemset.clear(); |
| 560 | StoreRefsForMemsetPattern.clear(); |
| 561 | StoreRefsForMemcpy.clear(); |
| 562 | for (Instruction &I : *BB) { |
| 563 | StoreInst *SI = dyn_cast<StoreInst>(&I); |
| 564 | if (!SI) |
| 565 | continue; |
| 566 | |
| 567 | // Make sure this is a strided store with a constant stride. |
| 568 | switch (isLegalStore(SI)) { |
| 569 | case LegalStoreKind::None: |
| 570 | // Nothing to do |
| 571 | break; |
| 572 | case LegalStoreKind::Memset: { |
| 573 | // Find the base pointer. |
| 574 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
| 575 | StoreRefsForMemset[Ptr].push_back(SI); |
| 576 | } break; |
| 577 | case LegalStoreKind::MemsetPattern: { |
| 578 | // Find the base pointer. |
| 579 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
| 580 | StoreRefsForMemsetPattern[Ptr].push_back(SI); |
| 581 | } break; |
| 582 | case LegalStoreKind::Memcpy: |
| 583 | case LegalStoreKind::UnorderedAtomicMemcpy: |
| 584 | StoreRefsForMemcpy.push_back(SI); |
| 585 | break; |
| 586 | default: |
| 587 | assert(false && "unhandled return value")((false && "unhandled return value") ? static_cast< void> (0) : __assert_fail ("false && \"unhandled return value\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 587, __PRETTY_FUNCTION__)); |
| 588 | break; |
| 589 | } |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
| 594 | /// with the specified backedge count. This block is known to be in the current |
| 595 | /// loop and not in any subloops. |
| 596 | bool LoopIdiomRecognize::runOnLoopBlock( |
| 597 | BasicBlock *BB, const SCEV *BECount, |
| 598 | SmallVectorImpl<BasicBlock *> &ExitBlocks) { |
| 599 | // We can only promote stores in this block if they are unconditionally |
| 600 | // executed in the loop. For a block to be unconditionally executed, it has |
| 601 | // to dominate all the exit blocks of the loop. Verify this now. |
| 602 | for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) |
| 603 | if (!DT->dominates(BB, ExitBlocks[i])) |
| 604 | return false; |
| 605 | |
| 606 | bool MadeChange = false; |
| 607 | // Look for store instructions, which may be optimized to memset/memcpy. |
| 608 | collectStores(BB); |
| 609 | |
| 610 | // Look for a single store or sets of stores with a common base, which can be |
| 611 | // optimized into a memset (memset_pattern). The latter most commonly happens |
| 612 | // with structs and handunrolled loops. |
| 613 | for (auto &SL : StoreRefsForMemset) |
| 614 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); |
| 615 | |
| 616 | for (auto &SL : StoreRefsForMemsetPattern) |
| 617 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); |
| 618 | |
| 619 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
| 620 | for (auto &SI : StoreRefsForMemcpy) |
| 621 | MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); |
| 622 | |
| 623 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
| 624 | Instruction *Inst = &*I++; |
| 625 | // Look for memset instructions, which may be optimized to a larger memset. |
| 626 | if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) { |
| 627 | WeakTrackingVH InstPtr(&*I); |
| 628 | if (!processLoopMemSet(MSI, BECount)) |
| 629 | continue; |
| 630 | MadeChange = true; |
| 631 | |
| 632 | // If processing the memset invalidated our iterator, start over from the |
| 633 | // top of the block. |
| 634 | if (!InstPtr) |
| 635 | I = BB->begin(); |
| 636 | continue; |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | return MadeChange; |
| 641 | } |
| 642 | |
| 643 | /// See if this store(s) can be promoted to a memset. |
| 644 | bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL, |
| 645 | const SCEV *BECount, ForMemset For) { |
| 646 | // Try to find consecutive stores that can be transformed into memsets. |
| 647 | SetVector<StoreInst *> Heads, Tails; |
| 648 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; |
| 649 | |
| 650 | // Do a quadratic search on all of the given stores and find |
| 651 | // all of the pairs of stores that follow each other. |
| 652 | SmallVector<unsigned, 16> IndexQueue; |
| 653 | for (unsigned i = 0, e = SL.size(); i < e; ++i) { |
| 654 | assert(SL[i]->isSimple() && "Expected only non-volatile stores.")((SL[i]->isSimple() && "Expected only non-volatile stores." ) ? static_cast<void> (0) : __assert_fail ("SL[i]->isSimple() && \"Expected only non-volatile stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 654, __PRETTY_FUNCTION__)); |
| 655 | |
| 656 | Value *FirstStoredVal = SL[i]->getValueOperand(); |
| 657 | Value *FirstStorePtr = SL[i]->getPointerOperand(); |
| 658 | const SCEVAddRecExpr *FirstStoreEv = |
| 659 | cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr)); |
| 660 | APInt FirstStride = getStoreStride(FirstStoreEv); |
| 661 | unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); |
| 662 | |
| 663 | // See if we can optimize just this store in isolation. |
| 664 | if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { |
| 665 | Heads.insert(SL[i]); |
| 666 | continue; |
| 667 | } |
| 668 | |
| 669 | Value *FirstSplatValue = nullptr; |
| 670 | Constant *FirstPatternValue = nullptr; |
| 671 | |
| 672 | if (For == ForMemset::Yes) |
| 673 | FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); |
| 674 | else |
| 675 | FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); |
| 676 | |
| 677 | assert((FirstSplatValue || FirstPatternValue) &&(((FirstSplatValue || FirstPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 678, __PRETTY_FUNCTION__)) |
| 678 | "Expected either splat value or pattern value.")(((FirstSplatValue || FirstPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 678, __PRETTY_FUNCTION__)); |
| 679 | |
| 680 | IndexQueue.clear(); |
| 681 | // If a store has multiple consecutive store candidates, search Stores |
| 682 | // array according to the sequence: from i+1 to e, then from i-1 to 0. |
| 683 | // This is because usually pairing with immediate succeeding or preceding |
| 684 | // candidate create the best chance to find memset opportunity. |
| 685 | unsigned j = 0; |
| 686 | for (j = i + 1; j < e; ++j) |
| 687 | IndexQueue.push_back(j); |
| 688 | for (j = i; j > 0; --j) |
| 689 | IndexQueue.push_back(j - 1); |
| 690 | |
| 691 | for (auto &k : IndexQueue) { |
| 692 | assert(SL[k]->isSimple() && "Expected only non-volatile stores.")((SL[k]->isSimple() && "Expected only non-volatile stores." ) ? static_cast<void> (0) : __assert_fail ("SL[k]->isSimple() && \"Expected only non-volatile stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 692, __PRETTY_FUNCTION__)); |
| 693 | Value *SecondStorePtr = SL[k]->getPointerOperand(); |
| 694 | const SCEVAddRecExpr *SecondStoreEv = |
| 695 | cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr)); |
| 696 | APInt SecondStride = getStoreStride(SecondStoreEv); |
| 697 | |
| 698 | if (FirstStride != SecondStride) |
| 699 | continue; |
| 700 | |
| 701 | Value *SecondStoredVal = SL[k]->getValueOperand(); |
| 702 | Value *SecondSplatValue = nullptr; |
| 703 | Constant *SecondPatternValue = nullptr; |
| 704 | |
| 705 | if (For == ForMemset::Yes) |
| 706 | SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); |
| 707 | else |
| 708 | SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); |
| 709 | |
| 710 | assert((SecondSplatValue || SecondPatternValue) &&(((SecondSplatValue || SecondPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 711, __PRETTY_FUNCTION__)) |
| 711 | "Expected either splat value or pattern value.")(((SecondSplatValue || SecondPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 711, __PRETTY_FUNCTION__)); |
| 712 | |
| 713 | if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { |
| 714 | if (For == ForMemset::Yes) { |
| 715 | if (isa<UndefValue>(FirstSplatValue)) |
| 716 | FirstSplatValue = SecondSplatValue; |
| 717 | if (FirstSplatValue != SecondSplatValue) |
| 718 | continue; |
| 719 | } else { |
| 720 | if (isa<UndefValue>(FirstPatternValue)) |
| 721 | FirstPatternValue = SecondPatternValue; |
| 722 | if (FirstPatternValue != SecondPatternValue) |
| 723 | continue; |
| 724 | } |
| 725 | Tails.insert(SL[k]); |
| 726 | Heads.insert(SL[i]); |
| 727 | ConsecutiveChain[SL[i]] = SL[k]; |
| 728 | break; |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | // We may run into multiple chains that merge into a single chain. We mark the |
| 734 | // stores that we transformed so that we don't visit the same store twice. |
| 735 | SmallPtrSet<Value *, 16> TransformedStores; |
| 736 | bool Changed = false; |
| 737 | |
| 738 | // For stores that start but don't end a link in the chain: |
| 739 | for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end(); |
| 740 | it != e; ++it) { |
| 741 | if (Tails.count(*it)) |
| 742 | continue; |
| 743 | |
| 744 | // We found a store instr that starts a chain. Now follow the chain and try |
| 745 | // to transform it. |
| 746 | SmallPtrSet<Instruction *, 8> AdjacentStores; |
| 747 | StoreInst *I = *it; |
| 748 | |
| 749 | StoreInst *HeadStore = I; |
| 750 | unsigned StoreSize = 0; |
| 751 | |
| 752 | // Collect the chain into a list. |
| 753 | while (Tails.count(I) || Heads.count(I)) { |
| 754 | if (TransformedStores.count(I)) |
| 755 | break; |
| 756 | AdjacentStores.insert(I); |
| 757 | |
| 758 | StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); |
| 759 | // Move to the next value in the chain. |
| 760 | I = ConsecutiveChain[I]; |
| 761 | } |
| 762 | |
| 763 | Value *StoredVal = HeadStore->getValueOperand(); |
| 764 | Value *StorePtr = HeadStore->getPointerOperand(); |
| 765 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 766 | APInt Stride = getStoreStride(StoreEv); |
| 767 | |
| 768 | // Check to see if the stride matches the size of the stores. If so, then |
| 769 | // we know that every byte is touched in the loop. |
| 770 | if (StoreSize != Stride && StoreSize != -Stride) |
| 771 | continue; |
| 772 | |
| 773 | bool NegStride = StoreSize == -Stride; |
| 774 | |
| 775 | if (processLoopStridedStore(StorePtr, StoreSize, |
| 776 | MaybeAlign(HeadStore->getAlignment()), |
| 777 | StoredVal, HeadStore, AdjacentStores, StoreEv, |
| 778 | BECount, NegStride)) { |
| 779 | TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); |
| 780 | Changed = true; |
| 781 | } |
| 782 | } |
| 783 | |
| 784 | return Changed; |
| 785 | } |
| 786 | |
| 787 | /// processLoopMemSet - See if this memset can be promoted to a large memset. |
| 788 | bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, |
| 789 | const SCEV *BECount) { |
| 790 | // We can only handle non-volatile memsets with a constant size. |
| 791 | if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) |
| 792 | return false; |
| 793 | |
| 794 | // If we're not allowed to hack on memset, we fail. |
| 795 | if (!HasMemset) |
| 796 | return false; |
| 797 | |
| 798 | Value *Pointer = MSI->getDest(); |
| 799 | |
| 800 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 801 | // loop, which indicates a strided store. If we have something else, it's a |
| 802 | // random store we can't handle. |
| 803 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer)); |
| 804 | if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine()) |
| 805 | return false; |
| 806 | |
| 807 | // Reject memsets that are so large that they overflow an unsigned. |
| 808 | uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); |
| 809 | if ((SizeInBytes >> 32) != 0) |
| 810 | return false; |
| 811 | |
| 812 | // Check to see if the stride matches the size of the memset. If so, then we |
| 813 | // know that every byte is touched in the loop. |
| 814 | const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
| 815 | if (!ConstStride) |
| 816 | return false; |
| 817 | |
| 818 | APInt Stride = ConstStride->getAPInt(); |
| 819 | if (SizeInBytes != Stride && SizeInBytes != -Stride) |
| 820 | return false; |
| 821 | |
| 822 | // Verify that the memset value is loop invariant. If not, we can't promote |
| 823 | // the memset. |
| 824 | Value *SplatValue = MSI->getValue(); |
| 825 | if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) |
| 826 | return false; |
| 827 | |
| 828 | SmallPtrSet<Instruction *, 1> MSIs; |
| 829 | MSIs.insert(MSI); |
| 830 | bool NegStride = SizeInBytes == -Stride; |
| 831 | return processLoopStridedStore( |
| 832 | Pointer, (unsigned)SizeInBytes, MaybeAlign(MSI->getDestAlignment()), |
| 833 | SplatValue, MSI, MSIs, Ev, BECount, NegStride, /*IsLoopMemset=*/true); |
| 834 | } |
| 835 | |
| 836 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
| 837 | /// specified pointer location, which is a loop-strided access. The 'Access' |
| 838 | /// argument specifies what the verboten forms of access are (read or write). |
| 839 | static bool |
| 840 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
| 841 | const SCEV *BECount, unsigned StoreSize, |
| 842 | AliasAnalysis &AA, |
| 843 | SmallPtrSetImpl<Instruction *> &IgnoredStores) { |
| 844 | // Get the location that may be stored across the loop. Since the access is |
| 845 | // strided positively through memory, we say that the modified location starts |
| 846 | // at the pointer and has infinite size. |
| 847 | LocationSize AccessSize = LocationSize::unknown(); |
| 848 | |
| 849 | // If the loop iterates a fixed number of times, we can refine the access size |
| 850 | // to be exactly the size of the memset, which is (BECount+1)*StoreSize |
| 851 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
| 852 | AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * |
| 853 | StoreSize); |
| 854 | |
| 855 | // TODO: For this to be really effective, we have to dive into the pointer |
| 856 | // operand in the store. Store to &A[i] of 100 will always return may alias |
| 857 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
| 858 | // which will then no-alias a store to &A[100]. |
| 859 | MemoryLocation StoreLoc(Ptr, AccessSize); |
| 860 | |
| 861 | for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; |
| 862 | ++BI) |
| 863 | for (Instruction &I : **BI) |
| 864 | if (IgnoredStores.count(&I) == 0 && |
| 865 | isModOrRefSet( |
| 866 | intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access))) |
| 867 | return true; |
| 868 | |
| 869 | return false; |
| 870 | } |
| 871 | |
| 872 | // If we have a negative stride, Start refers to the end of the memory location |
| 873 | // we're trying to memset. Therefore, we need to recompute the base pointer, |
| 874 | // which is just Start - BECount*Size. |
| 875 | static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, |
| 876 | Type *IntPtr, unsigned StoreSize, |
| 877 | ScalarEvolution *SE) { |
| 878 | const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); |
| 879 | if (StoreSize != 1) |
| 880 | Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize), |
| 881 | SCEV::FlagNUW); |
| 882 | return SE->getMinusSCEV(Start, Index); |
| 883 | } |
| 884 | |
| 885 | /// Compute the number of bytes as a SCEV from the backedge taken count. |
| 886 | /// |
| 887 | /// This also maps the SCEV into the provided type and tries to handle the |
| 888 | /// computation in a way that will fold cleanly. |
| 889 | static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, |
| 890 | unsigned StoreSize, Loop *CurLoop, |
| 891 | const DataLayout *DL, ScalarEvolution *SE) { |
| 892 | const SCEV *NumBytesS; |
| 893 | // The # stored bytes is (BECount+1)*Size. Expand the trip count out to |
| 894 | // pointer size if it isn't already. |
| 895 | // |
| 896 | // If we're going to need to zero extend the BE count, check if we can add |
| 897 | // one to it prior to zero extending without overflow. Provided this is safe, |
| 898 | // it allows better simplification of the +1. |
| 899 | if (DL->getTypeSizeInBits(BECount->getType()) < |
| 900 | DL->getTypeSizeInBits(IntPtr) && |
| 901 | SE->isLoopEntryGuardedByCond( |
| 902 | CurLoop, ICmpInst::ICMP_NE, BECount, |
| 903 | SE->getNegativeSCEV(SE->getOne(BECount->getType())))) { |
| 904 | NumBytesS = SE->getZeroExtendExpr( |
| 905 | SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW), |
| 906 | IntPtr); |
| 907 | } else { |
| 908 | NumBytesS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr), |
| 909 | SE->getOne(IntPtr), SCEV::FlagNUW); |
| 910 | } |
| 911 | |
| 912 | // And scale it based on the store size. |
| 913 | if (StoreSize != 1) { |
| 914 | NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), |
| 915 | SCEV::FlagNUW); |
| 916 | } |
| 917 | return NumBytesS; |
| 918 | } |
| 919 | |
| 920 | /// processLoopStridedStore - We see a strided store of some value. If we can |
| 921 | /// transform this into a memset or memset_pattern in the loop preheader, do so. |
| 922 | bool LoopIdiomRecognize::processLoopStridedStore( |
| 923 | Value *DestPtr, unsigned StoreSize, MaybeAlign StoreAlignment, |
| 924 | Value *StoredVal, Instruction *TheStore, |
| 925 | SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev, |
| 926 | const SCEV *BECount, bool NegStride, bool IsLoopMemset) { |
| 927 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
| 928 | Constant *PatternValue = nullptr; |
| 929 | |
| 930 | if (!SplatValue) |
| 931 | PatternValue = getMemSetPatternValue(StoredVal, DL); |
| 932 | |
| 933 | assert((SplatValue || PatternValue) &&(((SplatValue || PatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 934, __PRETTY_FUNCTION__)) |
| 934 | "Expected either splat value or pattern value.")(((SplatValue || PatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 934, __PRETTY_FUNCTION__)); |
| 935 | |
| 936 | // The trip count of the loop and the base pointer of the addrec SCEV is |
| 937 | // guaranteed to be loop invariant, which means that it should dominate the |
| 938 | // header. This allows us to insert code for it in the preheader. |
| 939 | unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); |
| 940 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
| 941 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 942 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
| 943 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
| 944 | |
| 945 | Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); |
| 946 | Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); |
| 947 | |
| 948 | bool Changed = false; |
| 949 | const SCEV *Start = Ev->getStart(); |
| 950 | // Handle negative strided loops. |
| 951 | if (NegStride) |
| 952 | Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSize, SE); |
| 953 | |
| 954 | // TODO: ideally we should still be able to generate memset if SCEV expander |
| 955 | // is taught to generate the dependencies at the latest point. |
| 956 | if (!isSafeToExpand(Start, *SE)) |
| 957 | return Changed; |
| 958 | |
| 959 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
| 960 | // this into a memset in the loop preheader now if we want. However, this |
| 961 | // would be unsafe to do if there is anything else in the loop that may read |
| 962 | // or write to the aliased location. Check for any overlap by generating the |
| 963 | // base pointer and checking the region. |
| 964 | Value *BasePtr = |
| 965 | Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); |
| 966 | |
| 967 | // From here on out, conservatively report to the pass manager that we've |
| 968 | // changed the IR, even if we later clean up these added instructions. There |
| 969 | // may be structural differences e.g. in the order of use lists not accounted |
| 970 | // for in just a textual dump of the IR. This is written as a variable, even |
| 971 | // though statically all the places this dominates could be replaced with |
| 972 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
| 973 | // the return value will read this comment, and leave them alone. |
| 974 | Changed = true; |
| 975 | |
| 976 | if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
| 977 | StoreSize, *AA, Stores)) |
| 978 | return Changed; |
| 979 | |
| 980 | if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) |
| 981 | return Changed; |
| 982 | |
| 983 | // Okay, everything looks good, insert the memset. |
| 984 | |
| 985 | const SCEV *NumBytesS = |
| 986 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
| 987 | |
| 988 | // TODO: ideally we should still be able to generate memset if SCEV expander |
| 989 | // is taught to generate the dependencies at the latest point. |
| 990 | if (!isSafeToExpand(NumBytesS, *SE)) |
| 991 | return Changed; |
| 992 | |
| 993 | Value *NumBytes = |
| 994 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
| 995 | |
| 996 | CallInst *NewCall; |
| 997 | if (SplatValue) { |
| 998 | NewCall = Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, |
| 999 | MaybeAlign(StoreAlignment)); |
| 1000 | } else { |
| 1001 | // Everything is emitted in default address space |
| 1002 | Type *Int8PtrTy = DestInt8PtrTy; |
| 1003 | |
| 1004 | Module *M = TheStore->getModule(); |
| 1005 | StringRef FuncName = "memset_pattern16"; |
| 1006 | FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(), |
| 1007 | Int8PtrTy, Int8PtrTy, IntIdxTy); |
| 1008 | inferLibFuncAttributes(M, FuncName, *TLI); |
| 1009 | |
| 1010 | // Otherwise we should form a memset_pattern16. PatternValue is known to be |
| 1011 | // an constant array of 16-bytes. Plop the value into a mergable global. |
| 1012 | GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, |
| 1013 | GlobalValue::PrivateLinkage, |
| 1014 | PatternValue, ".memset_pattern"); |
| 1015 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
| 1016 | GV->setAlignment(Align(16)); |
| 1017 | Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); |
| 1018 | NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); |
| 1019 | } |
| 1020 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
| 1021 | |
| 1022 | if (MSSAU) { |
| 1023 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
| 1024 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
| 1025 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
| 1026 | } |
| 1027 | |
| 1028 | LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false) |
| 1029 | << " from store to: " << *Ev << " at: " << *TheStoredo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false) |
| 1030 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false); |
| 1031 | |
| 1032 | ORE.emit([&]() { |
| 1033 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStridedStore", |
| 1034 | NewCall->getDebugLoc(), Preheader) |
| 1035 | << "Transformed loop-strided store into a call to " |
| 1036 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
| 1037 | << "() function"; |
| 1038 | }); |
| 1039 | |
| 1040 | // Okay, the memset has been formed. Zap the original store and anything that |
| 1041 | // feeds into it. |
| 1042 | for (auto *I : Stores) { |
| 1043 | if (MSSAU) |
| 1044 | MSSAU->removeMemoryAccess(I, true); |
| 1045 | deleteDeadInstruction(I); |
| 1046 | } |
| 1047 | if (MSSAU && VerifyMemorySSA) |
| 1048 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
| 1049 | ++NumMemSet; |
| 1050 | ExpCleaner.markResultUsed(); |
| 1051 | return true; |
| 1052 | } |
| 1053 | |
| 1054 | /// If the stored value is a strided load in the same loop with the same stride |
| 1055 | /// this may be transformable into a memcpy. This kicks in for stuff like |
| 1056 | /// for (i) A[i] = B[i]; |
| 1057 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, |
| 1058 | const SCEV *BECount) { |
| 1059 | assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.")((SI->isUnordered() && "Expected only non-volatile non-ordered stores." ) ? static_cast<void> (0) : __assert_fail ("SI->isUnordered() && \"Expected only non-volatile non-ordered stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1059, __PRETTY_FUNCTION__)); |
| 1060 | |
| 1061 | Value *StorePtr = SI->getPointerOperand(); |
| 1062 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 1063 | APInt Stride = getStoreStride(StoreEv); |
| 1064 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
| 1065 | bool NegStride = StoreSize == -Stride; |
| 1066 | |
| 1067 | // The store must be feeding a non-volatile load. |
| 1068 | LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); |
| 1069 | assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.")((LI->isUnordered() && "Expected only non-volatile non-ordered loads." ) ? static_cast<void> (0) : __assert_fail ("LI->isUnordered() && \"Expected only non-volatile non-ordered loads.\"" , "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1069, __PRETTY_FUNCTION__)); |
| 1070 | |
| 1071 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 1072 | // loop, which indicates a strided load. If we have something else, it's a |
| 1073 | // random load we can't handle. |
| 1074 | const SCEVAddRecExpr *LoadEv = |
| 1075 | cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
| 1076 | |
| 1077 | // The trip count of the loop and the base pointer of the addrec SCEV is |
| 1078 | // guaranteed to be loop invariant, which means that it should dominate the |
| 1079 | // header. This allows us to insert code for it in the preheader. |
| 1080 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
| 1081 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 1082 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
| 1083 | |
| 1084 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
| 1085 | |
| 1086 | bool Changed = false; |
| 1087 | const SCEV *StrStart = StoreEv->getStart(); |
| 1088 | unsigned StrAS = SI->getPointerAddressSpace(); |
| 1089 | Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); |
| 1090 | |
| 1091 | // Handle negative strided loops. |
| 1092 | if (NegStride) |
| 1093 | StrStart = getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSize, SE); |
| 1094 | |
| 1095 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
| 1096 | // this into a memcpy in the loop preheader now if we want. However, this |
| 1097 | // would be unsafe to do if there is anything else in the loop that may read |
| 1098 | // or write the memory region we're storing to. This includes the load that |
| 1099 | // feeds the stores. Check for an alias by generating the base address and |
| 1100 | // checking everything. |
| 1101 | Value *StoreBasePtr = Expander.expandCodeFor( |
| 1102 | StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); |
| 1103 | |
| 1104 | // From here on out, conservatively report to the pass manager that we've |
| 1105 | // changed the IR, even if we later clean up these added instructions. There |
| 1106 | // may be structural differences e.g. in the order of use lists not accounted |
| 1107 | // for in just a textual dump of the IR. This is written as a variable, even |
| 1108 | // though statically all the places this dominates could be replaced with |
| 1109 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
| 1110 | // the return value will read this comment, and leave them alone. |
| 1111 | Changed = true; |
| 1112 | |
| 1113 | SmallPtrSet<Instruction *, 1> Stores; |
| 1114 | Stores.insert(SI); |
| 1115 | if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
| 1116 | StoreSize, *AA, Stores)) |
| 1117 | return Changed; |
| 1118 | |
| 1119 | const SCEV *LdStart = LoadEv->getStart(); |
| 1120 | unsigned LdAS = LI->getPointerAddressSpace(); |
| 1121 | |
| 1122 | // Handle negative strided loops. |
| 1123 | if (NegStride) |
| 1124 | LdStart = getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSize, SE); |
| 1125 | |
| 1126 | // For a memcpy, we have to make sure that the input array is not being |
| 1127 | // mutated by the loop. |
| 1128 | Value *LoadBasePtr = Expander.expandCodeFor( |
| 1129 | LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); |
| 1130 | |
| 1131 | if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, |
| 1132 | StoreSize, *AA, Stores)) |
| 1133 | return Changed; |
| 1134 | |
| 1135 | if (avoidLIRForMultiBlockLoop()) |
| 1136 | return Changed; |
| 1137 | |
| 1138 | // Okay, everything is safe, we can transform this! |
| 1139 | |
| 1140 | const SCEV *NumBytesS = |
| 1141 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
| 1142 | |
| 1143 | Value *NumBytes = |
| 1144 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
| 1145 | |
| 1146 | CallInst *NewCall = nullptr; |
| 1147 | // Check whether to generate an unordered atomic memcpy: |
| 1148 | // If the load or store are atomic, then they must necessarily be unordered |
| 1149 | // by previous checks. |
| 1150 | if (!SI->isAtomic() && !LI->isAtomic()) |
| 1151 | NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr, |
| 1152 | LI->getAlign(), NumBytes); |
| 1153 | else { |
| 1154 | // We cannot allow unaligned ops for unordered load/store, so reject |
| 1155 | // anything where the alignment isn't at least the element size. |
| 1156 | const Align StoreAlign = SI->getAlign(); |
| 1157 | const Align LoadAlign = LI->getAlign(); |
| 1158 | if (StoreAlign < StoreSize || LoadAlign < StoreSize) |
| 1159 | return Changed; |
| 1160 | |
| 1161 | // If the element.atomic memcpy is not lowered into explicit |
| 1162 | // loads/stores later, then it will be lowered into an element-size |
| 1163 | // specific lib call. If the lib call doesn't exist for our store size, then |
| 1164 | // we shouldn't generate the memcpy. |
| 1165 | if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) |
| 1166 | return Changed; |
| 1167 | |
| 1168 | // Create the call. |
| 1169 | // Note that unordered atomic loads/stores are *required* by the spec to |
| 1170 | // have an alignment but non-atomic loads/stores may not. |
| 1171 | NewCall = Builder.CreateElementUnorderedAtomicMemCpy( |
| 1172 | StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, |
| 1173 | StoreSize); |
| 1174 | } |
| 1175 | NewCall->setDebugLoc(SI->getDebugLoc()); |
| 1176 | |
| 1177 | if (MSSAU) { |
| 1178 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
| 1179 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
| 1180 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
| 1181 | } |
| 1182 | |
| 1183 | LLVM_DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
| 1184 | << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
| 1185 | << " from store ptr=" << *StoreEv << " at: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
| 1186 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false); |
| 1187 | |
| 1188 | ORE.emit([&]() { |
| 1189 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStoreOfLoopLoad", |
| 1190 | NewCall->getDebugLoc(), Preheader) |
| 1191 | << "Formed a call to " |
| 1192 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
| 1193 | << "() function"; |
| 1194 | }); |
| 1195 | |
| 1196 | // Okay, the memcpy has been formed. Zap the original store and anything that |
| 1197 | // feeds into it. |
| 1198 | if (MSSAU) |
| 1199 | MSSAU->removeMemoryAccess(SI, true); |
| 1200 | deleteDeadInstruction(SI); |
| 1201 | if (MSSAU && VerifyMemorySSA) |
| 1202 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
| 1203 | ++NumMemCpy; |
| 1204 | ExpCleaner.markResultUsed(); |
| 1205 | return true; |
| 1206 | } |
| 1207 | |
| 1208 | // When compiling for codesize we avoid idiom recognition for a multi-block loop |
| 1209 | // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. |
| 1210 | // |
| 1211 | bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, |
| 1212 | bool IsLoopMemset) { |
| 1213 | if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { |
| 1214 | if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) { |
| 1215 | LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " " << CurLoop->getHeader ()->getParent()->getName() << " : LIR " << ( IsMemset ? "Memset" : "Memcpy") << " avoided: multi-block top-level loop\n" ; } } while (false) |
| 1216 | << " : LIR " << (IsMemset ? "Memset" : "Memcpy")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " " << CurLoop->getHeader ()->getParent()->getName() << " : LIR " << ( IsMemset ? "Memset" : "Memcpy") << " avoided: multi-block top-level loop\n" ; } } while (false) |
| 1217 | << " avoided: multi-block top-level loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " " << CurLoop->getHeader ()->getParent()->getName() << " : LIR " << ( IsMemset ? "Memset" : "Memcpy") << " avoided: multi-block top-level loop\n" ; } } while (false); |
| 1218 | return true; |
| 1219 | } |
| 1220 | } |
| 1221 | |
| 1222 | return false; |
| 1223 | } |
| 1224 | |
| 1225 | bool LoopIdiomRecognize::runOnNoncountableLoop() { |
| 1226 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Noncountable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 1227 | << CurLoop->getHeader()->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Noncountable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 1228 | << "] Noncountable Loop %"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Noncountable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
| 1229 | << CurLoop->getHeader()->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Noncountable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false); |
| 1230 | |
| 1231 | return recognizePopcount() || recognizeAndInsertFFS(); |
| 1232 | } |
| 1233 | |
| 1234 | /// Check if the given conditional branch is based on the comparison between |
| 1235 | /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is |
| 1236 | /// true), the control yields to the loop entry. If the branch matches the |
| 1237 | /// behavior, the variable involved in the comparison is returned. This function |
| 1238 | /// will be called to see if the precondition and postcondition of the loop are |
| 1239 | /// in desirable form. |
| 1240 | static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, |
| 1241 | bool JmpOnZero = false) { |
| 1242 | if (!BI || !BI->isConditional()) |
| 1243 | return nullptr; |
| 1244 | |
| 1245 | ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); |
| 1246 | if (!Cond) |
| 1247 | return nullptr; |
| 1248 | |
| 1249 | ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1)); |
| 1250 | if (!CmpZero || !CmpZero->isZero()) |
| 1251 | return nullptr; |
| 1252 | |
| 1253 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
| 1254 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
| 1255 | if (JmpOnZero) |
| 1256 | std::swap(TrueSucc, FalseSucc); |
| 1257 | |
| 1258 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
| 1259 | if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || |
| 1260 | (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) |
| 1261 | return Cond->getOperand(0); |
| 1262 | |
| 1263 | return nullptr; |
| 1264 | } |
| 1265 | |
| 1266 | // Check if the recurrence variable `VarX` is in the right form to create |
| 1267 | // the idiom. Returns the value coerced to a PHINode if so. |
| 1268 | static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, |
| 1269 | BasicBlock *LoopEntry) { |
| 1270 | auto *PhiX = dyn_cast<PHINode>(VarX); |
| 1271 | if (PhiX && PhiX->getParent() == LoopEntry && |
| 1272 | (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) |
| 1273 | return PhiX; |
| 1274 | return nullptr; |
| 1275 | } |
| 1276 | |
| 1277 | /// Return true iff the idiom is detected in the loop. |
| 1278 | /// |
| 1279 | /// Additionally: |
| 1280 | /// 1) \p CntInst is set to the instruction counting the population bit. |
| 1281 | /// 2) \p CntPhi is set to the corresponding phi node. |
| 1282 | /// 3) \p Var is set to the value whose population bits are being counted. |
| 1283 | /// |
| 1284 | /// The core idiom we are trying to detect is: |
| 1285 | /// \code |
| 1286 | /// if (x0 != 0) |
| 1287 | /// goto loop-exit // the precondition of the loop |
| 1288 | /// cnt0 = init-val; |
| 1289 | /// do { |
| 1290 | /// x1 = phi (x0, x2); |
| 1291 | /// cnt1 = phi(cnt0, cnt2); |
| 1292 | /// |
| 1293 | /// cnt2 = cnt1 + 1; |
| 1294 | /// ... |
| 1295 | /// x2 = x1 & (x1 - 1); |
| 1296 | /// ... |
| 1297 | /// } while(x != 0); |
| 1298 | /// |
| 1299 | /// loop-exit: |
| 1300 | /// \endcode |
| 1301 | static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, |
| 1302 | Instruction *&CntInst, PHINode *&CntPhi, |
| 1303 | Value *&Var) { |
| 1304 | // step 1: Check to see if the look-back branch match this pattern: |
| 1305 | // "if (a!=0) goto loop-entry". |
| 1306 | BasicBlock *LoopEntry; |
| 1307 | Instruction *DefX2, *CountInst; |
| 1308 | Value *VarX1, *VarX0; |
| 1309 | PHINode *PhiX, *CountPhi; |
| 1310 | |
| 1311 | DefX2 = CountInst = nullptr; |
| 1312 | VarX1 = VarX0 = nullptr; |
| 1313 | PhiX = CountPhi = nullptr; |
| 1314 | LoopEntry = *(CurLoop->block_begin()); |
| 1315 | |
| 1316 | // step 1: Check if the loop-back branch is in desirable form. |
| 1317 | { |
| 1318 | if (Value *T = matchCondition( |
| 1319 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
| 1320 | DefX2 = dyn_cast<Instruction>(T); |
| 1321 | else |
| 1322 | return false; |
| 1323 | } |
| 1324 | |
| 1325 | // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" |
| 1326 | { |
| 1327 | if (!DefX2 || DefX2->getOpcode() != Instruction::And) |
| 1328 | return false; |
| 1329 | |
| 1330 | BinaryOperator *SubOneOp; |
| 1331 | |
| 1332 | if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0)))) |
| 1333 | VarX1 = DefX2->getOperand(1); |
| 1334 | else { |
| 1335 | VarX1 = DefX2->getOperand(0); |
| 1336 | SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1)); |
| 1337 | } |
| 1338 | if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) |
| 1339 | return false; |
| 1340 | |
| 1341 | ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1)); |
| 1342 | if (!Dec || |
| 1343 | !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || |
| 1344 | (SubOneOp->getOpcode() == Instruction::Add && |
| 1345 | Dec->isMinusOne()))) { |
| 1346 | return false; |
| 1347 | } |
| 1348 | } |
| 1349 | |
| 1350 | // step 3: Check the recurrence of variable X |
| 1351 | PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); |
| 1352 | if (!PhiX) |
| 1353 | return false; |
| 1354 | |
| 1355 | // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 |
| 1356 | { |
| 1357 | CountInst = nullptr; |
| 1358 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
| 1359 | IterE = LoopEntry->end(); |
| 1360 | Iter != IterE; Iter++) { |
| 1361 | Instruction *Inst = &*Iter; |
| 1362 | if (Inst->getOpcode() != Instruction::Add) |
| 1363 | continue; |
| 1364 | |
| 1365 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
| 1366 | if (!Inc || !Inc->isOne()) |
| 1367 | continue; |
| 1368 | |
| 1369 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
| 1370 | if (!Phi) |
| 1371 | continue; |
| 1372 | |
| 1373 | // Check if the result of the instruction is live of the loop. |
| 1374 | bool LiveOutLoop = false; |
| 1375 | for (User *U : Inst->users()) { |
| 1376 | if ((cast<Instruction>(U))->getParent() != LoopEntry) { |
| 1377 | LiveOutLoop = true; |
| 1378 | break; |
| 1379 | } |
| 1380 | } |
| 1381 | |
| 1382 | if (LiveOutLoop) { |
| 1383 | CountInst = Inst; |
| 1384 | CountPhi = Phi; |
| 1385 | break; |
| 1386 | } |
| 1387 | } |
| 1388 | |
| 1389 | if (!CountInst) |
| 1390 | return false; |
| 1391 | } |
| 1392 | |
| 1393 | // step 5: check if the precondition is in this form: |
| 1394 | // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" |
| 1395 | { |
| 1396 | auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1397 | Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); |
| 1398 | if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) |
| 1399 | return false; |
| 1400 | |
| 1401 | CntInst = CountInst; |
| 1402 | CntPhi = CountPhi; |
| 1403 | Var = T; |
| 1404 | } |
| 1405 | |
| 1406 | return true; |
| 1407 | } |
| 1408 | |
| 1409 | /// Return true if the idiom is detected in the loop. |
| 1410 | /// |
| 1411 | /// Additionally: |
| 1412 | /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) |
| 1413 | /// or nullptr if there is no such. |
| 1414 | /// 2) \p CntPhi is set to the corresponding phi node |
| 1415 | /// or nullptr if there is no such. |
| 1416 | /// 3) \p Var is set to the value whose CTLZ could be used. |
| 1417 | /// 4) \p DefX is set to the instruction calculating Loop exit condition. |
| 1418 | /// |
| 1419 | /// The core idiom we are trying to detect is: |
| 1420 | /// \code |
| 1421 | /// if (x0 == 0) |
| 1422 | /// goto loop-exit // the precondition of the loop |
| 1423 | /// cnt0 = init-val; |
| 1424 | /// do { |
| 1425 | /// x = phi (x0, x.next); //PhiX |
| 1426 | /// cnt = phi(cnt0, cnt.next); |
| 1427 | /// |
| 1428 | /// cnt.next = cnt + 1; |
| 1429 | /// ... |
| 1430 | /// x.next = x >> 1; // DefX |
| 1431 | /// ... |
| 1432 | /// } while(x.next != 0); |
| 1433 | /// |
| 1434 | /// loop-exit: |
| 1435 | /// \endcode |
| 1436 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, |
| 1437 | Intrinsic::ID &IntrinID, Value *&InitX, |
| 1438 | Instruction *&CntInst, PHINode *&CntPhi, |
| 1439 | Instruction *&DefX) { |
| 1440 | BasicBlock *LoopEntry; |
| 1441 | Value *VarX = nullptr; |
| 1442 | |
| 1443 | DefX = nullptr; |
| 1444 | CntInst = nullptr; |
| 1445 | CntPhi = nullptr; |
| 1446 | LoopEntry = *(CurLoop->block_begin()); |
| 1447 | |
| 1448 | // step 1: Check if the loop-back branch is in desirable form. |
| 1449 | if (Value *T = matchCondition( |
| 1450 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
| 1451 | DefX = dyn_cast<Instruction>(T); |
| 1452 | else |
| 1453 | return false; |
| 1454 | |
| 1455 | // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" |
| 1456 | if (!DefX || !DefX->isShift()) |
| 1457 | return false; |
| 1458 | IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : |
| 1459 | Intrinsic::ctlz; |
| 1460 | ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)); |
| 1461 | if (!Shft || !Shft->isOne()) |
| 1462 | return false; |
| 1463 | VarX = DefX->getOperand(0); |
| 1464 | |
| 1465 | // step 3: Check the recurrence of variable X |
| 1466 | PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); |
| 1467 | if (!PhiX) |
| 1468 | return false; |
| 1469 | |
| 1470 | InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); |
| 1471 | |
| 1472 | // Make sure the initial value can't be negative otherwise the ashr in the |
| 1473 | // loop might never reach zero which would make the loop infinite. |
| 1474 | if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) |
| 1475 | return false; |
| 1476 | |
| 1477 | // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 |
| 1478 | // TODO: We can skip the step. If loop trip count is known (CTLZ), |
| 1479 | // then all uses of "cnt.next" could be optimized to the trip count |
| 1480 | // plus "cnt0". Currently it is not optimized. |
| 1481 | // This step could be used to detect POPCNT instruction: |
| 1482 | // cnt.next = cnt + (x.next & 1) |
| 1483 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
| 1484 | IterE = LoopEntry->end(); |
| 1485 | Iter != IterE; Iter++) { |
| 1486 | Instruction *Inst = &*Iter; |
| 1487 | if (Inst->getOpcode() != Instruction::Add) |
| 1488 | continue; |
| 1489 | |
| 1490 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
| 1491 | if (!Inc || !Inc->isOne()) |
| 1492 | continue; |
| 1493 | |
| 1494 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
| 1495 | if (!Phi) |
| 1496 | continue; |
| 1497 | |
| 1498 | CntInst = Inst; |
| 1499 | CntPhi = Phi; |
| 1500 | break; |
| 1501 | } |
| 1502 | if (!CntInst) |
| 1503 | return false; |
| 1504 | |
| 1505 | return true; |
| 1506 | } |
| 1507 | |
| 1508 | /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop |
| 1509 | /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new |
| 1510 | /// trip count returns true; otherwise, returns false. |
| 1511 | bool LoopIdiomRecognize::recognizeAndInsertFFS() { |
| 1512 | // Give up if the loop has multiple blocks or multiple backedges. |
| 1513 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
| 1514 | return false; |
| 1515 | |
| 1516 | Intrinsic::ID IntrinID; |
| 1517 | Value *InitX; |
| 1518 | Instruction *DefX = nullptr; |
| 1519 | PHINode *CntPhi = nullptr; |
| 1520 | Instruction *CntInst = nullptr; |
| 1521 | // Help decide if transformation is profitable. For ShiftUntilZero idiom, |
| 1522 | // this is always 6. |
| 1523 | size_t IdiomCanonicalSize = 6; |
| 1524 | |
| 1525 | if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, |
| 1526 | CntInst, CntPhi, DefX)) |
| 1527 | return false; |
| 1528 | |
| 1529 | bool IsCntPhiUsedOutsideLoop = false; |
| 1530 | for (User *U : CntPhi->users()) |
| 1531 | if (!CurLoop->contains(cast<Instruction>(U))) { |
| 1532 | IsCntPhiUsedOutsideLoop = true; |
| 1533 | break; |
| 1534 | } |
| 1535 | bool IsCntInstUsedOutsideLoop = false; |
| 1536 | for (User *U : CntInst->users()) |
| 1537 | if (!CurLoop->contains(cast<Instruction>(U))) { |
| 1538 | IsCntInstUsedOutsideLoop = true; |
| 1539 | break; |
| 1540 | } |
| 1541 | // If both CntInst and CntPhi are used outside the loop the profitability |
| 1542 | // is questionable. |
| 1543 | if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) |
| 1544 | return false; |
| 1545 | |
| 1546 | // For some CPUs result of CTLZ(X) intrinsic is undefined |
| 1547 | // when X is 0. If we can not guarantee X != 0, we need to check this |
| 1548 | // when expand. |
| 1549 | bool ZeroCheck = false; |
| 1550 | // It is safe to assume Preheader exist as it was checked in |
| 1551 | // parent function RunOnLoop. |
| 1552 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
| 1553 | |
| 1554 | // If we are using the count instruction outside the loop, make sure we |
| 1555 | // have a zero check as a precondition. Without the check the loop would run |
| 1556 | // one iteration for before any check of the input value. This means 0 and 1 |
| 1557 | // would have identical behavior in the original loop and thus |
| 1558 | if (!IsCntPhiUsedOutsideLoop) { |
| 1559 | auto *PreCondBB = PH->getSinglePredecessor(); |
| 1560 | if (!PreCondBB) |
| 1561 | return false; |
| 1562 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1563 | if (!PreCondBI) |
| 1564 | return false; |
| 1565 | if (matchCondition(PreCondBI, PH) != InitX) |
| 1566 | return false; |
| 1567 | ZeroCheck = true; |
| 1568 | } |
| 1569 | |
| 1570 | // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always |
| 1571 | // profitable if we delete the loop. |
| 1572 | |
| 1573 | // the loop has only 6 instructions: |
| 1574 | // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] |
| 1575 | // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] |
| 1576 | // %shr = ashr %n.addr.0, 1 |
| 1577 | // %tobool = icmp eq %shr, 0 |
| 1578 | // %inc = add nsw %i.0, 1 |
| 1579 | // br i1 %tobool |
| 1580 | |
| 1581 | const Value *Args[] = { |
| 1582 | InitX, ZeroCheck ? ConstantInt::getTrue(InitX->getContext()) |
| 1583 | : ConstantInt::getFalse(InitX->getContext())}; |
| 1584 | |
| 1585 | // @llvm.dbg doesn't count as they have no semantic effect. |
| 1586 | auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); |
| 1587 | uint32_t HeaderSize = |
| 1588 | std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); |
| 1589 | |
| 1590 | IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); |
| 1591 | int Cost = |
| 1592 | TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); |
| 1593 | if (HeaderSize != IdiomCanonicalSize && |
| 1594 | Cost > TargetTransformInfo::TCC_Basic) |
| 1595 | return false; |
| 1596 | |
| 1597 | transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, |
| 1598 | DefX->getDebugLoc(), ZeroCheck, |
| 1599 | IsCntPhiUsedOutsideLoop); |
| 1600 | return true; |
| 1601 | } |
| 1602 | |
| 1603 | /// Recognizes a population count idiom in a non-countable loop. |
| 1604 | /// |
| 1605 | /// If detected, transforms the relevant code to issue the popcount intrinsic |
| 1606 | /// function call, and returns true; otherwise, returns false. |
| 1607 | bool LoopIdiomRecognize::recognizePopcount() { |
| 1608 | if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) |
| 1609 | return false; |
| 1610 | |
| 1611 | // Counting population are usually conducted by few arithmetic instructions. |
| 1612 | // Such instructions can be easily "absorbed" by vacant slots in a |
| 1613 | // non-compact loop. Therefore, recognizing popcount idiom only makes sense |
| 1614 | // in a compact loop. |
| 1615 | |
| 1616 | // Give up if the loop has multiple blocks or multiple backedges. |
| 1617 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
| 1618 | return false; |
| 1619 | |
| 1620 | BasicBlock *LoopBody = *(CurLoop->block_begin()); |
| 1621 | if (LoopBody->size() >= 20) { |
| 1622 | // The loop is too big, bail out. |
| 1623 | return false; |
| 1624 | } |
| 1625 | |
| 1626 | // It should have a preheader containing nothing but an unconditional branch. |
| 1627 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
| 1628 | if (!PH || &PH->front() != PH->getTerminator()) |
| 1629 | return false; |
| 1630 | auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator()); |
| 1631 | if (!EntryBI || EntryBI->isConditional()) |
| 1632 | return false; |
| 1633 | |
| 1634 | // It should have a precondition block where the generated popcount intrinsic |
| 1635 | // function can be inserted. |
| 1636 | auto *PreCondBB = PH->getSinglePredecessor(); |
| 1637 | if (!PreCondBB) |
| 1638 | return false; |
| 1639 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1640 | if (!PreCondBI || PreCondBI->isUnconditional()) |
| 1641 | return false; |
| 1642 | |
| 1643 | Instruction *CntInst; |
| 1644 | PHINode *CntPhi; |
| 1645 | Value *Val; |
| 1646 | if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) |
| 1647 | return false; |
| 1648 | |
| 1649 | transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); |
| 1650 | return true; |
| 1651 | } |
| 1652 | |
| 1653 | static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
| 1654 | const DebugLoc &DL) { |
| 1655 | Value *Ops[] = {Val}; |
| 1656 | Type *Tys[] = {Val->getType()}; |
| 1657 | |
| 1658 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
| 1659 | Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); |
| 1660 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
| 1661 | CI->setDebugLoc(DL); |
| 1662 | |
| 1663 | return CI; |
| 1664 | } |
| 1665 | |
| 1666 | static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
| 1667 | const DebugLoc &DL, bool ZeroCheck, |
| 1668 | Intrinsic::ID IID) { |
| 1669 | Value *Ops[] = {Val, ZeroCheck ? IRBuilder.getTrue() : IRBuilder.getFalse()}; |
| 1670 | Type *Tys[] = {Val->getType()}; |
| 1671 | |
| 1672 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
| 1673 | Function *Func = Intrinsic::getDeclaration(M, IID, Tys); |
| 1674 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
| 1675 | CI->setDebugLoc(DL); |
| 1676 | |
| 1677 | return CI; |
| 1678 | } |
| 1679 | |
| 1680 | /// Transform the following loop (Using CTLZ, CTTZ is similar): |
| 1681 | /// loop: |
| 1682 | /// CntPhi = PHI [Cnt0, CntInst] |
| 1683 | /// PhiX = PHI [InitX, DefX] |
| 1684 | /// CntInst = CntPhi + 1 |
| 1685 | /// DefX = PhiX >> 1 |
| 1686 | /// LOOP_BODY |
| 1687 | /// Br: loop if (DefX != 0) |
| 1688 | /// Use(CntPhi) or Use(CntInst) |
| 1689 | /// |
| 1690 | /// Into: |
| 1691 | /// If CntPhi used outside the loop: |
| 1692 | /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) |
| 1693 | /// Count = CountPrev + 1 |
| 1694 | /// else |
| 1695 | /// Count = BitWidth(InitX) - CTLZ(InitX) |
| 1696 | /// loop: |
| 1697 | /// CntPhi = PHI [Cnt0, CntInst] |
| 1698 | /// PhiX = PHI [InitX, DefX] |
| 1699 | /// PhiCount = PHI [Count, Dec] |
| 1700 | /// CntInst = CntPhi + 1 |
| 1701 | /// DefX = PhiX >> 1 |
| 1702 | /// Dec = PhiCount - 1 |
| 1703 | /// LOOP_BODY |
| 1704 | /// Br: loop if (Dec != 0) |
| 1705 | /// Use(CountPrev + Cnt0) // Use(CntPhi) |
| 1706 | /// or |
| 1707 | /// Use(Count + Cnt0) // Use(CntInst) |
| 1708 | /// |
| 1709 | /// If LOOP_BODY is empty the loop will be deleted. |
| 1710 | /// If CntInst and DefX are not used in LOOP_BODY they will be removed. |
| 1711 | void LoopIdiomRecognize::transformLoopToCountable( |
| 1712 | Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, |
| 1713 | PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, |
| 1714 | bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { |
| 1715 | BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator()); |
| 1716 | |
| 1717 | // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block |
| 1718 | IRBuilder<> Builder(PreheaderBr); |
| 1719 | Builder.SetCurrentDebugLocation(DL); |
| 1720 | Value *FFS, *Count, *CountPrev, *NewCount, *InitXNext; |
| 1721 | |
| 1722 | // Count = BitWidth - CTLZ(InitX); |
| 1723 | // If there are uses of CntPhi create: |
| 1724 | // CountPrev = BitWidth - CTLZ(InitX >> 1); |
| 1725 | if (IsCntPhiUsedOutsideLoop) { |
| 1726 | if (DefX->getOpcode() == Instruction::AShr) |
| 1727 | InitXNext = |
| 1728 | Builder.CreateAShr(InitX, ConstantInt::get(InitX->getType(), 1)); |
| 1729 | else if (DefX->getOpcode() == Instruction::LShr) |
| 1730 | InitXNext = |
| 1731 | Builder.CreateLShr(InitX, ConstantInt::get(InitX->getType(), 1)); |
| 1732 | else if (DefX->getOpcode() == Instruction::Shl) // cttz |
| 1733 | InitXNext = |
| 1734 | Builder.CreateShl(InitX, ConstantInt::get(InitX->getType(), 1)); |
| 1735 | else |
| 1736 | llvm_unreachable("Unexpected opcode!")::llvm::llvm_unreachable_internal("Unexpected opcode!", "/build/llvm-toolchain-snapshot-12~++20200926111128+c6c5629f2fb/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1736); |
| 1737 | } else |
| 1738 | InitXNext = InitX; |
| 1739 | FFS = createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); |
| 1740 | Count = Builder.CreateSub( |
| 1741 | ConstantInt::get(FFS->getType(), |
| 1742 | FFS->getType()->getIntegerBitWidth()), |
| 1743 | FFS); |
| 1744 | if (IsCntPhiUsedOutsideLoop) { |
| 1745 | CountPrev = Count; |
| 1746 | Count = Builder.CreateAdd( |
| 1747 | CountPrev, |
| 1748 | ConstantInt::get(CountPrev->getType(), 1)); |
| 1749 | } |
| 1750 | |
| 1751 | NewCount = Builder.CreateZExtOrTrunc( |
| 1752 | IsCntPhiUsedOutsideLoop ? CountPrev : Count, |
| 1753 | cast<IntegerType>(CntInst->getType())); |
| 1754 | |
| 1755 | // If the counter's initial value is not zero, insert Add Inst. |
| 1756 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); |
| 1757 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
| 1758 | if (!InitConst || !InitConst->isZero()) |
| 1759 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
| 1760 | |
| 1761 | // Step 2: Insert new IV and loop condition: |
| 1762 | // loop: |
| 1763 | // ... |
| 1764 | // PhiCount = PHI [Count, Dec] |
| 1765 | // ... |
| 1766 | // Dec = PhiCount - 1 |
| 1767 | // ... |
| 1768 | // Br: loop if (Dec != 0) |
| 1769 | BasicBlock *Body = *(CurLoop->block_begin()); |
| 1770 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
| 1771 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
| 1772 | Type *Ty = Count->getType(); |
| 1773 | |
| 1774 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
| 1775 | |
| 1776 | Builder.SetInsertPoint(LbCond); |
| 1777 | Instruction *TcDec = cast<Instruction>( |
| 1778 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
| 1779 | "tcdec", false, true)); |
| 1780 | |
| 1781 | TcPhi->addIncoming(Count, Preheader); |
| 1782 | TcPhi->addIncoming(TcDec, Body); |
| 1783 | |
| 1784 | CmpInst::Predicate Pred = |
| 1785 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
| 1786 | LbCond->setPredicate(Pred); |
| 1787 | LbCond->setOperand(0, TcDec); |
| 1788 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
| 1789 | |
| 1790 | // Step 3: All the references to the original counter outside |
| 1791 | // the loop are replaced with the NewCount |
| 1792 | if (IsCntPhiUsedOutsideLoop) |
| 1793 | CntPhi->replaceUsesOutsideBlock(NewCount, Body); |
| 1794 | else |
| 1795 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
| 1796 | |
| 1797 | // step 4: Forget the "non-computable" trip-count SCEV associated with the |
| 1798 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 1799 | SE->forgetLoop(CurLoop); |
| 1800 | } |
| 1801 | |
| 1802 | void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, |
| 1803 | Instruction *CntInst, |
| 1804 | PHINode *CntPhi, Value *Var) { |
| 1805 | BasicBlock *PreHead = CurLoop->getLoopPreheader(); |
| 1806 | auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator()); |
| 1807 | const DebugLoc &DL = CntInst->getDebugLoc(); |
| 1808 | |
| 1809 | // Assuming before transformation, the loop is following: |
| 1810 | // if (x) // the precondition |
| 1811 | // do { cnt++; x &= x - 1; } while(x); |
| 1812 | |
| 1813 | // Step 1: Insert the ctpop instruction at the end of the precondition block |
| 1814 | IRBuilder<> Builder(PreCondBr); |
| 1815 | Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; |
| 1816 | { |
| 1817 | PopCnt = createPopcntIntrinsic(Builder, Var, DL); |
| 1818 | NewCount = PopCntZext = |
| 1819 | Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType())); |
| 1820 | |
| 1821 | if (NewCount != PopCnt) |
| 1822 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
| 1823 | |
| 1824 | // TripCnt is exactly the number of iterations the loop has |
| 1825 | TripCnt = NewCount; |
| 1826 | |
| 1827 | // If the population counter's initial value is not zero, insert Add Inst. |
| 1828 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); |
| 1829 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
| 1830 | if (!InitConst || !InitConst->isZero()) { |
| 1831 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
| 1832 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
| 1833 | } |
| 1834 | } |
| 1835 | |
| 1836 | // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to |
| 1837 | // "if (NewCount == 0) loop-exit". Without this change, the intrinsic |
| 1838 | // function would be partial dead code, and downstream passes will drag |
| 1839 | // it back from the precondition block to the preheader. |
| 1840 | { |
| 1841 | ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition()); |
| 1842 | |
| 1843 | Value *Opnd0 = PopCntZext; |
| 1844 | Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); |
| 1845 | if (PreCond->getOperand(0) != Var) |
| 1846 | std::swap(Opnd0, Opnd1); |
| 1847 | |
| 1848 | ICmpInst *NewPreCond = cast<ICmpInst>( |
| 1849 | Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); |
| 1850 | PreCondBr->setCondition(NewPreCond); |
| 1851 | |
| 1852 | RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); |
| 1853 | } |
| 1854 | |
| 1855 | // Step 3: Note that the population count is exactly the trip count of the |
| 1856 | // loop in question, which enable us to convert the loop from noncountable |
| 1857 | // loop into a countable one. The benefit is twofold: |
| 1858 | // |
| 1859 | // - If the loop only counts population, the entire loop becomes dead after |
| 1860 | // the transformation. It is a lot easier to prove a countable loop dead |
| 1861 | // than to prove a noncountable one. (In some C dialects, an infinite loop |
| 1862 | // isn't dead even if it computes nothing useful. In general, DCE needs |
| 1863 | // to prove a noncountable loop finite before safely delete it.) |
| 1864 | // |
| 1865 | // - If the loop also performs something else, it remains alive. |
| 1866 | // Since it is transformed to countable form, it can be aggressively |
| 1867 | // optimized by some optimizations which are in general not applicable |
| 1868 | // to a noncountable loop. |
| 1869 | // |
| 1870 | // After this step, this loop (conceptually) would look like following: |
| 1871 | // newcnt = __builtin_ctpop(x); |
| 1872 | // t = newcnt; |
| 1873 | // if (x) |
| 1874 | // do { cnt++; x &= x-1; t--) } while (t > 0); |
| 1875 | BasicBlock *Body = *(CurLoop->block_begin()); |
| 1876 | { |
| 1877 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
| 1878 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
| 1879 | Type *Ty = TripCnt->getType(); |
| 1880 | |
| 1881 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
| 1882 | |
| 1883 | Builder.SetInsertPoint(LbCond); |
| 1884 | Instruction *TcDec = cast<Instruction>( |
| 1885 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
| 1886 | "tcdec", false, true)); |
| 1887 | |
| 1888 | TcPhi->addIncoming(TripCnt, PreHead); |
| 1889 | TcPhi->addIncoming(TcDec, Body); |
| 1890 | |
| 1891 | CmpInst::Predicate Pred = |
| 1892 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; |
| 1893 | LbCond->setPredicate(Pred); |
| 1894 | LbCond->setOperand(0, TcDec); |
| 1895 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
| 1896 | } |
| 1897 | |
| 1898 | // Step 4: All the references to the original population counter outside |
| 1899 | // the loop are replaced with the NewCount -- the value returned from |
| 1900 | // __builtin_ctpop(). |
| 1901 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
| 1902 | |
| 1903 | // step 5: Forget the "non-computable" trip-count SCEV associated with the |
| 1904 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 1905 | SE->forgetLoop(CurLoop); |
| 1906 | } |