| File: | build/source/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp |
| Warning: | line 2613, column 5 Value stored to 'Pred' is never read |
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, 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/CmpInstAnalysis.h" |
| 51 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
| 52 | #include "llvm/Analysis/LoopInfo.h" |
| 53 | #include "llvm/Analysis/LoopPass.h" |
| 54 | #include "llvm/Analysis/MemoryLocation.h" |
| 55 | #include "llvm/Analysis/MemorySSA.h" |
| 56 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 57 | #include "llvm/Analysis/MustExecute.h" |
| 58 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 59 | #include "llvm/Analysis/ScalarEvolution.h" |
| 60 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 61 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 62 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 63 | #include "llvm/Analysis/ValueTracking.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/PatternMatch.h" |
| 83 | #include "llvm/IR/Type.h" |
| 84 | #include "llvm/IR/User.h" |
| 85 | #include "llvm/IR/Value.h" |
| 86 | #include "llvm/IR/ValueHandle.h" |
| 87 | #include "llvm/Support/Casting.h" |
| 88 | #include "llvm/Support/CommandLine.h" |
| 89 | #include "llvm/Support/Debug.h" |
| 90 | #include "llvm/Support/InstructionCost.h" |
| 91 | #include "llvm/Support/raw_ostream.h" |
| 92 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
| 93 | #include "llvm/Transforms/Utils/Local.h" |
| 94 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 95 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
| 96 | #include <algorithm> |
| 97 | #include <cassert> |
| 98 | #include <cstdint> |
| 99 | #include <utility> |
| 100 | #include <vector> |
| 101 | |
| 102 | using namespace llvm; |
| 103 | |
| 104 | #define DEBUG_TYPE"loop-idiom" "loop-idiom" |
| 105 | |
| 106 | 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"}; |
| 107 | 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"}; |
| 108 | STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores")static llvm::Statistic NumMemMove = {"loop-idiom", "NumMemMove" , "Number of memmove's formed from loop load+stores"}; |
| 109 | STATISTIC(static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
| 110 | NumShiftUntilBitTest,static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
| 111 | "Number of uncountable loops recognized as 'shift until bitttest' idiom")static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" }; |
| 112 | STATISTIC(NumShiftUntilZero,static llvm::Statistic NumShiftUntilZero = {"loop-idiom", "NumShiftUntilZero" , "Number of uncountable loops recognized as 'shift until zero' idiom" } |
| 113 | "Number of uncountable loops recognized as 'shift until zero' idiom")static llvm::Statistic NumShiftUntilZero = {"loop-idiom", "NumShiftUntilZero" , "Number of uncountable loops recognized as 'shift until zero' idiom" }; |
| 114 | |
| 115 | bool DisableLIRP::All; |
| 116 | static cl::opt<bool, true> |
| 117 | DisableLIRPAll("disable-" DEBUG_TYPE"loop-idiom" "-all", |
| 118 | cl::desc("Options to disable Loop Idiom Recognize Pass."), |
| 119 | cl::location(DisableLIRP::All), cl::init(false), |
| 120 | cl::ReallyHidden); |
| 121 | |
| 122 | bool DisableLIRP::Memset; |
| 123 | static cl::opt<bool, true> |
| 124 | DisableLIRPMemset("disable-" DEBUG_TYPE"loop-idiom" "-memset", |
| 125 | cl::desc("Proceed with loop idiom recognize pass, but do " |
| 126 | "not convert loop(s) to memset."), |
| 127 | cl::location(DisableLIRP::Memset), cl::init(false), |
| 128 | cl::ReallyHidden); |
| 129 | |
| 130 | bool DisableLIRP::Memcpy; |
| 131 | static cl::opt<bool, true> |
| 132 | DisableLIRPMemcpy("disable-" DEBUG_TYPE"loop-idiom" "-memcpy", |
| 133 | cl::desc("Proceed with loop idiom recognize pass, but do " |
| 134 | "not convert loop(s) to memcpy."), |
| 135 | cl::location(DisableLIRP::Memcpy), cl::init(false), |
| 136 | cl::ReallyHidden); |
| 137 | |
| 138 | static cl::opt<bool> UseLIRCodeSizeHeurs( |
| 139 | "use-lir-code-size-heurs", |
| 140 | cl::desc("Use loop idiom recognition code size heuristics when compiling" |
| 141 | "with -Os/-Oz"), |
| 142 | cl::init(true), cl::Hidden); |
| 143 | |
| 144 | namespace { |
| 145 | |
| 146 | class LoopIdiomRecognize { |
| 147 | Loop *CurLoop = nullptr; |
| 148 | AliasAnalysis *AA; |
| 149 | DominatorTree *DT; |
| 150 | LoopInfo *LI; |
| 151 | ScalarEvolution *SE; |
| 152 | TargetLibraryInfo *TLI; |
| 153 | const TargetTransformInfo *TTI; |
| 154 | const DataLayout *DL; |
| 155 | OptimizationRemarkEmitter &ORE; |
| 156 | bool ApplyCodeSizeHeuristics; |
| 157 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
| 158 | |
| 159 | public: |
| 160 | explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
| 161 | LoopInfo *LI, ScalarEvolution *SE, |
| 162 | TargetLibraryInfo *TLI, |
| 163 | const TargetTransformInfo *TTI, MemorySSA *MSSA, |
| 164 | const DataLayout *DL, |
| 165 | OptimizationRemarkEmitter &ORE) |
| 166 | : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { |
| 167 | if (MSSA) |
| 168 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); |
| 169 | } |
| 170 | |
| 171 | bool runOnLoop(Loop *L); |
| 172 | |
| 173 | private: |
| 174 | using StoreList = SmallVector<StoreInst *, 8>; |
| 175 | using StoreListMap = MapVector<Value *, StoreList>; |
| 176 | |
| 177 | StoreListMap StoreRefsForMemset; |
| 178 | StoreListMap StoreRefsForMemsetPattern; |
| 179 | StoreList StoreRefsForMemcpy; |
| 180 | bool HasMemset; |
| 181 | bool HasMemsetPattern; |
| 182 | bool HasMemcpy; |
| 183 | |
| 184 | /// Return code for isLegalStore() |
| 185 | enum LegalStoreKind { |
| 186 | None = 0, |
| 187 | Memset, |
| 188 | MemsetPattern, |
| 189 | Memcpy, |
| 190 | UnorderedAtomicMemcpy, |
| 191 | DontUse // Dummy retval never to be used. Allows catching errors in retval |
| 192 | // handling. |
| 193 | }; |
| 194 | |
| 195 | /// \name Countable Loop Idiom Handling |
| 196 | /// @{ |
| 197 | |
| 198 | bool runOnCountableLoop(); |
| 199 | bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, |
| 200 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
| 201 | |
| 202 | void collectStores(BasicBlock *BB); |
| 203 | LegalStoreKind isLegalStore(StoreInst *SI); |
| 204 | enum class ForMemset { No, Yes }; |
| 205 | bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount, |
| 206 | ForMemset For); |
| 207 | |
| 208 | template <typename MemInst> |
| 209 | bool processLoopMemIntrinsic( |
| 210 | BasicBlock *BB, |
| 211 | bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), |
| 212 | const SCEV *BECount); |
| 213 | bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount); |
| 214 | bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); |
| 215 | |
| 216 | bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV, |
| 217 | MaybeAlign StoreAlignment, Value *StoredVal, |
| 218 | Instruction *TheStore, |
| 219 | SmallPtrSetImpl<Instruction *> &Stores, |
| 220 | const SCEVAddRecExpr *Ev, const SCEV *BECount, |
| 221 | bool IsNegStride, bool IsLoopMemset = false); |
| 222 | bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); |
| 223 | bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr, |
| 224 | const SCEV *StoreSize, MaybeAlign StoreAlign, |
| 225 | MaybeAlign LoadAlign, Instruction *TheStore, |
| 226 | Instruction *TheLoad, |
| 227 | const SCEVAddRecExpr *StoreEv, |
| 228 | const SCEVAddRecExpr *LoadEv, |
| 229 | const SCEV *BECount); |
| 230 | bool avoidLIRForMultiBlockLoop(bool IsMemset = false, |
| 231 | bool IsLoopMemset = false); |
| 232 | |
| 233 | /// @} |
| 234 | /// \name Noncountable Loop Idiom Handling |
| 235 | /// @{ |
| 236 | |
| 237 | bool runOnNoncountableLoop(); |
| 238 | |
| 239 | bool recognizePopcount(); |
| 240 | void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, |
| 241 | PHINode *CntPhi, Value *Var); |
| 242 | bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz |
| 243 | void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, |
| 244 | Instruction *CntInst, PHINode *CntPhi, |
| 245 | Value *Var, Instruction *DefX, |
| 246 | const DebugLoc &DL, bool ZeroCheck, |
| 247 | bool IsCntPhiUsedOutsideLoop); |
| 248 | |
| 249 | bool recognizeShiftUntilBitTest(); |
| 250 | bool recognizeShiftUntilZero(); |
| 251 | |
| 252 | /// @} |
| 253 | }; |
| 254 | } // end anonymous namespace |
| 255 | |
| 256 | PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, |
| 257 | LoopStandardAnalysisResults &AR, |
| 258 | LPMUpdater &) { |
| 259 | if (DisableLIRP::All) |
| 260 | return PreservedAnalyses::all(); |
| 261 | |
| 262 | const auto *DL = &L.getHeader()->getModule()->getDataLayout(); |
| 263 | |
| 264 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
| 265 | // pass. Function analyses need to be preserved across loop transformations |
| 266 | // but ORE cannot be preserved (see comment before the pass definition). |
| 267 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
| 268 | |
| 269 | LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, |
| 270 | AR.MSSA, DL, ORE); |
| 271 | if (!LIR.runOnLoop(&L)) |
| 272 | return PreservedAnalyses::all(); |
| 273 | |
| 274 | auto PA = getLoopPassPreservedAnalyses(); |
| 275 | if (AR.MSSA) |
| 276 | PA.preserve<MemorySSAAnalysis>(); |
| 277 | return PA; |
| 278 | } |
| 279 | |
| 280 | static void deleteDeadInstruction(Instruction *I) { |
| 281 | I->replaceAllUsesWith(PoisonValue::get(I->getType())); |
| 282 | I->eraseFromParent(); |
| 283 | } |
| 284 | |
| 285 | //===----------------------------------------------------------------------===// |
| 286 | // |
| 287 | // Implementation of LoopIdiomRecognize |
| 288 | // |
| 289 | //===----------------------------------------------------------------------===// |
| 290 | |
| 291 | bool LoopIdiomRecognize::runOnLoop(Loop *L) { |
| 292 | CurLoop = L; |
| 293 | // If the loop could not be converted to canonical form, it must have an |
| 294 | // indirectbr in it, just give up. |
| 295 | if (!L->getLoopPreheader()) |
| 296 | return false; |
| 297 | |
| 298 | // Disable loop idiom recognition if the function's name is a common idiom. |
| 299 | StringRef Name = L->getHeader()->getParent()->getName(); |
| 300 | if (Name == "memset" || Name == "memcpy") |
| 301 | return false; |
| 302 | |
| 303 | // Determine if code size heuristics need to be applied. |
| 304 | ApplyCodeSizeHeuristics = |
| 305 | L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; |
| 306 | |
| 307 | HasMemset = TLI->has(LibFunc_memset); |
| 308 | HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); |
| 309 | HasMemcpy = TLI->has(LibFunc_memcpy); |
| 310 | |
| 311 | if (HasMemset || HasMemsetPattern || HasMemcpy) |
| 312 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
| 313 | return runOnCountableLoop(); |
| 314 | |
| 315 | return runOnNoncountableLoop(); |
| 316 | } |
| 317 | |
| 318 | bool LoopIdiomRecognize::runOnCountableLoop() { |
| 319 | const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); |
| 320 | assert(!isa<SCEVCouldNotCompute>(BECount) &&(static_cast <bool> (!isa<SCEVCouldNotCompute>(BECount ) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 322, __extension__ __PRETTY_FUNCTION__)) |
| 321 | "runOnCountableLoop() called on a loop without a predictable"(static_cast <bool> (!isa<SCEVCouldNotCompute>(BECount ) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 322, __extension__ __PRETTY_FUNCTION__)) |
| 322 | "backedge-taken count")(static_cast <bool> (!isa<SCEVCouldNotCompute>(BECount ) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 322, __extension__ __PRETTY_FUNCTION__)); |
| 323 | |
| 324 | // If this loop executes exactly one time, then it should be peeled, not |
| 325 | // optimized by this pass. |
| 326 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
| 327 | if (BECst->getAPInt() == 0) |
| 328 | return false; |
| 329 | |
| 330 | SmallVector<BasicBlock *, 8> ExitBlocks; |
| 331 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
| 332 | |
| 333 | 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) |
| 334 | << 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) |
| 335 | << "] 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) |
| 336 | << "\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); |
| 337 | |
| 338 | // The following transforms hoist stores/memsets into the loop pre-header. |
| 339 | // Give up if the loop has instructions that may throw. |
| 340 | SimpleLoopSafetyInfo SafetyInfo; |
| 341 | SafetyInfo.computeLoopSafetyInfo(CurLoop); |
| 342 | if (SafetyInfo.anyBlockMayThrow()) |
| 343 | return false; |
| 344 | |
| 345 | bool MadeChange = false; |
| 346 | |
| 347 | // Scan all the blocks in the loop that are not in subloops. |
| 348 | for (auto *BB : CurLoop->getBlocks()) { |
| 349 | // Ignore blocks in subloops. |
| 350 | if (LI->getLoopFor(BB) != CurLoop) |
| 351 | continue; |
| 352 | |
| 353 | MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); |
| 354 | } |
| 355 | return MadeChange; |
| 356 | } |
| 357 | |
| 358 | static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { |
| 359 | const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1)); |
| 360 | return ConstStride->getAPInt(); |
| 361 | } |
| 362 | |
| 363 | /// getMemSetPatternValue - If a strided store of the specified value is safe to |
| 364 | /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should |
| 365 | /// be passed in. Otherwise, return null. |
| 366 | /// |
| 367 | /// Note that we don't ever attempt to use memset_pattern8 or 4, because these |
| 368 | /// just replicate their input array and then pass on to memset_pattern16. |
| 369 | static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { |
| 370 | // FIXME: This could check for UndefValue because it can be merged into any |
| 371 | // other valid pattern. |
| 372 | |
| 373 | // If the value isn't a constant, we can't promote it to being in a constant |
| 374 | // array. We could theoretically do a store to an alloca or something, but |
| 375 | // that doesn't seem worthwhile. |
| 376 | Constant *C = dyn_cast<Constant>(V); |
| 377 | if (!C || isa<ConstantExpr>(C)) |
| 378 | return nullptr; |
| 379 | |
| 380 | // Only handle simple values that are a power of two bytes in size. |
| 381 | uint64_t Size = DL->getTypeSizeInBits(V->getType()); |
| 382 | if (Size == 0 || (Size & 7) || (Size & (Size - 1))) |
| 383 | return nullptr; |
| 384 | |
| 385 | // Don't care enough about darwin/ppc to implement this. |
| 386 | if (DL->isBigEndian()) |
| 387 | return nullptr; |
| 388 | |
| 389 | // Convert to size in bytes. |
| 390 | Size /= 8; |
| 391 | |
| 392 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
| 393 | // if the top and bottom are the same (e.g. for vectors and large integers). |
| 394 | if (Size > 16) |
| 395 | return nullptr; |
| 396 | |
| 397 | // If the constant is exactly 16 bytes, just use it. |
| 398 | if (Size == 16) |
| 399 | return C; |
| 400 | |
| 401 | // Otherwise, we'll use an array of the constants. |
| 402 | unsigned ArraySize = 16 / Size; |
| 403 | ArrayType *AT = ArrayType::get(V->getType(), ArraySize); |
| 404 | return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C)); |
| 405 | } |
| 406 | |
| 407 | LoopIdiomRecognize::LegalStoreKind |
| 408 | LoopIdiomRecognize::isLegalStore(StoreInst *SI) { |
| 409 | // Don't touch volatile stores. |
| 410 | if (SI->isVolatile()) |
| 411 | return LegalStoreKind::None; |
| 412 | // We only want simple or unordered-atomic stores. |
| 413 | if (!SI->isUnordered()) |
| 414 | return LegalStoreKind::None; |
| 415 | |
| 416 | // Avoid merging nontemporal stores. |
| 417 | if (SI->getMetadata(LLVMContext::MD_nontemporal)) |
| 418 | return LegalStoreKind::None; |
| 419 | |
| 420 | Value *StoredVal = SI->getValueOperand(); |
| 421 | Value *StorePtr = SI->getPointerOperand(); |
| 422 | |
| 423 | // Don't convert stores of non-integral pointer types to memsets (which stores |
| 424 | // integers). |
| 425 | if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType())) |
| 426 | return LegalStoreKind::None; |
| 427 | |
| 428 | // Reject stores that are so large that they overflow an unsigned. |
| 429 | // When storing out scalable vectors we bail out for now, since the code |
| 430 | // below currently only works for constant strides. |
| 431 | TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); |
| 432 | if (SizeInBits.isScalable() || (SizeInBits.getFixedValue() & 7) || |
| 433 | (SizeInBits.getFixedValue() >> 32) != 0) |
| 434 | return LegalStoreKind::None; |
| 435 | |
| 436 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 437 | // loop, which indicates a strided store. If we have something else, it's a |
| 438 | // random store we can't handle. |
| 439 | const SCEVAddRecExpr *StoreEv = |
| 440 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 441 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
| 442 | return LegalStoreKind::None; |
| 443 | |
| 444 | // Check to see if we have a constant stride. |
| 445 | if (!isa<SCEVConstant>(StoreEv->getOperand(1))) |
| 446 | return LegalStoreKind::None; |
| 447 | |
| 448 | // See if the store can be turned into a memset. |
| 449 | |
| 450 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
| 451 | // turned into a memset of i8 -1, assuming that all the consecutive bytes |
| 452 | // are stored. A store of i32 0x01020304 can never be turned into a memset, |
| 453 | // but it can be turned into memset_pattern if the target supports it. |
| 454 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
| 455 | |
| 456 | // Note: memset and memset_pattern on unordered-atomic is yet not supported |
| 457 | bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); |
| 458 | |
| 459 | // If we're allowed to form a memset, and the stored value would be |
| 460 | // acceptable for memset, use it. |
| 461 | if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && |
| 462 | // Verify that the stored value is loop invariant. If not, we can't |
| 463 | // promote the memset. |
| 464 | CurLoop->isLoopInvariant(SplatValue)) { |
| 465 | // It looks like we can use SplatValue. |
| 466 | return LegalStoreKind::Memset; |
| 467 | } |
| 468 | if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && |
| 469 | // Don't create memset_pattern16s with address spaces. |
| 470 | StorePtr->getType()->getPointerAddressSpace() == 0 && |
| 471 | getMemSetPatternValue(StoredVal, DL)) { |
| 472 | // It looks like we can use PatternValue! |
| 473 | return LegalStoreKind::MemsetPattern; |
| 474 | } |
| 475 | |
| 476 | // Otherwise, see if the store can be turned into a memcpy. |
| 477 | if (HasMemcpy && !DisableLIRP::Memcpy) { |
| 478 | // Check to see if the stride matches the size of the store. If so, then we |
| 479 | // know that every byte is touched in the loop. |
| 480 | APInt Stride = getStoreStride(StoreEv); |
| 481 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
| 482 | if (StoreSize != Stride && StoreSize != -Stride) |
| 483 | return LegalStoreKind::None; |
| 484 | |
| 485 | // The store must be feeding a non-volatile load. |
| 486 | LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); |
| 487 | |
| 488 | // Only allow non-volatile loads |
| 489 | if (!LI || LI->isVolatile()) |
| 490 | return LegalStoreKind::None; |
| 491 | // Only allow simple or unordered-atomic loads |
| 492 | if (!LI->isUnordered()) |
| 493 | return LegalStoreKind::None; |
| 494 | |
| 495 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 496 | // loop, which indicates a strided load. If we have something else, it's a |
| 497 | // random load we can't handle. |
| 498 | const SCEVAddRecExpr *LoadEv = |
| 499 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
| 500 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
| 501 | return LegalStoreKind::None; |
| 502 | |
| 503 | // The store and load must share the same stride. |
| 504 | if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) |
| 505 | return LegalStoreKind::None; |
| 506 | |
| 507 | // Success. This store can be converted into a memcpy. |
| 508 | UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); |
| 509 | return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy |
| 510 | : LegalStoreKind::Memcpy; |
| 511 | } |
| 512 | // This store can't be transformed into a memset/memcpy. |
| 513 | return LegalStoreKind::None; |
| 514 | } |
| 515 | |
| 516 | void LoopIdiomRecognize::collectStores(BasicBlock *BB) { |
| 517 | StoreRefsForMemset.clear(); |
| 518 | StoreRefsForMemsetPattern.clear(); |
| 519 | StoreRefsForMemcpy.clear(); |
| 520 | for (Instruction &I : *BB) { |
| 521 | StoreInst *SI = dyn_cast<StoreInst>(&I); |
| 522 | if (!SI) |
| 523 | continue; |
| 524 | |
| 525 | // Make sure this is a strided store with a constant stride. |
| 526 | switch (isLegalStore(SI)) { |
| 527 | case LegalStoreKind::None: |
| 528 | // Nothing to do |
| 529 | break; |
| 530 | case LegalStoreKind::Memset: { |
| 531 | // Find the base pointer. |
| 532 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
| 533 | StoreRefsForMemset[Ptr].push_back(SI); |
| 534 | } break; |
| 535 | case LegalStoreKind::MemsetPattern: { |
| 536 | // Find the base pointer. |
| 537 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
| 538 | StoreRefsForMemsetPattern[Ptr].push_back(SI); |
| 539 | } break; |
| 540 | case LegalStoreKind::Memcpy: |
| 541 | case LegalStoreKind::UnorderedAtomicMemcpy: |
| 542 | StoreRefsForMemcpy.push_back(SI); |
| 543 | break; |
| 544 | default: |
| 545 | assert(false && "unhandled return value")(static_cast <bool> (false && "unhandled return value" ) ? void (0) : __assert_fail ("false && \"unhandled return value\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 545, __extension__ __PRETTY_FUNCTION__)); |
| 546 | break; |
| 547 | } |
| 548 | } |
| 549 | } |
| 550 | |
| 551 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
| 552 | /// with the specified backedge count. This block is known to be in the current |
| 553 | /// loop and not in any subloops. |
| 554 | bool LoopIdiomRecognize::runOnLoopBlock( |
| 555 | BasicBlock *BB, const SCEV *BECount, |
| 556 | SmallVectorImpl<BasicBlock *> &ExitBlocks) { |
| 557 | // We can only promote stores in this block if they are unconditionally |
| 558 | // executed in the loop. For a block to be unconditionally executed, it has |
| 559 | // to dominate all the exit blocks of the loop. Verify this now. |
| 560 | for (BasicBlock *ExitBlock : ExitBlocks) |
| 561 | if (!DT->dominates(BB, ExitBlock)) |
| 562 | return false; |
| 563 | |
| 564 | bool MadeChange = false; |
| 565 | // Look for store instructions, which may be optimized to memset/memcpy. |
| 566 | collectStores(BB); |
| 567 | |
| 568 | // Look for a single store or sets of stores with a common base, which can be |
| 569 | // optimized into a memset (memset_pattern). The latter most commonly happens |
| 570 | // with structs and handunrolled loops. |
| 571 | for (auto &SL : StoreRefsForMemset) |
| 572 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); |
| 573 | |
| 574 | for (auto &SL : StoreRefsForMemsetPattern) |
| 575 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); |
| 576 | |
| 577 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
| 578 | for (auto &SI : StoreRefsForMemcpy) |
| 579 | MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); |
| 580 | |
| 581 | MadeChange |= processLoopMemIntrinsic<MemCpyInst>( |
| 582 | BB, &LoopIdiomRecognize::processLoopMemCpy, BECount); |
| 583 | MadeChange |= processLoopMemIntrinsic<MemSetInst>( |
| 584 | BB, &LoopIdiomRecognize::processLoopMemSet, BECount); |
| 585 | |
| 586 | return MadeChange; |
| 587 | } |
| 588 | |
| 589 | /// See if this store(s) can be promoted to a memset. |
| 590 | bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL, |
| 591 | const SCEV *BECount, ForMemset For) { |
| 592 | // Try to find consecutive stores that can be transformed into memsets. |
| 593 | SetVector<StoreInst *> Heads, Tails; |
| 594 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; |
| 595 | |
| 596 | // Do a quadratic search on all of the given stores and find |
| 597 | // all of the pairs of stores that follow each other. |
| 598 | SmallVector<unsigned, 16> IndexQueue; |
| 599 | for (unsigned i = 0, e = SL.size(); i < e; ++i) { |
| 600 | assert(SL[i]->isSimple() && "Expected only non-volatile stores.")(static_cast <bool> (SL[i]->isSimple() && "Expected only non-volatile stores." ) ? void (0) : __assert_fail ("SL[i]->isSimple() && \"Expected only non-volatile stores.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 600, __extension__ __PRETTY_FUNCTION__)); |
| 601 | |
| 602 | Value *FirstStoredVal = SL[i]->getValueOperand(); |
| 603 | Value *FirstStorePtr = SL[i]->getPointerOperand(); |
| 604 | const SCEVAddRecExpr *FirstStoreEv = |
| 605 | cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr)); |
| 606 | APInt FirstStride = getStoreStride(FirstStoreEv); |
| 607 | unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); |
| 608 | |
| 609 | // See if we can optimize just this store in isolation. |
| 610 | if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { |
| 611 | Heads.insert(SL[i]); |
| 612 | continue; |
| 613 | } |
| 614 | |
| 615 | Value *FirstSplatValue = nullptr; |
| 616 | Constant *FirstPatternValue = nullptr; |
| 617 | |
| 618 | if (For == ForMemset::Yes) |
| 619 | FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); |
| 620 | else |
| 621 | FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); |
| 622 | |
| 623 | assert((FirstSplatValue || FirstPatternValue) &&(static_cast <bool> ((FirstSplatValue || FirstPatternValue ) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 624, __extension__ __PRETTY_FUNCTION__)) |
| 624 | "Expected either splat value or pattern value.")(static_cast <bool> ((FirstSplatValue || FirstPatternValue ) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 624, __extension__ __PRETTY_FUNCTION__)); |
| 625 | |
| 626 | IndexQueue.clear(); |
| 627 | // If a store has multiple consecutive store candidates, search Stores |
| 628 | // array according to the sequence: from i+1 to e, then from i-1 to 0. |
| 629 | // This is because usually pairing with immediate succeeding or preceding |
| 630 | // candidate create the best chance to find memset opportunity. |
| 631 | unsigned j = 0; |
| 632 | for (j = i + 1; j < e; ++j) |
| 633 | IndexQueue.push_back(j); |
| 634 | for (j = i; j > 0; --j) |
| 635 | IndexQueue.push_back(j - 1); |
| 636 | |
| 637 | for (auto &k : IndexQueue) { |
| 638 | assert(SL[k]->isSimple() && "Expected only non-volatile stores.")(static_cast <bool> (SL[k]->isSimple() && "Expected only non-volatile stores." ) ? void (0) : __assert_fail ("SL[k]->isSimple() && \"Expected only non-volatile stores.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 638, __extension__ __PRETTY_FUNCTION__)); |
| 639 | Value *SecondStorePtr = SL[k]->getPointerOperand(); |
| 640 | const SCEVAddRecExpr *SecondStoreEv = |
| 641 | cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr)); |
| 642 | APInt SecondStride = getStoreStride(SecondStoreEv); |
| 643 | |
| 644 | if (FirstStride != SecondStride) |
| 645 | continue; |
| 646 | |
| 647 | Value *SecondStoredVal = SL[k]->getValueOperand(); |
| 648 | Value *SecondSplatValue = nullptr; |
| 649 | Constant *SecondPatternValue = nullptr; |
| 650 | |
| 651 | if (For == ForMemset::Yes) |
| 652 | SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); |
| 653 | else |
| 654 | SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); |
| 655 | |
| 656 | assert((SecondSplatValue || SecondPatternValue) &&(static_cast <bool> ((SecondSplatValue || SecondPatternValue ) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 657, __extension__ __PRETTY_FUNCTION__)) |
| 657 | "Expected either splat value or pattern value.")(static_cast <bool> ((SecondSplatValue || SecondPatternValue ) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 657, __extension__ __PRETTY_FUNCTION__)); |
| 658 | |
| 659 | if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { |
| 660 | if (For == ForMemset::Yes) { |
| 661 | if (isa<UndefValue>(FirstSplatValue)) |
| 662 | FirstSplatValue = SecondSplatValue; |
| 663 | if (FirstSplatValue != SecondSplatValue) |
| 664 | continue; |
| 665 | } else { |
| 666 | if (isa<UndefValue>(FirstPatternValue)) |
| 667 | FirstPatternValue = SecondPatternValue; |
| 668 | if (FirstPatternValue != SecondPatternValue) |
| 669 | continue; |
| 670 | } |
| 671 | Tails.insert(SL[k]); |
| 672 | Heads.insert(SL[i]); |
| 673 | ConsecutiveChain[SL[i]] = SL[k]; |
| 674 | break; |
| 675 | } |
| 676 | } |
| 677 | } |
| 678 | |
| 679 | // We may run into multiple chains that merge into a single chain. We mark the |
| 680 | // stores that we transformed so that we don't visit the same store twice. |
| 681 | SmallPtrSet<Value *, 16> TransformedStores; |
| 682 | bool Changed = false; |
| 683 | |
| 684 | // For stores that start but don't end a link in the chain: |
| 685 | for (StoreInst *I : Heads) { |
| 686 | if (Tails.count(I)) |
| 687 | continue; |
| 688 | |
| 689 | // We found a store instr that starts a chain. Now follow the chain and try |
| 690 | // to transform it. |
| 691 | SmallPtrSet<Instruction *, 8> AdjacentStores; |
| 692 | StoreInst *HeadStore = I; |
| 693 | unsigned StoreSize = 0; |
| 694 | |
| 695 | // Collect the chain into a list. |
| 696 | while (Tails.count(I) || Heads.count(I)) { |
| 697 | if (TransformedStores.count(I)) |
| 698 | break; |
| 699 | AdjacentStores.insert(I); |
| 700 | |
| 701 | StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); |
| 702 | // Move to the next value in the chain. |
| 703 | I = ConsecutiveChain[I]; |
| 704 | } |
| 705 | |
| 706 | Value *StoredVal = HeadStore->getValueOperand(); |
| 707 | Value *StorePtr = HeadStore->getPointerOperand(); |
| 708 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 709 | APInt Stride = getStoreStride(StoreEv); |
| 710 | |
| 711 | // Check to see if the stride matches the size of the stores. If so, then |
| 712 | // we know that every byte is touched in the loop. |
| 713 | if (StoreSize != Stride && StoreSize != -Stride) |
| 714 | continue; |
| 715 | |
| 716 | bool IsNegStride = StoreSize == -Stride; |
| 717 | |
| 718 | Type *IntIdxTy = DL->getIndexType(StorePtr->getType()); |
| 719 | const SCEV *StoreSizeSCEV = SE->getConstant(IntIdxTy, StoreSize); |
| 720 | if (processLoopStridedStore(StorePtr, StoreSizeSCEV, |
| 721 | MaybeAlign(HeadStore->getAlign()), StoredVal, |
| 722 | HeadStore, AdjacentStores, StoreEv, BECount, |
| 723 | IsNegStride)) { |
| 724 | TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); |
| 725 | Changed = true; |
| 726 | } |
| 727 | } |
| 728 | |
| 729 | return Changed; |
| 730 | } |
| 731 | |
| 732 | /// processLoopMemIntrinsic - Template function for calling different processor |
| 733 | /// functions based on mem intrinsic type. |
| 734 | template <typename MemInst> |
| 735 | bool LoopIdiomRecognize::processLoopMemIntrinsic( |
| 736 | BasicBlock *BB, |
| 737 | bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), |
| 738 | const SCEV *BECount) { |
| 739 | bool MadeChange = false; |
| 740 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
| 741 | Instruction *Inst = &*I++; |
| 742 | // Look for memory instructions, which may be optimized to a larger one. |
| 743 | if (MemInst *MI = dyn_cast<MemInst>(Inst)) { |
| 744 | WeakTrackingVH InstPtr(&*I); |
| 745 | if (!(this->*Processor)(MI, BECount)) |
| 746 | continue; |
| 747 | MadeChange = true; |
| 748 | |
| 749 | // If processing the instruction invalidated our iterator, start over from |
| 750 | // the top of the block. |
| 751 | if (!InstPtr) |
| 752 | I = BB->begin(); |
| 753 | } |
| 754 | } |
| 755 | return MadeChange; |
| 756 | } |
| 757 | |
| 758 | /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy |
| 759 | bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI, |
| 760 | const SCEV *BECount) { |
| 761 | // We can only handle non-volatile memcpys with a constant size. |
| 762 | if (MCI->isVolatile() || !isa<ConstantInt>(MCI->getLength())) |
| 763 | return false; |
| 764 | |
| 765 | // If we're not allowed to hack on memcpy, we fail. |
| 766 | if ((!HasMemcpy && !isa<MemCpyInlineInst>(MCI)) || DisableLIRP::Memcpy) |
| 767 | return false; |
| 768 | |
| 769 | Value *Dest = MCI->getDest(); |
| 770 | Value *Source = MCI->getSource(); |
| 771 | if (!Dest || !Source) |
| 772 | return false; |
| 773 | |
| 774 | // See if the load and store pointer expressions are AddRec like {base,+,1} on |
| 775 | // the current loop, which indicates a strided load and store. If we have |
| 776 | // something else, it's a random load or store we can't handle. |
| 777 | const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest)); |
| 778 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
| 779 | return false; |
| 780 | const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source)); |
| 781 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
| 782 | return false; |
| 783 | |
| 784 | // Reject memcpys that are so large that they overflow an unsigned. |
| 785 | uint64_t SizeInBytes = cast<ConstantInt>(MCI->getLength())->getZExtValue(); |
| 786 | if ((SizeInBytes >> 32) != 0) |
| 787 | return false; |
| 788 | |
| 789 | // Check if the stride matches the size of the memcpy. If so, then we know |
| 790 | // that every byte is touched in the loop. |
| 791 | const SCEVConstant *ConstStoreStride = |
| 792 | dyn_cast<SCEVConstant>(StoreEv->getOperand(1)); |
| 793 | const SCEVConstant *ConstLoadStride = |
| 794 | dyn_cast<SCEVConstant>(LoadEv->getOperand(1)); |
| 795 | if (!ConstStoreStride || !ConstLoadStride) |
| 796 | return false; |
| 797 | |
| 798 | APInt StoreStrideValue = ConstStoreStride->getAPInt(); |
| 799 | APInt LoadStrideValue = ConstLoadStride->getAPInt(); |
| 800 | // Huge stride value - give up |
| 801 | if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64) |
| 802 | return false; |
| 803 | |
| 804 | if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) { |
| 805 | ORE.emit([&]() { |
| 806 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "SizeStrideUnequal", MCI) |
| 807 | << ore::NV("Inst", "memcpy") << " in " |
| 808 | << ore::NV("Function", MCI->getFunction()) |
| 809 | << " function will not be hoisted: " |
| 810 | << ore::NV("Reason", "memcpy size is not equal to stride"); |
| 811 | }); |
| 812 | return false; |
| 813 | } |
| 814 | |
| 815 | int64_t StoreStrideInt = StoreStrideValue.getSExtValue(); |
| 816 | int64_t LoadStrideInt = LoadStrideValue.getSExtValue(); |
| 817 | // Check if the load stride matches the store stride. |
| 818 | if (StoreStrideInt != LoadStrideInt) |
| 819 | return false; |
| 820 | |
| 821 | return processLoopStoreOfLoopLoad( |
| 822 | Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes), |
| 823 | MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv, |
| 824 | BECount); |
| 825 | } |
| 826 | |
| 827 | /// processLoopMemSet - See if this memset can be promoted to a large memset. |
| 828 | bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, |
| 829 | const SCEV *BECount) { |
| 830 | // We can only handle non-volatile memsets. |
| 831 | if (MSI->isVolatile()) |
| 832 | return false; |
| 833 | |
| 834 | // If we're not allowed to hack on memset, we fail. |
| 835 | if (!HasMemset || DisableLIRP::Memset) |
| 836 | return false; |
| 837 | |
| 838 | Value *Pointer = MSI->getDest(); |
| 839 | |
| 840 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 841 | // loop, which indicates a strided store. If we have something else, it's a |
| 842 | // random store we can't handle. |
| 843 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer)); |
| 844 | if (!Ev || Ev->getLoop() != CurLoop) |
| 845 | return false; |
| 846 | if (!Ev->isAffine()) { |
| 847 | LLVM_DEBUG(dbgs() << " Pointer is not affine, abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Pointer is not affine, abort\n" ; } } while (false); |
| 848 | return false; |
| 849 | } |
| 850 | |
| 851 | const SCEV *PointerStrideSCEV = Ev->getOperand(1); |
| 852 | const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength()); |
| 853 | if (!PointerStrideSCEV || !MemsetSizeSCEV) |
| 854 | return false; |
| 855 | |
| 856 | bool IsNegStride = false; |
| 857 | const bool IsConstantSize = isa<ConstantInt>(MSI->getLength()); |
| 858 | |
| 859 | if (IsConstantSize) { |
| 860 | // Memset size is constant. |
| 861 | // Check if the pointer stride matches the memset size. If so, then |
| 862 | // we know that every byte is touched in the loop. |
| 863 | LLVM_DEBUG(dbgs() << " memset size is constant\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is constant\n" ; } } while (false); |
| 864 | uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); |
| 865 | const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
| 866 | if (!ConstStride) |
| 867 | return false; |
| 868 | |
| 869 | APInt Stride = ConstStride->getAPInt(); |
| 870 | if (SizeInBytes != Stride && SizeInBytes != -Stride) |
| 871 | return false; |
| 872 | |
| 873 | IsNegStride = SizeInBytes == -Stride; |
| 874 | } else { |
| 875 | // Memset size is non-constant. |
| 876 | // Check if the pointer stride matches the memset size. |
| 877 | // To be conservative, the pass would not promote pointers that aren't in |
| 878 | // address space zero. Also, the pass only handles memset length and stride |
| 879 | // that are invariant for the top level loop. |
| 880 | LLVM_DEBUG(dbgs() << " memset size is non-constant\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is non-constant\n" ; } } while (false); |
| 881 | if (Pointer->getType()->getPointerAddressSpace() != 0) { |
| 882 | LLVM_DEBUG(dbgs() << " pointer is not in address space zero, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " pointer is not in address space zero, " << "abort\n"; } } while (false) |
| 883 | << "abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " pointer is not in address space zero, " << "abort\n"; } } while (false); |
| 884 | return false; |
| 885 | } |
| 886 | if (!SE->isLoopInvariant(MemsetSizeSCEV, CurLoop)) { |
| 887 | LLVM_DEBUG(dbgs() << " memset size is not a loop-invariant, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is not a loop-invariant, " << "abort\n"; } } while (false) |
| 888 | << "abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is not a loop-invariant, " << "abort\n"; } } while (false); |
| 889 | return false; |
| 890 | } |
| 891 | |
| 892 | // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV |
| 893 | IsNegStride = PointerStrideSCEV->isNonConstantNegative(); |
| 894 | const SCEV *PositiveStrideSCEV = |
| 895 | IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV) |
| 896 | : PointerStrideSCEV; |
| 897 | LLVM_DEBUG(dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ) |
| 898 | << " PositiveStrideSCEV: " << *PositiveStrideSCEVdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ) |
| 899 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ); |
| 900 | |
| 901 | if (PositiveStrideSCEV != MemsetSizeSCEV) { |
| 902 | // If an expression is covered by the loop guard, compare again and |
| 903 | // proceed with optimization if equal. |
| 904 | const SCEV *FoldedPositiveStride = |
| 905 | SE->applyLoopGuards(PositiveStrideSCEV, CurLoop); |
| 906 | const SCEV *FoldedMemsetSize = |
| 907 | SE->applyLoopGuards(MemsetSizeSCEV, CurLoop); |
| 908 | |
| 909 | LLVM_DEBUG(dbgs() << " Try to fold SCEV based on loop guard\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Try to fold SCEV based on loop guard\n" << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n" << " FoldedPositiveStride: " << *FoldedPositiveStride << "\n"; } } while (false) |
| 910 | << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Try to fold SCEV based on loop guard\n" << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n" << " FoldedPositiveStride: " << *FoldedPositiveStride << "\n"; } } while (false) |
| 911 | << " FoldedPositiveStride: " << *FoldedPositiveStridedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Try to fold SCEV based on loop guard\n" << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n" << " FoldedPositiveStride: " << *FoldedPositiveStride << "\n"; } } while (false) |
| 912 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Try to fold SCEV based on loop guard\n" << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n" << " FoldedPositiveStride: " << *FoldedPositiveStride << "\n"; } } while (false); |
| 913 | |
| 914 | if (FoldedPositiveStride != FoldedMemsetSize) { |
| 915 | LLVM_DEBUG(dbgs() << " SCEV don't match, abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " SCEV don't match, abort\n" ; } } while (false); |
| 916 | return false; |
| 917 | } |
| 918 | } |
| 919 | } |
| 920 | |
| 921 | // Verify that the memset value is loop invariant. If not, we can't promote |
| 922 | // the memset. |
| 923 | Value *SplatValue = MSI->getValue(); |
| 924 | if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) |
| 925 | return false; |
| 926 | |
| 927 | SmallPtrSet<Instruction *, 1> MSIs; |
| 928 | MSIs.insert(MSI); |
| 929 | return processLoopStridedStore(Pointer, SE->getSCEV(MSI->getLength()), |
| 930 | MSI->getDestAlign(), SplatValue, MSI, MSIs, Ev, |
| 931 | BECount, IsNegStride, /*IsLoopMemset=*/true); |
| 932 | } |
| 933 | |
| 934 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
| 935 | /// specified pointer location, which is a loop-strided access. The 'Access' |
| 936 | /// argument specifies what the verboten forms of access are (read or write). |
| 937 | static bool |
| 938 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
| 939 | const SCEV *BECount, const SCEV *StoreSizeSCEV, |
| 940 | AliasAnalysis &AA, |
| 941 | SmallPtrSetImpl<Instruction *> &IgnoredInsts) { |
| 942 | // Get the location that may be stored across the loop. Since the access is |
| 943 | // strided positively through memory, we say that the modified location starts |
| 944 | // at the pointer and has infinite size. |
| 945 | LocationSize AccessSize = LocationSize::afterPointer(); |
| 946 | |
| 947 | // If the loop iterates a fixed number of times, we can refine the access size |
| 948 | // to be exactly the size of the memset, which is (BECount+1)*StoreSize |
| 949 | const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount); |
| 950 | const SCEVConstant *ConstSize = dyn_cast<SCEVConstant>(StoreSizeSCEV); |
| 951 | if (BECst && ConstSize) |
| 952 | AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * |
| 953 | ConstSize->getValue()->getZExtValue()); |
| 954 | |
| 955 | // TODO: For this to be really effective, we have to dive into the pointer |
| 956 | // operand in the store. Store to &A[i] of 100 will always return may alias |
| 957 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
| 958 | // which will then no-alias a store to &A[100]. |
| 959 | MemoryLocation StoreLoc(Ptr, AccessSize); |
| 960 | |
| 961 | for (BasicBlock *B : L->blocks()) |
| 962 | for (Instruction &I : *B) |
| 963 | if (!IgnoredInsts.contains(&I) && |
| 964 | isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access)) |
| 965 | return true; |
| 966 | return false; |
| 967 | } |
| 968 | |
| 969 | // If we have a negative stride, Start refers to the end of the memory location |
| 970 | // we're trying to memset. Therefore, we need to recompute the base pointer, |
| 971 | // which is just Start - BECount*Size. |
| 972 | static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, |
| 973 | Type *IntPtr, const SCEV *StoreSizeSCEV, |
| 974 | ScalarEvolution *SE) { |
| 975 | const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); |
| 976 | if (!StoreSizeSCEV->isOne()) { |
| 977 | // index = back edge count * store size |
| 978 | Index = SE->getMulExpr(Index, |
| 979 | SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), |
| 980 | SCEV::FlagNUW); |
| 981 | } |
| 982 | // base pointer = start - index * store size |
| 983 | return SE->getMinusSCEV(Start, Index); |
| 984 | } |
| 985 | |
| 986 | /// Compute the number of bytes as a SCEV from the backedge taken count. |
| 987 | /// |
| 988 | /// This also maps the SCEV into the provided type and tries to handle the |
| 989 | /// computation in a way that will fold cleanly. |
| 990 | static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, |
| 991 | const SCEV *StoreSizeSCEV, Loop *CurLoop, |
| 992 | const DataLayout *DL, ScalarEvolution *SE) { |
| 993 | const SCEV *TripCountSCEV = |
| 994 | SE->getTripCountFromExitCount(BECount, IntPtr, CurLoop); |
| 995 | return SE->getMulExpr(TripCountSCEV, |
| 996 | SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), |
| 997 | SCEV::FlagNUW); |
| 998 | } |
| 999 | |
| 1000 | /// processLoopStridedStore - We see a strided store of some value. If we can |
| 1001 | /// transform this into a memset or memset_pattern in the loop preheader, do so. |
| 1002 | bool LoopIdiomRecognize::processLoopStridedStore( |
| 1003 | Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment, |
| 1004 | Value *StoredVal, Instruction *TheStore, |
| 1005 | SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev, |
| 1006 | const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) { |
| 1007 | Module *M = TheStore->getModule(); |
| 1008 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
| 1009 | Constant *PatternValue = nullptr; |
| 1010 | |
| 1011 | if (!SplatValue) |
| 1012 | PatternValue = getMemSetPatternValue(StoredVal, DL); |
| 1013 | |
| 1014 | assert((SplatValue || PatternValue) &&(static_cast <bool> ((SplatValue || PatternValue) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1015, __extension__ __PRETTY_FUNCTION__)) |
| 1015 | "Expected either splat value or pattern value.")(static_cast <bool> ((SplatValue || PatternValue) && "Expected either splat value or pattern value.") ? void (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1015, __extension__ __PRETTY_FUNCTION__)); |
| 1016 | |
| 1017 | // The trip count of the loop and the base pointer of the addrec SCEV is |
| 1018 | // guaranteed to be loop invariant, which means that it should dominate the |
| 1019 | // header. This allows us to insert code for it in the preheader. |
| 1020 | unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); |
| 1021 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
| 1022 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 1023 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
| 1024 | SCEVExpanderCleaner ExpCleaner(Expander); |
| 1025 | |
| 1026 | Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); |
| 1027 | Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); |
| 1028 | |
| 1029 | bool Changed = false; |
| 1030 | const SCEV *Start = Ev->getStart(); |
| 1031 | // Handle negative strided loops. |
| 1032 | if (IsNegStride) |
| 1033 | Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSizeSCEV, SE); |
| 1034 | |
| 1035 | // TODO: ideally we should still be able to generate memset if SCEV expander |
| 1036 | // is taught to generate the dependencies at the latest point. |
| 1037 | if (!Expander.isSafeToExpand(Start)) |
| 1038 | return Changed; |
| 1039 | |
| 1040 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
| 1041 | // this into a memset in the loop preheader now if we want. However, this |
| 1042 | // would be unsafe to do if there is anything else in the loop that may read |
| 1043 | // or write to the aliased location. Check for any overlap by generating the |
| 1044 | // base pointer and checking the region. |
| 1045 | Value *BasePtr = |
| 1046 | Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); |
| 1047 | |
| 1048 | // From here on out, conservatively report to the pass manager that we've |
| 1049 | // changed the IR, even if we later clean up these added instructions. There |
| 1050 | // may be structural differences e.g. in the order of use lists not accounted |
| 1051 | // for in just a textual dump of the IR. This is written as a variable, even |
| 1052 | // though statically all the places this dominates could be replaced with |
| 1053 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
| 1054 | // the return value will read this comment, and leave them alone. |
| 1055 | Changed = true; |
| 1056 | |
| 1057 | if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
| 1058 | StoreSizeSCEV, *AA, Stores)) |
| 1059 | return Changed; |
| 1060 | |
| 1061 | if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) |
| 1062 | return Changed; |
| 1063 | |
| 1064 | // Okay, everything looks good, insert the memset. |
| 1065 | |
| 1066 | const SCEV *NumBytesS = |
| 1067 | getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); |
| 1068 | |
| 1069 | // TODO: ideally we should still be able to generate memset if SCEV expander |
| 1070 | // is taught to generate the dependencies at the latest point. |
| 1071 | if (!Expander.isSafeToExpand(NumBytesS)) |
| 1072 | return Changed; |
| 1073 | |
| 1074 | Value *NumBytes = |
| 1075 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
| 1076 | |
| 1077 | CallInst *NewCall; |
| 1078 | if (SplatValue) { |
| 1079 | AAMDNodes AATags = TheStore->getAAMetadata(); |
| 1080 | for (Instruction *Store : Stores) |
| 1081 | AATags = AATags.merge(Store->getAAMetadata()); |
| 1082 | if (auto CI = dyn_cast<ConstantInt>(NumBytes)) |
| 1083 | AATags = AATags.extendTo(CI->getZExtValue()); |
| 1084 | else |
| 1085 | AATags = AATags.extendTo(-1); |
| 1086 | |
| 1087 | NewCall = Builder.CreateMemSet( |
| 1088 | BasePtr, SplatValue, NumBytes, MaybeAlign(StoreAlignment), |
| 1089 | /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); |
| 1090 | } else if (isLibFuncEmittable(M, TLI, LibFunc_memset_pattern16)) { |
| 1091 | // Everything is emitted in default address space |
| 1092 | Type *Int8PtrTy = DestInt8PtrTy; |
| 1093 | |
| 1094 | StringRef FuncName = "memset_pattern16"; |
| 1095 | FunctionCallee MSP = getOrInsertLibFunc(M, *TLI, LibFunc_memset_pattern16, |
| 1096 | Builder.getVoidTy(), Int8PtrTy, Int8PtrTy, IntIdxTy); |
| 1097 | inferNonMandatoryLibFuncAttrs(M, FuncName, *TLI); |
| 1098 | |
| 1099 | // Otherwise we should form a memset_pattern16. PatternValue is known to be |
| 1100 | // an constant array of 16-bytes. Plop the value into a mergable global. |
| 1101 | GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, |
| 1102 | GlobalValue::PrivateLinkage, |
| 1103 | PatternValue, ".memset_pattern"); |
| 1104 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
| 1105 | GV->setAlignment(Align(16)); |
| 1106 | Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); |
| 1107 | NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); |
| 1108 | } else |
| 1109 | return Changed; |
| 1110 | |
| 1111 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
| 1112 | |
| 1113 | if (MSSAU) { |
| 1114 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
| 1115 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
| 1116 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
| 1117 | } |
| 1118 | |
| 1119 | 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) |
| 1120 | << " 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) |
| 1121 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false); |
| 1122 | |
| 1123 | ORE.emit([&]() { |
| 1124 | OptimizationRemark R(DEBUG_TYPE"loop-idiom", "ProcessLoopStridedStore", |
| 1125 | NewCall->getDebugLoc(), Preheader); |
| 1126 | R << "Transformed loop-strided store in " |
| 1127 | << ore::NV("Function", TheStore->getFunction()) |
| 1128 | << " function into a call to " |
| 1129 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
| 1130 | << "() intrinsic"; |
| 1131 | if (!Stores.empty()) |
| 1132 | R << ore::setExtraArgs(); |
| 1133 | for (auto *I : Stores) { |
| 1134 | R << ore::NV("FromBlock", I->getParent()->getName()) |
| 1135 | << ore::NV("ToBlock", Preheader->getName()); |
| 1136 | } |
| 1137 | return R; |
| 1138 | }); |
| 1139 | |
| 1140 | // Okay, the memset has been formed. Zap the original store and anything that |
| 1141 | // feeds into it. |
| 1142 | for (auto *I : Stores) { |
| 1143 | if (MSSAU) |
| 1144 | MSSAU->removeMemoryAccess(I, true); |
| 1145 | deleteDeadInstruction(I); |
| 1146 | } |
| 1147 | if (MSSAU && VerifyMemorySSA) |
| 1148 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
| 1149 | ++NumMemSet; |
| 1150 | ExpCleaner.markResultUsed(); |
| 1151 | return true; |
| 1152 | } |
| 1153 | |
| 1154 | /// If the stored value is a strided load in the same loop with the same stride |
| 1155 | /// this may be transformable into a memcpy. This kicks in for stuff like |
| 1156 | /// for (i) A[i] = B[i]; |
| 1157 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, |
| 1158 | const SCEV *BECount) { |
| 1159 | assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.")(static_cast <bool> (SI->isUnordered() && "Expected only non-volatile non-ordered stores." ) ? void (0) : __assert_fail ("SI->isUnordered() && \"Expected only non-volatile non-ordered stores.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1159, __extension__ __PRETTY_FUNCTION__)); |
| 1160 | |
| 1161 | Value *StorePtr = SI->getPointerOperand(); |
| 1162 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
| 1163 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
| 1164 | |
| 1165 | // The store must be feeding a non-volatile load. |
| 1166 | LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); |
| 1167 | assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.")(static_cast <bool> (LI->isUnordered() && "Expected only non-volatile non-ordered loads." ) ? void (0) : __assert_fail ("LI->isUnordered() && \"Expected only non-volatile non-ordered loads.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1167, __extension__ __PRETTY_FUNCTION__)); |
| 1168 | |
| 1169 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 1170 | // loop, which indicates a strided load. If we have something else, it's a |
| 1171 | // random load we can't handle. |
| 1172 | Value *LoadPtr = LI->getPointerOperand(); |
| 1173 | const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr)); |
| 1174 | |
| 1175 | const SCEV *StoreSizeSCEV = SE->getConstant(StorePtr->getType(), StoreSize); |
| 1176 | return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSizeSCEV, |
| 1177 | SI->getAlign(), LI->getAlign(), SI, LI, |
| 1178 | StoreEv, LoadEv, BECount); |
| 1179 | } |
| 1180 | |
| 1181 | namespace { |
| 1182 | class MemmoveVerifier { |
| 1183 | public: |
| 1184 | explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr, |
| 1185 | const DataLayout &DL) |
| 1186 | : DL(DL), BP1(llvm::GetPointerBaseWithConstantOffset( |
| 1187 | LoadBasePtr.stripPointerCasts(), LoadOff, DL)), |
| 1188 | BP2(llvm::GetPointerBaseWithConstantOffset( |
| 1189 | StoreBasePtr.stripPointerCasts(), StoreOff, DL)), |
| 1190 | IsSameObject(BP1 == BP2) {} |
| 1191 | |
| 1192 | bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride, |
| 1193 | const Instruction &TheLoad, |
| 1194 | bool IsMemCpy) const { |
| 1195 | if (IsMemCpy) { |
| 1196 | // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr |
| 1197 | // for negative stride. |
| 1198 | if ((!IsNegStride && LoadOff <= StoreOff) || |
| 1199 | (IsNegStride && LoadOff >= StoreOff)) |
| 1200 | return false; |
| 1201 | } else { |
| 1202 | // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr |
| 1203 | // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr. |
| 1204 | int64_t LoadSize = |
| 1205 | DL.getTypeSizeInBits(TheLoad.getType()).getFixedValue() / 8; |
| 1206 | if (BP1 != BP2 || LoadSize != int64_t(StoreSize)) |
| 1207 | return false; |
| 1208 | if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) || |
| 1209 | (IsNegStride && LoadOff + LoadSize > StoreOff)) |
| 1210 | return false; |
| 1211 | } |
| 1212 | return true; |
| 1213 | } |
| 1214 | |
| 1215 | private: |
| 1216 | const DataLayout &DL; |
| 1217 | int64_t LoadOff = 0; |
| 1218 | int64_t StoreOff = 0; |
| 1219 | const Value *BP1; |
| 1220 | const Value *BP2; |
| 1221 | |
| 1222 | public: |
| 1223 | const bool IsSameObject; |
| 1224 | }; |
| 1225 | } // namespace |
| 1226 | |
| 1227 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad( |
| 1228 | Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV, |
| 1229 | MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore, |
| 1230 | Instruction *TheLoad, const SCEVAddRecExpr *StoreEv, |
| 1231 | const SCEVAddRecExpr *LoadEv, const SCEV *BECount) { |
| 1232 | |
| 1233 | // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to |
| 1234 | // conservatively bail here, since otherwise we may have to transform |
| 1235 | // llvm.memcpy.inline into llvm.memcpy which is illegal. |
| 1236 | if (isa<MemCpyInlineInst>(TheStore)) |
| 1237 | return false; |
| 1238 | |
| 1239 | // The trip count of the loop and the base pointer of the addrec SCEV is |
| 1240 | // guaranteed to be loop invariant, which means that it should dominate the |
| 1241 | // header. This allows us to insert code for it in the preheader. |
| 1242 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
| 1243 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 1244 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
| 1245 | |
| 1246 | SCEVExpanderCleaner ExpCleaner(Expander); |
| 1247 | |
| 1248 | bool Changed = false; |
| 1249 | const SCEV *StrStart = StoreEv->getStart(); |
| 1250 | unsigned StrAS = DestPtr->getType()->getPointerAddressSpace(); |
| 1251 | Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); |
| 1252 | |
| 1253 | APInt Stride = getStoreStride(StoreEv); |
| 1254 | const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(StoreSizeSCEV); |
| 1255 | |
| 1256 | // TODO: Deal with non-constant size; Currently expect constant store size |
| 1257 | assert(ConstStoreSize && "store size is expected to be a constant")(static_cast <bool> (ConstStoreSize && "store size is expected to be a constant" ) ? void (0) : __assert_fail ("ConstStoreSize && \"store size is expected to be a constant\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1257, __extension__ __PRETTY_FUNCTION__)); |
| 1258 | |
| 1259 | int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue(); |
| 1260 | bool IsNegStride = StoreSize == -Stride; |
| 1261 | |
| 1262 | // Handle negative strided loops. |
| 1263 | if (IsNegStride) |
| 1264 | StrStart = |
| 1265 | getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSizeSCEV, SE); |
| 1266 | |
| 1267 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
| 1268 | // this into a memcpy in the loop preheader now if we want. However, this |
| 1269 | // would be unsafe to do if there is anything else in the loop that may read |
| 1270 | // or write the memory region we're storing to. This includes the load that |
| 1271 | // feeds the stores. Check for an alias by generating the base address and |
| 1272 | // checking everything. |
| 1273 | Value *StoreBasePtr = Expander.expandCodeFor( |
| 1274 | StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); |
| 1275 | |
| 1276 | // From here on out, conservatively report to the pass manager that we've |
| 1277 | // changed the IR, even if we later clean up these added instructions. There |
| 1278 | // may be structural differences e.g. in the order of use lists not accounted |
| 1279 | // for in just a textual dump of the IR. This is written as a variable, even |
| 1280 | // though statically all the places this dominates could be replaced with |
| 1281 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
| 1282 | // the return value will read this comment, and leave them alone. |
| 1283 | Changed = true; |
| 1284 | |
| 1285 | SmallPtrSet<Instruction *, 2> IgnoredInsts; |
| 1286 | IgnoredInsts.insert(TheStore); |
| 1287 | |
| 1288 | bool IsMemCpy = isa<MemCpyInst>(TheStore); |
| 1289 | const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store"; |
| 1290 | |
| 1291 | bool LoopAccessStore = |
| 1292 | mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
| 1293 | StoreSizeSCEV, *AA, IgnoredInsts); |
| 1294 | if (LoopAccessStore) { |
| 1295 | // For memmove case it's not enough to guarantee that loop doesn't access |
| 1296 | // TheStore and TheLoad. Additionally we need to make sure that TheStore is |
| 1297 | // the only user of TheLoad. |
| 1298 | if (!TheLoad->hasOneUse()) |
| 1299 | return Changed; |
| 1300 | IgnoredInsts.insert(TheLoad); |
| 1301 | if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, |
| 1302 | BECount, StoreSizeSCEV, *AA, IgnoredInsts)) { |
| 1303 | ORE.emit([&]() { |
| 1304 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "LoopMayAccessStore", |
| 1305 | TheStore) |
| 1306 | << ore::NV("Inst", InstRemark) << " in " |
| 1307 | << ore::NV("Function", TheStore->getFunction()) |
| 1308 | << " function will not be hoisted: " |
| 1309 | << ore::NV("Reason", "The loop may access store location"); |
| 1310 | }); |
| 1311 | return Changed; |
| 1312 | } |
| 1313 | IgnoredInsts.erase(TheLoad); |
| 1314 | } |
| 1315 | |
| 1316 | const SCEV *LdStart = LoadEv->getStart(); |
| 1317 | unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace(); |
| 1318 | |
| 1319 | // Handle negative strided loops. |
| 1320 | if (IsNegStride) |
| 1321 | LdStart = |
| 1322 | getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSizeSCEV, SE); |
| 1323 | |
| 1324 | // For a memcpy, we have to make sure that the input array is not being |
| 1325 | // mutated by the loop. |
| 1326 | Value *LoadBasePtr = Expander.expandCodeFor( |
| 1327 | LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); |
| 1328 | |
| 1329 | // If the store is a memcpy instruction, we must check if it will write to |
| 1330 | // the load memory locations. So remove it from the ignored stores. |
| 1331 | MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL); |
| 1332 | if (IsMemCpy && !Verifier.IsSameObject) |
| 1333 | IgnoredInsts.erase(TheStore); |
| 1334 | if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, |
| 1335 | StoreSizeSCEV, *AA, IgnoredInsts)) { |
| 1336 | ORE.emit([&]() { |
| 1337 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "LoopMayAccessLoad", TheLoad) |
| 1338 | << ore::NV("Inst", InstRemark) << " in " |
| 1339 | << ore::NV("Function", TheStore->getFunction()) |
| 1340 | << " function will not be hoisted: " |
| 1341 | << ore::NV("Reason", "The loop may access load location"); |
| 1342 | }); |
| 1343 | return Changed; |
| 1344 | } |
| 1345 | |
| 1346 | bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore; |
| 1347 | if (UseMemMove) |
| 1348 | if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, *TheLoad, |
| 1349 | IsMemCpy)) |
| 1350 | return Changed; |
| 1351 | |
| 1352 | if (avoidLIRForMultiBlockLoop()) |
| 1353 | return Changed; |
| 1354 | |
| 1355 | // Okay, everything is safe, we can transform this! |
| 1356 | |
| 1357 | const SCEV *NumBytesS = |
| 1358 | getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); |
| 1359 | |
| 1360 | Value *NumBytes = |
| 1361 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
| 1362 | |
| 1363 | AAMDNodes AATags = TheLoad->getAAMetadata(); |
| 1364 | AAMDNodes StoreAATags = TheStore->getAAMetadata(); |
| 1365 | AATags = AATags.merge(StoreAATags); |
| 1366 | if (auto CI = dyn_cast<ConstantInt>(NumBytes)) |
| 1367 | AATags = AATags.extendTo(CI->getZExtValue()); |
| 1368 | else |
| 1369 | AATags = AATags.extendTo(-1); |
| 1370 | |
| 1371 | CallInst *NewCall = nullptr; |
| 1372 | // Check whether to generate an unordered atomic memcpy: |
| 1373 | // If the load or store are atomic, then they must necessarily be unordered |
| 1374 | // by previous checks. |
| 1375 | if (!TheStore->isAtomic() && !TheLoad->isAtomic()) { |
| 1376 | if (UseMemMove) |
| 1377 | NewCall = Builder.CreateMemMove( |
| 1378 | StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, |
| 1379 | /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); |
| 1380 | else |
| 1381 | NewCall = |
| 1382 | Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, |
| 1383 | NumBytes, /*isVolatile=*/false, AATags.TBAA, |
| 1384 | AATags.TBAAStruct, AATags.Scope, AATags.NoAlias); |
| 1385 | } else { |
| 1386 | // For now don't support unordered atomic memmove. |
| 1387 | if (UseMemMove) |
| 1388 | return Changed; |
| 1389 | // We cannot allow unaligned ops for unordered load/store, so reject |
| 1390 | // anything where the alignment isn't at least the element size. |
| 1391 | assert((StoreAlign && LoadAlign) &&(static_cast <bool> ((StoreAlign && LoadAlign) && "Expect unordered load/store to have align.") ? void (0) : __assert_fail ("(StoreAlign && LoadAlign) && \"Expect unordered load/store to have align.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1392, __extension__ __PRETTY_FUNCTION__)) |
| 1392 | "Expect unordered load/store to have align.")(static_cast <bool> ((StoreAlign && LoadAlign) && "Expect unordered load/store to have align.") ? void (0) : __assert_fail ("(StoreAlign && LoadAlign) && \"Expect unordered load/store to have align.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1392, __extension__ __PRETTY_FUNCTION__)); |
| 1393 | if (*StoreAlign < StoreSize || *LoadAlign < StoreSize) |
| 1394 | return Changed; |
| 1395 | |
| 1396 | // If the element.atomic memcpy is not lowered into explicit |
| 1397 | // loads/stores later, then it will be lowered into an element-size |
| 1398 | // specific lib call. If the lib call doesn't exist for our store size, then |
| 1399 | // we shouldn't generate the memcpy. |
| 1400 | if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) |
| 1401 | return Changed; |
| 1402 | |
| 1403 | // Create the call. |
| 1404 | // Note that unordered atomic loads/stores are *required* by the spec to |
| 1405 | // have an alignment but non-atomic loads/stores may not. |
| 1406 | NewCall = Builder.CreateElementUnorderedAtomicMemCpy( |
| 1407 | StoreBasePtr, *StoreAlign, LoadBasePtr, *LoadAlign, NumBytes, StoreSize, |
| 1408 | AATags.TBAA, AATags.TBAAStruct, AATags.Scope, AATags.NoAlias); |
| 1409 | } |
| 1410 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
| 1411 | |
| 1412 | if (MSSAU) { |
| 1413 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
| 1414 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
| 1415 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
| 1416 | } |
| 1417 | |
| 1418 | LLVM_DEBUG(dbgs() << " Formed new call: " << *NewCall << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"; } } while (false) |
| 1419 | << " from load ptr=" << *LoadEv << " at: " << *TheLoaddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"; } } while (false) |
| 1420 | << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"; } } while (false) |
| 1421 | << " from store ptr=" << *StoreEv << " at: " << *TheStoredo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"; } } while (false) |
| 1422 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"; } } while (false); |
| 1423 | |
| 1424 | ORE.emit([&]() { |
| 1425 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStoreOfLoopLoad", |
| 1426 | NewCall->getDebugLoc(), Preheader) |
| 1427 | << "Formed a call to " |
| 1428 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
| 1429 | << "() intrinsic from " << ore::NV("Inst", InstRemark) |
| 1430 | << " instruction in " << ore::NV("Function", TheStore->getFunction()) |
| 1431 | << " function" |
| 1432 | << ore::setExtraArgs() |
| 1433 | << ore::NV("FromBlock", TheStore->getParent()->getName()) |
| 1434 | << ore::NV("ToBlock", Preheader->getName()); |
| 1435 | }); |
| 1436 | |
| 1437 | // Okay, a new call to memcpy/memmove has been formed. Zap the original store |
| 1438 | // and anything that feeds into it. |
| 1439 | if (MSSAU) |
| 1440 | MSSAU->removeMemoryAccess(TheStore, true); |
| 1441 | deleteDeadInstruction(TheStore); |
| 1442 | if (MSSAU && VerifyMemorySSA) |
| 1443 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
| 1444 | if (UseMemMove) |
| 1445 | ++NumMemMove; |
| 1446 | else |
| 1447 | ++NumMemCpy; |
| 1448 | ExpCleaner.markResultUsed(); |
| 1449 | return true; |
| 1450 | } |
| 1451 | |
| 1452 | // When compiling for codesize we avoid idiom recognition for a multi-block loop |
| 1453 | // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. |
| 1454 | // |
| 1455 | bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, |
| 1456 | bool IsLoopMemset) { |
| 1457 | if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { |
| 1458 | if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) { |
| 1459 | 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) |
| 1460 | << " : 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) |
| 1461 | << " 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); |
| 1462 | return true; |
| 1463 | } |
| 1464 | } |
| 1465 | |
| 1466 | return false; |
| 1467 | } |
| 1468 | |
| 1469 | bool LoopIdiomRecognize::runOnNoncountableLoop() { |
| 1470 | 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) |
| 1471 | << 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) |
| 1472 | << "] 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) |
| 1473 | << 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); |
| 1474 | |
| 1475 | return recognizePopcount() || recognizeAndInsertFFS() || |
| 1476 | recognizeShiftUntilBitTest() || recognizeShiftUntilZero(); |
| 1477 | } |
| 1478 | |
| 1479 | /// Check if the given conditional branch is based on the comparison between |
| 1480 | /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is |
| 1481 | /// true), the control yields to the loop entry. If the branch matches the |
| 1482 | /// behavior, the variable involved in the comparison is returned. This function |
| 1483 | /// will be called to see if the precondition and postcondition of the loop are |
| 1484 | /// in desirable form. |
| 1485 | static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, |
| 1486 | bool JmpOnZero = false) { |
| 1487 | if (!BI || !BI->isConditional()) |
| 1488 | return nullptr; |
| 1489 | |
| 1490 | ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); |
| 1491 | if (!Cond) |
| 1492 | return nullptr; |
| 1493 | |
| 1494 | ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1)); |
| 1495 | if (!CmpZero || !CmpZero->isZero()) |
| 1496 | return nullptr; |
| 1497 | |
| 1498 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
| 1499 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
| 1500 | if (JmpOnZero) |
| 1501 | std::swap(TrueSucc, FalseSucc); |
| 1502 | |
| 1503 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
| 1504 | if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || |
| 1505 | (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) |
| 1506 | return Cond->getOperand(0); |
| 1507 | |
| 1508 | return nullptr; |
| 1509 | } |
| 1510 | |
| 1511 | // Check if the recurrence variable `VarX` is in the right form to create |
| 1512 | // the idiom. Returns the value coerced to a PHINode if so. |
| 1513 | static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, |
| 1514 | BasicBlock *LoopEntry) { |
| 1515 | auto *PhiX = dyn_cast<PHINode>(VarX); |
| 1516 | if (PhiX && PhiX->getParent() == LoopEntry && |
| 1517 | (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) |
| 1518 | return PhiX; |
| 1519 | return nullptr; |
| 1520 | } |
| 1521 | |
| 1522 | /// Return true iff the idiom is detected in the loop. |
| 1523 | /// |
| 1524 | /// Additionally: |
| 1525 | /// 1) \p CntInst is set to the instruction counting the population bit. |
| 1526 | /// 2) \p CntPhi is set to the corresponding phi node. |
| 1527 | /// 3) \p Var is set to the value whose population bits are being counted. |
| 1528 | /// |
| 1529 | /// The core idiom we are trying to detect is: |
| 1530 | /// \code |
| 1531 | /// if (x0 != 0) |
| 1532 | /// goto loop-exit // the precondition of the loop |
| 1533 | /// cnt0 = init-val; |
| 1534 | /// do { |
| 1535 | /// x1 = phi (x0, x2); |
| 1536 | /// cnt1 = phi(cnt0, cnt2); |
| 1537 | /// |
| 1538 | /// cnt2 = cnt1 + 1; |
| 1539 | /// ... |
| 1540 | /// x2 = x1 & (x1 - 1); |
| 1541 | /// ... |
| 1542 | /// } while(x != 0); |
| 1543 | /// |
| 1544 | /// loop-exit: |
| 1545 | /// \endcode |
| 1546 | static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, |
| 1547 | Instruction *&CntInst, PHINode *&CntPhi, |
| 1548 | Value *&Var) { |
| 1549 | // step 1: Check to see if the look-back branch match this pattern: |
| 1550 | // "if (a!=0) goto loop-entry". |
| 1551 | BasicBlock *LoopEntry; |
| 1552 | Instruction *DefX2, *CountInst; |
| 1553 | Value *VarX1, *VarX0; |
| 1554 | PHINode *PhiX, *CountPhi; |
| 1555 | |
| 1556 | DefX2 = CountInst = nullptr; |
| 1557 | VarX1 = VarX0 = nullptr; |
| 1558 | PhiX = CountPhi = nullptr; |
| 1559 | LoopEntry = *(CurLoop->block_begin()); |
| 1560 | |
| 1561 | // step 1: Check if the loop-back branch is in desirable form. |
| 1562 | { |
| 1563 | if (Value *T = matchCondition( |
| 1564 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
| 1565 | DefX2 = dyn_cast<Instruction>(T); |
| 1566 | else |
| 1567 | return false; |
| 1568 | } |
| 1569 | |
| 1570 | // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" |
| 1571 | { |
| 1572 | if (!DefX2 || DefX2->getOpcode() != Instruction::And) |
| 1573 | return false; |
| 1574 | |
| 1575 | BinaryOperator *SubOneOp; |
| 1576 | |
| 1577 | if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0)))) |
| 1578 | VarX1 = DefX2->getOperand(1); |
| 1579 | else { |
| 1580 | VarX1 = DefX2->getOperand(0); |
| 1581 | SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1)); |
| 1582 | } |
| 1583 | if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) |
| 1584 | return false; |
| 1585 | |
| 1586 | ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1)); |
| 1587 | if (!Dec || |
| 1588 | !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || |
| 1589 | (SubOneOp->getOpcode() == Instruction::Add && |
| 1590 | Dec->isMinusOne()))) { |
| 1591 | return false; |
| 1592 | } |
| 1593 | } |
| 1594 | |
| 1595 | // step 3: Check the recurrence of variable X |
| 1596 | PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); |
| 1597 | if (!PhiX) |
| 1598 | return false; |
| 1599 | |
| 1600 | // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 |
| 1601 | { |
| 1602 | CountInst = nullptr; |
| 1603 | for (Instruction &Inst : llvm::make_range( |
| 1604 | LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { |
| 1605 | if (Inst.getOpcode() != Instruction::Add) |
| 1606 | continue; |
| 1607 | |
| 1608 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1)); |
| 1609 | if (!Inc || !Inc->isOne()) |
| 1610 | continue; |
| 1611 | |
| 1612 | PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); |
| 1613 | if (!Phi) |
| 1614 | continue; |
| 1615 | |
| 1616 | // Check if the result of the instruction is live of the loop. |
| 1617 | bool LiveOutLoop = false; |
| 1618 | for (User *U : Inst.users()) { |
| 1619 | if ((cast<Instruction>(U))->getParent() != LoopEntry) { |
| 1620 | LiveOutLoop = true; |
| 1621 | break; |
| 1622 | } |
| 1623 | } |
| 1624 | |
| 1625 | if (LiveOutLoop) { |
| 1626 | CountInst = &Inst; |
| 1627 | CountPhi = Phi; |
| 1628 | break; |
| 1629 | } |
| 1630 | } |
| 1631 | |
| 1632 | if (!CountInst) |
| 1633 | return false; |
| 1634 | } |
| 1635 | |
| 1636 | // step 5: check if the precondition is in this form: |
| 1637 | // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" |
| 1638 | { |
| 1639 | auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1640 | Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); |
| 1641 | if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) |
| 1642 | return false; |
| 1643 | |
| 1644 | CntInst = CountInst; |
| 1645 | CntPhi = CountPhi; |
| 1646 | Var = T; |
| 1647 | } |
| 1648 | |
| 1649 | return true; |
| 1650 | } |
| 1651 | |
| 1652 | /// Return true if the idiom is detected in the loop. |
| 1653 | /// |
| 1654 | /// Additionally: |
| 1655 | /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) |
| 1656 | /// or nullptr if there is no such. |
| 1657 | /// 2) \p CntPhi is set to the corresponding phi node |
| 1658 | /// or nullptr if there is no such. |
| 1659 | /// 3) \p Var is set to the value whose CTLZ could be used. |
| 1660 | /// 4) \p DefX is set to the instruction calculating Loop exit condition. |
| 1661 | /// |
| 1662 | /// The core idiom we are trying to detect is: |
| 1663 | /// \code |
| 1664 | /// if (x0 == 0) |
| 1665 | /// goto loop-exit // the precondition of the loop |
| 1666 | /// cnt0 = init-val; |
| 1667 | /// do { |
| 1668 | /// x = phi (x0, x.next); //PhiX |
| 1669 | /// cnt = phi(cnt0, cnt.next); |
| 1670 | /// |
| 1671 | /// cnt.next = cnt + 1; |
| 1672 | /// ... |
| 1673 | /// x.next = x >> 1; // DefX |
| 1674 | /// ... |
| 1675 | /// } while(x.next != 0); |
| 1676 | /// |
| 1677 | /// loop-exit: |
| 1678 | /// \endcode |
| 1679 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, |
| 1680 | Intrinsic::ID &IntrinID, Value *&InitX, |
| 1681 | Instruction *&CntInst, PHINode *&CntPhi, |
| 1682 | Instruction *&DefX) { |
| 1683 | BasicBlock *LoopEntry; |
| 1684 | Value *VarX = nullptr; |
| 1685 | |
| 1686 | DefX = nullptr; |
| 1687 | CntInst = nullptr; |
| 1688 | CntPhi = nullptr; |
| 1689 | LoopEntry = *(CurLoop->block_begin()); |
| 1690 | |
| 1691 | // step 1: Check if the loop-back branch is in desirable form. |
| 1692 | if (Value *T = matchCondition( |
| 1693 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
| 1694 | DefX = dyn_cast<Instruction>(T); |
| 1695 | else |
| 1696 | return false; |
| 1697 | |
| 1698 | // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" |
| 1699 | if (!DefX || !DefX->isShift()) |
| 1700 | return false; |
| 1701 | IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : |
| 1702 | Intrinsic::ctlz; |
| 1703 | ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)); |
| 1704 | if (!Shft || !Shft->isOne()) |
| 1705 | return false; |
| 1706 | VarX = DefX->getOperand(0); |
| 1707 | |
| 1708 | // step 3: Check the recurrence of variable X |
| 1709 | PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); |
| 1710 | if (!PhiX) |
| 1711 | return false; |
| 1712 | |
| 1713 | InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); |
| 1714 | |
| 1715 | // Make sure the initial value can't be negative otherwise the ashr in the |
| 1716 | // loop might never reach zero which would make the loop infinite. |
| 1717 | if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) |
| 1718 | return false; |
| 1719 | |
| 1720 | // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 |
| 1721 | // or cnt.next = cnt + -1. |
| 1722 | // TODO: We can skip the step. If loop trip count is known (CTLZ), |
| 1723 | // then all uses of "cnt.next" could be optimized to the trip count |
| 1724 | // plus "cnt0". Currently it is not optimized. |
| 1725 | // This step could be used to detect POPCNT instruction: |
| 1726 | // cnt.next = cnt + (x.next & 1) |
| 1727 | for (Instruction &Inst : llvm::make_range( |
| 1728 | LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { |
| 1729 | if (Inst.getOpcode() != Instruction::Add) |
| 1730 | continue; |
| 1731 | |
| 1732 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1)); |
| 1733 | if (!Inc || (!Inc->isOne() && !Inc->isMinusOne())) |
| 1734 | continue; |
| 1735 | |
| 1736 | PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); |
| 1737 | if (!Phi) |
| 1738 | continue; |
| 1739 | |
| 1740 | CntInst = &Inst; |
| 1741 | CntPhi = Phi; |
| 1742 | break; |
| 1743 | } |
| 1744 | if (!CntInst) |
| 1745 | return false; |
| 1746 | |
| 1747 | return true; |
| 1748 | } |
| 1749 | |
| 1750 | /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop |
| 1751 | /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new |
| 1752 | /// trip count returns true; otherwise, returns false. |
| 1753 | bool LoopIdiomRecognize::recognizeAndInsertFFS() { |
| 1754 | // Give up if the loop has multiple blocks or multiple backedges. |
| 1755 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
| 1756 | return false; |
| 1757 | |
| 1758 | Intrinsic::ID IntrinID; |
| 1759 | Value *InitX; |
| 1760 | Instruction *DefX = nullptr; |
| 1761 | PHINode *CntPhi = nullptr; |
| 1762 | Instruction *CntInst = nullptr; |
| 1763 | // Help decide if transformation is profitable. For ShiftUntilZero idiom, |
| 1764 | // this is always 6. |
| 1765 | size_t IdiomCanonicalSize = 6; |
| 1766 | |
| 1767 | if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, |
| 1768 | CntInst, CntPhi, DefX)) |
| 1769 | return false; |
| 1770 | |
| 1771 | bool IsCntPhiUsedOutsideLoop = false; |
| 1772 | for (User *U : CntPhi->users()) |
| 1773 | if (!CurLoop->contains(cast<Instruction>(U))) { |
| 1774 | IsCntPhiUsedOutsideLoop = true; |
| 1775 | break; |
| 1776 | } |
| 1777 | bool IsCntInstUsedOutsideLoop = false; |
| 1778 | for (User *U : CntInst->users()) |
| 1779 | if (!CurLoop->contains(cast<Instruction>(U))) { |
| 1780 | IsCntInstUsedOutsideLoop = true; |
| 1781 | break; |
| 1782 | } |
| 1783 | // If both CntInst and CntPhi are used outside the loop the profitability |
| 1784 | // is questionable. |
| 1785 | if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) |
| 1786 | return false; |
| 1787 | |
| 1788 | // For some CPUs result of CTLZ(X) intrinsic is undefined |
| 1789 | // when X is 0. If we can not guarantee X != 0, we need to check this |
| 1790 | // when expand. |
| 1791 | bool ZeroCheck = false; |
| 1792 | // It is safe to assume Preheader exist as it was checked in |
| 1793 | // parent function RunOnLoop. |
| 1794 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
| 1795 | |
| 1796 | // If we are using the count instruction outside the loop, make sure we |
| 1797 | // have a zero check as a precondition. Without the check the loop would run |
| 1798 | // one iteration for before any check of the input value. This means 0 and 1 |
| 1799 | // would have identical behavior in the original loop and thus |
| 1800 | if (!IsCntPhiUsedOutsideLoop) { |
| 1801 | auto *PreCondBB = PH->getSinglePredecessor(); |
| 1802 | if (!PreCondBB) |
| 1803 | return false; |
| 1804 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1805 | if (!PreCondBI) |
| 1806 | return false; |
| 1807 | if (matchCondition(PreCondBI, PH) != InitX) |
| 1808 | return false; |
| 1809 | ZeroCheck = true; |
| 1810 | } |
| 1811 | |
| 1812 | // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always |
| 1813 | // profitable if we delete the loop. |
| 1814 | |
| 1815 | // the loop has only 6 instructions: |
| 1816 | // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] |
| 1817 | // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] |
| 1818 | // %shr = ashr %n.addr.0, 1 |
| 1819 | // %tobool = icmp eq %shr, 0 |
| 1820 | // %inc = add nsw %i.0, 1 |
| 1821 | // br i1 %tobool |
| 1822 | |
| 1823 | const Value *Args[] = {InitX, |
| 1824 | ConstantInt::getBool(InitX->getContext(), ZeroCheck)}; |
| 1825 | |
| 1826 | // @llvm.dbg doesn't count as they have no semantic effect. |
| 1827 | auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); |
| 1828 | uint32_t HeaderSize = |
| 1829 | std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); |
| 1830 | |
| 1831 | IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); |
| 1832 | InstructionCost Cost = |
| 1833 | TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); |
| 1834 | if (HeaderSize != IdiomCanonicalSize && |
| 1835 | Cost > TargetTransformInfo::TCC_Basic) |
| 1836 | return false; |
| 1837 | |
| 1838 | transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, |
| 1839 | DefX->getDebugLoc(), ZeroCheck, |
| 1840 | IsCntPhiUsedOutsideLoop); |
| 1841 | return true; |
| 1842 | } |
| 1843 | |
| 1844 | /// Recognizes a population count idiom in a non-countable loop. |
| 1845 | /// |
| 1846 | /// If detected, transforms the relevant code to issue the popcount intrinsic |
| 1847 | /// function call, and returns true; otherwise, returns false. |
| 1848 | bool LoopIdiomRecognize::recognizePopcount() { |
| 1849 | if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) |
| 1850 | return false; |
| 1851 | |
| 1852 | // Counting population are usually conducted by few arithmetic instructions. |
| 1853 | // Such instructions can be easily "absorbed" by vacant slots in a |
| 1854 | // non-compact loop. Therefore, recognizing popcount idiom only makes sense |
| 1855 | // in a compact loop. |
| 1856 | |
| 1857 | // Give up if the loop has multiple blocks or multiple backedges. |
| 1858 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
| 1859 | return false; |
| 1860 | |
| 1861 | BasicBlock *LoopBody = *(CurLoop->block_begin()); |
| 1862 | if (LoopBody->size() >= 20) { |
| 1863 | // The loop is too big, bail out. |
| 1864 | return false; |
| 1865 | } |
| 1866 | |
| 1867 | // It should have a preheader containing nothing but an unconditional branch. |
| 1868 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
| 1869 | if (!PH || &PH->front() != PH->getTerminator()) |
| 1870 | return false; |
| 1871 | auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator()); |
| 1872 | if (!EntryBI || EntryBI->isConditional()) |
| 1873 | return false; |
| 1874 | |
| 1875 | // It should have a precondition block where the generated popcount intrinsic |
| 1876 | // function can be inserted. |
| 1877 | auto *PreCondBB = PH->getSinglePredecessor(); |
| 1878 | if (!PreCondBB) |
| 1879 | return false; |
| 1880 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
| 1881 | if (!PreCondBI || PreCondBI->isUnconditional()) |
| 1882 | return false; |
| 1883 | |
| 1884 | Instruction *CntInst; |
| 1885 | PHINode *CntPhi; |
| 1886 | Value *Val; |
| 1887 | if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) |
| 1888 | return false; |
| 1889 | |
| 1890 | transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); |
| 1891 | return true; |
| 1892 | } |
| 1893 | |
| 1894 | static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
| 1895 | const DebugLoc &DL) { |
| 1896 | Value *Ops[] = {Val}; |
| 1897 | Type *Tys[] = {Val->getType()}; |
| 1898 | |
| 1899 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
| 1900 | Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); |
| 1901 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
| 1902 | CI->setDebugLoc(DL); |
| 1903 | |
| 1904 | return CI; |
| 1905 | } |
| 1906 | |
| 1907 | static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
| 1908 | const DebugLoc &DL, bool ZeroCheck, |
| 1909 | Intrinsic::ID IID) { |
| 1910 | Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)}; |
| 1911 | Type *Tys[] = {Val->getType()}; |
| 1912 | |
| 1913 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
| 1914 | Function *Func = Intrinsic::getDeclaration(M, IID, Tys); |
| 1915 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
| 1916 | CI->setDebugLoc(DL); |
| 1917 | |
| 1918 | return CI; |
| 1919 | } |
| 1920 | |
| 1921 | /// Transform the following loop (Using CTLZ, CTTZ is similar): |
| 1922 | /// loop: |
| 1923 | /// CntPhi = PHI [Cnt0, CntInst] |
| 1924 | /// PhiX = PHI [InitX, DefX] |
| 1925 | /// CntInst = CntPhi + 1 |
| 1926 | /// DefX = PhiX >> 1 |
| 1927 | /// LOOP_BODY |
| 1928 | /// Br: loop if (DefX != 0) |
| 1929 | /// Use(CntPhi) or Use(CntInst) |
| 1930 | /// |
| 1931 | /// Into: |
| 1932 | /// If CntPhi used outside the loop: |
| 1933 | /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) |
| 1934 | /// Count = CountPrev + 1 |
| 1935 | /// else |
| 1936 | /// Count = BitWidth(InitX) - CTLZ(InitX) |
| 1937 | /// loop: |
| 1938 | /// CntPhi = PHI [Cnt0, CntInst] |
| 1939 | /// PhiX = PHI [InitX, DefX] |
| 1940 | /// PhiCount = PHI [Count, Dec] |
| 1941 | /// CntInst = CntPhi + 1 |
| 1942 | /// DefX = PhiX >> 1 |
| 1943 | /// Dec = PhiCount - 1 |
| 1944 | /// LOOP_BODY |
| 1945 | /// Br: loop if (Dec != 0) |
| 1946 | /// Use(CountPrev + Cnt0) // Use(CntPhi) |
| 1947 | /// or |
| 1948 | /// Use(Count + Cnt0) // Use(CntInst) |
| 1949 | /// |
| 1950 | /// If LOOP_BODY is empty the loop will be deleted. |
| 1951 | /// If CntInst and DefX are not used in LOOP_BODY they will be removed. |
| 1952 | void LoopIdiomRecognize::transformLoopToCountable( |
| 1953 | Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, |
| 1954 | PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, |
| 1955 | bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { |
| 1956 | BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator()); |
| 1957 | |
| 1958 | // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block |
| 1959 | IRBuilder<> Builder(PreheaderBr); |
| 1960 | Builder.SetCurrentDebugLocation(DL); |
| 1961 | |
| 1962 | // If there are no uses of CntPhi crate: |
| 1963 | // Count = BitWidth - CTLZ(InitX); |
| 1964 | // NewCount = Count; |
| 1965 | // If there are uses of CntPhi create: |
| 1966 | // NewCount = BitWidth - CTLZ(InitX >> 1); |
| 1967 | // Count = NewCount + 1; |
| 1968 | Value *InitXNext; |
| 1969 | if (IsCntPhiUsedOutsideLoop) { |
| 1970 | if (DefX->getOpcode() == Instruction::AShr) |
| 1971 | InitXNext = Builder.CreateAShr(InitX, 1); |
| 1972 | else if (DefX->getOpcode() == Instruction::LShr) |
| 1973 | InitXNext = Builder.CreateLShr(InitX, 1); |
| 1974 | else if (DefX->getOpcode() == Instruction::Shl) // cttz |
| 1975 | InitXNext = Builder.CreateShl(InitX, 1); |
| 1976 | else |
| 1977 | llvm_unreachable("Unexpected opcode!")::llvm::llvm_unreachable_internal("Unexpected opcode!", "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1977); |
| 1978 | } else |
| 1979 | InitXNext = InitX; |
| 1980 | Value *Count = |
| 1981 | createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); |
| 1982 | Type *CountTy = Count->getType(); |
| 1983 | Count = Builder.CreateSub( |
| 1984 | ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count); |
| 1985 | Value *NewCount = Count; |
| 1986 | if (IsCntPhiUsedOutsideLoop) |
| 1987 | Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1)); |
| 1988 | |
| 1989 | NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType()); |
| 1990 | |
| 1991 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); |
| 1992 | if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) { |
| 1993 | // If the counter was being incremented in the loop, add NewCount to the |
| 1994 | // counter's initial value, but only if the initial value is not zero. |
| 1995 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
| 1996 | if (!InitConst || !InitConst->isZero()) |
| 1997 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
| 1998 | } else { |
| 1999 | // If the count was being decremented in the loop, subtract NewCount from |
| 2000 | // the counter's initial value. |
| 2001 | NewCount = Builder.CreateSub(CntInitVal, NewCount); |
| 2002 | } |
| 2003 | |
| 2004 | // Step 2: Insert new IV and loop condition: |
| 2005 | // loop: |
| 2006 | // ... |
| 2007 | // PhiCount = PHI [Count, Dec] |
| 2008 | // ... |
| 2009 | // Dec = PhiCount - 1 |
| 2010 | // ... |
| 2011 | // Br: loop if (Dec != 0) |
| 2012 | BasicBlock *Body = *(CurLoop->block_begin()); |
| 2013 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
| 2014 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
| 2015 | |
| 2016 | PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front()); |
| 2017 | |
| 2018 | Builder.SetInsertPoint(LbCond); |
| 2019 | Instruction *TcDec = cast<Instruction>(Builder.CreateSub( |
| 2020 | TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true)); |
| 2021 | |
| 2022 | TcPhi->addIncoming(Count, Preheader); |
| 2023 | TcPhi->addIncoming(TcDec, Body); |
| 2024 | |
| 2025 | CmpInst::Predicate Pred = |
| 2026 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
| 2027 | LbCond->setPredicate(Pred); |
| 2028 | LbCond->setOperand(0, TcDec); |
| 2029 | LbCond->setOperand(1, ConstantInt::get(CountTy, 0)); |
| 2030 | |
| 2031 | // Step 3: All the references to the original counter outside |
| 2032 | // the loop are replaced with the NewCount |
| 2033 | if (IsCntPhiUsedOutsideLoop) |
| 2034 | CntPhi->replaceUsesOutsideBlock(NewCount, Body); |
| 2035 | else |
| 2036 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
| 2037 | |
| 2038 | // step 4: Forget the "non-computable" trip-count SCEV associated with the |
| 2039 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 2040 | SE->forgetLoop(CurLoop); |
| 2041 | } |
| 2042 | |
| 2043 | void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, |
| 2044 | Instruction *CntInst, |
| 2045 | PHINode *CntPhi, Value *Var) { |
| 2046 | BasicBlock *PreHead = CurLoop->getLoopPreheader(); |
| 2047 | auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator()); |
| 2048 | const DebugLoc &DL = CntInst->getDebugLoc(); |
| 2049 | |
| 2050 | // Assuming before transformation, the loop is following: |
| 2051 | // if (x) // the precondition |
| 2052 | // do { cnt++; x &= x - 1; } while(x); |
| 2053 | |
| 2054 | // Step 1: Insert the ctpop instruction at the end of the precondition block |
| 2055 | IRBuilder<> Builder(PreCondBr); |
| 2056 | Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; |
| 2057 | { |
| 2058 | PopCnt = createPopcntIntrinsic(Builder, Var, DL); |
| 2059 | NewCount = PopCntZext = |
| 2060 | Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType())); |
| 2061 | |
| 2062 | if (NewCount != PopCnt) |
| 2063 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
| 2064 | |
| 2065 | // TripCnt is exactly the number of iterations the loop has |
| 2066 | TripCnt = NewCount; |
| 2067 | |
| 2068 | // If the population counter's initial value is not zero, insert Add Inst. |
| 2069 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); |
| 2070 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
| 2071 | if (!InitConst || !InitConst->isZero()) { |
| 2072 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
| 2073 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
| 2074 | } |
| 2075 | } |
| 2076 | |
| 2077 | // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to |
| 2078 | // "if (NewCount == 0) loop-exit". Without this change, the intrinsic |
| 2079 | // function would be partial dead code, and downstream passes will drag |
| 2080 | // it back from the precondition block to the preheader. |
| 2081 | { |
| 2082 | ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition()); |
| 2083 | |
| 2084 | Value *Opnd0 = PopCntZext; |
| 2085 | Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); |
| 2086 | if (PreCond->getOperand(0) != Var) |
| 2087 | std::swap(Opnd0, Opnd1); |
| 2088 | |
| 2089 | ICmpInst *NewPreCond = cast<ICmpInst>( |
| 2090 | Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); |
| 2091 | PreCondBr->setCondition(NewPreCond); |
| 2092 | |
| 2093 | RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); |
| 2094 | } |
| 2095 | |
| 2096 | // Step 3: Note that the population count is exactly the trip count of the |
| 2097 | // loop in question, which enable us to convert the loop from noncountable |
| 2098 | // loop into a countable one. The benefit is twofold: |
| 2099 | // |
| 2100 | // - If the loop only counts population, the entire loop becomes dead after |
| 2101 | // the transformation. It is a lot easier to prove a countable loop dead |
| 2102 | // than to prove a noncountable one. (In some C dialects, an infinite loop |
| 2103 | // isn't dead even if it computes nothing useful. In general, DCE needs |
| 2104 | // to prove a noncountable loop finite before safely delete it.) |
| 2105 | // |
| 2106 | // - If the loop also performs something else, it remains alive. |
| 2107 | // Since it is transformed to countable form, it can be aggressively |
| 2108 | // optimized by some optimizations which are in general not applicable |
| 2109 | // to a noncountable loop. |
| 2110 | // |
| 2111 | // After this step, this loop (conceptually) would look like following: |
| 2112 | // newcnt = __builtin_ctpop(x); |
| 2113 | // t = newcnt; |
| 2114 | // if (x) |
| 2115 | // do { cnt++; x &= x-1; t--) } while (t > 0); |
| 2116 | BasicBlock *Body = *(CurLoop->block_begin()); |
| 2117 | { |
| 2118 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
| 2119 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
| 2120 | Type *Ty = TripCnt->getType(); |
| 2121 | |
| 2122 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
| 2123 | |
| 2124 | Builder.SetInsertPoint(LbCond); |
| 2125 | Instruction *TcDec = cast<Instruction>( |
| 2126 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
| 2127 | "tcdec", false, true)); |
| 2128 | |
| 2129 | TcPhi->addIncoming(TripCnt, PreHead); |
| 2130 | TcPhi->addIncoming(TcDec, Body); |
| 2131 | |
| 2132 | CmpInst::Predicate Pred = |
| 2133 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; |
| 2134 | LbCond->setPredicate(Pred); |
| 2135 | LbCond->setOperand(0, TcDec); |
| 2136 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
| 2137 | } |
| 2138 | |
| 2139 | // Step 4: All the references to the original population counter outside |
| 2140 | // the loop are replaced with the NewCount -- the value returned from |
| 2141 | // __builtin_ctpop(). |
| 2142 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
| 2143 | |
| 2144 | // step 5: Forget the "non-computable" trip-count SCEV associated with the |
| 2145 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 2146 | SE->forgetLoop(CurLoop); |
| 2147 | } |
| 2148 | |
| 2149 | /// Match loop-invariant value. |
| 2150 | template <typename SubPattern_t> struct match_LoopInvariant { |
| 2151 | SubPattern_t SubPattern; |
| 2152 | const Loop *L; |
| 2153 | |
| 2154 | match_LoopInvariant(const SubPattern_t &SP, const Loop *L) |
| 2155 | : SubPattern(SP), L(L) {} |
| 2156 | |
| 2157 | template <typename ITy> bool match(ITy *V) { |
| 2158 | return L->isLoopInvariant(V) && SubPattern.match(V); |
| 2159 | } |
| 2160 | }; |
| 2161 | |
| 2162 | /// Matches if the value is loop-invariant. |
| 2163 | template <typename Ty> |
| 2164 | inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) { |
| 2165 | return match_LoopInvariant<Ty>(M, L); |
| 2166 | } |
| 2167 | |
| 2168 | /// Return true if the idiom is detected in the loop. |
| 2169 | /// |
| 2170 | /// The core idiom we are trying to detect is: |
| 2171 | /// \code |
| 2172 | /// entry: |
| 2173 | /// <...> |
| 2174 | /// %bitmask = shl i32 1, %bitpos |
| 2175 | /// br label %loop |
| 2176 | /// |
| 2177 | /// loop: |
| 2178 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
| 2179 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
| 2180 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
| 2181 | /// %x.next = shl i32 %x.curr, 1 |
| 2182 | /// <...> |
| 2183 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
| 2184 | /// |
| 2185 | /// end: |
| 2186 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
| 2187 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
| 2188 | /// <...> |
| 2189 | /// \endcode |
| 2190 | static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX, |
| 2191 | Value *&BitMask, Value *&BitPos, |
| 2192 | Value *&CurrX, Instruction *&NextX) { |
| 2193 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-bittest idiom detection.\n" ; } } while (false) |
| 2194 | " Performing shift-until-bittest idiom detection.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-bittest idiom detection.\n" ; } } while (false); |
| 2195 | |
| 2196 | // Give up if the loop has multiple blocks or multiple backedges. |
| 2197 | if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { |
| 2198 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad block/backedge count.\n" ; } } while (false); |
| 2199 | return false; |
| 2200 | } |
| 2201 | |
| 2202 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
| 2203 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
| 2204 | assert(LoopPreheaderBB && "There is always a loop preheader.")(static_cast <bool> (LoopPreheaderBB && "There is always a loop preheader." ) ? void (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2204, __extension__ __PRETTY_FUNCTION__)); |
| 2205 | |
| 2206 | using namespace PatternMatch; |
| 2207 | |
| 2208 | // Step 1: Check if the loop backedge is in desirable form. |
| 2209 | |
| 2210 | ICmpInst::Predicate Pred; |
| 2211 | Value *CmpLHS, *CmpRHS; |
| 2212 | BasicBlock *TrueBB, *FalseBB; |
| 2213 | if (!match(LoopHeaderBB->getTerminator(), |
| 2214 | m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)), |
| 2215 | m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) { |
| 2216 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge structure.\n" ; } } while (false); |
| 2217 | return false; |
| 2218 | } |
| 2219 | |
| 2220 | // Step 2: Check if the backedge's condition is in desirable form. |
| 2221 | |
| 2222 | auto MatchVariableBitMask = [&]() { |
| 2223 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
| 2224 | match(CmpLHS, |
| 2225 | m_c_And(m_Value(CurrX), |
| 2226 | m_CombineAnd( |
| 2227 | m_Value(BitMask), |
| 2228 | m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)), |
| 2229 | CurLoop)))); |
| 2230 | }; |
| 2231 | auto MatchConstantBitMask = [&]() { |
| 2232 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
| 2233 | match(CmpLHS, m_And(m_Value(CurrX), |
| 2234 | m_CombineAnd(m_Value(BitMask), m_Power2()))) && |
| 2235 | (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask))); |
| 2236 | }; |
| 2237 | auto MatchDecomposableConstantBitMask = [&]() { |
| 2238 | APInt Mask; |
| 2239 | return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) && |
| 2240 | ICmpInst::isEquality(Pred) && Mask.isPowerOf2() && |
| 2241 | (BitMask = ConstantInt::get(CurrX->getType(), Mask)) && |
| 2242 | (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2())); |
| 2243 | }; |
| 2244 | |
| 2245 | if (!MatchVariableBitMask() && !MatchConstantBitMask() && |
| 2246 | !MatchDecomposableConstantBitMask()) { |
| 2247 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge comparison.\n" ; } } while (false); |
| 2248 | return false; |
| 2249 | } |
| 2250 | |
| 2251 | // Step 3: Check if the recurrence is in desirable form. |
| 2252 | auto *CurrXPN = dyn_cast<PHINode>(CurrX); |
| 2253 | if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) { |
| 2254 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Not an expected PHI node.\n" ; } } while (false); |
| 2255 | return false; |
| 2256 | } |
| 2257 | |
| 2258 | BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB); |
| 2259 | NextX = |
| 2260 | dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB)); |
| 2261 | |
| 2262 | assert(CurLoop->isLoopInvariant(BaseX) &&(static_cast <bool> (CurLoop->isLoopInvariant(BaseX) && "Expected BaseX to be avaliable in the preheader!" ) ? void (0) : __assert_fail ("CurLoop->isLoopInvariant(BaseX) && \"Expected BaseX to be avaliable in the preheader!\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2263, __extension__ __PRETTY_FUNCTION__)) |
| 2263 | "Expected BaseX to be avaliable in the preheader!")(static_cast <bool> (CurLoop->isLoopInvariant(BaseX) && "Expected BaseX to be avaliable in the preheader!" ) ? void (0) : __assert_fail ("CurLoop->isLoopInvariant(BaseX) && \"Expected BaseX to be avaliable in the preheader!\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2263, __extension__ __PRETTY_FUNCTION__)); |
| 2264 | |
| 2265 | if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) { |
| 2266 | // FIXME: support right-shift? |
| 2267 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad recurrence.\n" ; } } while (false); |
| 2268 | return false; |
| 2269 | } |
| 2270 | |
| 2271 | // Step 4: Check if the backedge's destinations are in desirable form. |
| 2272 | |
| 2273 | assert(ICmpInst::isEquality(Pred) &&(static_cast <bool> (ICmpInst::isEquality(Pred) && "Should only get equality predicates here.") ? void (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2274, __extension__ __PRETTY_FUNCTION__)) |
| 2274 | "Should only get equality predicates here.")(static_cast <bool> (ICmpInst::isEquality(Pred) && "Should only get equality predicates here.") ? void (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2274, __extension__ __PRETTY_FUNCTION__)); |
| 2275 | |
| 2276 | // cmp-br is commutative, so canonicalize to a single variant. |
| 2277 | if (Pred != ICmpInst::Predicate::ICMP_EQ) { |
| 2278 | Pred = ICmpInst::getInversePredicate(Pred); |
| 2279 | std::swap(TrueBB, FalseBB); |
| 2280 | } |
| 2281 | |
| 2282 | // We expect to exit loop when comparison yields false, |
| 2283 | // so when it yields true we should branch back to loop header. |
| 2284 | if (TrueBB != LoopHeaderBB) { |
| 2285 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge flow.\n" ; } } while (false); |
| 2286 | return false; |
| 2287 | } |
| 2288 | |
| 2289 | // Okay, idiom checks out. |
| 2290 | return true; |
| 2291 | } |
| 2292 | |
| 2293 | /// Look for the following loop: |
| 2294 | /// \code |
| 2295 | /// entry: |
| 2296 | /// <...> |
| 2297 | /// %bitmask = shl i32 1, %bitpos |
| 2298 | /// br label %loop |
| 2299 | /// |
| 2300 | /// loop: |
| 2301 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
| 2302 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
| 2303 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
| 2304 | /// %x.next = shl i32 %x.curr, 1 |
| 2305 | /// <...> |
| 2306 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
| 2307 | /// |
| 2308 | /// end: |
| 2309 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
| 2310 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
| 2311 | /// <...> |
| 2312 | /// \endcode |
| 2313 | /// |
| 2314 | /// And transform it into: |
| 2315 | /// \code |
| 2316 | /// entry: |
| 2317 | /// %bitmask = shl i32 1, %bitpos |
| 2318 | /// %lowbitmask = add i32 %bitmask, -1 |
| 2319 | /// %mask = or i32 %lowbitmask, %bitmask |
| 2320 | /// %x.masked = and i32 %x, %mask |
| 2321 | /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked, |
| 2322 | /// i1 true) |
| 2323 | /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros |
| 2324 | /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1 |
| 2325 | /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos |
| 2326 | /// %tripcount = add i32 %backedgetakencount, 1 |
| 2327 | /// %x.curr = shl i32 %x, %backedgetakencount |
| 2328 | /// %x.next = shl i32 %x, %tripcount |
| 2329 | /// br label %loop |
| 2330 | /// |
| 2331 | /// loop: |
| 2332 | /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ] |
| 2333 | /// %loop.iv.next = add nuw i32 %loop.iv, 1 |
| 2334 | /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount |
| 2335 | /// <...> |
| 2336 | /// br i1 %loop.ivcheck, label %end, label %loop |
| 2337 | /// |
| 2338 | /// end: |
| 2339 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
| 2340 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
| 2341 | /// <...> |
| 2342 | /// \endcode |
| 2343 | bool LoopIdiomRecognize::recognizeShiftUntilBitTest() { |
| 2344 | bool MadeChange = false; |
| 2345 | |
| 2346 | Value *X, *BitMask, *BitPos, *XCurr; |
| 2347 | Instruction *XNext; |
| 2348 | if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr, |
| 2349 | XNext)) { |
| 2350 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detection failed.\n" ; } } while (false) |
| 2351 | " shift-until-bittest idiom detection failed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detection failed.\n" ; } } while (false); |
| 2352 | return MadeChange; |
| 2353 | } |
| 2354 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detected!\n" ; } } while (false); |
| 2355 | |
| 2356 | // Ok, it is the idiom we were looking for, we *could* transform this loop, |
| 2357 | // but is it profitable to transform? |
| 2358 | |
| 2359 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
| 2360 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
| 2361 | assert(LoopPreheaderBB && "There is always a loop preheader.")(static_cast <bool> (LoopPreheaderBB && "There is always a loop preheader." ) ? void (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2361, __extension__ __PRETTY_FUNCTION__)); |
| 2362 | |
| 2363 | BasicBlock *SuccessorBB = CurLoop->getExitBlock(); |
| 2364 | assert(SuccessorBB && "There is only a single successor.")(static_cast <bool> (SuccessorBB && "There is only a single successor." ) ? void (0) : __assert_fail ("SuccessorBB && \"There is only a single successor.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2364, __extension__ __PRETTY_FUNCTION__)); |
| 2365 | |
| 2366 | IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); |
| 2367 | Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc()); |
| 2368 | |
| 2369 | Intrinsic::ID IntrID = Intrinsic::ctlz; |
| 2370 | Type *Ty = X->getType(); |
| 2371 | unsigned Bitwidth = Ty->getScalarSizeInBits(); |
| 2372 | |
| 2373 | TargetTransformInfo::TargetCostKind CostKind = |
| 2374 | TargetTransformInfo::TCK_SizeAndLatency; |
| 2375 | |
| 2376 | // The rewrite is considered to be unprofitable iff and only iff the |
| 2377 | // intrinsic/shift we'll use are not cheap. Note that we are okay with *just* |
| 2378 | // making the loop countable, even if nothing else changes. |
| 2379 | IntrinsicCostAttributes Attrs( |
| 2380 | IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getTrue()}); |
| 2381 | InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); |
| 2382 | if (Cost > TargetTransformInfo::TCC_Basic) { |
| 2383 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false) |
| 2384 | " Intrinsic is too costly, not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false); |
| 2385 | return MadeChange; |
| 2386 | } |
| 2387 | if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) > |
| 2388 | TargetTransformInfo::TCC_Basic) { |
| 2389 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Shift is too costly, not beneficial\n" ; } } while (false); |
| 2390 | return MadeChange; |
| 2391 | } |
| 2392 | |
| 2393 | // Ok, transform appears worthwhile. |
| 2394 | MadeChange = true; |
| 2395 | |
| 2396 | // Step 1: Compute the loop trip count. |
| 2397 | |
| 2398 | Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty), |
| 2399 | BitPos->getName() + ".lowbitmask"); |
| 2400 | Value *Mask = |
| 2401 | Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask"); |
| 2402 | Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked"); |
| 2403 | CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic( |
| 2404 | IntrID, Ty, {XMasked, /*is_zero_undef=*/Builder.getTrue()}, |
| 2405 | /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros"); |
| 2406 | Value *XMaskedNumActiveBits = Builder.CreateSub( |
| 2407 | ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros, |
| 2408 | XMasked->getName() + ".numactivebits", /*HasNUW=*/true, |
| 2409 | /*HasNSW=*/Bitwidth != 2); |
| 2410 | Value *XMaskedLeadingOnePos = |
| 2411 | Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty), |
| 2412 | XMasked->getName() + ".leadingonepos", /*HasNUW=*/false, |
| 2413 | /*HasNSW=*/Bitwidth > 2); |
| 2414 | |
| 2415 | Value *LoopBackedgeTakenCount = Builder.CreateSub( |
| 2416 | BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount", |
| 2417 | /*HasNUW=*/true, /*HasNSW=*/true); |
| 2418 | // We know loop's backedge-taken count, but what's loop's trip count? |
| 2419 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
| 2420 | Value *LoopTripCount = |
| 2421 | Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), |
| 2422 | CurLoop->getName() + ".tripcount", /*HasNUW=*/true, |
| 2423 | /*HasNSW=*/Bitwidth != 2); |
| 2424 | |
| 2425 | // Step 2: Compute the recurrence's final value without a loop. |
| 2426 | |
| 2427 | // NewX is always safe to compute, because `LoopBackedgeTakenCount` |
| 2428 | // will always be smaller than `bitwidth(X)`, i.e. we never get poison. |
| 2429 | Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount); |
| 2430 | NewX->takeName(XCurr); |
| 2431 | if (auto *I = dyn_cast<Instruction>(NewX)) |
| 2432 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
| 2433 | |
| 2434 | Value *NewXNext; |
| 2435 | // Rewriting XNext is more complicated, however, because `X << LoopTripCount` |
| 2436 | // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen |
| 2437 | // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know |
| 2438 | // that isn't the case, we'll need to emit an alternative, safe IR. |
| 2439 | if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() || |
| 2440 | PatternMatch::match( |
| 2441 | BitPos, PatternMatch::m_SpecificInt_ICMP( |
| 2442 | ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(), |
| 2443 | Ty->getScalarSizeInBits() - 1)))) |
| 2444 | NewXNext = Builder.CreateShl(X, LoopTripCount); |
| 2445 | else { |
| 2446 | // Otherwise, just additionally shift by one. It's the smallest solution, |
| 2447 | // alternatively, we could check that NewX is INT_MIN (or BitPos is ) |
| 2448 | // and select 0 instead. |
| 2449 | NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1)); |
| 2450 | } |
| 2451 | |
| 2452 | NewXNext->takeName(XNext); |
| 2453 | if (auto *I = dyn_cast<Instruction>(NewXNext)) |
| 2454 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
| 2455 | |
| 2456 | // Step 3: Adjust the successor basic block to recieve the computed |
| 2457 | // recurrence's final value instead of the recurrence itself. |
| 2458 | |
| 2459 | XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB); |
| 2460 | XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB); |
| 2461 | |
| 2462 | // Step 4: Rewrite the loop into a countable form, with canonical IV. |
| 2463 | |
| 2464 | // The new canonical induction variable. |
| 2465 | Builder.SetInsertPoint(&LoopHeaderBB->front()); |
| 2466 | auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); |
| 2467 | |
| 2468 | // The induction itself. |
| 2469 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
| 2470 | Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); |
| 2471 | auto *IVNext = |
| 2472 | Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next", |
| 2473 | /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); |
| 2474 | |
| 2475 | // The loop trip count check. |
| 2476 | auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount, |
| 2477 | CurLoop->getName() + ".ivcheck"); |
| 2478 | Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB); |
| 2479 | LoopHeaderBB->getTerminator()->eraseFromParent(); |
| 2480 | |
| 2481 | // Populate the IV PHI. |
| 2482 | IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); |
| 2483 | IV->addIncoming(IVNext, LoopHeaderBB); |
| 2484 | |
| 2485 | // Step 5: Forget the "non-computable" trip-count SCEV associated with the |
| 2486 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 2487 | |
| 2488 | SE->forgetLoop(CurLoop); |
| 2489 | |
| 2490 | // Other passes will take care of actually deleting the loop if possible. |
| 2491 | |
| 2492 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom optimized!\n" ; } } while (false); |
| 2493 | |
| 2494 | ++NumShiftUntilBitTest; |
| 2495 | return MadeChange; |
| 2496 | } |
| 2497 | |
| 2498 | /// Return true if the idiom is detected in the loop. |
| 2499 | /// |
| 2500 | /// The core idiom we are trying to detect is: |
| 2501 | /// \code |
| 2502 | /// entry: |
| 2503 | /// <...> |
| 2504 | /// %start = <...> |
| 2505 | /// %extraoffset = <...> |
| 2506 | /// <...> |
| 2507 | /// br label %for.cond |
| 2508 | /// |
| 2509 | /// loop: |
| 2510 | /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] |
| 2511 | /// %nbits = add nsw i8 %iv, %extraoffset |
| 2512 | /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits |
| 2513 | /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 |
| 2514 | /// %iv.next = add i8 %iv, 1 |
| 2515 | /// <...> |
| 2516 | /// br i1 %val.shifted.iszero, label %end, label %loop |
| 2517 | /// |
| 2518 | /// end: |
| 2519 | /// %iv.res = phi i8 [ %iv, %loop ] <...> |
| 2520 | /// %nbits.res = phi i8 [ %nbits, %loop ] <...> |
| 2521 | /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> |
| 2522 | /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> |
| 2523 | /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> |
| 2524 | /// <...> |
| 2525 | /// \endcode |
| 2526 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE, |
| 2527 | Instruction *&ValShiftedIsZero, |
| 2528 | Intrinsic::ID &IntrinID, Instruction *&IV, |
| 2529 | Value *&Start, Value *&Val, |
| 2530 | const SCEV *&ExtraOffsetExpr, |
| 2531 | bool &InvertedCond) { |
| 2532 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-zero idiom detection.\n" ; } } while (false) |
| 2533 | " Performing shift-until-zero idiom detection.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-zero idiom detection.\n" ; } } while (false); |
| 2534 | |
| 2535 | // Give up if the loop has multiple blocks or multiple backedges. |
| 2536 | if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { |
| 2537 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad block/backedge count.\n" ; } } while (false); |
| 2538 | return false; |
| 2539 | } |
| 2540 | |
| 2541 | Instruction *ValShifted, *NBits, *IVNext; |
| 2542 | Value *ExtraOffset; |
| 2543 | |
| 2544 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
| 2545 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
| 2546 | assert(LoopPreheaderBB && "There is always a loop preheader.")(static_cast <bool> (LoopPreheaderBB && "There is always a loop preheader." ) ? void (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2546, __extension__ __PRETTY_FUNCTION__)); |
| 2547 | |
| 2548 | using namespace PatternMatch; |
| 2549 | |
| 2550 | // Step 1: Check if the loop backedge, condition is in desirable form. |
| 2551 | |
| 2552 | ICmpInst::Predicate Pred; |
| 2553 | BasicBlock *TrueBB, *FalseBB; |
| 2554 | if (!match(LoopHeaderBB->getTerminator(), |
| 2555 | m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB), |
| 2556 | m_BasicBlock(FalseBB))) || |
| 2557 | !match(ValShiftedIsZero, |
| 2558 | m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) || |
| 2559 | !ICmpInst::isEquality(Pred)) { |
| 2560 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge structure.\n" ; } } while (false); |
| 2561 | return false; |
| 2562 | } |
| 2563 | |
| 2564 | // Step 2: Check if the comparison's operand is in desirable form. |
| 2565 | // FIXME: Val could be a one-input PHI node, which we should look past. |
| 2566 | if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop), |
| 2567 | m_Instruction(NBits)))) { |
| 2568 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad comparisons value computation.\n" ; } } while (false); |
| 2569 | return false; |
| 2570 | } |
| 2571 | IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz |
| 2572 | : Intrinsic::ctlz; |
| 2573 | |
| 2574 | // Step 3: Check if the shift amount is in desirable form. |
| 2575 | |
| 2576 | if (match(NBits, m_c_Add(m_Instruction(IV), |
| 2577 | m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && |
| 2578 | (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap())) |
| 2579 | ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset)); |
| 2580 | else if (match(NBits, |
| 2581 | m_Sub(m_Instruction(IV), |
| 2582 | m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && |
| 2583 | NBits->hasNoSignedWrap()) |
| 2584 | ExtraOffsetExpr = SE->getSCEV(ExtraOffset); |
| 2585 | else { |
| 2586 | IV = NBits; |
| 2587 | ExtraOffsetExpr = SE->getZero(NBits->getType()); |
| 2588 | } |
| 2589 | |
| 2590 | // Step 4: Check if the recurrence is in desirable form. |
| 2591 | auto *IVPN = dyn_cast<PHINode>(IV); |
| 2592 | if (!IVPN || IVPN->getParent() != LoopHeaderBB) { |
| 2593 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Not an expected PHI node.\n" ; } } while (false); |
| 2594 | return false; |
| 2595 | } |
| 2596 | |
| 2597 | Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB); |
| 2598 | IVNext = dyn_cast<Instruction>(IVPN->getIncomingValueForBlock(LoopHeaderBB)); |
| 2599 | |
| 2600 | if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) { |
| 2601 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad recurrence.\n" ; } } while (false); |
| 2602 | return false; |
| 2603 | } |
| 2604 | |
| 2605 | // Step 4: Check if the backedge's destinations are in desirable form. |
| 2606 | |
| 2607 | assert(ICmpInst::isEquality(Pred) &&(static_cast <bool> (ICmpInst::isEquality(Pred) && "Should only get equality predicates here.") ? void (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2608, __extension__ __PRETTY_FUNCTION__)) |
| 2608 | "Should only get equality predicates here.")(static_cast <bool> (ICmpInst::isEquality(Pred) && "Should only get equality predicates here.") ? void (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2608, __extension__ __PRETTY_FUNCTION__)); |
| 2609 | |
| 2610 | // cmp-br is commutative, so canonicalize to a single variant. |
| 2611 | InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ; |
| 2612 | if (InvertedCond) { |
| 2613 | Pred = ICmpInst::getInversePredicate(Pred); |
Value stored to 'Pred' is never read | |
| 2614 | std::swap(TrueBB, FalseBB); |
| 2615 | } |
| 2616 | |
| 2617 | // We expect to exit loop when comparison yields true, |
| 2618 | // so when it yields false we should branch back to loop header. |
| 2619 | if (FalseBB != LoopHeaderBB) { |
| 2620 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge flow.\n" ; } } while (false); |
| 2621 | return false; |
| 2622 | } |
| 2623 | |
| 2624 | // The new, countable, loop will certainly only run a known number of |
| 2625 | // iterations, It won't be infinite. But the old loop might be infinite |
| 2626 | // under certain conditions. For logical shifts, the value will become zero |
| 2627 | // after at most bitwidth(%Val) loop iterations. However, for arithmetic |
| 2628 | // right-shift, iff the sign bit was set, the value will never become zero, |
| 2629 | // and the loop may never finish. |
| 2630 | if (ValShifted->getOpcode() == Instruction::AShr && |
| 2631 | !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) { |
| 2632 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Can not prove the loop is finite.\n" ; } } while (false); |
| 2633 | return false; |
| 2634 | } |
| 2635 | |
| 2636 | // Okay, idiom checks out. |
| 2637 | return true; |
| 2638 | } |
| 2639 | |
| 2640 | /// Look for the following loop: |
| 2641 | /// \code |
| 2642 | /// entry: |
| 2643 | /// <...> |
| 2644 | /// %start = <...> |
| 2645 | /// %extraoffset = <...> |
| 2646 | /// <...> |
| 2647 | /// br label %for.cond |
| 2648 | /// |
| 2649 | /// loop: |
| 2650 | /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] |
| 2651 | /// %nbits = add nsw i8 %iv, %extraoffset |
| 2652 | /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits |
| 2653 | /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 |
| 2654 | /// %iv.next = add i8 %iv, 1 |
| 2655 | /// <...> |
| 2656 | /// br i1 %val.shifted.iszero, label %end, label %loop |
| 2657 | /// |
| 2658 | /// end: |
| 2659 | /// %iv.res = phi i8 [ %iv, %loop ] <...> |
| 2660 | /// %nbits.res = phi i8 [ %nbits, %loop ] <...> |
| 2661 | /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> |
| 2662 | /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> |
| 2663 | /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> |
| 2664 | /// <...> |
| 2665 | /// \endcode |
| 2666 | /// |
| 2667 | /// And transform it into: |
| 2668 | /// \code |
| 2669 | /// entry: |
| 2670 | /// <...> |
| 2671 | /// %start = <...> |
| 2672 | /// %extraoffset = <...> |
| 2673 | /// <...> |
| 2674 | /// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0) |
| 2675 | /// %val.numactivebits = sub i8 8, %val.numleadingzeros |
| 2676 | /// %extraoffset.neg = sub i8 0, %extraoffset |
| 2677 | /// %tmp = add i8 %val.numactivebits, %extraoffset.neg |
| 2678 | /// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start) |
| 2679 | /// %loop.tripcount = sub i8 %iv.final, %start |
| 2680 | /// br label %loop |
| 2681 | /// |
| 2682 | /// loop: |
| 2683 | /// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ] |
| 2684 | /// %loop.iv.next = add i8 %loop.iv, 1 |
| 2685 | /// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount |
| 2686 | /// %iv = add i8 %loop.iv, %start |
| 2687 | /// <...> |
| 2688 | /// br i1 %loop.ivcheck, label %end, label %loop |
| 2689 | /// |
| 2690 | /// end: |
| 2691 | /// %iv.res = phi i8 [ %iv.final, %loop ] <...> |
| 2692 | /// <...> |
| 2693 | /// \endcode |
| 2694 | bool LoopIdiomRecognize::recognizeShiftUntilZero() { |
| 2695 | bool MadeChange = false; |
| 2696 | |
| 2697 | Instruction *ValShiftedIsZero; |
| 2698 | Intrinsic::ID IntrID; |
| 2699 | Instruction *IV; |
| 2700 | Value *Start, *Val; |
| 2701 | const SCEV *ExtraOffsetExpr; |
| 2702 | bool InvertedCond; |
| 2703 | if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV, |
| 2704 | Start, Val, ExtraOffsetExpr, InvertedCond)) { |
| 2705 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-zero idiom detection failed.\n" ; } } while (false) |
| 2706 | " shift-until-zero idiom detection failed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-zero idiom detection failed.\n" ; } } while (false); |
| 2707 | return MadeChange; |
| 2708 | } |
| 2709 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-zero idiom detected!\n" ; } } while (false); |
| 2710 | |
| 2711 | // Ok, it is the idiom we were looking for, we *could* transform this loop, |
| 2712 | // but is it profitable to transform? |
| 2713 | |
| 2714 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
| 2715 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
| 2716 | assert(LoopPreheaderBB && "There is always a loop preheader.")(static_cast <bool> (LoopPreheaderBB && "There is always a loop preheader." ) ? void (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2716, __extension__ __PRETTY_FUNCTION__)); |
| 2717 | |
| 2718 | BasicBlock *SuccessorBB = CurLoop->getExitBlock(); |
| 2719 | assert(SuccessorBB && "There is only a single successor.")(static_cast <bool> (SuccessorBB && "There is only a single successor." ) ? void (0) : __assert_fail ("SuccessorBB && \"There is only a single successor.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 2719, __extension__ __PRETTY_FUNCTION__)); |
| 2720 | |
| 2721 | IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); |
| 2722 | Builder.SetCurrentDebugLocation(IV->getDebugLoc()); |
| 2723 | |
| 2724 | Type *Ty = Val->getType(); |
| 2725 | unsigned Bitwidth = Ty->getScalarSizeInBits(); |
| 2726 | |
| 2727 | TargetTransformInfo::TargetCostKind CostKind = |
| 2728 | TargetTransformInfo::TCK_SizeAndLatency; |
| 2729 | |
| 2730 | // The rewrite is considered to be unprofitable iff and only iff the |
| 2731 | // intrinsic we'll use are not cheap. Note that we are okay with *just* |
| 2732 | // making the loop countable, even if nothing else changes. |
| 2733 | IntrinsicCostAttributes Attrs( |
| 2734 | IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getFalse()}); |
| 2735 | InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); |
| 2736 | if (Cost > TargetTransformInfo::TCC_Basic) { |
| 2737 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false) |
| 2738 | " Intrinsic is too costly, not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false); |
| 2739 | return MadeChange; |
| 2740 | } |
| 2741 | |
| 2742 | // Ok, transform appears worthwhile. |
| 2743 | MadeChange = true; |
| 2744 | |
| 2745 | bool OffsetIsZero = false; |
| 2746 | if (auto *ExtraOffsetExprC = dyn_cast<SCEVConstant>(ExtraOffsetExpr)) |
| 2747 | OffsetIsZero = ExtraOffsetExprC->isZero(); |
| 2748 | |
| 2749 | // Step 1: Compute the loop's final IV value / trip count. |
| 2750 | |
| 2751 | CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic( |
| 2752 | IntrID, Ty, {Val, /*is_zero_undef=*/Builder.getFalse()}, |
| 2753 | /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros"); |
| 2754 | Value *ValNumActiveBits = Builder.CreateSub( |
| 2755 | ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros, |
| 2756 | Val->getName() + ".numactivebits", /*HasNUW=*/true, |
| 2757 | /*HasNSW=*/Bitwidth != 2); |
| 2758 | |
| 2759 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
| 2760 | Expander.setInsertPoint(&*Builder.GetInsertPoint()); |
| 2761 | Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr); |
| 2762 | |
| 2763 | Value *ValNumActiveBitsOffset = Builder.CreateAdd( |
| 2764 | ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset", |
| 2765 | /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true); |
| 2766 | Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty}, |
| 2767 | {ValNumActiveBitsOffset, Start}, |
| 2768 | /*FMFSource=*/nullptr, "iv.final"); |
| 2769 | |
| 2770 | auto *LoopBackedgeTakenCount = cast<Instruction>(Builder.CreateSub( |
| 2771 | IVFinal, Start, CurLoop->getName() + ".backedgetakencount", |
| 2772 | /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true)); |
| 2773 | // FIXME: or when the offset was `add nuw` |
| 2774 | |
| 2775 | // We know loop's backedge-taken count, but what's loop's trip count? |
| 2776 | Value *LoopTripCount = |
| 2777 | Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), |
| 2778 | CurLoop->getName() + ".tripcount", /*HasNUW=*/true, |
| 2779 | /*HasNSW=*/Bitwidth != 2); |
| 2780 | |
| 2781 | // Step 2: Adjust the successor basic block to recieve the original |
| 2782 | // induction variable's final value instead of the orig. IV itself. |
| 2783 | |
| 2784 | IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB); |
| 2785 | |
| 2786 | // Step 3: Rewrite the loop into a countable form, with canonical IV. |
| 2787 | |
| 2788 | // The new canonical induction variable. |
| 2789 | Builder.SetInsertPoint(&LoopHeaderBB->front()); |
| 2790 | auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); |
| 2791 | |
| 2792 | // The induction itself. |
| 2793 | Builder.SetInsertPoint(LoopHeaderBB->getFirstNonPHI()); |
| 2794 | auto *CIVNext = |
| 2795 | Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next", |
| 2796 | /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); |
| 2797 | |
| 2798 | // The loop trip count check. |
| 2799 | auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount, |
| 2800 | CurLoop->getName() + ".ivcheck"); |
| 2801 | auto *NewIVCheck = CIVCheck; |
| 2802 | if (InvertedCond) { |
| 2803 | NewIVCheck = Builder.CreateNot(CIVCheck); |
| 2804 | NewIVCheck->takeName(ValShiftedIsZero); |
| 2805 | } |
| 2806 | |
| 2807 | // The original IV, but rebased to be an offset to the CIV. |
| 2808 | auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false, |
| 2809 | /*HasNSW=*/true); // FIXME: what about NUW? |
| 2810 | IVDePHId->takeName(IV); |
| 2811 | |
| 2812 | // The loop terminator. |
| 2813 | Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); |
| 2814 | Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB); |
| 2815 | LoopHeaderBB->getTerminator()->eraseFromParent(); |
| 2816 | |
| 2817 | // Populate the IV PHI. |
| 2818 | CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); |
| 2819 | CIV->addIncoming(CIVNext, LoopHeaderBB); |
| 2820 | |
| 2821 | // Step 4: Forget the "non-computable" trip-count SCEV associated with the |
| 2822 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
| 2823 | |
| 2824 | SE->forgetLoop(CurLoop); |
| 2825 | |
| 2826 | // Step 5: Try to cleanup the loop's body somewhat. |
| 2827 | IV->replaceAllUsesWith(IVDePHId); |
| 2828 | IV->eraseFromParent(); |
| 2829 | |
| 2830 | ValShiftedIsZero->replaceAllUsesWith(NewIVCheck); |
| 2831 | ValShiftedIsZero->eraseFromParent(); |
| 2832 | |
| 2833 | // Other passes will take care of actually deleting the loop if possible. |
| 2834 | |
| 2835 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-zero idiom optimized!\n" ; } } while (false); |
| 2836 | |
| 2837 | ++NumShiftUntilZero; |
| 2838 | return MadeChange; |
| 2839 | } |