File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp |
Warning: | line 2713, 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/InitializePasses.h" |
88 | #include "llvm/Pass.h" |
89 | #include "llvm/Support/Casting.h" |
90 | #include "llvm/Support/CommandLine.h" |
91 | #include "llvm/Support/Debug.h" |
92 | #include "llvm/Support/InstructionCost.h" |
93 | #include "llvm/Support/raw_ostream.h" |
94 | #include "llvm/Transforms/Scalar.h" |
95 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
96 | #include "llvm/Transforms/Utils/Local.h" |
97 | #include "llvm/Transforms/Utils/LoopUtils.h" |
98 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
99 | #include <algorithm> |
100 | #include <cassert> |
101 | #include <cstdint> |
102 | #include <utility> |
103 | #include <vector> |
104 | |
105 | using namespace llvm; |
106 | |
107 | #define DEBUG_TYPE"loop-idiom" "loop-idiom" |
108 | |
109 | 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"}; |
110 | 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"}; |
111 | 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"}; |
112 | STATISTIC(static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
113 | NumShiftUntilBitTest,static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
114 | "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" }; |
115 | STATISTIC(NumShiftUntilZero,static llvm::Statistic NumShiftUntilZero = {"loop-idiom", "NumShiftUntilZero" , "Number of uncountable loops recognized as 'shift until zero' idiom" } |
116 | "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" }; |
117 | |
118 | bool DisableLIRP::All; |
119 | static cl::opt<bool, true> |
120 | DisableLIRPAll("disable-" DEBUG_TYPE"loop-idiom" "-all", |
121 | cl::desc("Options to disable Loop Idiom Recognize Pass."), |
122 | cl::location(DisableLIRP::All), cl::init(false), |
123 | cl::ReallyHidden); |
124 | |
125 | bool DisableLIRP::Memset; |
126 | static cl::opt<bool, true> |
127 | DisableLIRPMemset("disable-" DEBUG_TYPE"loop-idiom" "-memset", |
128 | cl::desc("Proceed with loop idiom recognize pass, but do " |
129 | "not convert loop(s) to memset."), |
130 | cl::location(DisableLIRP::Memset), cl::init(false), |
131 | cl::ReallyHidden); |
132 | |
133 | bool DisableLIRP::Memcpy; |
134 | static cl::opt<bool, true> |
135 | DisableLIRPMemcpy("disable-" DEBUG_TYPE"loop-idiom" "-memcpy", |
136 | cl::desc("Proceed with loop idiom recognize pass, but do " |
137 | "not convert loop(s) to memcpy."), |
138 | cl::location(DisableLIRP::Memcpy), cl::init(false), |
139 | cl::ReallyHidden); |
140 | |
141 | static cl::opt<bool> UseLIRCodeSizeHeurs( |
142 | "use-lir-code-size-heurs", |
143 | cl::desc("Use loop idiom recognition code size heuristics when compiling" |
144 | "with -Os/-Oz"), |
145 | cl::init(true), cl::Hidden); |
146 | |
147 | namespace { |
148 | |
149 | class LoopIdiomRecognize { |
150 | Loop *CurLoop = nullptr; |
151 | AliasAnalysis *AA; |
152 | DominatorTree *DT; |
153 | LoopInfo *LI; |
154 | ScalarEvolution *SE; |
155 | TargetLibraryInfo *TLI; |
156 | const TargetTransformInfo *TTI; |
157 | const DataLayout *DL; |
158 | OptimizationRemarkEmitter &ORE; |
159 | bool ApplyCodeSizeHeuristics; |
160 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
161 | |
162 | public: |
163 | explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
164 | LoopInfo *LI, ScalarEvolution *SE, |
165 | TargetLibraryInfo *TLI, |
166 | const TargetTransformInfo *TTI, MemorySSA *MSSA, |
167 | const DataLayout *DL, |
168 | OptimizationRemarkEmitter &ORE) |
169 | : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { |
170 | if (MSSA) |
171 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); |
172 | } |
173 | |
174 | bool runOnLoop(Loop *L); |
175 | |
176 | private: |
177 | using StoreList = SmallVector<StoreInst *, 8>; |
178 | using StoreListMap = MapVector<Value *, StoreList>; |
179 | |
180 | StoreListMap StoreRefsForMemset; |
181 | StoreListMap StoreRefsForMemsetPattern; |
182 | StoreList StoreRefsForMemcpy; |
183 | bool HasMemset; |
184 | bool HasMemsetPattern; |
185 | bool HasMemcpy; |
186 | |
187 | /// Return code for isLegalStore() |
188 | enum LegalStoreKind { |
189 | None = 0, |
190 | Memset, |
191 | MemsetPattern, |
192 | Memcpy, |
193 | UnorderedAtomicMemcpy, |
194 | DontUse // Dummy retval never to be used. Allows catching errors in retval |
195 | // handling. |
196 | }; |
197 | |
198 | /// \name Countable Loop Idiom Handling |
199 | /// @{ |
200 | |
201 | bool runOnCountableLoop(); |
202 | bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, |
203 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
204 | |
205 | void collectStores(BasicBlock *BB); |
206 | LegalStoreKind isLegalStore(StoreInst *SI); |
207 | enum class ForMemset { No, Yes }; |
208 | bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount, |
209 | ForMemset For); |
210 | |
211 | template <typename MemInst> |
212 | bool processLoopMemIntrinsic( |
213 | BasicBlock *BB, |
214 | bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), |
215 | const SCEV *BECount); |
216 | bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount); |
217 | bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); |
218 | |
219 | bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV, |
220 | MaybeAlign StoreAlignment, Value *StoredVal, |
221 | Instruction *TheStore, |
222 | SmallPtrSetImpl<Instruction *> &Stores, |
223 | const SCEVAddRecExpr *Ev, const SCEV *BECount, |
224 | bool IsNegStride, bool IsLoopMemset = false); |
225 | bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); |
226 | bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr, |
227 | const SCEV *StoreSize, MaybeAlign StoreAlign, |
228 | MaybeAlign LoadAlign, Instruction *TheStore, |
229 | Instruction *TheLoad, |
230 | const SCEVAddRecExpr *StoreEv, |
231 | const SCEVAddRecExpr *LoadEv, |
232 | const SCEV *BECount); |
233 | bool avoidLIRForMultiBlockLoop(bool IsMemset = false, |
234 | bool IsLoopMemset = false); |
235 | |
236 | /// @} |
237 | /// \name Noncountable Loop Idiom Handling |
238 | /// @{ |
239 | |
240 | bool runOnNoncountableLoop(); |
241 | |
242 | bool recognizePopcount(); |
243 | void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, |
244 | PHINode *CntPhi, Value *Var); |
245 | bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz |
246 | void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, |
247 | Instruction *CntInst, PHINode *CntPhi, |
248 | Value *Var, Instruction *DefX, |
249 | const DebugLoc &DL, bool ZeroCheck, |
250 | bool IsCntPhiUsedOutsideLoop); |
251 | |
252 | bool recognizeShiftUntilBitTest(); |
253 | bool recognizeShiftUntilZero(); |
254 | |
255 | /// @} |
256 | }; |
257 | |
258 | class LoopIdiomRecognizeLegacyPass : public LoopPass { |
259 | public: |
260 | static char ID; |
261 | |
262 | explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) { |
263 | initializeLoopIdiomRecognizeLegacyPassPass( |
264 | *PassRegistry::getPassRegistry()); |
265 | } |
266 | |
267 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { |
268 | if (DisableLIRP::All) |
269 | return false; |
270 | |
271 | if (skipLoop(L)) |
272 | return false; |
273 | |
274 | AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
275 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
276 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
277 | ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
278 | TargetLibraryInfo *TLI = |
279 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
280 | *L->getHeader()->getParent()); |
281 | const TargetTransformInfo *TTI = |
282 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( |
283 | *L->getHeader()->getParent()); |
284 | const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout(); |
285 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
286 | MemorySSA *MSSA = nullptr; |
287 | if (MSSAAnalysis) |
288 | MSSA = &MSSAAnalysis->getMSSA(); |
289 | |
290 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis |
291 | // pass. Function analyses need to be preserved across loop transformations |
292 | // but ORE cannot be preserved (see comment before the pass definition). |
293 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); |
294 | |
295 | LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE); |
296 | return LIR.runOnLoop(L); |
297 | } |
298 | |
299 | /// This transformation requires natural loop information & requires that |
300 | /// loop preheaders be inserted into the CFG. |
301 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
302 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
303 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
304 | AU.addPreserved<MemorySSAWrapperPass>(); |
305 | getLoopAnalysisUsage(AU); |
306 | } |
307 | }; |
308 | |
309 | } // end anonymous namespace |
310 | |
311 | char LoopIdiomRecognizeLegacyPass::ID = 0; |
312 | |
313 | PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, |
314 | LoopStandardAnalysisResults &AR, |
315 | LPMUpdater &) { |
316 | if (DisableLIRP::All) |
317 | return PreservedAnalyses::all(); |
318 | |
319 | const auto *DL = &L.getHeader()->getModule()->getDataLayout(); |
320 | |
321 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
322 | // pass. Function analyses need to be preserved across loop transformations |
323 | // but ORE cannot be preserved (see comment before the pass definition). |
324 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
325 | |
326 | LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, |
327 | AR.MSSA, DL, ORE); |
328 | if (!LIR.runOnLoop(&L)) |
329 | return PreservedAnalyses::all(); |
330 | |
331 | auto PA = getLoopPassPreservedAnalyses(); |
332 | if (AR.MSSA) |
333 | PA.preserve<MemorySSAAnalysis>(); |
334 | return PA; |
335 | } |
336 | |
337 | INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
338 | "Recognize loop idioms", false, false)static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
339 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); |
340 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
341 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); |
342 | INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
343 | "Recognize loop idioms", false, false)PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
344 | |
345 | Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); } |
346 | |
347 | static void deleteDeadInstruction(Instruction *I) { |
348 | I->replaceAllUsesWith(UndefValue::get(I->getType())); |
349 | I->eraseFromParent(); |
350 | } |
351 | |
352 | //===----------------------------------------------------------------------===// |
353 | // |
354 | // Implementation of LoopIdiomRecognize |
355 | // |
356 | //===----------------------------------------------------------------------===// |
357 | |
358 | bool LoopIdiomRecognize::runOnLoop(Loop *L) { |
359 | CurLoop = L; |
360 | // If the loop could not be converted to canonical form, it must have an |
361 | // indirectbr in it, just give up. |
362 | if (!L->getLoopPreheader()) |
363 | return false; |
364 | |
365 | // Disable loop idiom recognition if the function's name is a common idiom. |
366 | StringRef Name = L->getHeader()->getParent()->getName(); |
367 | if (Name == "memset" || Name == "memcpy") |
368 | return false; |
369 | |
370 | // Determine if code size heuristics need to be applied. |
371 | ApplyCodeSizeHeuristics = |
372 | L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; |
373 | |
374 | HasMemset = TLI->has(LibFunc_memset); |
375 | HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); |
376 | HasMemcpy = TLI->has(LibFunc_memcpy); |
377 | |
378 | if (HasMemset || HasMemsetPattern || HasMemcpy) |
379 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
380 | return runOnCountableLoop(); |
381 | |
382 | return runOnNoncountableLoop(); |
383 | } |
384 | |
385 | bool LoopIdiomRecognize::runOnCountableLoop() { |
386 | const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); |
387 | 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", 389, __extension__ __PRETTY_FUNCTION__)) |
388 | "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", 389, __extension__ __PRETTY_FUNCTION__)) |
389 | "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", 389, __extension__ __PRETTY_FUNCTION__)); |
390 | |
391 | // If this loop executes exactly one time, then it should be peeled, not |
392 | // optimized by this pass. |
393 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
394 | if (BECst->getAPInt() == 0) |
395 | return false; |
396 | |
397 | SmallVector<BasicBlock *, 8> ExitBlocks; |
398 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
399 | |
400 | 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) |
401 | << 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) |
402 | << "] 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) |
403 | << "\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); |
404 | |
405 | // The following transforms hoist stores/memsets into the loop pre-header. |
406 | // Give up if the loop has instructions that may throw. |
407 | SimpleLoopSafetyInfo SafetyInfo; |
408 | SafetyInfo.computeLoopSafetyInfo(CurLoop); |
409 | if (SafetyInfo.anyBlockMayThrow()) |
410 | return false; |
411 | |
412 | bool MadeChange = false; |
413 | |
414 | // Scan all the blocks in the loop that are not in subloops. |
415 | for (auto *BB : CurLoop->getBlocks()) { |
416 | // Ignore blocks in subloops. |
417 | if (LI->getLoopFor(BB) != CurLoop) |
418 | continue; |
419 | |
420 | MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); |
421 | } |
422 | return MadeChange; |
423 | } |
424 | |
425 | static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { |
426 | const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1)); |
427 | return ConstStride->getAPInt(); |
428 | } |
429 | |
430 | /// getMemSetPatternValue - If a strided store of the specified value is safe to |
431 | /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should |
432 | /// be passed in. Otherwise, return null. |
433 | /// |
434 | /// Note that we don't ever attempt to use memset_pattern8 or 4, because these |
435 | /// just replicate their input array and then pass on to memset_pattern16. |
436 | static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { |
437 | // FIXME: This could check for UndefValue because it can be merged into any |
438 | // other valid pattern. |
439 | |
440 | // If the value isn't a constant, we can't promote it to being in a constant |
441 | // array. We could theoretically do a store to an alloca or something, but |
442 | // that doesn't seem worthwhile. |
443 | Constant *C = dyn_cast<Constant>(V); |
444 | if (!C) |
445 | return nullptr; |
446 | |
447 | // Only handle simple values that are a power of two bytes in size. |
448 | uint64_t Size = DL->getTypeSizeInBits(V->getType()); |
449 | if (Size == 0 || (Size & 7) || (Size & (Size - 1))) |
450 | return nullptr; |
451 | |
452 | // Don't care enough about darwin/ppc to implement this. |
453 | if (DL->isBigEndian()) |
454 | return nullptr; |
455 | |
456 | // Convert to size in bytes. |
457 | Size /= 8; |
458 | |
459 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
460 | // if the top and bottom are the same (e.g. for vectors and large integers). |
461 | if (Size > 16) |
462 | return nullptr; |
463 | |
464 | // If the constant is exactly 16 bytes, just use it. |
465 | if (Size == 16) |
466 | return C; |
467 | |
468 | // Otherwise, we'll use an array of the constants. |
469 | unsigned ArraySize = 16 / Size; |
470 | ArrayType *AT = ArrayType::get(V->getType(), ArraySize); |
471 | return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C)); |
472 | } |
473 | |
474 | LoopIdiomRecognize::LegalStoreKind |
475 | LoopIdiomRecognize::isLegalStore(StoreInst *SI) { |
476 | // Don't touch volatile stores. |
477 | if (SI->isVolatile()) |
478 | return LegalStoreKind::None; |
479 | // We only want simple or unordered-atomic stores. |
480 | if (!SI->isUnordered()) |
481 | return LegalStoreKind::None; |
482 | |
483 | // Avoid merging nontemporal stores. |
484 | if (SI->getMetadata(LLVMContext::MD_nontemporal)) |
485 | return LegalStoreKind::None; |
486 | |
487 | Value *StoredVal = SI->getValueOperand(); |
488 | Value *StorePtr = SI->getPointerOperand(); |
489 | |
490 | // Don't convert stores of non-integral pointer types to memsets (which stores |
491 | // integers). |
492 | if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType())) |
493 | return LegalStoreKind::None; |
494 | |
495 | // Reject stores that are so large that they overflow an unsigned. |
496 | // When storing out scalable vectors we bail out for now, since the code |
497 | // below currently only works for constant strides. |
498 | TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); |
499 | if (SizeInBits.isScalable() || (SizeInBits.getFixedSize() & 7) || |
500 | (SizeInBits.getFixedSize() >> 32) != 0) |
501 | return LegalStoreKind::None; |
502 | |
503 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
504 | // loop, which indicates a strided store. If we have something else, it's a |
505 | // random store we can't handle. |
506 | const SCEVAddRecExpr *StoreEv = |
507 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
508 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
509 | return LegalStoreKind::None; |
510 | |
511 | // Check to see if we have a constant stride. |
512 | if (!isa<SCEVConstant>(StoreEv->getOperand(1))) |
513 | return LegalStoreKind::None; |
514 | |
515 | // See if the store can be turned into a memset. |
516 | |
517 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
518 | // turned into a memset of i8 -1, assuming that all the consecutive bytes |
519 | // are stored. A store of i32 0x01020304 can never be turned into a memset, |
520 | // but it can be turned into memset_pattern if the target supports it. |
521 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
522 | |
523 | // Note: memset and memset_pattern on unordered-atomic is yet not supported |
524 | bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); |
525 | |
526 | // If we're allowed to form a memset, and the stored value would be |
527 | // acceptable for memset, use it. |
528 | if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && |
529 | // Verify that the stored value is loop invariant. If not, we can't |
530 | // promote the memset. |
531 | CurLoop->isLoopInvariant(SplatValue)) { |
532 | // It looks like we can use SplatValue. |
533 | return LegalStoreKind::Memset; |
534 | } |
535 | if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && |
536 | // Don't create memset_pattern16s with address spaces. |
537 | StorePtr->getType()->getPointerAddressSpace() == 0 && |
538 | getMemSetPatternValue(StoredVal, DL)) { |
539 | // It looks like we can use PatternValue! |
540 | return LegalStoreKind::MemsetPattern; |
541 | } |
542 | |
543 | // Otherwise, see if the store can be turned into a memcpy. |
544 | if (HasMemcpy && !DisableLIRP::Memcpy) { |
545 | // Check to see if the stride matches the size of the store. If so, then we |
546 | // know that every byte is touched in the loop. |
547 | APInt Stride = getStoreStride(StoreEv); |
548 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
549 | if (StoreSize != Stride && StoreSize != -Stride) |
550 | return LegalStoreKind::None; |
551 | |
552 | // The store must be feeding a non-volatile load. |
553 | LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); |
554 | |
555 | // Only allow non-volatile loads |
556 | if (!LI || LI->isVolatile()) |
557 | return LegalStoreKind::None; |
558 | // Only allow simple or unordered-atomic loads |
559 | if (!LI->isUnordered()) |
560 | return LegalStoreKind::None; |
561 | |
562 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
563 | // loop, which indicates a strided load. If we have something else, it's a |
564 | // random load we can't handle. |
565 | const SCEVAddRecExpr *LoadEv = |
566 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
567 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
568 | return LegalStoreKind::None; |
569 | |
570 | // The store and load must share the same stride. |
571 | if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) |
572 | return LegalStoreKind::None; |
573 | |
574 | // Success. This store can be converted into a memcpy. |
575 | UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); |
576 | return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy |
577 | : LegalStoreKind::Memcpy; |
578 | } |
579 | // This store can't be transformed into a memset/memcpy. |
580 | return LegalStoreKind::None; |
581 | } |
582 | |
583 | void LoopIdiomRecognize::collectStores(BasicBlock *BB) { |
584 | StoreRefsForMemset.clear(); |
585 | StoreRefsForMemsetPattern.clear(); |
586 | StoreRefsForMemcpy.clear(); |
587 | for (Instruction &I : *BB) { |
588 | StoreInst *SI = dyn_cast<StoreInst>(&I); |
589 | if (!SI) |
590 | continue; |
591 | |
592 | // Make sure this is a strided store with a constant stride. |
593 | switch (isLegalStore(SI)) { |
594 | case LegalStoreKind::None: |
595 | // Nothing to do |
596 | break; |
597 | case LegalStoreKind::Memset: { |
598 | // Find the base pointer. |
599 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
600 | StoreRefsForMemset[Ptr].push_back(SI); |
601 | } break; |
602 | case LegalStoreKind::MemsetPattern: { |
603 | // Find the base pointer. |
604 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
605 | StoreRefsForMemsetPattern[Ptr].push_back(SI); |
606 | } break; |
607 | case LegalStoreKind::Memcpy: |
608 | case LegalStoreKind::UnorderedAtomicMemcpy: |
609 | StoreRefsForMemcpy.push_back(SI); |
610 | break; |
611 | default: |
612 | 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", 612, __extension__ __PRETTY_FUNCTION__)); |
613 | break; |
614 | } |
615 | } |
616 | } |
617 | |
618 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
619 | /// with the specified backedge count. This block is known to be in the current |
620 | /// loop and not in any subloops. |
621 | bool LoopIdiomRecognize::runOnLoopBlock( |
622 | BasicBlock *BB, const SCEV *BECount, |
623 | SmallVectorImpl<BasicBlock *> &ExitBlocks) { |
624 | // We can only promote stores in this block if they are unconditionally |
625 | // executed in the loop. For a block to be unconditionally executed, it has |
626 | // to dominate all the exit blocks of the loop. Verify this now. |
627 | for (BasicBlock *ExitBlock : ExitBlocks) |
628 | if (!DT->dominates(BB, ExitBlock)) |
629 | return false; |
630 | |
631 | bool MadeChange = false; |
632 | // Look for store instructions, which may be optimized to memset/memcpy. |
633 | collectStores(BB); |
634 | |
635 | // Look for a single store or sets of stores with a common base, which can be |
636 | // optimized into a memset (memset_pattern). The latter most commonly happens |
637 | // with structs and handunrolled loops. |
638 | for (auto &SL : StoreRefsForMemset) |
639 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); |
640 | |
641 | for (auto &SL : StoreRefsForMemsetPattern) |
642 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); |
643 | |
644 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
645 | for (auto &SI : StoreRefsForMemcpy) |
646 | MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); |
647 | |
648 | MadeChange |= processLoopMemIntrinsic<MemCpyInst>( |
649 | BB, &LoopIdiomRecognize::processLoopMemCpy, BECount); |
650 | MadeChange |= processLoopMemIntrinsic<MemSetInst>( |
651 | BB, &LoopIdiomRecognize::processLoopMemSet, BECount); |
652 | |
653 | return MadeChange; |
654 | } |
655 | |
656 | /// See if this store(s) can be promoted to a memset. |
657 | bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL, |
658 | const SCEV *BECount, ForMemset For) { |
659 | // Try to find consecutive stores that can be transformed into memsets. |
660 | SetVector<StoreInst *> Heads, Tails; |
661 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; |
662 | |
663 | // Do a quadratic search on all of the given stores and find |
664 | // all of the pairs of stores that follow each other. |
665 | SmallVector<unsigned, 16> IndexQueue; |
666 | for (unsigned i = 0, e = SL.size(); i < e; ++i) { |
667 | 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", 667, __extension__ __PRETTY_FUNCTION__)); |
668 | |
669 | Value *FirstStoredVal = SL[i]->getValueOperand(); |
670 | Value *FirstStorePtr = SL[i]->getPointerOperand(); |
671 | const SCEVAddRecExpr *FirstStoreEv = |
672 | cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr)); |
673 | APInt FirstStride = getStoreStride(FirstStoreEv); |
674 | unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); |
675 | |
676 | // See if we can optimize just this store in isolation. |
677 | if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { |
678 | Heads.insert(SL[i]); |
679 | continue; |
680 | } |
681 | |
682 | Value *FirstSplatValue = nullptr; |
683 | Constant *FirstPatternValue = nullptr; |
684 | |
685 | if (For == ForMemset::Yes) |
686 | FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); |
687 | else |
688 | FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); |
689 | |
690 | 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", 691, __extension__ __PRETTY_FUNCTION__)) |
691 | "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", 691, __extension__ __PRETTY_FUNCTION__)); |
692 | |
693 | IndexQueue.clear(); |
694 | // If a store has multiple consecutive store candidates, search Stores |
695 | // array according to the sequence: from i+1 to e, then from i-1 to 0. |
696 | // This is because usually pairing with immediate succeeding or preceding |
697 | // candidate create the best chance to find memset opportunity. |
698 | unsigned j = 0; |
699 | for (j = i + 1; j < e; ++j) |
700 | IndexQueue.push_back(j); |
701 | for (j = i; j > 0; --j) |
702 | IndexQueue.push_back(j - 1); |
703 | |
704 | for (auto &k : IndexQueue) { |
705 | 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", 705, __extension__ __PRETTY_FUNCTION__)); |
706 | Value *SecondStorePtr = SL[k]->getPointerOperand(); |
707 | const SCEVAddRecExpr *SecondStoreEv = |
708 | cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr)); |
709 | APInt SecondStride = getStoreStride(SecondStoreEv); |
710 | |
711 | if (FirstStride != SecondStride) |
712 | continue; |
713 | |
714 | Value *SecondStoredVal = SL[k]->getValueOperand(); |
715 | Value *SecondSplatValue = nullptr; |
716 | Constant *SecondPatternValue = nullptr; |
717 | |
718 | if (For == ForMemset::Yes) |
719 | SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); |
720 | else |
721 | SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); |
722 | |
723 | 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", 724, __extension__ __PRETTY_FUNCTION__)) |
724 | "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", 724, __extension__ __PRETTY_FUNCTION__)); |
725 | |
726 | if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { |
727 | if (For == ForMemset::Yes) { |
728 | if (isa<UndefValue>(FirstSplatValue)) |
729 | FirstSplatValue = SecondSplatValue; |
730 | if (FirstSplatValue != SecondSplatValue) |
731 | continue; |
732 | } else { |
733 | if (isa<UndefValue>(FirstPatternValue)) |
734 | FirstPatternValue = SecondPatternValue; |
735 | if (FirstPatternValue != SecondPatternValue) |
736 | continue; |
737 | } |
738 | Tails.insert(SL[k]); |
739 | Heads.insert(SL[i]); |
740 | ConsecutiveChain[SL[i]] = SL[k]; |
741 | break; |
742 | } |
743 | } |
744 | } |
745 | |
746 | // We may run into multiple chains that merge into a single chain. We mark the |
747 | // stores that we transformed so that we don't visit the same store twice. |
748 | SmallPtrSet<Value *, 16> TransformedStores; |
749 | bool Changed = false; |
750 | |
751 | // For stores that start but don't end a link in the chain: |
752 | for (StoreInst *I : Heads) { |
753 | if (Tails.count(I)) |
754 | continue; |
755 | |
756 | // We found a store instr that starts a chain. Now follow the chain and try |
757 | // to transform it. |
758 | SmallPtrSet<Instruction *, 8> AdjacentStores; |
759 | StoreInst *HeadStore = I; |
760 | unsigned StoreSize = 0; |
761 | |
762 | // Collect the chain into a list. |
763 | while (Tails.count(I) || Heads.count(I)) { |
764 | if (TransformedStores.count(I)) |
765 | break; |
766 | AdjacentStores.insert(I); |
767 | |
768 | StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); |
769 | // Move to the next value in the chain. |
770 | I = ConsecutiveChain[I]; |
771 | } |
772 | |
773 | Value *StoredVal = HeadStore->getValueOperand(); |
774 | Value *StorePtr = HeadStore->getPointerOperand(); |
775 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
776 | APInt Stride = getStoreStride(StoreEv); |
777 | |
778 | // Check to see if the stride matches the size of the stores. If so, then |
779 | // we know that every byte is touched in the loop. |
780 | if (StoreSize != Stride && StoreSize != -Stride) |
781 | continue; |
782 | |
783 | bool IsNegStride = StoreSize == -Stride; |
784 | |
785 | Type *IntIdxTy = DL->getIndexType(StorePtr->getType()); |
786 | const SCEV *StoreSizeSCEV = SE->getConstant(IntIdxTy, StoreSize); |
787 | if (processLoopStridedStore(StorePtr, StoreSizeSCEV, |
788 | MaybeAlign(HeadStore->getAlign()), StoredVal, |
789 | HeadStore, AdjacentStores, StoreEv, BECount, |
790 | IsNegStride)) { |
791 | TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); |
792 | Changed = true; |
793 | } |
794 | } |
795 | |
796 | return Changed; |
797 | } |
798 | |
799 | /// processLoopMemIntrinsic - Template function for calling different processor |
800 | /// functions based on mem instrinsic type. |
801 | template <typename MemInst> |
802 | bool LoopIdiomRecognize::processLoopMemIntrinsic( |
803 | BasicBlock *BB, |
804 | bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), |
805 | const SCEV *BECount) { |
806 | bool MadeChange = false; |
807 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
808 | Instruction *Inst = &*I++; |
809 | // Look for memory instructions, which may be optimized to a larger one. |
810 | if (MemInst *MI = dyn_cast<MemInst>(Inst)) { |
811 | WeakTrackingVH InstPtr(&*I); |
812 | if (!(this->*Processor)(MI, BECount)) |
813 | continue; |
814 | MadeChange = true; |
815 | |
816 | // If processing the instruction invalidated our iterator, start over from |
817 | // the top of the block. |
818 | if (!InstPtr) |
819 | I = BB->begin(); |
820 | } |
821 | } |
822 | return MadeChange; |
823 | } |
824 | |
825 | /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy |
826 | bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI, |
827 | const SCEV *BECount) { |
828 | // We can only handle non-volatile memcpys with a constant size. |
829 | if (MCI->isVolatile() || !isa<ConstantInt>(MCI->getLength())) |
830 | return false; |
831 | |
832 | // If we're not allowed to hack on memcpy, we fail. |
833 | if ((!HasMemcpy && !isa<MemCpyInlineInst>(MCI)) || DisableLIRP::Memcpy) |
834 | return false; |
835 | |
836 | Value *Dest = MCI->getDest(); |
837 | Value *Source = MCI->getSource(); |
838 | if (!Dest || !Source) |
839 | return false; |
840 | |
841 | // See if the load and store pointer expressions are AddRec like {base,+,1} on |
842 | // the current loop, which indicates a strided load and store. If we have |
843 | // something else, it's a random load or store we can't handle. |
844 | const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest)); |
845 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
846 | return false; |
847 | const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source)); |
848 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
849 | return false; |
850 | |
851 | // Reject memcpys that are so large that they overflow an unsigned. |
852 | uint64_t SizeInBytes = cast<ConstantInt>(MCI->getLength())->getZExtValue(); |
853 | if ((SizeInBytes >> 32) != 0) |
854 | return false; |
855 | |
856 | // Check if the stride matches the size of the memcpy. If so, then we know |
857 | // that every byte is touched in the loop. |
858 | const SCEVConstant *ConstStoreStride = |
859 | dyn_cast<SCEVConstant>(StoreEv->getOperand(1)); |
860 | const SCEVConstant *ConstLoadStride = |
861 | dyn_cast<SCEVConstant>(LoadEv->getOperand(1)); |
862 | if (!ConstStoreStride || !ConstLoadStride) |
863 | return false; |
864 | |
865 | APInt StoreStrideValue = ConstStoreStride->getAPInt(); |
866 | APInt LoadStrideValue = ConstLoadStride->getAPInt(); |
867 | // Huge stride value - give up |
868 | if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64) |
869 | return false; |
870 | |
871 | if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) { |
872 | ORE.emit([&]() { |
873 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "SizeStrideUnequal", MCI) |
874 | << ore::NV("Inst", "memcpy") << " in " |
875 | << ore::NV("Function", MCI->getFunction()) |
876 | << " function will not be hoisted: " |
877 | << ore::NV("Reason", "memcpy size is not equal to stride"); |
878 | }); |
879 | return false; |
880 | } |
881 | |
882 | int64_t StoreStrideInt = StoreStrideValue.getSExtValue(); |
883 | int64_t LoadStrideInt = LoadStrideValue.getSExtValue(); |
884 | // Check if the load stride matches the store stride. |
885 | if (StoreStrideInt != LoadStrideInt) |
886 | return false; |
887 | |
888 | return processLoopStoreOfLoopLoad( |
889 | Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes), |
890 | MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv, |
891 | BECount); |
892 | } |
893 | |
894 | /// processLoopMemSet - See if this memset can be promoted to a large memset. |
895 | bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, |
896 | const SCEV *BECount) { |
897 | // We can only handle non-volatile memsets. |
898 | if (MSI->isVolatile()) |
899 | return false; |
900 | |
901 | // If we're not allowed to hack on memset, we fail. |
902 | if (!HasMemset || DisableLIRP::Memset) |
903 | return false; |
904 | |
905 | Value *Pointer = MSI->getDest(); |
906 | |
907 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
908 | // loop, which indicates a strided store. If we have something else, it's a |
909 | // random store we can't handle. |
910 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer)); |
911 | if (!Ev || Ev->getLoop() != CurLoop) |
912 | return false; |
913 | if (!Ev->isAffine()) { |
914 | 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); |
915 | return false; |
916 | } |
917 | |
918 | const SCEV *PointerStrideSCEV = Ev->getOperand(1); |
919 | const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength()); |
920 | if (!PointerStrideSCEV || !MemsetSizeSCEV) |
921 | return false; |
922 | |
923 | bool IsNegStride = false; |
924 | const bool IsConstantSize = isa<ConstantInt>(MSI->getLength()); |
925 | |
926 | if (IsConstantSize) { |
927 | // Memset size is constant. |
928 | // Check if the pointer stride matches the memset size. If so, then |
929 | // we know that every byte is touched in the loop. |
930 | LLVM_DEBUG(dbgs() << " memset size is constant\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is constant\n" ; } } while (false); |
931 | uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); |
932 | const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
933 | if (!ConstStride) |
934 | return false; |
935 | |
936 | APInt Stride = ConstStride->getAPInt(); |
937 | if (SizeInBytes != Stride && SizeInBytes != -Stride) |
938 | return false; |
939 | |
940 | IsNegStride = SizeInBytes == -Stride; |
941 | } else { |
942 | // Memset size is non-constant. |
943 | // Check if the pointer stride matches the memset size. |
944 | // To be conservative, the pass would not promote pointers that aren't in |
945 | // address space zero. Also, the pass only handles memset length and stride |
946 | // that are invariant for the top level loop. |
947 | 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); |
948 | if (Pointer->getType()->getPointerAddressSpace() != 0) { |
949 | 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) |
950 | << "abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " pointer is not in address space zero, " << "abort\n"; } } while (false); |
951 | return false; |
952 | } |
953 | if (!SE->isLoopInvariant(MemsetSizeSCEV, CurLoop)) { |
954 | 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) |
955 | << "abort\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " memset size is not a loop-invariant, " << "abort\n"; } } while (false); |
956 | return false; |
957 | } |
958 | |
959 | // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV |
960 | IsNegStride = PointerStrideSCEV->isNonConstantNegative(); |
961 | const SCEV *PositiveStrideSCEV = |
962 | IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV) |
963 | : PointerStrideSCEV; |
964 | LLVM_DEBUG(dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ) |
965 | << " PositiveStrideSCEV: " << *PositiveStrideSCEVdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ) |
966 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"; } } while (false ); |
967 | |
968 | if (PositiveStrideSCEV != MemsetSizeSCEV) { |
969 | // If an expression is covered by the loop guard, compare again and |
970 | // proceed with optimization if equal. |
971 | const SCEV *FoldedPositiveStride = |
972 | SE->applyLoopGuards(PositiveStrideSCEV, CurLoop); |
973 | const SCEV *FoldedMemsetSize = |
974 | SE->applyLoopGuards(MemsetSizeSCEV, CurLoop); |
975 | |
976 | 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) |
977 | << " 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) |
978 | << " 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) |
979 | << "\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); |
980 | |
981 | if (FoldedPositiveStride != FoldedMemsetSize) { |
982 | 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); |
983 | return false; |
984 | } |
985 | } |
986 | } |
987 | |
988 | // Verify that the memset value is loop invariant. If not, we can't promote |
989 | // the memset. |
990 | Value *SplatValue = MSI->getValue(); |
991 | if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) |
992 | return false; |
993 | |
994 | SmallPtrSet<Instruction *, 1> MSIs; |
995 | MSIs.insert(MSI); |
996 | return processLoopStridedStore(Pointer, SE->getSCEV(MSI->getLength()), |
997 | MaybeAlign(MSI->getDestAlignment()), |
998 | SplatValue, MSI, MSIs, Ev, BECount, |
999 | IsNegStride, /*IsLoopMemset=*/true); |
1000 | } |
1001 | |
1002 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
1003 | /// specified pointer location, which is a loop-strided access. The 'Access' |
1004 | /// argument specifies what the verboten forms of access are (read or write). |
1005 | static bool |
1006 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
1007 | const SCEV *BECount, const SCEV *StoreSizeSCEV, |
1008 | AliasAnalysis &AA, |
1009 | SmallPtrSetImpl<Instruction *> &IgnoredInsts) { |
1010 | // Get the location that may be stored across the loop. Since the access is |
1011 | // strided positively through memory, we say that the modified location starts |
1012 | // at the pointer and has infinite size. |
1013 | LocationSize AccessSize = LocationSize::afterPointer(); |
1014 | |
1015 | // If the loop iterates a fixed number of times, we can refine the access size |
1016 | // to be exactly the size of the memset, which is (BECount+1)*StoreSize |
1017 | const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount); |
1018 | const SCEVConstant *ConstSize = dyn_cast<SCEVConstant>(StoreSizeSCEV); |
1019 | if (BECst && ConstSize) |
1020 | AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * |
1021 | ConstSize->getValue()->getZExtValue()); |
1022 | |
1023 | // TODO: For this to be really effective, we have to dive into the pointer |
1024 | // operand in the store. Store to &A[i] of 100 will always return may alias |
1025 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
1026 | // which will then no-alias a store to &A[100]. |
1027 | MemoryLocation StoreLoc(Ptr, AccessSize); |
1028 | |
1029 | for (BasicBlock *B : L->blocks()) |
1030 | for (Instruction &I : *B) |
1031 | if (!IgnoredInsts.contains(&I) && |
1032 | isModOrRefSet( |
1033 | intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access))) |
1034 | return true; |
1035 | return false; |
1036 | } |
1037 | |
1038 | // If we have a negative stride, Start refers to the end of the memory location |
1039 | // we're trying to memset. Therefore, we need to recompute the base pointer, |
1040 | // which is just Start - BECount*Size. |
1041 | static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, |
1042 | Type *IntPtr, const SCEV *StoreSizeSCEV, |
1043 | ScalarEvolution *SE) { |
1044 | const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); |
1045 | if (!StoreSizeSCEV->isOne()) { |
1046 | // index = back edge count * store size |
1047 | Index = SE->getMulExpr(Index, |
1048 | SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), |
1049 | SCEV::FlagNUW); |
1050 | } |
1051 | // base pointer = start - index * store size |
1052 | return SE->getMinusSCEV(Start, Index); |
1053 | } |
1054 | |
1055 | /// Compute trip count from the backedge taken count. |
1056 | static const SCEV *getTripCount(const SCEV *BECount, Type *IntPtr, |
1057 | Loop *CurLoop, const DataLayout *DL, |
1058 | ScalarEvolution *SE) { |
1059 | const SCEV *TripCountS = nullptr; |
1060 | // The # stored bytes is (BECount+1). Expand the trip count out to |
1061 | // pointer size if it isn't already. |
1062 | // |
1063 | // If we're going to need to zero extend the BE count, check if we can add |
1064 | // one to it prior to zero extending without overflow. Provided this is safe, |
1065 | // it allows better simplification of the +1. |
1066 | if (DL->getTypeSizeInBits(BECount->getType()) < |
1067 | DL->getTypeSizeInBits(IntPtr) && |
1068 | SE->isLoopEntryGuardedByCond( |
1069 | CurLoop, ICmpInst::ICMP_NE, BECount, |
1070 | SE->getNegativeSCEV(SE->getOne(BECount->getType())))) { |
1071 | TripCountS = SE->getZeroExtendExpr( |
1072 | SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW), |
1073 | IntPtr); |
1074 | } else { |
1075 | TripCountS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr), |
1076 | SE->getOne(IntPtr), SCEV::FlagNUW); |
1077 | } |
1078 | |
1079 | return TripCountS; |
1080 | } |
1081 | |
1082 | /// Compute the number of bytes as a SCEV from the backedge taken count. |
1083 | /// |
1084 | /// This also maps the SCEV into the provided type and tries to handle the |
1085 | /// computation in a way that will fold cleanly. |
1086 | static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, |
1087 | const SCEV *StoreSizeSCEV, Loop *CurLoop, |
1088 | const DataLayout *DL, ScalarEvolution *SE) { |
1089 | const SCEV *TripCountSCEV = getTripCount(BECount, IntPtr, CurLoop, DL, SE); |
1090 | |
1091 | return SE->getMulExpr(TripCountSCEV, |
1092 | SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), |
1093 | SCEV::FlagNUW); |
1094 | } |
1095 | |
1096 | /// processLoopStridedStore - We see a strided store of some value. If we can |
1097 | /// transform this into a memset or memset_pattern in the loop preheader, do so. |
1098 | bool LoopIdiomRecognize::processLoopStridedStore( |
1099 | Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment, |
1100 | Value *StoredVal, Instruction *TheStore, |
1101 | SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev, |
1102 | const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) { |
1103 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
1104 | Constant *PatternValue = nullptr; |
1105 | |
1106 | if (!SplatValue) |
1107 | PatternValue = getMemSetPatternValue(StoredVal, DL); |
1108 | |
1109 | 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", 1110, __extension__ __PRETTY_FUNCTION__)) |
1110 | "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", 1110, __extension__ __PRETTY_FUNCTION__)); |
1111 | |
1112 | // The trip count of the loop and the base pointer of the addrec SCEV is |
1113 | // guaranteed to be loop invariant, which means that it should dominate the |
1114 | // header. This allows us to insert code for it in the preheader. |
1115 | unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); |
1116 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
1117 | IRBuilder<> Builder(Preheader->getTerminator()); |
1118 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
1119 | SCEVExpanderCleaner ExpCleaner(Expander); |
1120 | |
1121 | Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); |
1122 | Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); |
1123 | |
1124 | bool Changed = false; |
1125 | const SCEV *Start = Ev->getStart(); |
1126 | // Handle negative strided loops. |
1127 | if (IsNegStride) |
1128 | Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSizeSCEV, SE); |
1129 | |
1130 | // TODO: ideally we should still be able to generate memset if SCEV expander |
1131 | // is taught to generate the dependencies at the latest point. |
1132 | if (!isSafeToExpand(Start, *SE)) |
1133 | return Changed; |
1134 | |
1135 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
1136 | // this into a memset in the loop preheader now if we want. However, this |
1137 | // would be unsafe to do if there is anything else in the loop that may read |
1138 | // or write to the aliased location. Check for any overlap by generating the |
1139 | // base pointer and checking the region. |
1140 | Value *BasePtr = |
1141 | Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); |
1142 | |
1143 | // From here on out, conservatively report to the pass manager that we've |
1144 | // changed the IR, even if we later clean up these added instructions. There |
1145 | // may be structural differences e.g. in the order of use lists not accounted |
1146 | // for in just a textual dump of the IR. This is written as a variable, even |
1147 | // though statically all the places this dominates could be replaced with |
1148 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
1149 | // the return value will read this comment, and leave them alone. |
1150 | Changed = true; |
1151 | |
1152 | if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
1153 | StoreSizeSCEV, *AA, Stores)) |
1154 | return Changed; |
1155 | |
1156 | if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) |
1157 | return Changed; |
1158 | |
1159 | // Okay, everything looks good, insert the memset. |
1160 | |
1161 | const SCEV *NumBytesS = |
1162 | getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); |
1163 | |
1164 | // TODO: ideally we should still be able to generate memset if SCEV expander |
1165 | // is taught to generate the dependencies at the latest point. |
1166 | if (!isSafeToExpand(NumBytesS, *SE)) |
1167 | return Changed; |
1168 | |
1169 | Value *NumBytes = |
1170 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
1171 | |
1172 | CallInst *NewCall; |
1173 | if (SplatValue) { |
1174 | AAMDNodes AATags = TheStore->getAAMetadata(); |
1175 | for (Instruction *Store : Stores) |
1176 | AATags = AATags.merge(Store->getAAMetadata()); |
1177 | if (auto CI = dyn_cast<ConstantInt>(NumBytes)) |
1178 | AATags = AATags.extendTo(CI->getZExtValue()); |
1179 | else |
1180 | AATags = AATags.extendTo(-1); |
1181 | |
1182 | NewCall = Builder.CreateMemSet( |
1183 | BasePtr, SplatValue, NumBytes, MaybeAlign(StoreAlignment), |
1184 | /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); |
1185 | } else { |
1186 | // Everything is emitted in default address space |
1187 | Type *Int8PtrTy = DestInt8PtrTy; |
1188 | |
1189 | Module *M = TheStore->getModule(); |
1190 | StringRef FuncName = "memset_pattern16"; |
1191 | FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(), |
1192 | Int8PtrTy, Int8PtrTy, IntIdxTy); |
1193 | inferLibFuncAttributes(M, FuncName, *TLI); |
1194 | |
1195 | // Otherwise we should form a memset_pattern16. PatternValue is known to be |
1196 | // an constant array of 16-bytes. Plop the value into a mergable global. |
1197 | GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, |
1198 | GlobalValue::PrivateLinkage, |
1199 | PatternValue, ".memset_pattern"); |
1200 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
1201 | GV->setAlignment(Align(16)); |
1202 | Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); |
1203 | NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); |
1204 | } |
1205 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
1206 | |
1207 | if (MSSAU) { |
1208 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1209 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1210 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1211 | } |
1212 | |
1213 | 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) |
1214 | << " 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) |
1215 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false); |
1216 | |
1217 | ORE.emit([&]() { |
1218 | OptimizationRemark R(DEBUG_TYPE"loop-idiom", "ProcessLoopStridedStore", |
1219 | NewCall->getDebugLoc(), Preheader); |
1220 | R << "Transformed loop-strided store in " |
1221 | << ore::NV("Function", TheStore->getFunction()) |
1222 | << " function into a call to " |
1223 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1224 | << "() intrinsic"; |
1225 | if (!Stores.empty()) |
1226 | R << ore::setExtraArgs(); |
1227 | for (auto *I : Stores) { |
1228 | R << ore::NV("FromBlock", I->getParent()->getName()) |
1229 | << ore::NV("ToBlock", Preheader->getName()); |
1230 | } |
1231 | return R; |
1232 | }); |
1233 | |
1234 | // Okay, the memset has been formed. Zap the original store and anything that |
1235 | // feeds into it. |
1236 | for (auto *I : Stores) { |
1237 | if (MSSAU) |
1238 | MSSAU->removeMemoryAccess(I, true); |
1239 | deleteDeadInstruction(I); |
1240 | } |
1241 | if (MSSAU && VerifyMemorySSA) |
1242 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1243 | ++NumMemSet; |
1244 | ExpCleaner.markResultUsed(); |
1245 | return true; |
1246 | } |
1247 | |
1248 | /// If the stored value is a strided load in the same loop with the same stride |
1249 | /// this may be transformable into a memcpy. This kicks in for stuff like |
1250 | /// for (i) A[i] = B[i]; |
1251 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, |
1252 | const SCEV *BECount) { |
1253 | 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", 1253, __extension__ __PRETTY_FUNCTION__)); |
1254 | |
1255 | Value *StorePtr = SI->getPointerOperand(); |
1256 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
1257 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
1258 | |
1259 | // The store must be feeding a non-volatile load. |
1260 | LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); |
1261 | 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", 1261, __extension__ __PRETTY_FUNCTION__)); |
1262 | |
1263 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
1264 | // loop, which indicates a strided load. If we have something else, it's a |
1265 | // random load we can't handle. |
1266 | Value *LoadPtr = LI->getPointerOperand(); |
1267 | const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr)); |
1268 | |
1269 | const SCEV *StoreSizeSCEV = SE->getConstant(StorePtr->getType(), StoreSize); |
1270 | return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSizeSCEV, |
1271 | SI->getAlign(), LI->getAlign(), SI, LI, |
1272 | StoreEv, LoadEv, BECount); |
1273 | } |
1274 | |
1275 | class MemmoveVerifier { |
1276 | public: |
1277 | explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr, |
1278 | const DataLayout &DL) |
1279 | : DL(DL), BP1(llvm::GetPointerBaseWithConstantOffset( |
1280 | LoadBasePtr.stripPointerCasts(), LoadOff, DL)), |
1281 | BP2(llvm::GetPointerBaseWithConstantOffset( |
1282 | StoreBasePtr.stripPointerCasts(), StoreOff, DL)), |
1283 | IsSameObject(BP1 == BP2) {} |
1284 | |
1285 | bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride, |
1286 | const Instruction &TheLoad, |
1287 | bool IsMemCpy) const { |
1288 | if (IsMemCpy) { |
1289 | // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr |
1290 | // for negative stride. |
1291 | if ((!IsNegStride && LoadOff <= StoreOff) || |
1292 | (IsNegStride && LoadOff >= StoreOff)) |
1293 | return false; |
1294 | } else { |
1295 | // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr |
1296 | // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr. |
1297 | int64_t LoadSize = |
1298 | DL.getTypeSizeInBits(TheLoad.getType()).getFixedSize() / 8; |
1299 | if (BP1 != BP2 || LoadSize != int64_t(StoreSize)) |
1300 | return false; |
1301 | if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) || |
1302 | (IsNegStride && LoadOff + LoadSize > StoreOff)) |
1303 | return false; |
1304 | } |
1305 | return true; |
1306 | } |
1307 | |
1308 | private: |
1309 | const DataLayout &DL; |
1310 | int64_t LoadOff = 0; |
1311 | int64_t StoreOff = 0; |
1312 | const Value *BP1; |
1313 | const Value *BP2; |
1314 | |
1315 | public: |
1316 | const bool IsSameObject; |
1317 | }; |
1318 | |
1319 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad( |
1320 | Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV, |
1321 | MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore, |
1322 | Instruction *TheLoad, const SCEVAddRecExpr *StoreEv, |
1323 | const SCEVAddRecExpr *LoadEv, const SCEV *BECount) { |
1324 | |
1325 | // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to |
1326 | // conservatively bail here, since otherwise we may have to transform |
1327 | // llvm.memcpy.inline into llvm.memcpy which is illegal. |
1328 | if (isa<MemCpyInlineInst>(TheStore)) |
1329 | return false; |
1330 | |
1331 | // The trip count of the loop and the base pointer of the addrec SCEV is |
1332 | // guaranteed to be loop invariant, which means that it should dominate the |
1333 | // header. This allows us to insert code for it in the preheader. |
1334 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
1335 | IRBuilder<> Builder(Preheader->getTerminator()); |
1336 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
1337 | |
1338 | SCEVExpanderCleaner ExpCleaner(Expander); |
1339 | |
1340 | bool Changed = false; |
1341 | const SCEV *StrStart = StoreEv->getStart(); |
1342 | unsigned StrAS = DestPtr->getType()->getPointerAddressSpace(); |
1343 | Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); |
1344 | |
1345 | APInt Stride = getStoreStride(StoreEv); |
1346 | const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(StoreSizeSCEV); |
1347 | |
1348 | // TODO: Deal with non-constant size; Currently expect constant store size |
1349 | 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", 1349, __extension__ __PRETTY_FUNCTION__)); |
1350 | |
1351 | int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue(); |
1352 | bool IsNegStride = StoreSize == -Stride; |
1353 | |
1354 | // Handle negative strided loops. |
1355 | if (IsNegStride) |
1356 | StrStart = |
1357 | getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSizeSCEV, SE); |
1358 | |
1359 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
1360 | // this into a memcpy in the loop preheader now if we want. However, this |
1361 | // would be unsafe to do if there is anything else in the loop that may read |
1362 | // or write the memory region we're storing to. This includes the load that |
1363 | // feeds the stores. Check for an alias by generating the base address and |
1364 | // checking everything. |
1365 | Value *StoreBasePtr = Expander.expandCodeFor( |
1366 | StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); |
1367 | |
1368 | // From here on out, conservatively report to the pass manager that we've |
1369 | // changed the IR, even if we later clean up these added instructions. There |
1370 | // may be structural differences e.g. in the order of use lists not accounted |
1371 | // for in just a textual dump of the IR. This is written as a variable, even |
1372 | // though statically all the places this dominates could be replaced with |
1373 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
1374 | // the return value will read this comment, and leave them alone. |
1375 | Changed = true; |
1376 | |
1377 | SmallPtrSet<Instruction *, 2> IgnoredInsts; |
1378 | IgnoredInsts.insert(TheStore); |
1379 | |
1380 | bool IsMemCpy = isa<MemCpyInst>(TheStore); |
1381 | const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store"; |
1382 | |
1383 | bool LoopAccessStore = |
1384 | mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
1385 | StoreSizeSCEV, *AA, IgnoredInsts); |
1386 | if (LoopAccessStore) { |
1387 | // For memmove case it's not enough to guarantee that loop doesn't access |
1388 | // TheStore and TheLoad. Additionally we need to make sure that TheStore is |
1389 | // the only user of TheLoad. |
1390 | if (!TheLoad->hasOneUse()) |
1391 | return Changed; |
1392 | IgnoredInsts.insert(TheLoad); |
1393 | if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, |
1394 | BECount, StoreSizeSCEV, *AA, IgnoredInsts)) { |
1395 | ORE.emit([&]() { |
1396 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "LoopMayAccessStore", |
1397 | TheStore) |
1398 | << ore::NV("Inst", InstRemark) << " in " |
1399 | << ore::NV("Function", TheStore->getFunction()) |
1400 | << " function will not be hoisted: " |
1401 | << ore::NV("Reason", "The loop may access store location"); |
1402 | }); |
1403 | return Changed; |
1404 | } |
1405 | IgnoredInsts.erase(TheLoad); |
1406 | } |
1407 | |
1408 | const SCEV *LdStart = LoadEv->getStart(); |
1409 | unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace(); |
1410 | |
1411 | // Handle negative strided loops. |
1412 | if (IsNegStride) |
1413 | LdStart = |
1414 | getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSizeSCEV, SE); |
1415 | |
1416 | // For a memcpy, we have to make sure that the input array is not being |
1417 | // mutated by the loop. |
1418 | Value *LoadBasePtr = Expander.expandCodeFor( |
1419 | LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); |
1420 | |
1421 | // If the store is a memcpy instruction, we must check if it will write to |
1422 | // the load memory locations. So remove it from the ignored stores. |
1423 | if (IsMemCpy) |
1424 | IgnoredInsts.erase(TheStore); |
1425 | MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL); |
1426 | if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, |
1427 | StoreSizeSCEV, *AA, IgnoredInsts)) { |
1428 | if (!IsMemCpy) { |
1429 | ORE.emit([&]() { |
1430 | return OptimizationRemarkMissed(DEBUG_TYPE"loop-idiom", "LoopMayAccessLoad", |
1431 | TheLoad) |
1432 | << ore::NV("Inst", InstRemark) << " in " |
1433 | << ore::NV("Function", TheStore->getFunction()) |
1434 | << " function will not be hoisted: " |
1435 | << ore::NV("Reason", "The loop may access load location"); |
1436 | }); |
1437 | return Changed; |
1438 | } |
1439 | // At this point loop may access load only for memcpy in same underlying |
1440 | // object. If that's not the case bail out. |
1441 | if (!Verifier.IsSameObject) |
1442 | return Changed; |
1443 | } |
1444 | |
1445 | bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore; |
1446 | if (UseMemMove) |
1447 | if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, *TheLoad, |
1448 | IsMemCpy)) |
1449 | return Changed; |
1450 | |
1451 | if (avoidLIRForMultiBlockLoop()) |
1452 | return Changed; |
1453 | |
1454 | // Okay, everything is safe, we can transform this! |
1455 | |
1456 | const SCEV *NumBytesS = |
1457 | getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); |
1458 | |
1459 | Value *NumBytes = |
1460 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
1461 | |
1462 | AAMDNodes AATags = TheLoad->getAAMetadata(); |
1463 | AAMDNodes StoreAATags = TheStore->getAAMetadata(); |
1464 | AATags = AATags.merge(StoreAATags); |
1465 | if (auto CI = dyn_cast<ConstantInt>(NumBytes)) |
1466 | AATags = AATags.extendTo(CI->getZExtValue()); |
1467 | else |
1468 | AATags = AATags.extendTo(-1); |
1469 | |
1470 | CallInst *NewCall = nullptr; |
1471 | // Check whether to generate an unordered atomic memcpy: |
1472 | // If the load or store are atomic, then they must necessarily be unordered |
1473 | // by previous checks. |
1474 | if (!TheStore->isAtomic() && !TheLoad->isAtomic()) { |
1475 | if (UseMemMove) |
1476 | NewCall = Builder.CreateMemMove( |
1477 | StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, |
1478 | /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); |
1479 | else |
1480 | NewCall = |
1481 | Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, |
1482 | NumBytes, /*isVolatile=*/false, AATags.TBAA, |
1483 | AATags.TBAAStruct, AATags.Scope, AATags.NoAlias); |
1484 | } else { |
1485 | // For now don't support unordered atomic memmove. |
1486 | if (UseMemMove) |
1487 | return Changed; |
1488 | // We cannot allow unaligned ops for unordered load/store, so reject |
1489 | // anything where the alignment isn't at least the element size. |
1490 | assert((StoreAlign.hasValue() && LoadAlign.hasValue()) &&(static_cast <bool> ((StoreAlign.hasValue() && LoadAlign .hasValue()) && "Expect unordered load/store to have align." ) ? void (0) : __assert_fail ("(StoreAlign.hasValue() && LoadAlign.hasValue()) && \"Expect unordered load/store to have align.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1491, __extension__ __PRETTY_FUNCTION__)) |
1491 | "Expect unordered load/store to have align.")(static_cast <bool> ((StoreAlign.hasValue() && LoadAlign .hasValue()) && "Expect unordered load/store to have align." ) ? void (0) : __assert_fail ("(StoreAlign.hasValue() && LoadAlign.hasValue()) && \"Expect unordered load/store to have align.\"" , "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp", 1491, __extension__ __PRETTY_FUNCTION__)); |
1492 | if (StoreAlign.getValue() < StoreSize || LoadAlign.getValue() < StoreSize) |
1493 | return Changed; |
1494 | |
1495 | // If the element.atomic memcpy is not lowered into explicit |
1496 | // loads/stores later, then it will be lowered into an element-size |
1497 | // specific lib call. If the lib call doesn't exist for our store size, then |
1498 | // we shouldn't generate the memcpy. |
1499 | if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) |
1500 | return Changed; |
1501 | |
1502 | // Create the call. |
1503 | // Note that unordered atomic loads/stores are *required* by the spec to |
1504 | // have an alignment but non-atomic loads/stores may not. |
1505 | NewCall = Builder.CreateElementUnorderedAtomicMemCpy( |
1506 | StoreBasePtr, StoreAlign.getValue(), LoadBasePtr, LoadAlign.getValue(), |
1507 | NumBytes, StoreSize, AATags.TBAA, AATags.TBAAStruct, AATags.Scope, |
1508 | AATags.NoAlias); |
1509 | } |
1510 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
1511 | |
1512 | if (MSSAU) { |
1513 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1514 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1515 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1516 | } |
1517 | |
1518 | 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) |
1519 | << " 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) |
1520 | << "\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) |
1521 | << " 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) |
1522 | << "\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); |
1523 | |
1524 | ORE.emit([&]() { |
1525 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStoreOfLoopLoad", |
1526 | NewCall->getDebugLoc(), Preheader) |
1527 | << "Formed a call to " |
1528 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1529 | << "() intrinsic from " << ore::NV("Inst", InstRemark) |
1530 | << " instruction in " << ore::NV("Function", TheStore->getFunction()) |
1531 | << " function" |
1532 | << ore::setExtraArgs() |
1533 | << ore::NV("FromBlock", TheStore->getParent()->getName()) |
1534 | << ore::NV("ToBlock", Preheader->getName()); |
1535 | }); |
1536 | |
1537 | // Okay, a new call to memcpy/memmove has been formed. Zap the original store |
1538 | // and anything that feeds into it. |
1539 | if (MSSAU) |
1540 | MSSAU->removeMemoryAccess(TheStore, true); |
1541 | deleteDeadInstruction(TheStore); |
1542 | if (MSSAU && VerifyMemorySSA) |
1543 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1544 | if (UseMemMove) |
1545 | ++NumMemMove; |
1546 | else |
1547 | ++NumMemCpy; |
1548 | ExpCleaner.markResultUsed(); |
1549 | return true; |
1550 | } |
1551 | |
1552 | // When compiling for codesize we avoid idiom recognition for a multi-block loop |
1553 | // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. |
1554 | // |
1555 | bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, |
1556 | bool IsLoopMemset) { |
1557 | if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { |
1558 | if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) { |
1559 | 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) |
1560 | << " : 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) |
1561 | << " 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); |
1562 | return true; |
1563 | } |
1564 | } |
1565 | |
1566 | return false; |
1567 | } |
1568 | |
1569 | bool LoopIdiomRecognize::runOnNoncountableLoop() { |
1570 | 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) |
1571 | << 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) |
1572 | << "] 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) |
1573 | << 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); |
1574 | |
1575 | return recognizePopcount() || recognizeAndInsertFFS() || |
1576 | recognizeShiftUntilBitTest() || recognizeShiftUntilZero(); |
1577 | } |
1578 | |
1579 | /// Check if the given conditional branch is based on the comparison between |
1580 | /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is |
1581 | /// true), the control yields to the loop entry. If the branch matches the |
1582 | /// behavior, the variable involved in the comparison is returned. This function |
1583 | /// will be called to see if the precondition and postcondition of the loop are |
1584 | /// in desirable form. |
1585 | static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, |
1586 | bool JmpOnZero = false) { |
1587 | if (!BI || !BI->isConditional()) |
1588 | return nullptr; |
1589 | |
1590 | ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); |
1591 | if (!Cond) |
1592 | return nullptr; |
1593 | |
1594 | ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1)); |
1595 | if (!CmpZero || !CmpZero->isZero()) |
1596 | return nullptr; |
1597 | |
1598 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
1599 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
1600 | if (JmpOnZero) |
1601 | std::swap(TrueSucc, FalseSucc); |
1602 | |
1603 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
1604 | if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || |
1605 | (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) |
1606 | return Cond->getOperand(0); |
1607 | |
1608 | return nullptr; |
1609 | } |
1610 | |
1611 | // Check if the recurrence variable `VarX` is in the right form to create |
1612 | // the idiom. Returns the value coerced to a PHINode if so. |
1613 | static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, |
1614 | BasicBlock *LoopEntry) { |
1615 | auto *PhiX = dyn_cast<PHINode>(VarX); |
1616 | if (PhiX && PhiX->getParent() == LoopEntry && |
1617 | (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) |
1618 | return PhiX; |
1619 | return nullptr; |
1620 | } |
1621 | |
1622 | /// Return true iff the idiom is detected in the loop. |
1623 | /// |
1624 | /// Additionally: |
1625 | /// 1) \p CntInst is set to the instruction counting the population bit. |
1626 | /// 2) \p CntPhi is set to the corresponding phi node. |
1627 | /// 3) \p Var is set to the value whose population bits are being counted. |
1628 | /// |
1629 | /// The core idiom we are trying to detect is: |
1630 | /// \code |
1631 | /// if (x0 != 0) |
1632 | /// goto loop-exit // the precondition of the loop |
1633 | /// cnt0 = init-val; |
1634 | /// do { |
1635 | /// x1 = phi (x0, x2); |
1636 | /// cnt1 = phi(cnt0, cnt2); |
1637 | /// |
1638 | /// cnt2 = cnt1 + 1; |
1639 | /// ... |
1640 | /// x2 = x1 & (x1 - 1); |
1641 | /// ... |
1642 | /// } while(x != 0); |
1643 | /// |
1644 | /// loop-exit: |
1645 | /// \endcode |
1646 | static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, |
1647 | Instruction *&CntInst, PHINode *&CntPhi, |
1648 | Value *&Var) { |
1649 | // step 1: Check to see if the look-back branch match this pattern: |
1650 | // "if (a!=0) goto loop-entry". |
1651 | BasicBlock *LoopEntry; |
1652 | Instruction *DefX2, *CountInst; |
1653 | Value *VarX1, *VarX0; |
1654 | PHINode *PhiX, *CountPhi; |
1655 | |
1656 | DefX2 = CountInst = nullptr; |
1657 | VarX1 = VarX0 = nullptr; |
1658 | PhiX = CountPhi = nullptr; |
1659 | LoopEntry = *(CurLoop->block_begin()); |
1660 | |
1661 | // step 1: Check if the loop-back branch is in desirable form. |
1662 | { |
1663 | if (Value *T = matchCondition( |
1664 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1665 | DefX2 = dyn_cast<Instruction>(T); |
1666 | else |
1667 | return false; |
1668 | } |
1669 | |
1670 | // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" |
1671 | { |
1672 | if (!DefX2 || DefX2->getOpcode() != Instruction::And) |
1673 | return false; |
1674 | |
1675 | BinaryOperator *SubOneOp; |
1676 | |
1677 | if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0)))) |
1678 | VarX1 = DefX2->getOperand(1); |
1679 | else { |
1680 | VarX1 = DefX2->getOperand(0); |
1681 | SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1)); |
1682 | } |
1683 | if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) |
1684 | return false; |
1685 | |
1686 | ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1)); |
1687 | if (!Dec || |
1688 | !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || |
1689 | (SubOneOp->getOpcode() == Instruction::Add && |
1690 | Dec->isMinusOne()))) { |
1691 | return false; |
1692 | } |
1693 | } |
1694 | |
1695 | // step 3: Check the recurrence of variable X |
1696 | PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); |
1697 | if (!PhiX) |
1698 | return false; |
1699 | |
1700 | // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 |
1701 | { |
1702 | CountInst = nullptr; |
1703 | for (Instruction &Inst : llvm::make_range( |
1704 | LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { |
1705 | if (Inst.getOpcode() != Instruction::Add) |
1706 | continue; |
1707 | |
1708 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1)); |
1709 | if (!Inc || !Inc->isOne()) |
1710 | continue; |
1711 | |
1712 | PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); |
1713 | if (!Phi) |
1714 | continue; |
1715 | |
1716 | // Check if the result of the instruction is live of the loop. |
1717 | bool LiveOutLoop = false; |
1718 | for (User *U : Inst.users()) { |
1719 | if ((cast<Instruction>(U))->getParent() != LoopEntry) { |
1720 | LiveOutLoop = true; |
1721 | break; |
1722 | } |
1723 | } |
1724 | |
1725 | if (LiveOutLoop) { |
1726 | CountInst = &Inst; |
1727 | CountPhi = Phi; |
1728 | break; |
1729 | } |
1730 | } |
1731 | |
1732 | if (!CountInst) |
1733 | return false; |
1734 | } |
1735 | |
1736 | // step 5: check if the precondition is in this form: |
1737 | // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" |
1738 | { |
1739 | auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1740 | Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); |
1741 | if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) |
1742 | return false; |
1743 | |
1744 | CntInst = CountInst; |
1745 | CntPhi = CountPhi; |
1746 | Var = T; |
1747 | } |
1748 | |
1749 | return true; |
1750 | } |
1751 | |
1752 | /// Return true if the idiom is detected in the loop. |
1753 | /// |
1754 | /// Additionally: |
1755 | /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) |
1756 | /// or nullptr if there is no such. |
1757 | /// 2) \p CntPhi is set to the corresponding phi node |
1758 | /// or nullptr if there is no such. |
1759 | /// 3) \p Var is set to the value whose CTLZ could be used. |
1760 | /// 4) \p DefX is set to the instruction calculating Loop exit condition. |
1761 | /// |
1762 | /// The core idiom we are trying to detect is: |
1763 | /// \code |
1764 | /// if (x0 == 0) |
1765 | /// goto loop-exit // the precondition of the loop |
1766 | /// cnt0 = init-val; |
1767 | /// do { |
1768 | /// x = phi (x0, x.next); //PhiX |
1769 | /// cnt = phi(cnt0, cnt.next); |
1770 | /// |
1771 | /// cnt.next = cnt + 1; |
1772 | /// ... |
1773 | /// x.next = x >> 1; // DefX |
1774 | /// ... |
1775 | /// } while(x.next != 0); |
1776 | /// |
1777 | /// loop-exit: |
1778 | /// \endcode |
1779 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, |
1780 | Intrinsic::ID &IntrinID, Value *&InitX, |
1781 | Instruction *&CntInst, PHINode *&CntPhi, |
1782 | Instruction *&DefX) { |
1783 | BasicBlock *LoopEntry; |
1784 | Value *VarX = nullptr; |
1785 | |
1786 | DefX = nullptr; |
1787 | CntInst = nullptr; |
1788 | CntPhi = nullptr; |
1789 | LoopEntry = *(CurLoop->block_begin()); |
1790 | |
1791 | // step 1: Check if the loop-back branch is in desirable form. |
1792 | if (Value *T = matchCondition( |
1793 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1794 | DefX = dyn_cast<Instruction>(T); |
1795 | else |
1796 | return false; |
1797 | |
1798 | // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" |
1799 | if (!DefX || !DefX->isShift()) |
1800 | return false; |
1801 | IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : |
1802 | Intrinsic::ctlz; |
1803 | ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)); |
1804 | if (!Shft || !Shft->isOne()) |
1805 | return false; |
1806 | VarX = DefX->getOperand(0); |
1807 | |
1808 | // step 3: Check the recurrence of variable X |
1809 | PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); |
1810 | if (!PhiX) |
1811 | return false; |
1812 | |
1813 | InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); |
1814 | |
1815 | // Make sure the initial value can't be negative otherwise the ashr in the |
1816 | // loop might never reach zero which would make the loop infinite. |
1817 | if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) |
1818 | return false; |
1819 | |
1820 | // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 |
1821 | // or cnt.next = cnt + -1. |
1822 | // TODO: We can skip the step. If loop trip count is known (CTLZ), |
1823 | // then all uses of "cnt.next" could be optimized to the trip count |
1824 | // plus "cnt0". Currently it is not optimized. |
1825 | // This step could be used to detect POPCNT instruction: |
1826 | // cnt.next = cnt + (x.next & 1) |
1827 | for (Instruction &Inst : llvm::make_range( |
1828 | LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { |
1829 | if (Inst.getOpcode() != Instruction::Add) |
1830 | continue; |
1831 | |
1832 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1)); |
1833 | if (!Inc || (!Inc->isOne() && !Inc->isMinusOne())) |
1834 | continue; |
1835 | |
1836 | PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); |
1837 | if (!Phi) |
1838 | continue; |
1839 | |
1840 | CntInst = &Inst; |
1841 | CntPhi = Phi; |
1842 | break; |
1843 | } |
1844 | if (!CntInst) |
1845 | return false; |
1846 | |
1847 | return true; |
1848 | } |
1849 | |
1850 | /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop |
1851 | /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new |
1852 | /// trip count returns true; otherwise, returns false. |
1853 | bool LoopIdiomRecognize::recognizeAndInsertFFS() { |
1854 | // Give up if the loop has multiple blocks or multiple backedges. |
1855 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1856 | return false; |
1857 | |
1858 | Intrinsic::ID IntrinID; |
1859 | Value *InitX; |
1860 | Instruction *DefX = nullptr; |
1861 | PHINode *CntPhi = nullptr; |
1862 | Instruction *CntInst = nullptr; |
1863 | // Help decide if transformation is profitable. For ShiftUntilZero idiom, |
1864 | // this is always 6. |
1865 | size_t IdiomCanonicalSize = 6; |
1866 | |
1867 | if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, |
1868 | CntInst, CntPhi, DefX)) |
1869 | return false; |
1870 | |
1871 | bool IsCntPhiUsedOutsideLoop = false; |
1872 | for (User *U : CntPhi->users()) |
1873 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1874 | IsCntPhiUsedOutsideLoop = true; |
1875 | break; |
1876 | } |
1877 | bool IsCntInstUsedOutsideLoop = false; |
1878 | for (User *U : CntInst->users()) |
1879 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1880 | IsCntInstUsedOutsideLoop = true; |
1881 | break; |
1882 | } |
1883 | // If both CntInst and CntPhi are used outside the loop the profitability |
1884 | // is questionable. |
1885 | if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) |
1886 | return false; |
1887 | |
1888 | // For some CPUs result of CTLZ(X) intrinsic is undefined |
1889 | // when X is 0. If we can not guarantee X != 0, we need to check this |
1890 | // when expand. |
1891 | bool ZeroCheck = false; |
1892 | // It is safe to assume Preheader exist as it was checked in |
1893 | // parent function RunOnLoop. |
1894 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1895 | |
1896 | // If we are using the count instruction outside the loop, make sure we |
1897 | // have a zero check as a precondition. Without the check the loop would run |
1898 | // one iteration for before any check of the input value. This means 0 and 1 |
1899 | // would have identical behavior in the original loop and thus |
1900 | if (!IsCntPhiUsedOutsideLoop) { |
1901 | auto *PreCondBB = PH->getSinglePredecessor(); |
1902 | if (!PreCondBB) |
1903 | return false; |
1904 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1905 | if (!PreCondBI) |
1906 | return false; |
1907 | if (matchCondition(PreCondBI, PH) != InitX) |
1908 | return false; |
1909 | ZeroCheck = true; |
1910 | } |
1911 | |
1912 | // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always |
1913 | // profitable if we delete the loop. |
1914 | |
1915 | // the loop has only 6 instructions: |
1916 | // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] |
1917 | // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] |
1918 | // %shr = ashr %n.addr.0, 1 |
1919 | // %tobool = icmp eq %shr, 0 |
1920 | // %inc = add nsw %i.0, 1 |
1921 | // br i1 %tobool |
1922 | |
1923 | const Value *Args[] = {InitX, |
1924 | ConstantInt::getBool(InitX->getContext(), ZeroCheck)}; |
1925 | |
1926 | // @llvm.dbg doesn't count as they have no semantic effect. |
1927 | auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); |
1928 | uint32_t HeaderSize = |
1929 | std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); |
1930 | |
1931 | IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); |
1932 | InstructionCost Cost = |
1933 | TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); |
1934 | if (HeaderSize != IdiomCanonicalSize && |
1935 | Cost > TargetTransformInfo::TCC_Basic) |
1936 | return false; |
1937 | |
1938 | transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, |
1939 | DefX->getDebugLoc(), ZeroCheck, |
1940 | IsCntPhiUsedOutsideLoop); |
1941 | return true; |
1942 | } |
1943 | |
1944 | /// Recognizes a population count idiom in a non-countable loop. |
1945 | /// |
1946 | /// If detected, transforms the relevant code to issue the popcount intrinsic |
1947 | /// function call, and returns true; otherwise, returns false. |
1948 | bool LoopIdiomRecognize::recognizePopcount() { |
1949 | if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) |
1950 | return false; |
1951 | |
1952 | // Counting population are usually conducted by few arithmetic instructions. |
1953 | // Such instructions can be easily "absorbed" by vacant slots in a |
1954 | // non-compact loop. Therefore, recognizing popcount idiom only makes sense |
1955 | // in a compact loop. |
1956 | |
1957 | // Give up if the loop has multiple blocks or multiple backedges. |
1958 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1959 | return false; |
1960 | |
1961 | BasicBlock *LoopBody = *(CurLoop->block_begin()); |
1962 | if (LoopBody->size() >= 20) { |
1963 | // The loop is too big, bail out. |
1964 | return false; |
1965 | } |
1966 | |
1967 | // It should have a preheader containing nothing but an unconditional branch. |
1968 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1969 | if (!PH || &PH->front() != PH->getTerminator()) |
1970 | return false; |
1971 | auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator()); |
1972 | if (!EntryBI || EntryBI->isConditional()) |
1973 | return false; |
1974 | |
1975 | // It should have a precondition block where the generated popcount intrinsic |
1976 | // function can be inserted. |
1977 | auto *PreCondBB = PH->getSinglePredecessor(); |
1978 | if (!PreCondBB) |
1979 | return false; |
1980 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1981 | if (!PreCondBI || PreCondBI->isUnconditional()) |
1982 | return false; |
1983 | |
1984 | Instruction *CntInst; |
1985 | PHINode *CntPhi; |
1986 | Value *Val; |
1987 | if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) |
1988 | return false; |
1989 | |
1990 | transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); |
1991 | return true; |
1992 | } |
1993 | |
1994 | static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
1995 | const DebugLoc &DL) { |
1996 | Value *Ops[] = {Val}; |
1997 | Type *Tys[] = {Val->getType()}; |
1998 | |
1999 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
2000 | Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); |
2001 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
2002 | CI->setDebugLoc(DL); |
2003 | |
2004 | return CI; |
2005 | } |
2006 | |
2007 | static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
2008 | const DebugLoc &DL, bool ZeroCheck, |
2009 | Intrinsic::ID IID) { |
2010 | Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)}; |
2011 | Type *Tys[] = {Val->getType()}; |
2012 | |
2013 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
2014 | Function *Func = Intrinsic::getDeclaration(M, IID, Tys); |
2015 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
2016 | CI->setDebugLoc(DL); |
2017 | |
2018 | return CI; |
2019 | } |
2020 | |
2021 | /// Transform the following loop (Using CTLZ, CTTZ is similar): |
2022 | /// loop: |
2023 | /// CntPhi = PHI [Cnt0, CntInst] |
2024 | /// PhiX = PHI [InitX, DefX] |
2025 | /// CntInst = CntPhi + 1 |
2026 | /// DefX = PhiX >> 1 |
2027 | /// LOOP_BODY |
2028 | /// Br: loop if (DefX != 0) |
2029 | /// Use(CntPhi) or Use(CntInst) |
2030 | /// |
2031 | /// Into: |
2032 | /// If CntPhi used outside the loop: |
2033 | /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) |
2034 | /// Count = CountPrev + 1 |
2035 | /// else |
2036 | /// Count = BitWidth(InitX) - CTLZ(InitX) |
2037 | /// loop: |
2038 | /// CntPhi = PHI [Cnt0, CntInst] |
2039 | /// PhiX = PHI [InitX, DefX] |
2040 | /// PhiCount = PHI [Count, Dec] |
2041 | /// CntInst = CntPhi + 1 |
2042 | /// DefX = PhiX >> 1 |
2043 | /// Dec = PhiCount - 1 |
2044 | /// LOOP_BODY |
2045 | /// Br: loop if (Dec != 0) |
2046 | /// Use(CountPrev + Cnt0) // Use(CntPhi) |
2047 | /// or |
2048 | /// Use(Count + Cnt0) // Use(CntInst) |
2049 | /// |
2050 | /// If LOOP_BODY is empty the loop will be deleted. |
2051 | /// If CntInst and DefX are not used in LOOP_BODY they will be removed. |
2052 | void LoopIdiomRecognize::transformLoopToCountable( |
2053 | Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, |
2054 | PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, |
2055 | bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { |
2056 | BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator()); |
2057 | |
2058 | // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block |
2059 | IRBuilder<> Builder(PreheaderBr); |
2060 | Builder.SetCurrentDebugLocation(DL); |
2061 | |
2062 | // If there are no uses of CntPhi crate: |
2063 | // Count = BitWidth - CTLZ(InitX); |
2064 | // NewCount = Count; |
2065 | // If there are uses of CntPhi create: |
2066 | // NewCount = BitWidth - CTLZ(InitX >> 1); |
2067 | // Count = NewCount + 1; |
2068 | Value *InitXNext; |
2069 | if (IsCntPhiUsedOutsideLoop) { |
2070 | if (DefX->getOpcode() == Instruction::AShr) |
2071 | InitXNext = Builder.CreateAShr(InitX, 1); |
2072 | else if (DefX->getOpcode() == Instruction::LShr) |
2073 | InitXNext = Builder.CreateLShr(InitX, 1); |
2074 | else if (DefX->getOpcode() == Instruction::Shl) // cttz |
2075 | InitXNext = Builder.CreateShl(InitX, 1); |
2076 | else |
2077 | llvm_unreachable("Unexpected opcode!")::llvm::llvm_unreachable_internal("Unexpected opcode!", "llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 2077); |
2078 | } else |
2079 | InitXNext = InitX; |
2080 | Value *Count = |
2081 | createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); |
2082 | Type *CountTy = Count->getType(); |
2083 | Count = Builder.CreateSub( |
2084 | ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count); |
2085 | Value *NewCount = Count; |
2086 | if (IsCntPhiUsedOutsideLoop) |
2087 | Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1)); |
2088 | |
2089 | NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType()); |
2090 | |
2091 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); |
2092 | if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) { |
2093 | // If the counter was being incremented in the loop, add NewCount to the |
2094 | // counter's initial value, but only if the initial value is not zero. |
2095 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
2096 | if (!InitConst || !InitConst->isZero()) |
2097 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
2098 | } else { |
2099 | // If the count was being decremented in the loop, subtract NewCount from |
2100 | // the counter's initial value. |
2101 | NewCount = Builder.CreateSub(CntInitVal, NewCount); |
2102 | } |
2103 | |
2104 | // Step 2: Insert new IV and loop condition: |
2105 | // loop: |
2106 | // ... |
2107 | // PhiCount = PHI [Count, Dec] |
2108 | // ... |
2109 | // Dec = PhiCount - 1 |
2110 | // ... |
2111 | // Br: loop if (Dec != 0) |
2112 | BasicBlock *Body = *(CurLoop->block_begin()); |
2113 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
2114 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
2115 | |
2116 | PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front()); |
2117 | |
2118 | Builder.SetInsertPoint(LbCond); |
2119 | Instruction *TcDec = cast<Instruction>(Builder.CreateSub( |
2120 | TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true)); |
2121 | |
2122 | TcPhi->addIncoming(Count, Preheader); |
2123 | TcPhi->addIncoming(TcDec, Body); |
2124 | |
2125 | CmpInst::Predicate Pred = |
2126 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
2127 | LbCond->setPredicate(Pred); |
2128 | LbCond->setOperand(0, TcDec); |
2129 | LbCond->setOperand(1, ConstantInt::get(CountTy, 0)); |
2130 | |
2131 | // Step 3: All the references to the original counter outside |
2132 | // the loop are replaced with the NewCount |
2133 | if (IsCntPhiUsedOutsideLoop) |
2134 | CntPhi->replaceUsesOutsideBlock(NewCount, Body); |
2135 | else |
2136 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
2137 | |
2138 | // step 4: Forget the "non-computable" trip-count SCEV associated with the |
2139 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
2140 | SE->forgetLoop(CurLoop); |
2141 | } |
2142 | |
2143 | void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, |
2144 | Instruction *CntInst, |
2145 | PHINode *CntPhi, Value *Var) { |
2146 | BasicBlock *PreHead = CurLoop->getLoopPreheader(); |
2147 | auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator()); |
2148 | const DebugLoc &DL = CntInst->getDebugLoc(); |
2149 | |
2150 | // Assuming before transformation, the loop is following: |
2151 | // if (x) // the precondition |
2152 | // do { cnt++; x &= x - 1; } while(x); |
2153 | |
2154 | // Step 1: Insert the ctpop instruction at the end of the precondition block |
2155 | IRBuilder<> Builder(PreCondBr); |
2156 | Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; |
2157 | { |
2158 | PopCnt = createPopcntIntrinsic(Builder, Var, DL); |
2159 | NewCount = PopCntZext = |
2160 | Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType())); |
2161 | |
2162 | if (NewCount != PopCnt) |
2163 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
2164 | |
2165 | // TripCnt is exactly the number of iterations the loop has |
2166 | TripCnt = NewCount; |
2167 | |
2168 | // If the population counter's initial value is not zero, insert Add Inst. |
2169 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); |
2170 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
2171 | if (!InitConst || !InitConst->isZero()) { |
2172 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
2173 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
2174 | } |
2175 | } |
2176 | |
2177 | // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to |
2178 | // "if (NewCount == 0) loop-exit". Without this change, the intrinsic |
2179 | // function would be partial dead code, and downstream passes will drag |
2180 | // it back from the precondition block to the preheader. |
2181 | { |
2182 | ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition()); |
2183 | |
2184 | Value *Opnd0 = PopCntZext; |
2185 | Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); |
2186 | if (PreCond->getOperand(0) != Var) |
2187 | std::swap(Opnd0, Opnd1); |
2188 | |
2189 | ICmpInst *NewPreCond = cast<ICmpInst>( |
2190 | Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); |
2191 | PreCondBr->setCondition(NewPreCond); |
2192 | |
2193 | RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); |
2194 | } |
2195 | |
2196 | // Step 3: Note that the population count is exactly the trip count of the |
2197 | // loop in question, which enable us to convert the loop from noncountable |
2198 | // loop into a countable one. The benefit is twofold: |
2199 | // |
2200 | // - If the loop only counts population, the entire loop becomes dead after |
2201 | // the transformation. It is a lot easier to prove a countable loop dead |
2202 | // than to prove a noncountable one. (In some C dialects, an infinite loop |
2203 | // isn't dead even if it computes nothing useful. In general, DCE needs |
2204 | // to prove a noncountable loop finite before safely delete it.) |
2205 | // |
2206 | // - If the loop also performs something else, it remains alive. |
2207 | // Since it is transformed to countable form, it can be aggressively |
2208 | // optimized by some optimizations which are in general not applicable |
2209 | // to a noncountable loop. |
2210 | // |
2211 | // After this step, this loop (conceptually) would look like following: |
2212 | // newcnt = __builtin_ctpop(x); |
2213 | // t = newcnt; |
2214 | // if (x) |
2215 | // do { cnt++; x &= x-1; t--) } while (t > 0); |
2216 | BasicBlock *Body = *(CurLoop->block_begin()); |
2217 | { |
2218 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
2219 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
2220 | Type *Ty = TripCnt->getType(); |
2221 | |
2222 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
2223 | |
2224 | Builder.SetInsertPoint(LbCond); |
2225 | Instruction *TcDec = cast<Instruction>( |
2226 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
2227 | "tcdec", false, true)); |
2228 | |
2229 | TcPhi->addIncoming(TripCnt, PreHead); |
2230 | TcPhi->addIncoming(TcDec, Body); |
2231 | |
2232 | CmpInst::Predicate Pred = |
2233 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; |
2234 | LbCond->setPredicate(Pred); |
2235 | LbCond->setOperand(0, TcDec); |
2236 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
2237 | } |
2238 | |
2239 | // Step 4: All the references to the original population counter outside |
2240 | // the loop are replaced with the NewCount -- the value returned from |
2241 | // __builtin_ctpop(). |
2242 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
2243 | |
2244 | // step 5: Forget the "non-computable" trip-count SCEV associated with the |
2245 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
2246 | SE->forgetLoop(CurLoop); |
2247 | } |
2248 | |
2249 | /// Match loop-invariant value. |
2250 | template <typename SubPattern_t> struct match_LoopInvariant { |
2251 | SubPattern_t SubPattern; |
2252 | const Loop *L; |
2253 | |
2254 | match_LoopInvariant(const SubPattern_t &SP, const Loop *L) |
2255 | : SubPattern(SP), L(L) {} |
2256 | |
2257 | template <typename ITy> bool match(ITy *V) { |
2258 | return L->isLoopInvariant(V) && SubPattern.match(V); |
2259 | } |
2260 | }; |
2261 | |
2262 | /// Matches if the value is loop-invariant. |
2263 | template <typename Ty> |
2264 | inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) { |
2265 | return match_LoopInvariant<Ty>(M, L); |
2266 | } |
2267 | |
2268 | /// Return true if the idiom is detected in the loop. |
2269 | /// |
2270 | /// The core idiom we are trying to detect is: |
2271 | /// \code |
2272 | /// entry: |
2273 | /// <...> |
2274 | /// %bitmask = shl i32 1, %bitpos |
2275 | /// br label %loop |
2276 | /// |
2277 | /// loop: |
2278 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
2279 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
2280 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
2281 | /// %x.next = shl i32 %x.curr, 1 |
2282 | /// <...> |
2283 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
2284 | /// |
2285 | /// end: |
2286 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
2287 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
2288 | /// <...> |
2289 | /// \endcode |
2290 | static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX, |
2291 | Value *&BitMask, Value *&BitPos, |
2292 | Value *&CurrX, Instruction *&NextX) { |
2293 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-bittest idiom detection.\n" ; } } while (false) |
2294 | " 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); |
2295 | |
2296 | // Give up if the loop has multiple blocks or multiple backedges. |
2297 | if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { |
2298 | 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); |
2299 | return false; |
2300 | } |
2301 | |
2302 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
2303 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
2304 | 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", 2304, __extension__ __PRETTY_FUNCTION__)); |
2305 | |
2306 | using namespace PatternMatch; |
2307 | |
2308 | // Step 1: Check if the loop backedge is in desirable form. |
2309 | |
2310 | ICmpInst::Predicate Pred; |
2311 | Value *CmpLHS, *CmpRHS; |
2312 | BasicBlock *TrueBB, *FalseBB; |
2313 | if (!match(LoopHeaderBB->getTerminator(), |
2314 | m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)), |
2315 | m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) { |
2316 | 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); |
2317 | return false; |
2318 | } |
2319 | |
2320 | // Step 2: Check if the backedge's condition is in desirable form. |
2321 | |
2322 | auto MatchVariableBitMask = [&]() { |
2323 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
2324 | match(CmpLHS, |
2325 | m_c_And(m_Value(CurrX), |
2326 | m_CombineAnd( |
2327 | m_Value(BitMask), |
2328 | m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)), |
2329 | CurLoop)))); |
2330 | }; |
2331 | auto MatchConstantBitMask = [&]() { |
2332 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
2333 | match(CmpLHS, m_And(m_Value(CurrX), |
2334 | m_CombineAnd(m_Value(BitMask), m_Power2()))) && |
2335 | (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask))); |
2336 | }; |
2337 | auto MatchDecomposableConstantBitMask = [&]() { |
2338 | APInt Mask; |
2339 | return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) && |
2340 | ICmpInst::isEquality(Pred) && Mask.isPowerOf2() && |
2341 | (BitMask = ConstantInt::get(CurrX->getType(), Mask)) && |
2342 | (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2())); |
2343 | }; |
2344 | |
2345 | if (!MatchVariableBitMask() && !MatchConstantBitMask() && |
2346 | !MatchDecomposableConstantBitMask()) { |
2347 | 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); |
2348 | return false; |
2349 | } |
2350 | |
2351 | // Step 3: Check if the recurrence is in desirable form. |
2352 | auto *CurrXPN = dyn_cast<PHINode>(CurrX); |
2353 | if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) { |
2354 | 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); |
2355 | return false; |
2356 | } |
2357 | |
2358 | BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB); |
2359 | NextX = |
2360 | dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB)); |
2361 | |
2362 | 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", 2363, __extension__ __PRETTY_FUNCTION__)) |
2363 | "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", 2363, __extension__ __PRETTY_FUNCTION__)); |
2364 | |
2365 | if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) { |
2366 | // FIXME: support right-shift? |
2367 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad recurrence.\n" ; } } while (false); |
2368 | return false; |
2369 | } |
2370 | |
2371 | // Step 4: Check if the backedge's destinations are in desirable form. |
2372 | |
2373 | 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", 2374, __extension__ __PRETTY_FUNCTION__)) |
2374 | "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", 2374, __extension__ __PRETTY_FUNCTION__)); |
2375 | |
2376 | // cmp-br is commutative, so canonicalize to a single variant. |
2377 | if (Pred != ICmpInst::Predicate::ICMP_EQ) { |
2378 | Pred = ICmpInst::getInversePredicate(Pred); |
2379 | std::swap(TrueBB, FalseBB); |
2380 | } |
2381 | |
2382 | // We expect to exit loop when comparison yields false, |
2383 | // so when it yields true we should branch back to loop header. |
2384 | if (TrueBB != LoopHeaderBB) { |
2385 | 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); |
2386 | return false; |
2387 | } |
2388 | |
2389 | // Okay, idiom checks out. |
2390 | return true; |
2391 | } |
2392 | |
2393 | /// Look for the following loop: |
2394 | /// \code |
2395 | /// entry: |
2396 | /// <...> |
2397 | /// %bitmask = shl i32 1, %bitpos |
2398 | /// br label %loop |
2399 | /// |
2400 | /// loop: |
2401 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
2402 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
2403 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
2404 | /// %x.next = shl i32 %x.curr, 1 |
2405 | /// <...> |
2406 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
2407 | /// |
2408 | /// end: |
2409 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
2410 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
2411 | /// <...> |
2412 | /// \endcode |
2413 | /// |
2414 | /// And transform it into: |
2415 | /// \code |
2416 | /// entry: |
2417 | /// %bitmask = shl i32 1, %bitpos |
2418 | /// %lowbitmask = add i32 %bitmask, -1 |
2419 | /// %mask = or i32 %lowbitmask, %bitmask |
2420 | /// %x.masked = and i32 %x, %mask |
2421 | /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked, |
2422 | /// i1 true) |
2423 | /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros |
2424 | /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1 |
2425 | /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos |
2426 | /// %tripcount = add i32 %backedgetakencount, 1 |
2427 | /// %x.curr = shl i32 %x, %backedgetakencount |
2428 | /// %x.next = shl i32 %x, %tripcount |
2429 | /// br label %loop |
2430 | /// |
2431 | /// loop: |
2432 | /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ] |
2433 | /// %loop.iv.next = add nuw i32 %loop.iv, 1 |
2434 | /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount |
2435 | /// <...> |
2436 | /// br i1 %loop.ivcheck, label %end, label %loop |
2437 | /// |
2438 | /// end: |
2439 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
2440 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
2441 | /// <...> |
2442 | /// \endcode |
2443 | bool LoopIdiomRecognize::recognizeShiftUntilBitTest() { |
2444 | bool MadeChange = false; |
2445 | |
2446 | Value *X, *BitMask, *BitPos, *XCurr; |
2447 | Instruction *XNext; |
2448 | if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr, |
2449 | XNext)) { |
2450 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detection failed.\n" ; } } while (false) |
2451 | " 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); |
2452 | return MadeChange; |
2453 | } |
2454 | 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); |
2455 | |
2456 | // Ok, it is the idiom we were looking for, we *could* transform this loop, |
2457 | // but is it profitable to transform? |
2458 | |
2459 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
2460 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
2461 | 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", 2461, __extension__ __PRETTY_FUNCTION__)); |
2462 | |
2463 | BasicBlock *SuccessorBB = CurLoop->getExitBlock(); |
2464 | 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", 2464, __extension__ __PRETTY_FUNCTION__)); |
2465 | |
2466 | IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); |
2467 | Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc()); |
2468 | |
2469 | Intrinsic::ID IntrID = Intrinsic::ctlz; |
2470 | Type *Ty = X->getType(); |
2471 | unsigned Bitwidth = Ty->getScalarSizeInBits(); |
2472 | |
2473 | TargetTransformInfo::TargetCostKind CostKind = |
2474 | TargetTransformInfo::TCK_SizeAndLatency; |
2475 | |
2476 | // The rewrite is considered to be unprofitable iff and only iff the |
2477 | // intrinsic/shift we'll use are not cheap. Note that we are okay with *just* |
2478 | // making the loop countable, even if nothing else changes. |
2479 | IntrinsicCostAttributes Attrs( |
2480 | IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getTrue()}); |
2481 | InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); |
2482 | if (Cost > TargetTransformInfo::TCC_Basic) { |
2483 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false) |
2484 | " 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); |
2485 | return MadeChange; |
2486 | } |
2487 | if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) > |
2488 | TargetTransformInfo::TCC_Basic) { |
2489 | 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); |
2490 | return MadeChange; |
2491 | } |
2492 | |
2493 | // Ok, transform appears worthwhile. |
2494 | MadeChange = true; |
2495 | |
2496 | // Step 1: Compute the loop trip count. |
2497 | |
2498 | Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty), |
2499 | BitPos->getName() + ".lowbitmask"); |
2500 | Value *Mask = |
2501 | Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask"); |
2502 | Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked"); |
2503 | CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic( |
2504 | IntrID, Ty, {XMasked, /*is_zero_undef=*/Builder.getTrue()}, |
2505 | /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros"); |
2506 | Value *XMaskedNumActiveBits = Builder.CreateSub( |
2507 | ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros, |
2508 | XMasked->getName() + ".numactivebits", /*HasNUW=*/true, |
2509 | /*HasNSW=*/Bitwidth != 2); |
2510 | Value *XMaskedLeadingOnePos = |
2511 | Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty), |
2512 | XMasked->getName() + ".leadingonepos", /*HasNUW=*/false, |
2513 | /*HasNSW=*/Bitwidth > 2); |
2514 | |
2515 | Value *LoopBackedgeTakenCount = Builder.CreateSub( |
2516 | BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount", |
2517 | /*HasNUW=*/true, /*HasNSW=*/true); |
2518 | // We know loop's backedge-taken count, but what's loop's trip count? |
2519 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
2520 | Value *LoopTripCount = |
2521 | Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), |
2522 | CurLoop->getName() + ".tripcount", /*HasNUW=*/true, |
2523 | /*HasNSW=*/Bitwidth != 2); |
2524 | |
2525 | // Step 2: Compute the recurrence's final value without a loop. |
2526 | |
2527 | // NewX is always safe to compute, because `LoopBackedgeTakenCount` |
2528 | // will always be smaller than `bitwidth(X)`, i.e. we never get poison. |
2529 | Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount); |
2530 | NewX->takeName(XCurr); |
2531 | if (auto *I = dyn_cast<Instruction>(NewX)) |
2532 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
2533 | |
2534 | Value *NewXNext; |
2535 | // Rewriting XNext is more complicated, however, because `X << LoopTripCount` |
2536 | // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen |
2537 | // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know |
2538 | // that isn't the case, we'll need to emit an alternative, safe IR. |
2539 | if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() || |
2540 | PatternMatch::match( |
2541 | BitPos, PatternMatch::m_SpecificInt_ICMP( |
2542 | ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(), |
2543 | Ty->getScalarSizeInBits() - 1)))) |
2544 | NewXNext = Builder.CreateShl(X, LoopTripCount); |
2545 | else { |
2546 | // Otherwise, just additionally shift by one. It's the smallest solution, |
2547 | // alternatively, we could check that NewX is INT_MIN (or BitPos is ) |
2548 | // and select 0 instead. |
2549 | NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1)); |
2550 | } |
2551 | |
2552 | NewXNext->takeName(XNext); |
2553 | if (auto *I = dyn_cast<Instruction>(NewXNext)) |
2554 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
2555 | |
2556 | // Step 3: Adjust the successor basic block to recieve the computed |
2557 | // recurrence's final value instead of the recurrence itself. |
2558 | |
2559 | XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB); |
2560 | XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB); |
2561 | |
2562 | // Step 4: Rewrite the loop into a countable form, with canonical IV. |
2563 | |
2564 | // The new canonical induction variable. |
2565 | Builder.SetInsertPoint(&LoopHeaderBB->front()); |
2566 | auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); |
2567 | |
2568 | // The induction itself. |
2569 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
2570 | Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); |
2571 | auto *IVNext = |
2572 | Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next", |
2573 | /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); |
2574 | |
2575 | // The loop trip count check. |
2576 | auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount, |
2577 | CurLoop->getName() + ".ivcheck"); |
2578 | Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB); |
2579 | LoopHeaderBB->getTerminator()->eraseFromParent(); |
2580 | |
2581 | // Populate the IV PHI. |
2582 | IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); |
2583 | IV->addIncoming(IVNext, LoopHeaderBB); |
2584 | |
2585 | // Step 5: Forget the "non-computable" trip-count SCEV associated with the |
2586 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
2587 | |
2588 | SE->forgetLoop(CurLoop); |
2589 | |
2590 | // Other passes will take care of actually deleting the loop if possible. |
2591 | |
2592 | 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); |
2593 | |
2594 | ++NumShiftUntilBitTest; |
2595 | return MadeChange; |
2596 | } |
2597 | |
2598 | /// Return true if the idiom is detected in the loop. |
2599 | /// |
2600 | /// The core idiom we are trying to detect is: |
2601 | /// \code |
2602 | /// entry: |
2603 | /// <...> |
2604 | /// %start = <...> |
2605 | /// %extraoffset = <...> |
2606 | /// <...> |
2607 | /// br label %for.cond |
2608 | /// |
2609 | /// loop: |
2610 | /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] |
2611 | /// %nbits = add nsw i8 %iv, %extraoffset |
2612 | /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits |
2613 | /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 |
2614 | /// %iv.next = add i8 %iv, 1 |
2615 | /// <...> |
2616 | /// br i1 %val.shifted.iszero, label %end, label %loop |
2617 | /// |
2618 | /// end: |
2619 | /// %iv.res = phi i8 [ %iv, %loop ] <...> |
2620 | /// %nbits.res = phi i8 [ %nbits, %loop ] <...> |
2621 | /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> |
2622 | /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> |
2623 | /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> |
2624 | /// <...> |
2625 | /// \endcode |
2626 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE, |
2627 | Instruction *&ValShiftedIsZero, |
2628 | Intrinsic::ID &IntrinID, Instruction *&IV, |
2629 | Value *&Start, Value *&Val, |
2630 | const SCEV *&ExtraOffsetExpr, |
2631 | bool &InvertedCond) { |
2632 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-zero idiom detection.\n" ; } } while (false) |
2633 | " 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); |
2634 | |
2635 | // Give up if the loop has multiple blocks or multiple backedges. |
2636 | if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { |
2637 | 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); |
2638 | return false; |
2639 | } |
2640 | |
2641 | Instruction *ValShifted, *NBits, *IVNext; |
2642 | Value *ExtraOffset; |
2643 | |
2644 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
2645 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
2646 | 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", 2646, __extension__ __PRETTY_FUNCTION__)); |
2647 | |
2648 | using namespace PatternMatch; |
2649 | |
2650 | // Step 1: Check if the loop backedge, condition is in desirable form. |
2651 | |
2652 | ICmpInst::Predicate Pred; |
2653 | BasicBlock *TrueBB, *FalseBB; |
2654 | if (!match(LoopHeaderBB->getTerminator(), |
2655 | m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB), |
2656 | m_BasicBlock(FalseBB))) || |
2657 | !match(ValShiftedIsZero, |
2658 | m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) || |
2659 | !ICmpInst::isEquality(Pred)) { |
2660 | 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); |
2661 | return false; |
2662 | } |
2663 | |
2664 | // Step 2: Check if the comparison's operand is in desirable form. |
2665 | // FIXME: Val could be a one-input PHI node, which we should look past. |
2666 | if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop), |
2667 | m_Instruction(NBits)))) { |
2668 | 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); |
2669 | return false; |
2670 | } |
2671 | IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz |
2672 | : Intrinsic::ctlz; |
2673 | |
2674 | // Step 3: Check if the shift amount is in desirable form. |
2675 | |
2676 | if (match(NBits, m_c_Add(m_Instruction(IV), |
2677 | m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && |
2678 | (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap())) |
2679 | ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset)); |
2680 | else if (match(NBits, |
2681 | m_Sub(m_Instruction(IV), |
2682 | m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && |
2683 | NBits->hasNoSignedWrap()) |
2684 | ExtraOffsetExpr = SE->getSCEV(ExtraOffset); |
2685 | else { |
2686 | IV = NBits; |
2687 | ExtraOffsetExpr = SE->getZero(NBits->getType()); |
2688 | } |
2689 | |
2690 | // Step 4: Check if the recurrence is in desirable form. |
2691 | auto *IVPN = dyn_cast<PHINode>(IV); |
2692 | if (!IVPN || IVPN->getParent() != LoopHeaderBB) { |
2693 | 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); |
2694 | return false; |
2695 | } |
2696 | |
2697 | Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB); |
2698 | IVNext = dyn_cast<Instruction>(IVPN->getIncomingValueForBlock(LoopHeaderBB)); |
2699 | |
2700 | if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) { |
2701 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad recurrence.\n" ; } } while (false); |
2702 | return false; |
2703 | } |
2704 | |
2705 | // Step 4: Check if the backedge's destinations are in desirable form. |
2706 | |
2707 | 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", 2708, __extension__ __PRETTY_FUNCTION__)) |
2708 | "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", 2708, __extension__ __PRETTY_FUNCTION__)); |
2709 | |
2710 | // cmp-br is commutative, so canonicalize to a single variant. |
2711 | InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ; |
2712 | if (InvertedCond) { |
2713 | Pred = ICmpInst::getInversePredicate(Pred); |
Value stored to 'Pred' is never read | |
2714 | std::swap(TrueBB, FalseBB); |
2715 | } |
2716 | |
2717 | // We expect to exit loop when comparison yields true, |
2718 | // so when it yields false we should branch back to loop header. |
2719 | if (FalseBB != LoopHeaderBB) { |
2720 | 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); |
2721 | return false; |
2722 | } |
2723 | |
2724 | // The new, countable, loop will certainly only run a known number of |
2725 | // iterations, It won't be infinite. But the old loop might be infinite |
2726 | // under certain conditions. For logical shifts, the value will become zero |
2727 | // after at most bitwidth(%Val) loop iterations. However, for arithmetic |
2728 | // right-shift, iff the sign bit was set, the value will never become zero, |
2729 | // and the loop may never finish. |
2730 | if (ValShifted->getOpcode() == Instruction::AShr && |
2731 | !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) { |
2732 | 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); |
2733 | return false; |
2734 | } |
2735 | |
2736 | // Okay, idiom checks out. |
2737 | return true; |
2738 | } |
2739 | |
2740 | /// Look for the following loop: |
2741 | /// \code |
2742 | /// entry: |
2743 | /// <...> |
2744 | /// %start = <...> |
2745 | /// %extraoffset = <...> |
2746 | /// <...> |
2747 | /// br label %for.cond |
2748 | /// |
2749 | /// loop: |
2750 | /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] |
2751 | /// %nbits = add nsw i8 %iv, %extraoffset |
2752 | /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits |
2753 | /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 |
2754 | /// %iv.next = add i8 %iv, 1 |
2755 | /// <...> |
2756 | /// br i1 %val.shifted.iszero, label %end, label %loop |
2757 | /// |
2758 | /// end: |
2759 | /// %iv.res = phi i8 [ %iv, %loop ] <...> |
2760 | /// %nbits.res = phi i8 [ %nbits, %loop ] <...> |
2761 | /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> |
2762 | /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> |
2763 | /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> |
2764 | /// <...> |
2765 | /// \endcode |
2766 | /// |
2767 | /// And transform it into: |
2768 | /// \code |
2769 | /// entry: |
2770 | /// <...> |
2771 | /// %start = <...> |
2772 | /// %extraoffset = <...> |
2773 | /// <...> |
2774 | /// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0) |
2775 | /// %val.numactivebits = sub i8 8, %val.numleadingzeros |
2776 | /// %extraoffset.neg = sub i8 0, %extraoffset |
2777 | /// %tmp = add i8 %val.numactivebits, %extraoffset.neg |
2778 | /// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start) |
2779 | /// %loop.tripcount = sub i8 %iv.final, %start |
2780 | /// br label %loop |
2781 | /// |
2782 | /// loop: |
2783 | /// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ] |
2784 | /// %loop.iv.next = add i8 %loop.iv, 1 |
2785 | /// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount |
2786 | /// %iv = add i8 %loop.iv, %start |
2787 | /// <...> |
2788 | /// br i1 %loop.ivcheck, label %end, label %loop |
2789 | /// |
2790 | /// end: |
2791 | /// %iv.res = phi i8 [ %iv.final, %loop ] <...> |
2792 | /// <...> |
2793 | /// \endcode |
2794 | bool LoopIdiomRecognize::recognizeShiftUntilZero() { |
2795 | bool MadeChange = false; |
2796 | |
2797 | Instruction *ValShiftedIsZero; |
2798 | Intrinsic::ID IntrID; |
2799 | Instruction *IV; |
2800 | Value *Start, *Val; |
2801 | const SCEV *ExtraOffsetExpr; |
2802 | bool InvertedCond; |
2803 | if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV, |
2804 | Start, Val, ExtraOffsetExpr, InvertedCond)) { |
2805 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-zero idiom detection failed.\n" ; } } while (false) |
2806 | " 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); |
2807 | return MadeChange; |
2808 | } |
2809 | 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); |
2810 | |
2811 | // Ok, it is the idiom we were looking for, we *could* transform this loop, |
2812 | // but is it profitable to transform? |
2813 | |
2814 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
2815 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
2816 | 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", 2816, __extension__ __PRETTY_FUNCTION__)); |
2817 | |
2818 | BasicBlock *SuccessorBB = CurLoop->getExitBlock(); |
2819 | 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", 2819, __extension__ __PRETTY_FUNCTION__)); |
2820 | |
2821 | IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); |
2822 | Builder.SetCurrentDebugLocation(IV->getDebugLoc()); |
2823 | |
2824 | Type *Ty = Val->getType(); |
2825 | unsigned Bitwidth = Ty->getScalarSizeInBits(); |
2826 | |
2827 | TargetTransformInfo::TargetCostKind CostKind = |
2828 | TargetTransformInfo::TCK_SizeAndLatency; |
2829 | |
2830 | // The rewrite is considered to be unprofitable iff and only iff the |
2831 | // intrinsic we'll use are not cheap. Note that we are okay with *just* |
2832 | // making the loop countable, even if nothing else changes. |
2833 | IntrinsicCostAttributes Attrs( |
2834 | IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getFalse()}); |
2835 | InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); |
2836 | if (Cost > TargetTransformInfo::TCC_Basic) { |
2837 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false) |
2838 | " 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); |
2839 | return MadeChange; |
2840 | } |
2841 | |
2842 | // Ok, transform appears worthwhile. |
2843 | MadeChange = true; |
2844 | |
2845 | bool OffsetIsZero = false; |
2846 | if (auto *ExtraOffsetExprC = dyn_cast<SCEVConstant>(ExtraOffsetExpr)) |
2847 | OffsetIsZero = ExtraOffsetExprC->isZero(); |
2848 | |
2849 | // Step 1: Compute the loop's final IV value / trip count. |
2850 | |
2851 | CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic( |
2852 | IntrID, Ty, {Val, /*is_zero_undef=*/Builder.getFalse()}, |
2853 | /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros"); |
2854 | Value *ValNumActiveBits = Builder.CreateSub( |
2855 | ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros, |
2856 | Val->getName() + ".numactivebits", /*HasNUW=*/true, |
2857 | /*HasNSW=*/Bitwidth != 2); |
2858 | |
2859 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
2860 | Expander.setInsertPoint(&*Builder.GetInsertPoint()); |
2861 | Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr); |
2862 | |
2863 | Value *ValNumActiveBitsOffset = Builder.CreateAdd( |
2864 | ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset", |
2865 | /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true); |
2866 | Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty}, |
2867 | {ValNumActiveBitsOffset, Start}, |
2868 | /*FMFSource=*/nullptr, "iv.final"); |
2869 | |
2870 | auto *LoopBackedgeTakenCount = cast<Instruction>(Builder.CreateSub( |
2871 | IVFinal, Start, CurLoop->getName() + ".backedgetakencount", |
2872 | /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true)); |
2873 | // FIXME: or when the offset was `add nuw` |
2874 | |
2875 | // We know loop's backedge-taken count, but what's loop's trip count? |
2876 | Value *LoopTripCount = |
2877 | Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), |
2878 | CurLoop->getName() + ".tripcount", /*HasNUW=*/true, |
2879 | /*HasNSW=*/Bitwidth != 2); |
2880 | |
2881 | // Step 2: Adjust the successor basic block to recieve the original |
2882 | // induction variable's final value instead of the orig. IV itself. |
2883 | |
2884 | IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB); |
2885 | |
2886 | // Step 3: Rewrite the loop into a countable form, with canonical IV. |
2887 | |
2888 | // The new canonical induction variable. |
2889 | Builder.SetInsertPoint(&LoopHeaderBB->front()); |
2890 | auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); |
2891 | |
2892 | // The induction itself. |
2893 | Builder.SetInsertPoint(LoopHeaderBB->getFirstNonPHI()); |
2894 | auto *CIVNext = |
2895 | Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next", |
2896 | /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); |
2897 | |
2898 | // The loop trip count check. |
2899 | auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount, |
2900 | CurLoop->getName() + ".ivcheck"); |
2901 | auto *NewIVCheck = CIVCheck; |
2902 | if (InvertedCond) { |
2903 | NewIVCheck = Builder.CreateNot(CIVCheck); |
2904 | NewIVCheck->takeName(ValShiftedIsZero); |
2905 | } |
2906 | |
2907 | // The original IV, but rebased to be an offset to the CIV. |
2908 | auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false, |
2909 | /*HasNSW=*/true); // FIXME: what about NUW? |
2910 | IVDePHId->takeName(IV); |
2911 | |
2912 | // The loop terminator. |
2913 | Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); |
2914 | Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB); |
2915 | LoopHeaderBB->getTerminator()->eraseFromParent(); |
2916 | |
2917 | // Populate the IV PHI. |
2918 | CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); |
2919 | CIV->addIncoming(CIVNext, LoopHeaderBB); |
2920 | |
2921 | // Step 4: Forget the "non-computable" trip-count SCEV associated with the |
2922 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
2923 | |
2924 | SE->forgetLoop(CurLoop); |
2925 | |
2926 | // Step 5: Try to cleanup the loop's body somewhat. |
2927 | IV->replaceAllUsesWith(IVDePHId); |
2928 | IV->eraseFromParent(); |
2929 | |
2930 | ValShiftedIsZero->replaceAllUsesWith(NewIVCheck); |
2931 | ValShiftedIsZero->eraseFromParent(); |
2932 | |
2933 | // Other passes will take care of actually deleting the loop if possible. |
2934 | |
2935 | 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); |
2936 | |
2937 | ++NumShiftUntilZero; |
2938 | return MadeChange; |
2939 | } |