File: | llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp |
Warning: | line 510, column 15 Although the value stored to 'PatternValue' is used in the enclosing expression, the value is never actually read from 'PatternValue' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass implements an idiom recognizer that transforms simple loops into a |
10 | // non-loop form. In cases that this kicks in, it can be a significant |
11 | // performance win. |
12 | // |
13 | // If compiling for code size we avoid idiom recognition if the resulting |
14 | // code could be larger than the code for the original loop. One way this could |
15 | // happen is if the loop is not removable after idiom recognition due to the |
16 | // presence of non-idiom instructions. The initial implementation of the |
17 | // heuristics applies to idioms in multi-block loops. |
18 | // |
19 | //===----------------------------------------------------------------------===// |
20 | // |
21 | // TODO List: |
22 | // |
23 | // Future loop memory idioms to recognize: |
24 | // memcmp, memmove, strlen, etc. |
25 | // Future floating point idioms to recognize in -ffast-math mode: |
26 | // fpowi |
27 | // Future integer operation idioms to recognize: |
28 | // ctpop |
29 | // |
30 | // Beware that isel's default lowering for ctpop is highly inefficient for |
31 | // i64 and larger types when i64 is legal and the value has few bits set. It |
32 | // would be good to enhance isel to emit a loop for ctpop in this case. |
33 | // |
34 | // This could recognize common matrix multiplies and dot product idioms and |
35 | // replace them with calls to BLAS (if linked in??). |
36 | // |
37 | //===----------------------------------------------------------------------===// |
38 | |
39 | #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h" |
40 | #include "llvm/ADT/APInt.h" |
41 | #include "llvm/ADT/ArrayRef.h" |
42 | #include "llvm/ADT/DenseMap.h" |
43 | #include "llvm/ADT/MapVector.h" |
44 | #include "llvm/ADT/SetVector.h" |
45 | #include "llvm/ADT/SmallPtrSet.h" |
46 | #include "llvm/ADT/SmallVector.h" |
47 | #include "llvm/ADT/Statistic.h" |
48 | #include "llvm/ADT/StringRef.h" |
49 | #include "llvm/Analysis/AliasAnalysis.h" |
50 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
51 | #include "llvm/Analysis/LoopInfo.h" |
52 | #include "llvm/Analysis/LoopPass.h" |
53 | #include "llvm/Analysis/MemoryLocation.h" |
54 | #include "llvm/Analysis/MemorySSA.h" |
55 | #include "llvm/Analysis/MemorySSAUpdater.h" |
56 | #include "llvm/Analysis/MustExecute.h" |
57 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
58 | #include "llvm/Analysis/ScalarEvolution.h" |
59 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
60 | #include "llvm/Analysis/TargetLibraryInfo.h" |
61 | #include "llvm/Analysis/TargetTransformInfo.h" |
62 | #include "llvm/Analysis/ValueTracking.h" |
63 | #include "llvm/IR/Attributes.h" |
64 | #include "llvm/IR/BasicBlock.h" |
65 | #include "llvm/IR/Constant.h" |
66 | #include "llvm/IR/Constants.h" |
67 | #include "llvm/IR/DataLayout.h" |
68 | #include "llvm/IR/DebugLoc.h" |
69 | #include "llvm/IR/DerivedTypes.h" |
70 | #include "llvm/IR/Dominators.h" |
71 | #include "llvm/IR/GlobalValue.h" |
72 | #include "llvm/IR/GlobalVariable.h" |
73 | #include "llvm/IR/IRBuilder.h" |
74 | #include "llvm/IR/InstrTypes.h" |
75 | #include "llvm/IR/Instruction.h" |
76 | #include "llvm/IR/Instructions.h" |
77 | #include "llvm/IR/IntrinsicInst.h" |
78 | #include "llvm/IR/Intrinsics.h" |
79 | #include "llvm/IR/LLVMContext.h" |
80 | #include "llvm/IR/Module.h" |
81 | #include "llvm/IR/PassManager.h" |
82 | #include "llvm/IR/Type.h" |
83 | #include "llvm/IR/User.h" |
84 | #include "llvm/IR/Value.h" |
85 | #include "llvm/IR/ValueHandle.h" |
86 | #include "llvm/InitializePasses.h" |
87 | #include "llvm/Pass.h" |
88 | #include "llvm/Support/Casting.h" |
89 | #include "llvm/Support/CommandLine.h" |
90 | #include "llvm/Support/Debug.h" |
91 | #include "llvm/Support/raw_ostream.h" |
92 | #include "llvm/Transforms/Scalar.h" |
93 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
94 | #include "llvm/Transforms/Utils/Local.h" |
95 | #include "llvm/Transforms/Utils/LoopUtils.h" |
96 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
97 | #include <algorithm> |
98 | #include <cassert> |
99 | #include <cstdint> |
100 | #include <utility> |
101 | #include <vector> |
102 | |
103 | using namespace llvm; |
104 | |
105 | #define DEBUG_TYPE"loop-idiom" "loop-idiom" |
106 | |
107 | STATISTIC(NumMemSet, "Number of memset's formed from loop stores")static llvm::Statistic NumMemSet = {"loop-idiom", "NumMemSet" , "Number of memset's formed from loop stores"}; |
108 | STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores")static llvm::Statistic NumMemCpy = {"loop-idiom", "NumMemCpy" , "Number of memcpy's formed from loop load+stores"}; |
109 | |
110 | bool DisableLIRP::All; |
111 | static cl::opt<bool, true> |
112 | DisableLIRPAll("disable-" DEBUG_TYPE"loop-idiom" "-all", |
113 | cl::desc("Options to disable Loop Idiom Recognize Pass."), |
114 | cl::location(DisableLIRP::All), cl::init(false), |
115 | cl::ReallyHidden); |
116 | |
117 | bool DisableLIRP::Memset; |
118 | static cl::opt<bool, true> |
119 | DisableLIRPMemset("disable-" DEBUG_TYPE"loop-idiom" "-memset", |
120 | cl::desc("Proceed with loop idiom recognize pass, but do " |
121 | "not convert loop(s) to memset."), |
122 | cl::location(DisableLIRP::Memset), cl::init(false), |
123 | cl::ReallyHidden); |
124 | |
125 | bool DisableLIRP::Memcpy; |
126 | static cl::opt<bool, true> |
127 | DisableLIRPMemcpy("disable-" DEBUG_TYPE"loop-idiom" "-memcpy", |
128 | cl::desc("Proceed with loop idiom recognize pass, but do " |
129 | "not convert loop(s) to memcpy."), |
130 | cl::location(DisableLIRP::Memcpy), cl::init(false), |
131 | cl::ReallyHidden); |
132 | |
133 | static cl::opt<bool> UseLIRCodeSizeHeurs( |
134 | "use-lir-code-size-heurs", |
135 | cl::desc("Use loop idiom recognition code size heuristics when compiling" |
136 | "with -Os/-Oz"), |
137 | cl::init(true), cl::Hidden); |
138 | |
139 | namespace { |
140 | |
141 | class LoopIdiomRecognize { |
142 | Loop *CurLoop = nullptr; |
143 | AliasAnalysis *AA; |
144 | DominatorTree *DT; |
145 | LoopInfo *LI; |
146 | ScalarEvolution *SE; |
147 | TargetLibraryInfo *TLI; |
148 | const TargetTransformInfo *TTI; |
149 | const DataLayout *DL; |
150 | OptimizationRemarkEmitter &ORE; |
151 | bool ApplyCodeSizeHeuristics; |
152 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
153 | |
154 | public: |
155 | explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
156 | LoopInfo *LI, ScalarEvolution *SE, |
157 | TargetLibraryInfo *TLI, |
158 | const TargetTransformInfo *TTI, MemorySSA *MSSA, |
159 | const DataLayout *DL, |
160 | OptimizationRemarkEmitter &ORE) |
161 | : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { |
162 | if (MSSA) |
163 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); |
164 | } |
165 | |
166 | bool runOnLoop(Loop *L); |
167 | |
168 | private: |
169 | using StoreList = SmallVector<StoreInst *, 8>; |
170 | using StoreListMap = MapVector<Value *, StoreList>; |
171 | |
172 | StoreListMap StoreRefsForMemset; |
173 | StoreListMap StoreRefsForMemsetPattern; |
174 | StoreList StoreRefsForMemcpy; |
175 | bool HasMemset; |
176 | bool HasMemsetPattern; |
177 | bool HasMemcpy; |
178 | |
179 | /// Return code for isLegalStore() |
180 | enum LegalStoreKind { |
181 | None = 0, |
182 | Memset, |
183 | MemsetPattern, |
184 | Memcpy, |
185 | UnorderedAtomicMemcpy, |
186 | DontUse // Dummy retval never to be used. Allows catching errors in retval |
187 | // handling. |
188 | }; |
189 | |
190 | /// \name Countable Loop Idiom Handling |
191 | /// @{ |
192 | |
193 | bool runOnCountableLoop(); |
194 | bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, |
195 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
196 | |
197 | void collectStores(BasicBlock *BB); |
198 | LegalStoreKind isLegalStore(StoreInst *SI); |
199 | enum class ForMemset { No, Yes }; |
200 | bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount, |
201 | ForMemset For); |
202 | bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); |
203 | |
204 | bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize, |
205 | MaybeAlign StoreAlignment, Value *StoredVal, |
206 | Instruction *TheStore, |
207 | SmallPtrSetImpl<Instruction *> &Stores, |
208 | const SCEVAddRecExpr *Ev, const SCEV *BECount, |
209 | bool NegStride, bool IsLoopMemset = false); |
210 | bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); |
211 | bool avoidLIRForMultiBlockLoop(bool IsMemset = false, |
212 | bool IsLoopMemset = false); |
213 | |
214 | /// @} |
215 | /// \name Noncountable Loop Idiom Handling |
216 | /// @{ |
217 | |
218 | bool runOnNoncountableLoop(); |
219 | |
220 | bool recognizePopcount(); |
221 | void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, |
222 | PHINode *CntPhi, Value *Var); |
223 | bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz |
224 | void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, |
225 | Instruction *CntInst, PHINode *CntPhi, |
226 | Value *Var, Instruction *DefX, |
227 | const DebugLoc &DL, bool ZeroCheck, |
228 | bool IsCntPhiUsedOutsideLoop); |
229 | |
230 | /// @} |
231 | }; |
232 | |
233 | class LoopIdiomRecognizeLegacyPass : public LoopPass { |
234 | public: |
235 | static char ID; |
236 | |
237 | explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) { |
238 | initializeLoopIdiomRecognizeLegacyPassPass( |
239 | *PassRegistry::getPassRegistry()); |
240 | } |
241 | |
242 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { |
243 | if (DisableLIRP::All) |
244 | return false; |
245 | |
246 | if (skipLoop(L)) |
247 | return false; |
248 | |
249 | AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
250 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
251 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
252 | ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
253 | TargetLibraryInfo *TLI = |
254 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
255 | *L->getHeader()->getParent()); |
256 | const TargetTransformInfo *TTI = |
257 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( |
258 | *L->getHeader()->getParent()); |
259 | const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout(); |
260 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
261 | MemorySSA *MSSA = nullptr; |
262 | if (MSSAAnalysis) |
263 | MSSA = &MSSAAnalysis->getMSSA(); |
264 | |
265 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis |
266 | // pass. Function analyses need to be preserved across loop transformations |
267 | // but ORE cannot be preserved (see comment before the pass definition). |
268 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); |
269 | |
270 | LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE); |
271 | return LIR.runOnLoop(L); |
272 | } |
273 | |
274 | /// This transformation requires natural loop information & requires that |
275 | /// loop preheaders be inserted into the CFG. |
276 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
277 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
278 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
279 | AU.addPreserved<MemorySSAWrapperPass>(); |
280 | getLoopAnalysisUsage(AU); |
281 | } |
282 | }; |
283 | |
284 | } // end anonymous namespace |
285 | |
286 | char LoopIdiomRecognizeLegacyPass::ID = 0; |
287 | |
288 | PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, |
289 | LoopStandardAnalysisResults &AR, |
290 | LPMUpdater &) { |
291 | if (DisableLIRP::All) |
292 | return PreservedAnalyses::all(); |
293 | |
294 | const auto *DL = &L.getHeader()->getModule()->getDataLayout(); |
295 | |
296 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
297 | // pass. Function analyses need to be preserved across loop transformations |
298 | // but ORE cannot be preserved (see comment before the pass definition). |
299 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
300 | |
301 | LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, |
302 | AR.MSSA, DL, ORE); |
303 | if (!LIR.runOnLoop(&L)) |
304 | return PreservedAnalyses::all(); |
305 | |
306 | auto PA = getLoopPassPreservedAnalyses(); |
307 | if (AR.MSSA) |
308 | PA.preserve<MemorySSAAnalysis>(); |
309 | return PA; |
310 | } |
311 | |
312 | INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
313 | "Recognize loop idioms", false, false)static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
314 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); |
315 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
316 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); |
317 | INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
318 | "Recognize loop idioms", false, false)PassInfo *PI = new PassInfo( "Recognize loop idioms", "loop-idiom" , &LoopIdiomRecognizeLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<LoopIdiomRecognizeLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopIdiomRecognizeLegacyPassPassFlag ; void llvm::initializeLoopIdiomRecognizeLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopIdiomRecognizeLegacyPassPassFlag , initializeLoopIdiomRecognizeLegacyPassPassOnce, std::ref(Registry )); } |
319 | |
320 | Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); } |
321 | |
322 | static void deleteDeadInstruction(Instruction *I) { |
323 | I->replaceAllUsesWith(UndefValue::get(I->getType())); |
324 | I->eraseFromParent(); |
325 | } |
326 | |
327 | //===----------------------------------------------------------------------===// |
328 | // |
329 | // Implementation of LoopIdiomRecognize |
330 | // |
331 | //===----------------------------------------------------------------------===// |
332 | |
333 | bool LoopIdiomRecognize::runOnLoop(Loop *L) { |
334 | CurLoop = L; |
335 | // If the loop could not be converted to canonical form, it must have an |
336 | // indirectbr in it, just give up. |
337 | if (!L->getLoopPreheader()) |
338 | return false; |
339 | |
340 | // Disable loop idiom recognition if the function's name is a common idiom. |
341 | StringRef Name = L->getHeader()->getParent()->getName(); |
342 | if (Name == "memset" || Name == "memcpy") |
343 | return false; |
344 | |
345 | // Determine if code size heuristics need to be applied. |
346 | ApplyCodeSizeHeuristics = |
347 | L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; |
348 | |
349 | HasMemset = TLI->has(LibFunc_memset); |
350 | HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); |
351 | HasMemcpy = TLI->has(LibFunc_memcpy); |
352 | |
353 | if (HasMemset || HasMemsetPattern || HasMemcpy) |
354 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
355 | return runOnCountableLoop(); |
356 | |
357 | return runOnNoncountableLoop(); |
358 | } |
359 | |
360 | bool LoopIdiomRecognize::runOnCountableLoop() { |
361 | const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); |
362 | assert(!isa<SCEVCouldNotCompute>(BECount) &&((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)) |
363 | "runOnCountableLoop() called on a loop without a predictable"((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)) |
364 | "backedge-taken count")((!isa<SCEVCouldNotCompute>(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count") ? static_cast<void> (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BECount) && \"runOnCountableLoop() called on a loop without a predictable\" \"backedge-taken count\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 364, __PRETTY_FUNCTION__)); |
365 | |
366 | // If this loop executes exactly one time, then it should be peeled, not |
367 | // optimized by this pass. |
368 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
369 | if (BECst->getAPInt() == 0) |
370 | return false; |
371 | |
372 | SmallVector<BasicBlock *, 8> ExitBlocks; |
373 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
374 | |
375 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
376 | << CurLoop->getHeader()->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
377 | << "] Countable Loop %" << CurLoop->getHeader()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false) |
378 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Scanning: F[" << CurLoop->getHeader()->getParent()->getName () << "] Countable Loop %" << CurLoop->getHeader ()->getName() << "\n"; } } while (false); |
379 | |
380 | // The following transforms hoist stores/memsets into the loop pre-header. |
381 | // Give up if the loop has instructions that may throw. |
382 | SimpleLoopSafetyInfo SafetyInfo; |
383 | SafetyInfo.computeLoopSafetyInfo(CurLoop); |
384 | if (SafetyInfo.anyBlockMayThrow()) |
385 | return false; |
386 | |
387 | bool MadeChange = false; |
388 | |
389 | // Scan all the blocks in the loop that are not in subloops. |
390 | for (auto *BB : CurLoop->getBlocks()) { |
391 | // Ignore blocks in subloops. |
392 | if (LI->getLoopFor(BB) != CurLoop) |
393 | continue; |
394 | |
395 | MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); |
396 | } |
397 | return MadeChange; |
398 | } |
399 | |
400 | static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { |
401 | const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1)); |
402 | return ConstStride->getAPInt(); |
403 | } |
404 | |
405 | /// getMemSetPatternValue - If a strided store of the specified value is safe to |
406 | /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should |
407 | /// be passed in. Otherwise, return null. |
408 | /// |
409 | /// Note that we don't ever attempt to use memset_pattern8 or 4, because these |
410 | /// just replicate their input array and then pass on to memset_pattern16. |
411 | static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { |
412 | // FIXME: This could check for UndefValue because it can be merged into any |
413 | // other valid pattern. |
414 | |
415 | // If the value isn't a constant, we can't promote it to being in a constant |
416 | // array. We could theoretically do a store to an alloca or something, but |
417 | // that doesn't seem worthwhile. |
418 | Constant *C = dyn_cast<Constant>(V); |
419 | if (!C) |
420 | return nullptr; |
421 | |
422 | // Only handle simple values that are a power of two bytes in size. |
423 | uint64_t Size = DL->getTypeSizeInBits(V->getType()); |
424 | if (Size == 0 || (Size & 7) || (Size & (Size - 1))) |
425 | return nullptr; |
426 | |
427 | // Don't care enough about darwin/ppc to implement this. |
428 | if (DL->isBigEndian()) |
429 | return nullptr; |
430 | |
431 | // Convert to size in bytes. |
432 | Size /= 8; |
433 | |
434 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
435 | // if the top and bottom are the same (e.g. for vectors and large integers). |
436 | if (Size > 16) |
437 | return nullptr; |
438 | |
439 | // If the constant is exactly 16 bytes, just use it. |
440 | if (Size == 16) |
441 | return C; |
442 | |
443 | // Otherwise, we'll use an array of the constants. |
444 | unsigned ArraySize = 16 / Size; |
445 | ArrayType *AT = ArrayType::get(V->getType(), ArraySize); |
446 | return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C)); |
447 | } |
448 | |
449 | LoopIdiomRecognize::LegalStoreKind |
450 | LoopIdiomRecognize::isLegalStore(StoreInst *SI) { |
451 | // Don't touch volatile stores. |
452 | if (SI->isVolatile()) |
453 | return LegalStoreKind::None; |
454 | // We only want simple or unordered-atomic stores. |
455 | if (!SI->isUnordered()) |
456 | return LegalStoreKind::None; |
457 | |
458 | // Don't convert stores of non-integral pointer types to memsets (which stores |
459 | // integers). |
460 | if (DL->isNonIntegralPointerType(SI->getValueOperand()->getType())) |
461 | return LegalStoreKind::None; |
462 | |
463 | // Avoid merging nontemporal stores. |
464 | if (SI->getMetadata(LLVMContext::MD_nontemporal)) |
465 | return LegalStoreKind::None; |
466 | |
467 | Value *StoredVal = SI->getValueOperand(); |
468 | Value *StorePtr = SI->getPointerOperand(); |
469 | |
470 | // Reject stores that are so large that they overflow an unsigned. |
471 | uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); |
472 | if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) |
473 | return LegalStoreKind::None; |
474 | |
475 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
476 | // loop, which indicates a strided store. If we have something else, it's a |
477 | // random store we can't handle. |
478 | const SCEVAddRecExpr *StoreEv = |
479 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
480 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
481 | return LegalStoreKind::None; |
482 | |
483 | // Check to see if we have a constant stride. |
484 | if (!isa<SCEVConstant>(StoreEv->getOperand(1))) |
485 | return LegalStoreKind::None; |
486 | |
487 | // See if the store can be turned into a memset. |
488 | |
489 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
490 | // turned into a memset of i8 -1, assuming that all the consecutive bytes |
491 | // are stored. A store of i32 0x01020304 can never be turned into a memset, |
492 | // but it can be turned into memset_pattern if the target supports it. |
493 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
494 | Constant *PatternValue = nullptr; |
495 | |
496 | // Note: memset and memset_pattern on unordered-atomic is yet not supported |
497 | bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); |
498 | |
499 | // If we're allowed to form a memset, and the stored value would be |
500 | // acceptable for memset, use it. |
501 | if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && |
502 | // Verify that the stored value is loop invariant. If not, we can't |
503 | // promote the memset. |
504 | CurLoop->isLoopInvariant(SplatValue)) { |
505 | // It looks like we can use SplatValue. |
506 | return LegalStoreKind::Memset; |
507 | } else if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && |
508 | // Don't create memset_pattern16s with address spaces. |
509 | StorePtr->getType()->getPointerAddressSpace() == 0 && |
510 | (PatternValue = getMemSetPatternValue(StoredVal, DL))) { |
Although the value stored to 'PatternValue' is used in the enclosing expression, the value is never actually read from 'PatternValue' | |
511 | // It looks like we can use PatternValue! |
512 | return LegalStoreKind::MemsetPattern; |
513 | } |
514 | |
515 | // Otherwise, see if the store can be turned into a memcpy. |
516 | if (HasMemcpy && !DisableLIRP::Memcpy) { |
517 | // Check to see if the stride matches the size of the store. If so, then we |
518 | // know that every byte is touched in the loop. |
519 | APInt Stride = getStoreStride(StoreEv); |
520 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
521 | if (StoreSize != Stride && StoreSize != -Stride) |
522 | return LegalStoreKind::None; |
523 | |
524 | // The store must be feeding a non-volatile load. |
525 | LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); |
526 | |
527 | // Only allow non-volatile loads |
528 | if (!LI || LI->isVolatile()) |
529 | return LegalStoreKind::None; |
530 | // Only allow simple or unordered-atomic loads |
531 | if (!LI->isUnordered()) |
532 | return LegalStoreKind::None; |
533 | |
534 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
535 | // loop, which indicates a strided load. If we have something else, it's a |
536 | // random load we can't handle. |
537 | const SCEVAddRecExpr *LoadEv = |
538 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
539 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
540 | return LegalStoreKind::None; |
541 | |
542 | // The store and load must share the same stride. |
543 | if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) |
544 | return LegalStoreKind::None; |
545 | |
546 | // Success. This store can be converted into a memcpy. |
547 | UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); |
548 | return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy |
549 | : LegalStoreKind::Memcpy; |
550 | } |
551 | // This store can't be transformed into a memset/memcpy. |
552 | return LegalStoreKind::None; |
553 | } |
554 | |
555 | void LoopIdiomRecognize::collectStores(BasicBlock *BB) { |
556 | StoreRefsForMemset.clear(); |
557 | StoreRefsForMemsetPattern.clear(); |
558 | StoreRefsForMemcpy.clear(); |
559 | for (Instruction &I : *BB) { |
560 | StoreInst *SI = dyn_cast<StoreInst>(&I); |
561 | if (!SI) |
562 | continue; |
563 | |
564 | // Make sure this is a strided store with a constant stride. |
565 | switch (isLegalStore(SI)) { |
566 | case LegalStoreKind::None: |
567 | // Nothing to do |
568 | break; |
569 | case LegalStoreKind::Memset: { |
570 | // Find the base pointer. |
571 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
572 | StoreRefsForMemset[Ptr].push_back(SI); |
573 | } break; |
574 | case LegalStoreKind::MemsetPattern: { |
575 | // Find the base pointer. |
576 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
577 | StoreRefsForMemsetPattern[Ptr].push_back(SI); |
578 | } break; |
579 | case LegalStoreKind::Memcpy: |
580 | case LegalStoreKind::UnorderedAtomicMemcpy: |
581 | StoreRefsForMemcpy.push_back(SI); |
582 | break; |
583 | default: |
584 | assert(false && "unhandled return value")((false && "unhandled return value") ? static_cast< void> (0) : __assert_fail ("false && \"unhandled return value\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 584, __PRETTY_FUNCTION__)); |
585 | break; |
586 | } |
587 | } |
588 | } |
589 | |
590 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
591 | /// with the specified backedge count. This block is known to be in the current |
592 | /// loop and not in any subloops. |
593 | bool LoopIdiomRecognize::runOnLoopBlock( |
594 | BasicBlock *BB, const SCEV *BECount, |
595 | SmallVectorImpl<BasicBlock *> &ExitBlocks) { |
596 | // We can only promote stores in this block if they are unconditionally |
597 | // executed in the loop. For a block to be unconditionally executed, it has |
598 | // to dominate all the exit blocks of the loop. Verify this now. |
599 | for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) |
600 | if (!DT->dominates(BB, ExitBlocks[i])) |
601 | return false; |
602 | |
603 | bool MadeChange = false; |
604 | // Look for store instructions, which may be optimized to memset/memcpy. |
605 | collectStores(BB); |
606 | |
607 | // Look for a single store or sets of stores with a common base, which can be |
608 | // optimized into a memset (memset_pattern). The latter most commonly happens |
609 | // with structs and handunrolled loops. |
610 | for (auto &SL : StoreRefsForMemset) |
611 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); |
612 | |
613 | for (auto &SL : StoreRefsForMemsetPattern) |
614 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); |
615 | |
616 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
617 | for (auto &SI : StoreRefsForMemcpy) |
618 | MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); |
619 | |
620 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
621 | Instruction *Inst = &*I++; |
622 | // Look for memset instructions, which may be optimized to a larger memset. |
623 | if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) { |
624 | WeakTrackingVH InstPtr(&*I); |
625 | if (!processLoopMemSet(MSI, BECount)) |
626 | continue; |
627 | MadeChange = true; |
628 | |
629 | // If processing the memset invalidated our iterator, start over from the |
630 | // top of the block. |
631 | if (!InstPtr) |
632 | I = BB->begin(); |
633 | continue; |
634 | } |
635 | } |
636 | |
637 | return MadeChange; |
638 | } |
639 | |
640 | /// See if this store(s) can be promoted to a memset. |
641 | bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL, |
642 | const SCEV *BECount, ForMemset For) { |
643 | // Try to find consecutive stores that can be transformed into memsets. |
644 | SetVector<StoreInst *> Heads, Tails; |
645 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; |
646 | |
647 | // Do a quadratic search on all of the given stores and find |
648 | // all of the pairs of stores that follow each other. |
649 | SmallVector<unsigned, 16> IndexQueue; |
650 | for (unsigned i = 0, e = SL.size(); i < e; ++i) { |
651 | assert(SL[i]->isSimple() && "Expected only non-volatile stores.")((SL[i]->isSimple() && "Expected only non-volatile stores." ) ? static_cast<void> (0) : __assert_fail ("SL[i]->isSimple() && \"Expected only non-volatile stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 651, __PRETTY_FUNCTION__)); |
652 | |
653 | Value *FirstStoredVal = SL[i]->getValueOperand(); |
654 | Value *FirstStorePtr = SL[i]->getPointerOperand(); |
655 | const SCEVAddRecExpr *FirstStoreEv = |
656 | cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr)); |
657 | APInt FirstStride = getStoreStride(FirstStoreEv); |
658 | unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); |
659 | |
660 | // See if we can optimize just this store in isolation. |
661 | if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { |
662 | Heads.insert(SL[i]); |
663 | continue; |
664 | } |
665 | |
666 | Value *FirstSplatValue = nullptr; |
667 | Constant *FirstPatternValue = nullptr; |
668 | |
669 | if (For == ForMemset::Yes) |
670 | FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); |
671 | else |
672 | FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); |
673 | |
674 | assert((FirstSplatValue || FirstPatternValue) &&(((FirstSplatValue || FirstPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 675, __PRETTY_FUNCTION__)) |
675 | "Expected either splat value or pattern value.")(((FirstSplatValue || FirstPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(FirstSplatValue || FirstPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 675, __PRETTY_FUNCTION__)); |
676 | |
677 | IndexQueue.clear(); |
678 | // If a store has multiple consecutive store candidates, search Stores |
679 | // array according to the sequence: from i+1 to e, then from i-1 to 0. |
680 | // This is because usually pairing with immediate succeeding or preceding |
681 | // candidate create the best chance to find memset opportunity. |
682 | unsigned j = 0; |
683 | for (j = i + 1; j < e; ++j) |
684 | IndexQueue.push_back(j); |
685 | for (j = i; j > 0; --j) |
686 | IndexQueue.push_back(j - 1); |
687 | |
688 | for (auto &k : IndexQueue) { |
689 | assert(SL[k]->isSimple() && "Expected only non-volatile stores.")((SL[k]->isSimple() && "Expected only non-volatile stores." ) ? static_cast<void> (0) : __assert_fail ("SL[k]->isSimple() && \"Expected only non-volatile stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 689, __PRETTY_FUNCTION__)); |
690 | Value *SecondStorePtr = SL[k]->getPointerOperand(); |
691 | const SCEVAddRecExpr *SecondStoreEv = |
692 | cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr)); |
693 | APInt SecondStride = getStoreStride(SecondStoreEv); |
694 | |
695 | if (FirstStride != SecondStride) |
696 | continue; |
697 | |
698 | Value *SecondStoredVal = SL[k]->getValueOperand(); |
699 | Value *SecondSplatValue = nullptr; |
700 | Constant *SecondPatternValue = nullptr; |
701 | |
702 | if (For == ForMemset::Yes) |
703 | SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); |
704 | else |
705 | SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); |
706 | |
707 | assert((SecondSplatValue || SecondPatternValue) &&(((SecondSplatValue || SecondPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 708, __PRETTY_FUNCTION__)) |
708 | "Expected either splat value or pattern value.")(((SecondSplatValue || SecondPatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SecondSplatValue || SecondPatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 708, __PRETTY_FUNCTION__)); |
709 | |
710 | if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { |
711 | if (For == ForMemset::Yes) { |
712 | if (isa<UndefValue>(FirstSplatValue)) |
713 | FirstSplatValue = SecondSplatValue; |
714 | if (FirstSplatValue != SecondSplatValue) |
715 | continue; |
716 | } else { |
717 | if (isa<UndefValue>(FirstPatternValue)) |
718 | FirstPatternValue = SecondPatternValue; |
719 | if (FirstPatternValue != SecondPatternValue) |
720 | continue; |
721 | } |
722 | Tails.insert(SL[k]); |
723 | Heads.insert(SL[i]); |
724 | ConsecutiveChain[SL[i]] = SL[k]; |
725 | break; |
726 | } |
727 | } |
728 | } |
729 | |
730 | // We may run into multiple chains that merge into a single chain. We mark the |
731 | // stores that we transformed so that we don't visit the same store twice. |
732 | SmallPtrSet<Value *, 16> TransformedStores; |
733 | bool Changed = false; |
734 | |
735 | // For stores that start but don't end a link in the chain: |
736 | for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end(); |
737 | it != e; ++it) { |
738 | if (Tails.count(*it)) |
739 | continue; |
740 | |
741 | // We found a store instr that starts a chain. Now follow the chain and try |
742 | // to transform it. |
743 | SmallPtrSet<Instruction *, 8> AdjacentStores; |
744 | StoreInst *I = *it; |
745 | |
746 | StoreInst *HeadStore = I; |
747 | unsigned StoreSize = 0; |
748 | |
749 | // Collect the chain into a list. |
750 | while (Tails.count(I) || Heads.count(I)) { |
751 | if (TransformedStores.count(I)) |
752 | break; |
753 | AdjacentStores.insert(I); |
754 | |
755 | StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); |
756 | // Move to the next value in the chain. |
757 | I = ConsecutiveChain[I]; |
758 | } |
759 | |
760 | Value *StoredVal = HeadStore->getValueOperand(); |
761 | Value *StorePtr = HeadStore->getPointerOperand(); |
762 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
763 | APInt Stride = getStoreStride(StoreEv); |
764 | |
765 | // Check to see if the stride matches the size of the stores. If so, then |
766 | // we know that every byte is touched in the loop. |
767 | if (StoreSize != Stride && StoreSize != -Stride) |
768 | continue; |
769 | |
770 | bool NegStride = StoreSize == -Stride; |
771 | |
772 | if (processLoopStridedStore(StorePtr, StoreSize, |
773 | MaybeAlign(HeadStore->getAlignment()), |
774 | StoredVal, HeadStore, AdjacentStores, StoreEv, |
775 | BECount, NegStride)) { |
776 | TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); |
777 | Changed = true; |
778 | } |
779 | } |
780 | |
781 | return Changed; |
782 | } |
783 | |
784 | /// processLoopMemSet - See if this memset can be promoted to a large memset. |
785 | bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, |
786 | const SCEV *BECount) { |
787 | // We can only handle non-volatile memsets with a constant size. |
788 | if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) |
789 | return false; |
790 | |
791 | // If we're not allowed to hack on memset, we fail. |
792 | if (!HasMemset) |
793 | return false; |
794 | |
795 | Value *Pointer = MSI->getDest(); |
796 | |
797 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
798 | // loop, which indicates a strided store. If we have something else, it's a |
799 | // random store we can't handle. |
800 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer)); |
801 | if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine()) |
802 | return false; |
803 | |
804 | // Reject memsets that are so large that they overflow an unsigned. |
805 | uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); |
806 | if ((SizeInBytes >> 32) != 0) |
807 | return false; |
808 | |
809 | // Check to see if the stride matches the size of the memset. If so, then we |
810 | // know that every byte is touched in the loop. |
811 | const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
812 | if (!ConstStride) |
813 | return false; |
814 | |
815 | APInt Stride = ConstStride->getAPInt(); |
816 | if (SizeInBytes != Stride && SizeInBytes != -Stride) |
817 | return false; |
818 | |
819 | // Verify that the memset value is loop invariant. If not, we can't promote |
820 | // the memset. |
821 | Value *SplatValue = MSI->getValue(); |
822 | if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) |
823 | return false; |
824 | |
825 | SmallPtrSet<Instruction *, 1> MSIs; |
826 | MSIs.insert(MSI); |
827 | bool NegStride = SizeInBytes == -Stride; |
828 | return processLoopStridedStore( |
829 | Pointer, (unsigned)SizeInBytes, MaybeAlign(MSI->getDestAlignment()), |
830 | SplatValue, MSI, MSIs, Ev, BECount, NegStride, /*IsLoopMemset=*/true); |
831 | } |
832 | |
833 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
834 | /// specified pointer location, which is a loop-strided access. The 'Access' |
835 | /// argument specifies what the verboten forms of access are (read or write). |
836 | static bool |
837 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
838 | const SCEV *BECount, unsigned StoreSize, |
839 | AliasAnalysis &AA, |
840 | SmallPtrSetImpl<Instruction *> &IgnoredStores) { |
841 | // Get the location that may be stored across the loop. Since the access is |
842 | // strided positively through memory, we say that the modified location starts |
843 | // at the pointer and has infinite size. |
844 | LocationSize AccessSize = LocationSize::unknown(); |
845 | |
846 | // If the loop iterates a fixed number of times, we can refine the access size |
847 | // to be exactly the size of the memset, which is (BECount+1)*StoreSize |
848 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
849 | AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * |
850 | StoreSize); |
851 | |
852 | // TODO: For this to be really effective, we have to dive into the pointer |
853 | // operand in the store. Store to &A[i] of 100 will always return may alias |
854 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
855 | // which will then no-alias a store to &A[100]. |
856 | MemoryLocation StoreLoc(Ptr, AccessSize); |
857 | |
858 | for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; |
859 | ++BI) |
860 | for (Instruction &I : **BI) |
861 | if (IgnoredStores.count(&I) == 0 && |
862 | isModOrRefSet( |
863 | intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access))) |
864 | return true; |
865 | |
866 | return false; |
867 | } |
868 | |
869 | // If we have a negative stride, Start refers to the end of the memory location |
870 | // we're trying to memset. Therefore, we need to recompute the base pointer, |
871 | // which is just Start - BECount*Size. |
872 | static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, |
873 | Type *IntPtr, unsigned StoreSize, |
874 | ScalarEvolution *SE) { |
875 | const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); |
876 | if (StoreSize != 1) |
877 | Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize), |
878 | SCEV::FlagNUW); |
879 | return SE->getMinusSCEV(Start, Index); |
880 | } |
881 | |
882 | /// Compute the number of bytes as a SCEV from the backedge taken count. |
883 | /// |
884 | /// This also maps the SCEV into the provided type and tries to handle the |
885 | /// computation in a way that will fold cleanly. |
886 | static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, |
887 | unsigned StoreSize, Loop *CurLoop, |
888 | const DataLayout *DL, ScalarEvolution *SE) { |
889 | const SCEV *NumBytesS; |
890 | // The # stored bytes is (BECount+1)*Size. Expand the trip count out to |
891 | // pointer size if it isn't already. |
892 | // |
893 | // If we're going to need to zero extend the BE count, check if we can add |
894 | // one to it prior to zero extending without overflow. Provided this is safe, |
895 | // it allows better simplification of the +1. |
896 | if (DL->getTypeSizeInBits(BECount->getType()) < |
897 | DL->getTypeSizeInBits(IntPtr) && |
898 | SE->isLoopEntryGuardedByCond( |
899 | CurLoop, ICmpInst::ICMP_NE, BECount, |
900 | SE->getNegativeSCEV(SE->getOne(BECount->getType())))) { |
901 | NumBytesS = SE->getZeroExtendExpr( |
902 | SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW), |
903 | IntPtr); |
904 | } else { |
905 | NumBytesS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr), |
906 | SE->getOne(IntPtr), SCEV::FlagNUW); |
907 | } |
908 | |
909 | // And scale it based on the store size. |
910 | if (StoreSize != 1) { |
911 | NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), |
912 | SCEV::FlagNUW); |
913 | } |
914 | return NumBytesS; |
915 | } |
916 | |
917 | /// processLoopStridedStore - We see a strided store of some value. If we can |
918 | /// transform this into a memset or memset_pattern in the loop preheader, do so. |
919 | bool LoopIdiomRecognize::processLoopStridedStore( |
920 | Value *DestPtr, unsigned StoreSize, MaybeAlign StoreAlignment, |
921 | Value *StoredVal, Instruction *TheStore, |
922 | SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev, |
923 | const SCEV *BECount, bool NegStride, bool IsLoopMemset) { |
924 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
925 | Constant *PatternValue = nullptr; |
926 | |
927 | if (!SplatValue) |
928 | PatternValue = getMemSetPatternValue(StoredVal, DL); |
929 | |
930 | assert((SplatValue || PatternValue) &&(((SplatValue || PatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 931, __PRETTY_FUNCTION__)) |
931 | "Expected either splat value or pattern value.")(((SplatValue || PatternValue) && "Expected either splat value or pattern value." ) ? static_cast<void> (0) : __assert_fail ("(SplatValue || PatternValue) && \"Expected either splat value or pattern value.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 931, __PRETTY_FUNCTION__)); |
932 | |
933 | // The trip count of the loop and the base pointer of the addrec SCEV is |
934 | // guaranteed to be loop invariant, which means that it should dominate the |
935 | // header. This allows us to insert code for it in the preheader. |
936 | unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); |
937 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
938 | IRBuilder<> Builder(Preheader->getTerminator()); |
939 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
940 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
941 | |
942 | Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); |
943 | Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); |
944 | |
945 | bool Changed = false; |
946 | const SCEV *Start = Ev->getStart(); |
947 | // Handle negative strided loops. |
948 | if (NegStride) |
949 | Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSize, SE); |
950 | |
951 | // TODO: ideally we should still be able to generate memset if SCEV expander |
952 | // is taught to generate the dependencies at the latest point. |
953 | if (!isSafeToExpand(Start, *SE)) |
954 | return Changed; |
955 | |
956 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
957 | // this into a memset in the loop preheader now if we want. However, this |
958 | // would be unsafe to do if there is anything else in the loop that may read |
959 | // or write to the aliased location. Check for any overlap by generating the |
960 | // base pointer and checking the region. |
961 | Value *BasePtr = |
962 | Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); |
963 | |
964 | // From here on out, conservatively report to the pass manager that we've |
965 | // changed the IR, even if we later clean up these added instructions. There |
966 | // may be structural differences e.g. in the order of use lists not accounted |
967 | // for in just a textual dump of the IR. This is written as a variable, even |
968 | // though statically all the places this dominates could be replaced with |
969 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
970 | // the return value will read this comment, and leave them alone. |
971 | Changed = true; |
972 | |
973 | if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
974 | StoreSize, *AA, Stores)) |
975 | return Changed; |
976 | |
977 | if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) |
978 | return Changed; |
979 | |
980 | // Okay, everything looks good, insert the memset. |
981 | |
982 | const SCEV *NumBytesS = |
983 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
984 | |
985 | // TODO: ideally we should still be able to generate memset if SCEV expander |
986 | // is taught to generate the dependencies at the latest point. |
987 | if (!isSafeToExpand(NumBytesS, *SE)) |
988 | return Changed; |
989 | |
990 | Value *NumBytes = |
991 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
992 | |
993 | CallInst *NewCall; |
994 | if (SplatValue) { |
995 | NewCall = Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, |
996 | MaybeAlign(StoreAlignment)); |
997 | } else { |
998 | // Everything is emitted in default address space |
999 | Type *Int8PtrTy = DestInt8PtrTy; |
1000 | |
1001 | Module *M = TheStore->getModule(); |
1002 | StringRef FuncName = "memset_pattern16"; |
1003 | FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(), |
1004 | Int8PtrTy, Int8PtrTy, IntIdxTy); |
1005 | inferLibFuncAttributes(M, FuncName, *TLI); |
1006 | |
1007 | // Otherwise we should form a memset_pattern16. PatternValue is known to be |
1008 | // an constant array of 16-bytes. Plop the value into a mergable global. |
1009 | GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, |
1010 | GlobalValue::PrivateLinkage, |
1011 | PatternValue, ".memset_pattern"); |
1012 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
1013 | GV->setAlignment(Align(16)); |
1014 | Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); |
1015 | NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); |
1016 | } |
1017 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
1018 | |
1019 | if (MSSAU) { |
1020 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1021 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1022 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1023 | } |
1024 | |
1025 | 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) |
1026 | << " 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) |
1027 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false); |
1028 | |
1029 | ORE.emit([&]() { |
1030 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStridedStore", |
1031 | NewCall->getDebugLoc(), Preheader) |
1032 | << "Transformed loop-strided store into a call to " |
1033 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1034 | << "() function"; |
1035 | }); |
1036 | |
1037 | // Okay, the memset has been formed. Zap the original store and anything that |
1038 | // feeds into it. |
1039 | for (auto *I : Stores) { |
1040 | if (MSSAU) |
1041 | MSSAU->removeMemoryAccess(I, true); |
1042 | deleteDeadInstruction(I); |
1043 | } |
1044 | if (MSSAU && VerifyMemorySSA) |
1045 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1046 | ++NumMemSet; |
1047 | ExpCleaner.markResultUsed(); |
1048 | return true; |
1049 | } |
1050 | |
1051 | /// If the stored value is a strided load in the same loop with the same stride |
1052 | /// this may be transformable into a memcpy. This kicks in for stuff like |
1053 | /// for (i) A[i] = B[i]; |
1054 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, |
1055 | const SCEV *BECount) { |
1056 | assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.")((SI->isUnordered() && "Expected only non-volatile non-ordered stores." ) ? static_cast<void> (0) : __assert_fail ("SI->isUnordered() && \"Expected only non-volatile non-ordered stores.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1056, __PRETTY_FUNCTION__)); |
1057 | |
1058 | Value *StorePtr = SI->getPointerOperand(); |
1059 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
1060 | APInt Stride = getStoreStride(StoreEv); |
1061 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
1062 | bool NegStride = StoreSize == -Stride; |
1063 | |
1064 | // The store must be feeding a non-volatile load. |
1065 | LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); |
1066 | assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.")((LI->isUnordered() && "Expected only non-volatile non-ordered loads." ) ? static_cast<void> (0) : __assert_fail ("LI->isUnordered() && \"Expected only non-volatile non-ordered loads.\"" , "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1066, __PRETTY_FUNCTION__)); |
1067 | |
1068 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
1069 | // loop, which indicates a strided load. If we have something else, it's a |
1070 | // random load we can't handle. |
1071 | const SCEVAddRecExpr *LoadEv = |
1072 | cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
1073 | |
1074 | // The trip count of the loop and the base pointer of the addrec SCEV is |
1075 | // guaranteed to be loop invariant, which means that it should dominate the |
1076 | // header. This allows us to insert code for it in the preheader. |
1077 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
1078 | IRBuilder<> Builder(Preheader->getTerminator()); |
1079 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
1080 | |
1081 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
1082 | |
1083 | bool Changed = false; |
1084 | const SCEV *StrStart = StoreEv->getStart(); |
1085 | unsigned StrAS = SI->getPointerAddressSpace(); |
1086 | Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); |
1087 | |
1088 | // Handle negative strided loops. |
1089 | if (NegStride) |
1090 | StrStart = getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSize, SE); |
1091 | |
1092 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
1093 | // this into a memcpy in the loop preheader now if we want. However, this |
1094 | // would be unsafe to do if there is anything else in the loop that may read |
1095 | // or write the memory region we're storing to. This includes the load that |
1096 | // feeds the stores. Check for an alias by generating the base address and |
1097 | // checking everything. |
1098 | Value *StoreBasePtr = Expander.expandCodeFor( |
1099 | StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); |
1100 | |
1101 | // From here on out, conservatively report to the pass manager that we've |
1102 | // changed the IR, even if we later clean up these added instructions. There |
1103 | // may be structural differences e.g. in the order of use lists not accounted |
1104 | // for in just a textual dump of the IR. This is written as a variable, even |
1105 | // though statically all the places this dominates could be replaced with |
1106 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
1107 | // the return value will read this comment, and leave them alone. |
1108 | Changed = true; |
1109 | |
1110 | SmallPtrSet<Instruction *, 1> Stores; |
1111 | Stores.insert(SI); |
1112 | if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
1113 | StoreSize, *AA, Stores)) |
1114 | return Changed; |
1115 | |
1116 | const SCEV *LdStart = LoadEv->getStart(); |
1117 | unsigned LdAS = LI->getPointerAddressSpace(); |
1118 | |
1119 | // Handle negative strided loops. |
1120 | if (NegStride) |
1121 | LdStart = getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSize, SE); |
1122 | |
1123 | // For a memcpy, we have to make sure that the input array is not being |
1124 | // mutated by the loop. |
1125 | Value *LoadBasePtr = Expander.expandCodeFor( |
1126 | LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); |
1127 | |
1128 | if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, |
1129 | StoreSize, *AA, Stores)) |
1130 | return Changed; |
1131 | |
1132 | if (avoidLIRForMultiBlockLoop()) |
1133 | return Changed; |
1134 | |
1135 | // Okay, everything is safe, we can transform this! |
1136 | |
1137 | const SCEV *NumBytesS = |
1138 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
1139 | |
1140 | Value *NumBytes = |
1141 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
1142 | |
1143 | CallInst *NewCall = nullptr; |
1144 | // Check whether to generate an unordered atomic memcpy: |
1145 | // If the load or store are atomic, then they must necessarily be unordered |
1146 | // by previous checks. |
1147 | if (!SI->isAtomic() && !LI->isAtomic()) |
1148 | NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr, |
1149 | LI->getAlign(), NumBytes); |
1150 | else { |
1151 | // We cannot allow unaligned ops for unordered load/store, so reject |
1152 | // anything where the alignment isn't at least the element size. |
1153 | const Align StoreAlign = SI->getAlign(); |
1154 | const Align LoadAlign = LI->getAlign(); |
1155 | if (StoreAlign < StoreSize || LoadAlign < StoreSize) |
1156 | return Changed; |
1157 | |
1158 | // If the element.atomic memcpy is not lowered into explicit |
1159 | // loads/stores later, then it will be lowered into an element-size |
1160 | // specific lib call. If the lib call doesn't exist for our store size, then |
1161 | // we shouldn't generate the memcpy. |
1162 | if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) |
1163 | return Changed; |
1164 | |
1165 | // Create the call. |
1166 | // Note that unordered atomic loads/stores are *required* by the spec to |
1167 | // have an alignment but non-atomic loads/stores may not. |
1168 | NewCall = Builder.CreateElementUnorderedAtomicMemCpy( |
1169 | StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, |
1170 | StoreSize); |
1171 | } |
1172 | NewCall->setDebugLoc(SI->getDebugLoc()); |
1173 | |
1174 | if (MSSAU) { |
1175 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1176 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1177 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1178 | } |
1179 | |
1180 | LLVM_DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
1181 | << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
1182 | << " from store ptr=" << *StoreEv << " at: " << *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false) |
1183 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"; } } while (false); |
1184 | |
1185 | ORE.emit([&]() { |
1186 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStoreOfLoopLoad", |
1187 | NewCall->getDebugLoc(), Preheader) |
1188 | << "Formed a call to " |
1189 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1190 | << "() function"; |
1191 | }); |
1192 | |
1193 | // Okay, the memcpy has been formed. Zap the original store and anything that |
1194 | // feeds into it. |
1195 | if (MSSAU) |
1196 | MSSAU->removeMemoryAccess(SI, true); |
1197 | deleteDeadInstruction(SI); |
1198 | if (MSSAU && VerifyMemorySSA) |
1199 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1200 | ++NumMemCpy; |
1201 | ExpCleaner.markResultUsed(); |
1202 | return true; |
1203 | } |
1204 | |
1205 | // When compiling for codesize we avoid idiom recognition for a multi-block loop |
1206 | // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. |
1207 | // |
1208 | bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, |
1209 | bool IsLoopMemset) { |
1210 | if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { |
1211 | if (!CurLoop->getParentLoop() && (!IsMemset || !IsLoopMemset)) { |
1212 | 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) |
1213 | << " : 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) |
1214 | << " 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); |
1215 | return true; |
1216 | } |
1217 | } |
1218 | |
1219 | return false; |
1220 | } |
1221 | |
1222 | bool LoopIdiomRecognize::runOnNoncountableLoop() { |
1223 | 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) |
1224 | << 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) |
1225 | << "] 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) |
1226 | << 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); |
1227 | |
1228 | return recognizePopcount() || recognizeAndInsertFFS(); |
1229 | } |
1230 | |
1231 | /// Check if the given conditional branch is based on the comparison between |
1232 | /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is |
1233 | /// true), the control yields to the loop entry. If the branch matches the |
1234 | /// behavior, the variable involved in the comparison is returned. This function |
1235 | /// will be called to see if the precondition and postcondition of the loop are |
1236 | /// in desirable form. |
1237 | static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, |
1238 | bool JmpOnZero = false) { |
1239 | if (!BI || !BI->isConditional()) |
1240 | return nullptr; |
1241 | |
1242 | ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); |
1243 | if (!Cond) |
1244 | return nullptr; |
1245 | |
1246 | ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1)); |
1247 | if (!CmpZero || !CmpZero->isZero()) |
1248 | return nullptr; |
1249 | |
1250 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
1251 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
1252 | if (JmpOnZero) |
1253 | std::swap(TrueSucc, FalseSucc); |
1254 | |
1255 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
1256 | if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || |
1257 | (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) |
1258 | return Cond->getOperand(0); |
1259 | |
1260 | return nullptr; |
1261 | } |
1262 | |
1263 | // Check if the recurrence variable `VarX` is in the right form to create |
1264 | // the idiom. Returns the value coerced to a PHINode if so. |
1265 | static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, |
1266 | BasicBlock *LoopEntry) { |
1267 | auto *PhiX = dyn_cast<PHINode>(VarX); |
1268 | if (PhiX && PhiX->getParent() == LoopEntry && |
1269 | (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) |
1270 | return PhiX; |
1271 | return nullptr; |
1272 | } |
1273 | |
1274 | /// Return true iff the idiom is detected in the loop. |
1275 | /// |
1276 | /// Additionally: |
1277 | /// 1) \p CntInst is set to the instruction counting the population bit. |
1278 | /// 2) \p CntPhi is set to the corresponding phi node. |
1279 | /// 3) \p Var is set to the value whose population bits are being counted. |
1280 | /// |
1281 | /// The core idiom we are trying to detect is: |
1282 | /// \code |
1283 | /// if (x0 != 0) |
1284 | /// goto loop-exit // the precondition of the loop |
1285 | /// cnt0 = init-val; |
1286 | /// do { |
1287 | /// x1 = phi (x0, x2); |
1288 | /// cnt1 = phi(cnt0, cnt2); |
1289 | /// |
1290 | /// cnt2 = cnt1 + 1; |
1291 | /// ... |
1292 | /// x2 = x1 & (x1 - 1); |
1293 | /// ... |
1294 | /// } while(x != 0); |
1295 | /// |
1296 | /// loop-exit: |
1297 | /// \endcode |
1298 | static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, |
1299 | Instruction *&CntInst, PHINode *&CntPhi, |
1300 | Value *&Var) { |
1301 | // step 1: Check to see if the look-back branch match this pattern: |
1302 | // "if (a!=0) goto loop-entry". |
1303 | BasicBlock *LoopEntry; |
1304 | Instruction *DefX2, *CountInst; |
1305 | Value *VarX1, *VarX0; |
1306 | PHINode *PhiX, *CountPhi; |
1307 | |
1308 | DefX2 = CountInst = nullptr; |
1309 | VarX1 = VarX0 = nullptr; |
1310 | PhiX = CountPhi = nullptr; |
1311 | LoopEntry = *(CurLoop->block_begin()); |
1312 | |
1313 | // step 1: Check if the loop-back branch is in desirable form. |
1314 | { |
1315 | if (Value *T = matchCondition( |
1316 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1317 | DefX2 = dyn_cast<Instruction>(T); |
1318 | else |
1319 | return false; |
1320 | } |
1321 | |
1322 | // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" |
1323 | { |
1324 | if (!DefX2 || DefX2->getOpcode() != Instruction::And) |
1325 | return false; |
1326 | |
1327 | BinaryOperator *SubOneOp; |
1328 | |
1329 | if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0)))) |
1330 | VarX1 = DefX2->getOperand(1); |
1331 | else { |
1332 | VarX1 = DefX2->getOperand(0); |
1333 | SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1)); |
1334 | } |
1335 | if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) |
1336 | return false; |
1337 | |
1338 | ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1)); |
1339 | if (!Dec || |
1340 | !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || |
1341 | (SubOneOp->getOpcode() == Instruction::Add && |
1342 | Dec->isMinusOne()))) { |
1343 | return false; |
1344 | } |
1345 | } |
1346 | |
1347 | // step 3: Check the recurrence of variable X |
1348 | PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); |
1349 | if (!PhiX) |
1350 | return false; |
1351 | |
1352 | // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 |
1353 | { |
1354 | CountInst = nullptr; |
1355 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
1356 | IterE = LoopEntry->end(); |
1357 | Iter != IterE; Iter++) { |
1358 | Instruction *Inst = &*Iter; |
1359 | if (Inst->getOpcode() != Instruction::Add) |
1360 | continue; |
1361 | |
1362 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
1363 | if (!Inc || !Inc->isOne()) |
1364 | continue; |
1365 | |
1366 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
1367 | if (!Phi) |
1368 | continue; |
1369 | |
1370 | // Check if the result of the instruction is live of the loop. |
1371 | bool LiveOutLoop = false; |
1372 | for (User *U : Inst->users()) { |
1373 | if ((cast<Instruction>(U))->getParent() != LoopEntry) { |
1374 | LiveOutLoop = true; |
1375 | break; |
1376 | } |
1377 | } |
1378 | |
1379 | if (LiveOutLoop) { |
1380 | CountInst = Inst; |
1381 | CountPhi = Phi; |
1382 | break; |
1383 | } |
1384 | } |
1385 | |
1386 | if (!CountInst) |
1387 | return false; |
1388 | } |
1389 | |
1390 | // step 5: check if the precondition is in this form: |
1391 | // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" |
1392 | { |
1393 | auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1394 | Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); |
1395 | if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) |
1396 | return false; |
1397 | |
1398 | CntInst = CountInst; |
1399 | CntPhi = CountPhi; |
1400 | Var = T; |
1401 | } |
1402 | |
1403 | return true; |
1404 | } |
1405 | |
1406 | /// Return true if the idiom is detected in the loop. |
1407 | /// |
1408 | /// Additionally: |
1409 | /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) |
1410 | /// or nullptr if there is no such. |
1411 | /// 2) \p CntPhi is set to the corresponding phi node |
1412 | /// or nullptr if there is no such. |
1413 | /// 3) \p Var is set to the value whose CTLZ could be used. |
1414 | /// 4) \p DefX is set to the instruction calculating Loop exit condition. |
1415 | /// |
1416 | /// The core idiom we are trying to detect is: |
1417 | /// \code |
1418 | /// if (x0 == 0) |
1419 | /// goto loop-exit // the precondition of the loop |
1420 | /// cnt0 = init-val; |
1421 | /// do { |
1422 | /// x = phi (x0, x.next); //PhiX |
1423 | /// cnt = phi(cnt0, cnt.next); |
1424 | /// |
1425 | /// cnt.next = cnt + 1; |
1426 | /// ... |
1427 | /// x.next = x >> 1; // DefX |
1428 | /// ... |
1429 | /// } while(x.next != 0); |
1430 | /// |
1431 | /// loop-exit: |
1432 | /// \endcode |
1433 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, |
1434 | Intrinsic::ID &IntrinID, Value *&InitX, |
1435 | Instruction *&CntInst, PHINode *&CntPhi, |
1436 | Instruction *&DefX) { |
1437 | BasicBlock *LoopEntry; |
1438 | Value *VarX = nullptr; |
1439 | |
1440 | DefX = nullptr; |
1441 | CntInst = nullptr; |
1442 | CntPhi = nullptr; |
1443 | LoopEntry = *(CurLoop->block_begin()); |
1444 | |
1445 | // step 1: Check if the loop-back branch is in desirable form. |
1446 | if (Value *T = matchCondition( |
1447 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1448 | DefX = dyn_cast<Instruction>(T); |
1449 | else |
1450 | return false; |
1451 | |
1452 | // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" |
1453 | if (!DefX || !DefX->isShift()) |
1454 | return false; |
1455 | IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : |
1456 | Intrinsic::ctlz; |
1457 | ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)); |
1458 | if (!Shft || !Shft->isOne()) |
1459 | return false; |
1460 | VarX = DefX->getOperand(0); |
1461 | |
1462 | // step 3: Check the recurrence of variable X |
1463 | PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); |
1464 | if (!PhiX) |
1465 | return false; |
1466 | |
1467 | InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); |
1468 | |
1469 | // Make sure the initial value can't be negative otherwise the ashr in the |
1470 | // loop might never reach zero which would make the loop infinite. |
1471 | if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) |
1472 | return false; |
1473 | |
1474 | // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 |
1475 | // TODO: We can skip the step. If loop trip count is known (CTLZ), |
1476 | // then all uses of "cnt.next" could be optimized to the trip count |
1477 | // plus "cnt0". Currently it is not optimized. |
1478 | // This step could be used to detect POPCNT instruction: |
1479 | // cnt.next = cnt + (x.next & 1) |
1480 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
1481 | IterE = LoopEntry->end(); |
1482 | Iter != IterE; Iter++) { |
1483 | Instruction *Inst = &*Iter; |
1484 | if (Inst->getOpcode() != Instruction::Add) |
1485 | continue; |
1486 | |
1487 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
1488 | if (!Inc || !Inc->isOne()) |
1489 | continue; |
1490 | |
1491 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
1492 | if (!Phi) |
1493 | continue; |
1494 | |
1495 | CntInst = Inst; |
1496 | CntPhi = Phi; |
1497 | break; |
1498 | } |
1499 | if (!CntInst) |
1500 | return false; |
1501 | |
1502 | return true; |
1503 | } |
1504 | |
1505 | /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop |
1506 | /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new |
1507 | /// trip count returns true; otherwise, returns false. |
1508 | bool LoopIdiomRecognize::recognizeAndInsertFFS() { |
1509 | // Give up if the loop has multiple blocks or multiple backedges. |
1510 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1511 | return false; |
1512 | |
1513 | Intrinsic::ID IntrinID; |
1514 | Value *InitX; |
1515 | Instruction *DefX = nullptr; |
1516 | PHINode *CntPhi = nullptr; |
1517 | Instruction *CntInst = nullptr; |
1518 | // Help decide if transformation is profitable. For ShiftUntilZero idiom, |
1519 | // this is always 6. |
1520 | size_t IdiomCanonicalSize = 6; |
1521 | |
1522 | if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, |
1523 | CntInst, CntPhi, DefX)) |
1524 | return false; |
1525 | |
1526 | bool IsCntPhiUsedOutsideLoop = false; |
1527 | for (User *U : CntPhi->users()) |
1528 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1529 | IsCntPhiUsedOutsideLoop = true; |
1530 | break; |
1531 | } |
1532 | bool IsCntInstUsedOutsideLoop = false; |
1533 | for (User *U : CntInst->users()) |
1534 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1535 | IsCntInstUsedOutsideLoop = true; |
1536 | break; |
1537 | } |
1538 | // If both CntInst and CntPhi are used outside the loop the profitability |
1539 | // is questionable. |
1540 | if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) |
1541 | return false; |
1542 | |
1543 | // For some CPUs result of CTLZ(X) intrinsic is undefined |
1544 | // when X is 0. If we can not guarantee X != 0, we need to check this |
1545 | // when expand. |
1546 | bool ZeroCheck = false; |
1547 | // It is safe to assume Preheader exist as it was checked in |
1548 | // parent function RunOnLoop. |
1549 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1550 | |
1551 | // If we are using the count instruction outside the loop, make sure we |
1552 | // have a zero check as a precondition. Without the check the loop would run |
1553 | // one iteration for before any check of the input value. This means 0 and 1 |
1554 | // would have identical behavior in the original loop and thus |
1555 | if (!IsCntPhiUsedOutsideLoop) { |
1556 | auto *PreCondBB = PH->getSinglePredecessor(); |
1557 | if (!PreCondBB) |
1558 | return false; |
1559 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1560 | if (!PreCondBI) |
1561 | return false; |
1562 | if (matchCondition(PreCondBI, PH) != InitX) |
1563 | return false; |
1564 | ZeroCheck = true; |
1565 | } |
1566 | |
1567 | // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always |
1568 | // profitable if we delete the loop. |
1569 | |
1570 | // the loop has only 6 instructions: |
1571 | // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] |
1572 | // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] |
1573 | // %shr = ashr %n.addr.0, 1 |
1574 | // %tobool = icmp eq %shr, 0 |
1575 | // %inc = add nsw %i.0, 1 |
1576 | // br i1 %tobool |
1577 | |
1578 | const Value *Args[] = { |
1579 | InitX, ZeroCheck ? ConstantInt::getTrue(InitX->getContext()) |
1580 | : ConstantInt::getFalse(InitX->getContext())}; |
1581 | |
1582 | // @llvm.dbg doesn't count as they have no semantic effect. |
1583 | auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); |
1584 | uint32_t HeaderSize = |
1585 | std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); |
1586 | |
1587 | IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); |
1588 | int Cost = |
1589 | TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); |
1590 | if (HeaderSize != IdiomCanonicalSize && |
1591 | Cost > TargetTransformInfo::TCC_Basic) |
1592 | return false; |
1593 | |
1594 | transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, |
1595 | DefX->getDebugLoc(), ZeroCheck, |
1596 | IsCntPhiUsedOutsideLoop); |
1597 | return true; |
1598 | } |
1599 | |
1600 | /// Recognizes a population count idiom in a non-countable loop. |
1601 | /// |
1602 | /// If detected, transforms the relevant code to issue the popcount intrinsic |
1603 | /// function call, and returns true; otherwise, returns false. |
1604 | bool LoopIdiomRecognize::recognizePopcount() { |
1605 | if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) |
1606 | return false; |
1607 | |
1608 | // Counting population are usually conducted by few arithmetic instructions. |
1609 | // Such instructions can be easily "absorbed" by vacant slots in a |
1610 | // non-compact loop. Therefore, recognizing popcount idiom only makes sense |
1611 | // in a compact loop. |
1612 | |
1613 | // Give up if the loop has multiple blocks or multiple backedges. |
1614 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1615 | return false; |
1616 | |
1617 | BasicBlock *LoopBody = *(CurLoop->block_begin()); |
1618 | if (LoopBody->size() >= 20) { |
1619 | // The loop is too big, bail out. |
1620 | return false; |
1621 | } |
1622 | |
1623 | // It should have a preheader containing nothing but an unconditional branch. |
1624 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1625 | if (!PH || &PH->front() != PH->getTerminator()) |
1626 | return false; |
1627 | auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator()); |
1628 | if (!EntryBI || EntryBI->isConditional()) |
1629 | return false; |
1630 | |
1631 | // It should have a precondition block where the generated popcount intrinsic |
1632 | // function can be inserted. |
1633 | auto *PreCondBB = PH->getSinglePredecessor(); |
1634 | if (!PreCondBB) |
1635 | return false; |
1636 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1637 | if (!PreCondBI || PreCondBI->isUnconditional()) |
1638 | return false; |
1639 | |
1640 | Instruction *CntInst; |
1641 | PHINode *CntPhi; |
1642 | Value *Val; |
1643 | if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) |
1644 | return false; |
1645 | |
1646 | transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); |
1647 | return true; |
1648 | } |
1649 | |
1650 | static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
1651 | const DebugLoc &DL) { |
1652 | Value *Ops[] = {Val}; |
1653 | Type *Tys[] = {Val->getType()}; |
1654 | |
1655 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
1656 | Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); |
1657 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
1658 | CI->setDebugLoc(DL); |
1659 | |
1660 | return CI; |
1661 | } |
1662 | |
1663 | static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
1664 | const DebugLoc &DL, bool ZeroCheck, |
1665 | Intrinsic::ID IID) { |
1666 | Value *Ops[] = {Val, ZeroCheck ? IRBuilder.getTrue() : IRBuilder.getFalse()}; |
1667 | Type *Tys[] = {Val->getType()}; |
1668 | |
1669 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
1670 | Function *Func = Intrinsic::getDeclaration(M, IID, Tys); |
1671 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
1672 | CI->setDebugLoc(DL); |
1673 | |
1674 | return CI; |
1675 | } |
1676 | |
1677 | /// Transform the following loop (Using CTLZ, CTTZ is similar): |
1678 | /// loop: |
1679 | /// CntPhi = PHI [Cnt0, CntInst] |
1680 | /// PhiX = PHI [InitX, DefX] |
1681 | /// CntInst = CntPhi + 1 |
1682 | /// DefX = PhiX >> 1 |
1683 | /// LOOP_BODY |
1684 | /// Br: loop if (DefX != 0) |
1685 | /// Use(CntPhi) or Use(CntInst) |
1686 | /// |
1687 | /// Into: |
1688 | /// If CntPhi used outside the loop: |
1689 | /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) |
1690 | /// Count = CountPrev + 1 |
1691 | /// else |
1692 | /// Count = BitWidth(InitX) - CTLZ(InitX) |
1693 | /// loop: |
1694 | /// CntPhi = PHI [Cnt0, CntInst] |
1695 | /// PhiX = PHI [InitX, DefX] |
1696 | /// PhiCount = PHI [Count, Dec] |
1697 | /// CntInst = CntPhi + 1 |
1698 | /// DefX = PhiX >> 1 |
1699 | /// Dec = PhiCount - 1 |
1700 | /// LOOP_BODY |
1701 | /// Br: loop if (Dec != 0) |
1702 | /// Use(CountPrev + Cnt0) // Use(CntPhi) |
1703 | /// or |
1704 | /// Use(Count + Cnt0) // Use(CntInst) |
1705 | /// |
1706 | /// If LOOP_BODY is empty the loop will be deleted. |
1707 | /// If CntInst and DefX are not used in LOOP_BODY they will be removed. |
1708 | void LoopIdiomRecognize::transformLoopToCountable( |
1709 | Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, |
1710 | PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, |
1711 | bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { |
1712 | BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator()); |
1713 | |
1714 | // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block |
1715 | IRBuilder<> Builder(PreheaderBr); |
1716 | Builder.SetCurrentDebugLocation(DL); |
1717 | Value *FFS, *Count, *CountPrev, *NewCount, *InitXNext; |
1718 | |
1719 | // Count = BitWidth - CTLZ(InitX); |
1720 | // If there are uses of CntPhi create: |
1721 | // CountPrev = BitWidth - CTLZ(InitX >> 1); |
1722 | if (IsCntPhiUsedOutsideLoop) { |
1723 | if (DefX->getOpcode() == Instruction::AShr) |
1724 | InitXNext = |
1725 | Builder.CreateAShr(InitX, ConstantInt::get(InitX->getType(), 1)); |
1726 | else if (DefX->getOpcode() == Instruction::LShr) |
1727 | InitXNext = |
1728 | Builder.CreateLShr(InitX, ConstantInt::get(InitX->getType(), 1)); |
1729 | else if (DefX->getOpcode() == Instruction::Shl) // cttz |
1730 | InitXNext = |
1731 | Builder.CreateShl(InitX, ConstantInt::get(InitX->getType(), 1)); |
1732 | else |
1733 | llvm_unreachable("Unexpected opcode!")::llvm::llvm_unreachable_internal("Unexpected opcode!", "/build/llvm-toolchain-snapshot-12~++20200911111112+c0825fa5fc3/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1733); |
1734 | } else |
1735 | InitXNext = InitX; |
1736 | FFS = createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); |
1737 | Count = Builder.CreateSub( |
1738 | ConstantInt::get(FFS->getType(), |
1739 | FFS->getType()->getIntegerBitWidth()), |
1740 | FFS); |
1741 | if (IsCntPhiUsedOutsideLoop) { |
1742 | CountPrev = Count; |
1743 | Count = Builder.CreateAdd( |
1744 | CountPrev, |
1745 | ConstantInt::get(CountPrev->getType(), 1)); |
1746 | } |
1747 | |
1748 | NewCount = Builder.CreateZExtOrTrunc( |
1749 | IsCntPhiUsedOutsideLoop ? CountPrev : Count, |
1750 | cast<IntegerType>(CntInst->getType())); |
1751 | |
1752 | // If the counter's initial value is not zero, insert Add Inst. |
1753 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); |
1754 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
1755 | if (!InitConst || !InitConst->isZero()) |
1756 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
1757 | |
1758 | // Step 2: Insert new IV and loop condition: |
1759 | // loop: |
1760 | // ... |
1761 | // PhiCount = PHI [Count, Dec] |
1762 | // ... |
1763 | // Dec = PhiCount - 1 |
1764 | // ... |
1765 | // Br: loop if (Dec != 0) |
1766 | BasicBlock *Body = *(CurLoop->block_begin()); |
1767 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
1768 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
1769 | Type *Ty = Count->getType(); |
1770 | |
1771 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
1772 | |
1773 | Builder.SetInsertPoint(LbCond); |
1774 | Instruction *TcDec = cast<Instruction>( |
1775 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
1776 | "tcdec", false, true)); |
1777 | |
1778 | TcPhi->addIncoming(Count, Preheader); |
1779 | TcPhi->addIncoming(TcDec, Body); |
1780 | |
1781 | CmpInst::Predicate Pred = |
1782 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
1783 | LbCond->setPredicate(Pred); |
1784 | LbCond->setOperand(0, TcDec); |
1785 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
1786 | |
1787 | // Step 3: All the references to the original counter outside |
1788 | // the loop are replaced with the NewCount |
1789 | if (IsCntPhiUsedOutsideLoop) |
1790 | CntPhi->replaceUsesOutsideBlock(NewCount, Body); |
1791 | else |
1792 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
1793 | |
1794 | // step 4: Forget the "non-computable" trip-count SCEV associated with the |
1795 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
1796 | SE->forgetLoop(CurLoop); |
1797 | } |
1798 | |
1799 | void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, |
1800 | Instruction *CntInst, |
1801 | PHINode *CntPhi, Value *Var) { |
1802 | BasicBlock *PreHead = CurLoop->getLoopPreheader(); |
1803 | auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator()); |
1804 | const DebugLoc &DL = CntInst->getDebugLoc(); |
1805 | |
1806 | // Assuming before transformation, the loop is following: |
1807 | // if (x) // the precondition |
1808 | // do { cnt++; x &= x - 1; } while(x); |
1809 | |
1810 | // Step 1: Insert the ctpop instruction at the end of the precondition block |
1811 | IRBuilder<> Builder(PreCondBr); |
1812 | Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; |
1813 | { |
1814 | PopCnt = createPopcntIntrinsic(Builder, Var, DL); |
1815 | NewCount = PopCntZext = |
1816 | Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType())); |
1817 | |
1818 | if (NewCount != PopCnt) |
1819 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
1820 | |
1821 | // TripCnt is exactly the number of iterations the loop has |
1822 | TripCnt = NewCount; |
1823 | |
1824 | // If the population counter's initial value is not zero, insert Add Inst. |
1825 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); |
1826 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
1827 | if (!InitConst || !InitConst->isZero()) { |
1828 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
1829 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
1830 | } |
1831 | } |
1832 | |
1833 | // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to |
1834 | // "if (NewCount == 0) loop-exit". Without this change, the intrinsic |
1835 | // function would be partial dead code, and downstream passes will drag |
1836 | // it back from the precondition block to the preheader. |
1837 | { |
1838 | ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition()); |
1839 | |
1840 | Value *Opnd0 = PopCntZext; |
1841 | Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); |
1842 | if (PreCond->getOperand(0) != Var) |
1843 | std::swap(Opnd0, Opnd1); |
1844 | |
1845 | ICmpInst *NewPreCond = cast<ICmpInst>( |
1846 | Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); |
1847 | PreCondBr->setCondition(NewPreCond); |
1848 | |
1849 | RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); |
1850 | } |
1851 | |
1852 | // Step 3: Note that the population count is exactly the trip count of the |
1853 | // loop in question, which enable us to convert the loop from noncountable |
1854 | // loop into a countable one. The benefit is twofold: |
1855 | // |
1856 | // - If the loop only counts population, the entire loop becomes dead after |
1857 | // the transformation. It is a lot easier to prove a countable loop dead |
1858 | // than to prove a noncountable one. (In some C dialects, an infinite loop |
1859 | // isn't dead even if it computes nothing useful. In general, DCE needs |
1860 | // to prove a noncountable loop finite before safely delete it.) |
1861 | // |
1862 | // - If the loop also performs something else, it remains alive. |
1863 | // Since it is transformed to countable form, it can be aggressively |
1864 | // optimized by some optimizations which are in general not applicable |
1865 | // to a noncountable loop. |
1866 | // |
1867 | // After this step, this loop (conceptually) would look like following: |
1868 | // newcnt = __builtin_ctpop(x); |
1869 | // t = newcnt; |
1870 | // if (x) |
1871 | // do { cnt++; x &= x-1; t--) } while (t > 0); |
1872 | BasicBlock *Body = *(CurLoop->block_begin()); |
1873 | { |
1874 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
1875 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
1876 | Type *Ty = TripCnt->getType(); |
1877 | |
1878 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
1879 | |
1880 | Builder.SetInsertPoint(LbCond); |
1881 | Instruction *TcDec = cast<Instruction>( |
1882 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
1883 | "tcdec", false, true)); |
1884 | |
1885 | TcPhi->addIncoming(TripCnt, PreHead); |
1886 | TcPhi->addIncoming(TcDec, Body); |
1887 | |
1888 | CmpInst::Predicate Pred = |
1889 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; |
1890 | LbCond->setPredicate(Pred); |
1891 | LbCond->setOperand(0, TcDec); |
1892 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
1893 | } |
1894 | |
1895 | // Step 4: All the references to the original population counter outside |
1896 | // the loop are replaced with the NewCount -- the value returned from |
1897 | // __builtin_ctpop(). |
1898 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
1899 | |
1900 | // step 5: Forget the "non-computable" trip-count SCEV associated with the |
1901 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
1902 | SE->forgetLoop(CurLoop); |
1903 | } |