File: | llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp |
Warning: | line 521, 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/CmpInstAnalysis.h" |
51 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
52 | #include "llvm/Analysis/LoopInfo.h" |
53 | #include "llvm/Analysis/LoopPass.h" |
54 | #include "llvm/Analysis/MemoryLocation.h" |
55 | #include "llvm/Analysis/MemorySSA.h" |
56 | #include "llvm/Analysis/MemorySSAUpdater.h" |
57 | #include "llvm/Analysis/MustExecute.h" |
58 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
59 | #include "llvm/Analysis/ScalarEvolution.h" |
60 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
61 | #include "llvm/Analysis/TargetLibraryInfo.h" |
62 | #include "llvm/Analysis/TargetTransformInfo.h" |
63 | #include "llvm/Analysis/ValueTracking.h" |
64 | #include "llvm/IR/Attributes.h" |
65 | #include "llvm/IR/BasicBlock.h" |
66 | #include "llvm/IR/Constant.h" |
67 | #include "llvm/IR/Constants.h" |
68 | #include "llvm/IR/DataLayout.h" |
69 | #include "llvm/IR/DebugLoc.h" |
70 | #include "llvm/IR/DerivedTypes.h" |
71 | #include "llvm/IR/Dominators.h" |
72 | #include "llvm/IR/GlobalValue.h" |
73 | #include "llvm/IR/GlobalVariable.h" |
74 | #include "llvm/IR/IRBuilder.h" |
75 | #include "llvm/IR/InstrTypes.h" |
76 | #include "llvm/IR/Instruction.h" |
77 | #include "llvm/IR/Instructions.h" |
78 | #include "llvm/IR/IntrinsicInst.h" |
79 | #include "llvm/IR/Intrinsics.h" |
80 | #include "llvm/IR/LLVMContext.h" |
81 | #include "llvm/IR/Module.h" |
82 | #include "llvm/IR/PassManager.h" |
83 | #include "llvm/IR/PatternMatch.h" |
84 | #include "llvm/IR/Type.h" |
85 | #include "llvm/IR/User.h" |
86 | #include "llvm/IR/Value.h" |
87 | #include "llvm/IR/ValueHandle.h" |
88 | #include "llvm/InitializePasses.h" |
89 | #include "llvm/Pass.h" |
90 | #include "llvm/Support/Casting.h" |
91 | #include "llvm/Support/CommandLine.h" |
92 | #include "llvm/Support/Debug.h" |
93 | #include "llvm/Support/InstructionCost.h" |
94 | #include "llvm/Support/raw_ostream.h" |
95 | #include "llvm/Transforms/Scalar.h" |
96 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
97 | #include "llvm/Transforms/Utils/Local.h" |
98 | #include "llvm/Transforms/Utils/LoopUtils.h" |
99 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
100 | #include <algorithm> |
101 | #include <cassert> |
102 | #include <cstdint> |
103 | #include <utility> |
104 | #include <vector> |
105 | |
106 | using namespace llvm; |
107 | |
108 | #define DEBUG_TYPE"loop-idiom" "loop-idiom" |
109 | |
110 | 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"}; |
111 | 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"}; |
112 | STATISTIC(static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
113 | NumShiftUntilBitTest,static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" } |
114 | "Number of uncountable loops recognized as 'shift until bitttest' idiom")static llvm::Statistic NumShiftUntilBitTest = {"loop-idiom", "NumShiftUntilBitTest" , "Number of uncountable loops recognized as 'shift until bitttest' idiom" }; |
115 | |
116 | bool DisableLIRP::All; |
117 | static cl::opt<bool, true> |
118 | DisableLIRPAll("disable-" DEBUG_TYPE"loop-idiom" "-all", |
119 | cl::desc("Options to disable Loop Idiom Recognize Pass."), |
120 | cl::location(DisableLIRP::All), cl::init(false), |
121 | cl::ReallyHidden); |
122 | |
123 | bool DisableLIRP::Memset; |
124 | static cl::opt<bool, true> |
125 | DisableLIRPMemset("disable-" DEBUG_TYPE"loop-idiom" "-memset", |
126 | cl::desc("Proceed with loop idiom recognize pass, but do " |
127 | "not convert loop(s) to memset."), |
128 | cl::location(DisableLIRP::Memset), cl::init(false), |
129 | cl::ReallyHidden); |
130 | |
131 | bool DisableLIRP::Memcpy; |
132 | static cl::opt<bool, true> |
133 | DisableLIRPMemcpy("disable-" DEBUG_TYPE"loop-idiom" "-memcpy", |
134 | cl::desc("Proceed with loop idiom recognize pass, but do " |
135 | "not convert loop(s) to memcpy."), |
136 | cl::location(DisableLIRP::Memcpy), cl::init(false), |
137 | cl::ReallyHidden); |
138 | |
139 | static cl::opt<bool> UseLIRCodeSizeHeurs( |
140 | "use-lir-code-size-heurs", |
141 | cl::desc("Use loop idiom recognition code size heuristics when compiling" |
142 | "with -Os/-Oz"), |
143 | cl::init(true), cl::Hidden); |
144 | |
145 | namespace { |
146 | |
147 | class LoopIdiomRecognize { |
148 | Loop *CurLoop = nullptr; |
149 | AliasAnalysis *AA; |
150 | DominatorTree *DT; |
151 | LoopInfo *LI; |
152 | ScalarEvolution *SE; |
153 | TargetLibraryInfo *TLI; |
154 | const TargetTransformInfo *TTI; |
155 | const DataLayout *DL; |
156 | OptimizationRemarkEmitter &ORE; |
157 | bool ApplyCodeSizeHeuristics; |
158 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
159 | |
160 | public: |
161 | explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, |
162 | LoopInfo *LI, ScalarEvolution *SE, |
163 | TargetLibraryInfo *TLI, |
164 | const TargetTransformInfo *TTI, MemorySSA *MSSA, |
165 | const DataLayout *DL, |
166 | OptimizationRemarkEmitter &ORE) |
167 | : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { |
168 | if (MSSA) |
169 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); |
170 | } |
171 | |
172 | bool runOnLoop(Loop *L); |
173 | |
174 | private: |
175 | using StoreList = SmallVector<StoreInst *, 8>; |
176 | using StoreListMap = MapVector<Value *, StoreList>; |
177 | |
178 | StoreListMap StoreRefsForMemset; |
179 | StoreListMap StoreRefsForMemsetPattern; |
180 | StoreList StoreRefsForMemcpy; |
181 | bool HasMemset; |
182 | bool HasMemsetPattern; |
183 | bool HasMemcpy; |
184 | |
185 | /// Return code for isLegalStore() |
186 | enum LegalStoreKind { |
187 | None = 0, |
188 | Memset, |
189 | MemsetPattern, |
190 | Memcpy, |
191 | UnorderedAtomicMemcpy, |
192 | DontUse // Dummy retval never to be used. Allows catching errors in retval |
193 | // handling. |
194 | }; |
195 | |
196 | /// \name Countable Loop Idiom Handling |
197 | /// @{ |
198 | |
199 | bool runOnCountableLoop(); |
200 | bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, |
201 | SmallVectorImpl<BasicBlock *> &ExitBlocks); |
202 | |
203 | void collectStores(BasicBlock *BB); |
204 | LegalStoreKind isLegalStore(StoreInst *SI); |
205 | enum class ForMemset { No, Yes }; |
206 | bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount, |
207 | ForMemset For); |
208 | bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); |
209 | |
210 | bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize, |
211 | MaybeAlign StoreAlignment, Value *StoredVal, |
212 | Instruction *TheStore, |
213 | SmallPtrSetImpl<Instruction *> &Stores, |
214 | const SCEVAddRecExpr *Ev, const SCEV *BECount, |
215 | bool NegStride, bool IsLoopMemset = false); |
216 | bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); |
217 | bool avoidLIRForMultiBlockLoop(bool IsMemset = false, |
218 | bool IsLoopMemset = false); |
219 | |
220 | /// @} |
221 | /// \name Noncountable Loop Idiom Handling |
222 | /// @{ |
223 | |
224 | bool runOnNoncountableLoop(); |
225 | |
226 | bool recognizePopcount(); |
227 | void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, |
228 | PHINode *CntPhi, Value *Var); |
229 | bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz |
230 | void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, |
231 | Instruction *CntInst, PHINode *CntPhi, |
232 | Value *Var, Instruction *DefX, |
233 | const DebugLoc &DL, bool ZeroCheck, |
234 | bool IsCntPhiUsedOutsideLoop); |
235 | |
236 | bool recognizeShiftUntilBitTest(); |
237 | |
238 | /// @} |
239 | }; |
240 | |
241 | class LoopIdiomRecognizeLegacyPass : public LoopPass { |
242 | public: |
243 | static char ID; |
244 | |
245 | explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) { |
246 | initializeLoopIdiomRecognizeLegacyPassPass( |
247 | *PassRegistry::getPassRegistry()); |
248 | } |
249 | |
250 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { |
251 | if (DisableLIRP::All) |
252 | return false; |
253 | |
254 | if (skipLoop(L)) |
255 | return false; |
256 | |
257 | AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
258 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
259 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
260 | ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
261 | TargetLibraryInfo *TLI = |
262 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
263 | *L->getHeader()->getParent()); |
264 | const TargetTransformInfo *TTI = |
265 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( |
266 | *L->getHeader()->getParent()); |
267 | const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout(); |
268 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
269 | MemorySSA *MSSA = nullptr; |
270 | if (MSSAAnalysis) |
271 | MSSA = &MSSAAnalysis->getMSSA(); |
272 | |
273 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis |
274 | // pass. Function analyses need to be preserved across loop transformations |
275 | // but ORE cannot be preserved (see comment before the pass definition). |
276 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); |
277 | |
278 | LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE); |
279 | return LIR.runOnLoop(L); |
280 | } |
281 | |
282 | /// This transformation requires natural loop information & requires that |
283 | /// loop preheaders be inserted into the CFG. |
284 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
285 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
286 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
287 | AU.addPreserved<MemorySSAWrapperPass>(); |
288 | getLoopAnalysisUsage(AU); |
289 | } |
290 | }; |
291 | |
292 | } // end anonymous namespace |
293 | |
294 | char LoopIdiomRecognizeLegacyPass::ID = 0; |
295 | |
296 | PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, |
297 | LoopStandardAnalysisResults &AR, |
298 | LPMUpdater &) { |
299 | if (DisableLIRP::All) |
300 | return PreservedAnalyses::all(); |
301 | |
302 | const auto *DL = &L.getHeader()->getModule()->getDataLayout(); |
303 | |
304 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
305 | // pass. Function analyses need to be preserved across loop transformations |
306 | // but ORE cannot be preserved (see comment before the pass definition). |
307 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
308 | |
309 | LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, |
310 | AR.MSSA, DL, ORE); |
311 | if (!LIR.runOnLoop(&L)) |
312 | return PreservedAnalyses::all(); |
313 | |
314 | auto PA = getLoopPassPreservedAnalyses(); |
315 | if (AR.MSSA) |
316 | PA.preserve<MemorySSAAnalysis>(); |
317 | return PA; |
318 | } |
319 | |
320 | INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
321 | "Recognize loop idioms", false, false)static void *initializeLoopIdiomRecognizeLegacyPassPassOnce(PassRegistry &Registry) { |
322 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); |
323 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
324 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); |
325 | 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 )); } |
326 | "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 )); } |
327 | |
328 | Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); } |
329 | |
330 | static void deleteDeadInstruction(Instruction *I) { |
331 | I->replaceAllUsesWith(UndefValue::get(I->getType())); |
332 | I->eraseFromParent(); |
333 | } |
334 | |
335 | //===----------------------------------------------------------------------===// |
336 | // |
337 | // Implementation of LoopIdiomRecognize |
338 | // |
339 | //===----------------------------------------------------------------------===// |
340 | |
341 | bool LoopIdiomRecognize::runOnLoop(Loop *L) { |
342 | CurLoop = L; |
343 | // If the loop could not be converted to canonical form, it must have an |
344 | // indirectbr in it, just give up. |
345 | if (!L->getLoopPreheader()) |
346 | return false; |
347 | |
348 | // Disable loop idiom recognition if the function's name is a common idiom. |
349 | StringRef Name = L->getHeader()->getParent()->getName(); |
350 | if (Name == "memset" || Name == "memcpy") |
351 | return false; |
352 | |
353 | // Determine if code size heuristics need to be applied. |
354 | ApplyCodeSizeHeuristics = |
355 | L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; |
356 | |
357 | HasMemset = TLI->has(LibFunc_memset); |
358 | HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); |
359 | HasMemcpy = TLI->has(LibFunc_memcpy); |
360 | |
361 | if (HasMemset || HasMemsetPattern || HasMemcpy) |
362 | if (SE->hasLoopInvariantBackedgeTakenCount(L)) |
363 | return runOnCountableLoop(); |
364 | |
365 | return runOnNoncountableLoop(); |
366 | } |
367 | |
368 | bool LoopIdiomRecognize::runOnCountableLoop() { |
369 | const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); |
370 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 372, __PRETTY_FUNCTION__)) |
371 | "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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 372, __PRETTY_FUNCTION__)) |
372 | "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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 372, __PRETTY_FUNCTION__)); |
373 | |
374 | // If this loop executes exactly one time, then it should be peeled, not |
375 | // optimized by this pass. |
376 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
377 | if (BECst->getAPInt() == 0) |
378 | return false; |
379 | |
380 | SmallVector<BasicBlock *, 8> ExitBlocks; |
381 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
382 | |
383 | 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) |
384 | << 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) |
385 | << "] 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) |
386 | << "\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); |
387 | |
388 | // The following transforms hoist stores/memsets into the loop pre-header. |
389 | // Give up if the loop has instructions that may throw. |
390 | SimpleLoopSafetyInfo SafetyInfo; |
391 | SafetyInfo.computeLoopSafetyInfo(CurLoop); |
392 | if (SafetyInfo.anyBlockMayThrow()) |
393 | return false; |
394 | |
395 | bool MadeChange = false; |
396 | |
397 | // Scan all the blocks in the loop that are not in subloops. |
398 | for (auto *BB : CurLoop->getBlocks()) { |
399 | // Ignore blocks in subloops. |
400 | if (LI->getLoopFor(BB) != CurLoop) |
401 | continue; |
402 | |
403 | MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); |
404 | } |
405 | return MadeChange; |
406 | } |
407 | |
408 | static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { |
409 | const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1)); |
410 | return ConstStride->getAPInt(); |
411 | } |
412 | |
413 | /// getMemSetPatternValue - If a strided store of the specified value is safe to |
414 | /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should |
415 | /// be passed in. Otherwise, return null. |
416 | /// |
417 | /// Note that we don't ever attempt to use memset_pattern8 or 4, because these |
418 | /// just replicate their input array and then pass on to memset_pattern16. |
419 | static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { |
420 | // FIXME: This could check for UndefValue because it can be merged into any |
421 | // other valid pattern. |
422 | |
423 | // If the value isn't a constant, we can't promote it to being in a constant |
424 | // array. We could theoretically do a store to an alloca or something, but |
425 | // that doesn't seem worthwhile. |
426 | Constant *C = dyn_cast<Constant>(V); |
427 | if (!C) |
428 | return nullptr; |
429 | |
430 | // Only handle simple values that are a power of two bytes in size. |
431 | uint64_t Size = DL->getTypeSizeInBits(V->getType()); |
432 | if (Size == 0 || (Size & 7) || (Size & (Size - 1))) |
433 | return nullptr; |
434 | |
435 | // Don't care enough about darwin/ppc to implement this. |
436 | if (DL->isBigEndian()) |
437 | return nullptr; |
438 | |
439 | // Convert to size in bytes. |
440 | Size /= 8; |
441 | |
442 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
443 | // if the top and bottom are the same (e.g. for vectors and large integers). |
444 | if (Size > 16) |
445 | return nullptr; |
446 | |
447 | // If the constant is exactly 16 bytes, just use it. |
448 | if (Size == 16) |
449 | return C; |
450 | |
451 | // Otherwise, we'll use an array of the constants. |
452 | unsigned ArraySize = 16 / Size; |
453 | ArrayType *AT = ArrayType::get(V->getType(), ArraySize); |
454 | return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C)); |
455 | } |
456 | |
457 | LoopIdiomRecognize::LegalStoreKind |
458 | LoopIdiomRecognize::isLegalStore(StoreInst *SI) { |
459 | // Don't touch volatile stores. |
460 | if (SI->isVolatile()) |
461 | return LegalStoreKind::None; |
462 | // We only want simple or unordered-atomic stores. |
463 | if (!SI->isUnordered()) |
464 | return LegalStoreKind::None; |
465 | |
466 | // Avoid merging nontemporal stores. |
467 | if (SI->getMetadata(LLVMContext::MD_nontemporal)) |
468 | return LegalStoreKind::None; |
469 | |
470 | Value *StoredVal = SI->getValueOperand(); |
471 | Value *StorePtr = SI->getPointerOperand(); |
472 | |
473 | // Don't convert stores of non-integral pointer types to memsets (which stores |
474 | // integers). |
475 | if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType())) |
476 | return LegalStoreKind::None; |
477 | |
478 | // Reject stores that are so large that they overflow an unsigned. |
479 | // When storing out scalable vectors we bail out for now, since the code |
480 | // below currently only works for constant strides. |
481 | TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); |
482 | if (SizeInBits.isScalable() || (SizeInBits.getFixedSize() & 7) || |
483 | (SizeInBits.getFixedSize() >> 32) != 0) |
484 | return LegalStoreKind::None; |
485 | |
486 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
487 | // loop, which indicates a strided store. If we have something else, it's a |
488 | // random store we can't handle. |
489 | const SCEVAddRecExpr *StoreEv = |
490 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
491 | if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) |
492 | return LegalStoreKind::None; |
493 | |
494 | // Check to see if we have a constant stride. |
495 | if (!isa<SCEVConstant>(StoreEv->getOperand(1))) |
496 | return LegalStoreKind::None; |
497 | |
498 | // See if the store can be turned into a memset. |
499 | |
500 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
501 | // turned into a memset of i8 -1, assuming that all the consecutive bytes |
502 | // are stored. A store of i32 0x01020304 can never be turned into a memset, |
503 | // but it can be turned into memset_pattern if the target supports it. |
504 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
505 | Constant *PatternValue = nullptr; |
506 | |
507 | // Note: memset and memset_pattern on unordered-atomic is yet not supported |
508 | bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); |
509 | |
510 | // If we're allowed to form a memset, and the stored value would be |
511 | // acceptable for memset, use it. |
512 | if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && |
513 | // Verify that the stored value is loop invariant. If not, we can't |
514 | // promote the memset. |
515 | CurLoop->isLoopInvariant(SplatValue)) { |
516 | // It looks like we can use SplatValue. |
517 | return LegalStoreKind::Memset; |
518 | } else if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && |
519 | // Don't create memset_pattern16s with address spaces. |
520 | StorePtr->getType()->getPointerAddressSpace() == 0 && |
521 | (PatternValue = getMemSetPatternValue(StoredVal, DL))) { |
Although the value stored to 'PatternValue' is used in the enclosing expression, the value is never actually read from 'PatternValue' | |
522 | // It looks like we can use PatternValue! |
523 | return LegalStoreKind::MemsetPattern; |
524 | } |
525 | |
526 | // Otherwise, see if the store can be turned into a memcpy. |
527 | if (HasMemcpy && !DisableLIRP::Memcpy) { |
528 | // Check to see if the stride matches the size of the store. If so, then we |
529 | // know that every byte is touched in the loop. |
530 | APInt Stride = getStoreStride(StoreEv); |
531 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
532 | if (StoreSize != Stride && StoreSize != -Stride) |
533 | return LegalStoreKind::None; |
534 | |
535 | // The store must be feeding a non-volatile load. |
536 | LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); |
537 | |
538 | // Only allow non-volatile loads |
539 | if (!LI || LI->isVolatile()) |
540 | return LegalStoreKind::None; |
541 | // Only allow simple or unordered-atomic loads |
542 | if (!LI->isUnordered()) |
543 | return LegalStoreKind::None; |
544 | |
545 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
546 | // loop, which indicates a strided load. If we have something else, it's a |
547 | // random load we can't handle. |
548 | const SCEVAddRecExpr *LoadEv = |
549 | dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
550 | if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) |
551 | return LegalStoreKind::None; |
552 | |
553 | // The store and load must share the same stride. |
554 | if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) |
555 | return LegalStoreKind::None; |
556 | |
557 | // Success. This store can be converted into a memcpy. |
558 | UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); |
559 | return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy |
560 | : LegalStoreKind::Memcpy; |
561 | } |
562 | // This store can't be transformed into a memset/memcpy. |
563 | return LegalStoreKind::None; |
564 | } |
565 | |
566 | void LoopIdiomRecognize::collectStores(BasicBlock *BB) { |
567 | StoreRefsForMemset.clear(); |
568 | StoreRefsForMemsetPattern.clear(); |
569 | StoreRefsForMemcpy.clear(); |
570 | for (Instruction &I : *BB) { |
571 | StoreInst *SI = dyn_cast<StoreInst>(&I); |
572 | if (!SI) |
573 | continue; |
574 | |
575 | // Make sure this is a strided store with a constant stride. |
576 | switch (isLegalStore(SI)) { |
577 | case LegalStoreKind::None: |
578 | // Nothing to do |
579 | break; |
580 | case LegalStoreKind::Memset: { |
581 | // Find the base pointer. |
582 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
583 | StoreRefsForMemset[Ptr].push_back(SI); |
584 | } break; |
585 | case LegalStoreKind::MemsetPattern: { |
586 | // Find the base pointer. |
587 | Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); |
588 | StoreRefsForMemsetPattern[Ptr].push_back(SI); |
589 | } break; |
590 | case LegalStoreKind::Memcpy: |
591 | case LegalStoreKind::UnorderedAtomicMemcpy: |
592 | StoreRefsForMemcpy.push_back(SI); |
593 | break; |
594 | default: |
595 | assert(false && "unhandled return value")((false && "unhandled return value") ? static_cast< void> (0) : __assert_fail ("false && \"unhandled return value\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 595, __PRETTY_FUNCTION__)); |
596 | break; |
597 | } |
598 | } |
599 | } |
600 | |
601 | /// runOnLoopBlock - Process the specified block, which lives in a counted loop |
602 | /// with the specified backedge count. This block is known to be in the current |
603 | /// loop and not in any subloops. |
604 | bool LoopIdiomRecognize::runOnLoopBlock( |
605 | BasicBlock *BB, const SCEV *BECount, |
606 | SmallVectorImpl<BasicBlock *> &ExitBlocks) { |
607 | // We can only promote stores in this block if they are unconditionally |
608 | // executed in the loop. For a block to be unconditionally executed, it has |
609 | // to dominate all the exit blocks of the loop. Verify this now. |
610 | for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) |
611 | if (!DT->dominates(BB, ExitBlocks[i])) |
612 | return false; |
613 | |
614 | bool MadeChange = false; |
615 | // Look for store instructions, which may be optimized to memset/memcpy. |
616 | collectStores(BB); |
617 | |
618 | // Look for a single store or sets of stores with a common base, which can be |
619 | // optimized into a memset (memset_pattern). The latter most commonly happens |
620 | // with structs and handunrolled loops. |
621 | for (auto &SL : StoreRefsForMemset) |
622 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); |
623 | |
624 | for (auto &SL : StoreRefsForMemsetPattern) |
625 | MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); |
626 | |
627 | // Optimize the store into a memcpy, if it feeds an similarly strided load. |
628 | for (auto &SI : StoreRefsForMemcpy) |
629 | MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); |
630 | |
631 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { |
632 | Instruction *Inst = &*I++; |
633 | // Look for memset instructions, which may be optimized to a larger memset. |
634 | if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) { |
635 | WeakTrackingVH InstPtr(&*I); |
636 | if (!processLoopMemSet(MSI, BECount)) |
637 | continue; |
638 | MadeChange = true; |
639 | |
640 | // If processing the memset invalidated our iterator, start over from the |
641 | // top of the block. |
642 | if (!InstPtr) |
643 | I = BB->begin(); |
644 | continue; |
645 | } |
646 | } |
647 | |
648 | return MadeChange; |
649 | } |
650 | |
651 | /// See if this store(s) can be promoted to a memset. |
652 | bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL, |
653 | const SCEV *BECount, ForMemset For) { |
654 | // Try to find consecutive stores that can be transformed into memsets. |
655 | SetVector<StoreInst *> Heads, Tails; |
656 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; |
657 | |
658 | // Do a quadratic search on all of the given stores and find |
659 | // all of the pairs of stores that follow each other. |
660 | SmallVector<unsigned, 16> IndexQueue; |
661 | for (unsigned i = 0, e = SL.size(); i < e; ++i) { |
662 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 662, __PRETTY_FUNCTION__)); |
663 | |
664 | Value *FirstStoredVal = SL[i]->getValueOperand(); |
665 | Value *FirstStorePtr = SL[i]->getPointerOperand(); |
666 | const SCEVAddRecExpr *FirstStoreEv = |
667 | cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr)); |
668 | APInt FirstStride = getStoreStride(FirstStoreEv); |
669 | unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); |
670 | |
671 | // See if we can optimize just this store in isolation. |
672 | if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { |
673 | Heads.insert(SL[i]); |
674 | continue; |
675 | } |
676 | |
677 | Value *FirstSplatValue = nullptr; |
678 | Constant *FirstPatternValue = nullptr; |
679 | |
680 | if (For == ForMemset::Yes) |
681 | FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); |
682 | else |
683 | FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); |
684 | |
685 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 686, __PRETTY_FUNCTION__)) |
686 | "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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 686, __PRETTY_FUNCTION__)); |
687 | |
688 | IndexQueue.clear(); |
689 | // If a store has multiple consecutive store candidates, search Stores |
690 | // array according to the sequence: from i+1 to e, then from i-1 to 0. |
691 | // This is because usually pairing with immediate succeeding or preceding |
692 | // candidate create the best chance to find memset opportunity. |
693 | unsigned j = 0; |
694 | for (j = i + 1; j < e; ++j) |
695 | IndexQueue.push_back(j); |
696 | for (j = i; j > 0; --j) |
697 | IndexQueue.push_back(j - 1); |
698 | |
699 | for (auto &k : IndexQueue) { |
700 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 700, __PRETTY_FUNCTION__)); |
701 | Value *SecondStorePtr = SL[k]->getPointerOperand(); |
702 | const SCEVAddRecExpr *SecondStoreEv = |
703 | cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr)); |
704 | APInt SecondStride = getStoreStride(SecondStoreEv); |
705 | |
706 | if (FirstStride != SecondStride) |
707 | continue; |
708 | |
709 | Value *SecondStoredVal = SL[k]->getValueOperand(); |
710 | Value *SecondSplatValue = nullptr; |
711 | Constant *SecondPatternValue = nullptr; |
712 | |
713 | if (For == ForMemset::Yes) |
714 | SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); |
715 | else |
716 | SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); |
717 | |
718 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 719, __PRETTY_FUNCTION__)) |
719 | "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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 719, __PRETTY_FUNCTION__)); |
720 | |
721 | if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { |
722 | if (For == ForMemset::Yes) { |
723 | if (isa<UndefValue>(FirstSplatValue)) |
724 | FirstSplatValue = SecondSplatValue; |
725 | if (FirstSplatValue != SecondSplatValue) |
726 | continue; |
727 | } else { |
728 | if (isa<UndefValue>(FirstPatternValue)) |
729 | FirstPatternValue = SecondPatternValue; |
730 | if (FirstPatternValue != SecondPatternValue) |
731 | continue; |
732 | } |
733 | Tails.insert(SL[k]); |
734 | Heads.insert(SL[i]); |
735 | ConsecutiveChain[SL[i]] = SL[k]; |
736 | break; |
737 | } |
738 | } |
739 | } |
740 | |
741 | // We may run into multiple chains that merge into a single chain. We mark the |
742 | // stores that we transformed so that we don't visit the same store twice. |
743 | SmallPtrSet<Value *, 16> TransformedStores; |
744 | bool Changed = false; |
745 | |
746 | // For stores that start but don't end a link in the chain: |
747 | for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end(); |
748 | it != e; ++it) { |
749 | if (Tails.count(*it)) |
750 | continue; |
751 | |
752 | // We found a store instr that starts a chain. Now follow the chain and try |
753 | // to transform it. |
754 | SmallPtrSet<Instruction *, 8> AdjacentStores; |
755 | StoreInst *I = *it; |
756 | |
757 | StoreInst *HeadStore = I; |
758 | unsigned StoreSize = 0; |
759 | |
760 | // Collect the chain into a list. |
761 | while (Tails.count(I) || Heads.count(I)) { |
762 | if (TransformedStores.count(I)) |
763 | break; |
764 | AdjacentStores.insert(I); |
765 | |
766 | StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); |
767 | // Move to the next value in the chain. |
768 | I = ConsecutiveChain[I]; |
769 | } |
770 | |
771 | Value *StoredVal = HeadStore->getValueOperand(); |
772 | Value *StorePtr = HeadStore->getPointerOperand(); |
773 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
774 | APInt Stride = getStoreStride(StoreEv); |
775 | |
776 | // Check to see if the stride matches the size of the stores. If so, then |
777 | // we know that every byte is touched in the loop. |
778 | if (StoreSize != Stride && StoreSize != -Stride) |
779 | continue; |
780 | |
781 | bool NegStride = StoreSize == -Stride; |
782 | |
783 | if (processLoopStridedStore(StorePtr, StoreSize, |
784 | MaybeAlign(HeadStore->getAlignment()), |
785 | StoredVal, HeadStore, AdjacentStores, StoreEv, |
786 | BECount, NegStride)) { |
787 | TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); |
788 | Changed = true; |
789 | } |
790 | } |
791 | |
792 | return Changed; |
793 | } |
794 | |
795 | /// processLoopMemSet - See if this memset can be promoted to a large memset. |
796 | bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, |
797 | const SCEV *BECount) { |
798 | // We can only handle non-volatile memsets with a constant size. |
799 | if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) |
800 | return false; |
801 | |
802 | // If we're not allowed to hack on memset, we fail. |
803 | if (!HasMemset) |
804 | return false; |
805 | |
806 | Value *Pointer = MSI->getDest(); |
807 | |
808 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
809 | // loop, which indicates a strided store. If we have something else, it's a |
810 | // random store we can't handle. |
811 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer)); |
812 | if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine()) |
813 | return false; |
814 | |
815 | // Reject memsets that are so large that they overflow an unsigned. |
816 | uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); |
817 | if ((SizeInBytes >> 32) != 0) |
818 | return false; |
819 | |
820 | // Check to see if the stride matches the size of the memset. If so, then we |
821 | // know that every byte is touched in the loop. |
822 | const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
823 | if (!ConstStride) |
824 | return false; |
825 | |
826 | APInt Stride = ConstStride->getAPInt(); |
827 | if (SizeInBytes != Stride && SizeInBytes != -Stride) |
828 | return false; |
829 | |
830 | // Verify that the memset value is loop invariant. If not, we can't promote |
831 | // the memset. |
832 | Value *SplatValue = MSI->getValue(); |
833 | if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) |
834 | return false; |
835 | |
836 | SmallPtrSet<Instruction *, 1> MSIs; |
837 | MSIs.insert(MSI); |
838 | bool NegStride = SizeInBytes == -Stride; |
839 | return processLoopStridedStore( |
840 | Pointer, (unsigned)SizeInBytes, MaybeAlign(MSI->getDestAlignment()), |
841 | SplatValue, MSI, MSIs, Ev, BECount, NegStride, /*IsLoopMemset=*/true); |
842 | } |
843 | |
844 | /// mayLoopAccessLocation - Return true if the specified loop might access the |
845 | /// specified pointer location, which is a loop-strided access. The 'Access' |
846 | /// argument specifies what the verboten forms of access are (read or write). |
847 | static bool |
848 | mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, |
849 | const SCEV *BECount, unsigned StoreSize, |
850 | AliasAnalysis &AA, |
851 | SmallPtrSetImpl<Instruction *> &IgnoredStores) { |
852 | // Get the location that may be stored across the loop. Since the access is |
853 | // strided positively through memory, we say that the modified location starts |
854 | // at the pointer and has infinite size. |
855 | LocationSize AccessSize = LocationSize::afterPointer(); |
856 | |
857 | // If the loop iterates a fixed number of times, we can refine the access size |
858 | // to be exactly the size of the memset, which is (BECount+1)*StoreSize |
859 | if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) |
860 | AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * |
861 | StoreSize); |
862 | |
863 | // TODO: For this to be really effective, we have to dive into the pointer |
864 | // operand in the store. Store to &A[i] of 100 will always return may alias |
865 | // with store of &A[100], we need to StoreLoc to be "A" with size of 100, |
866 | // which will then no-alias a store to &A[100]. |
867 | MemoryLocation StoreLoc(Ptr, AccessSize); |
868 | |
869 | for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; |
870 | ++BI) |
871 | for (Instruction &I : **BI) |
872 | if (IgnoredStores.count(&I) == 0 && |
873 | isModOrRefSet( |
874 | intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access))) |
875 | return true; |
876 | |
877 | return false; |
878 | } |
879 | |
880 | // If we have a negative stride, Start refers to the end of the memory location |
881 | // we're trying to memset. Therefore, we need to recompute the base pointer, |
882 | // which is just Start - BECount*Size. |
883 | static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, |
884 | Type *IntPtr, unsigned StoreSize, |
885 | ScalarEvolution *SE) { |
886 | const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); |
887 | if (StoreSize != 1) |
888 | Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize), |
889 | SCEV::FlagNUW); |
890 | return SE->getMinusSCEV(Start, Index); |
891 | } |
892 | |
893 | /// Compute the number of bytes as a SCEV from the backedge taken count. |
894 | /// |
895 | /// This also maps the SCEV into the provided type and tries to handle the |
896 | /// computation in a way that will fold cleanly. |
897 | static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, |
898 | unsigned StoreSize, Loop *CurLoop, |
899 | const DataLayout *DL, ScalarEvolution *SE) { |
900 | const SCEV *NumBytesS; |
901 | // The # stored bytes is (BECount+1)*Size. Expand the trip count out to |
902 | // pointer size if it isn't already. |
903 | // |
904 | // If we're going to need to zero extend the BE count, check if we can add |
905 | // one to it prior to zero extending without overflow. Provided this is safe, |
906 | // it allows better simplification of the +1. |
907 | if (DL->getTypeSizeInBits(BECount->getType()).getFixedSize() < |
908 | DL->getTypeSizeInBits(IntPtr).getFixedSize() && |
909 | SE->isLoopEntryGuardedByCond( |
910 | CurLoop, ICmpInst::ICMP_NE, BECount, |
911 | SE->getNegativeSCEV(SE->getOne(BECount->getType())))) { |
912 | NumBytesS = SE->getZeroExtendExpr( |
913 | SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW), |
914 | IntPtr); |
915 | } else { |
916 | NumBytesS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr), |
917 | SE->getOne(IntPtr), SCEV::FlagNUW); |
918 | } |
919 | |
920 | // And scale it based on the store size. |
921 | if (StoreSize != 1) { |
922 | NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), |
923 | SCEV::FlagNUW); |
924 | } |
925 | return NumBytesS; |
926 | } |
927 | |
928 | /// processLoopStridedStore - We see a strided store of some value. If we can |
929 | /// transform this into a memset or memset_pattern in the loop preheader, do so. |
930 | bool LoopIdiomRecognize::processLoopStridedStore( |
931 | Value *DestPtr, unsigned StoreSize, MaybeAlign StoreAlignment, |
932 | Value *StoredVal, Instruction *TheStore, |
933 | SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev, |
934 | const SCEV *BECount, bool NegStride, bool IsLoopMemset) { |
935 | Value *SplatValue = isBytewiseValue(StoredVal, *DL); |
936 | Constant *PatternValue = nullptr; |
937 | |
938 | if (!SplatValue) |
939 | PatternValue = getMemSetPatternValue(StoredVal, DL); |
940 | |
941 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 942, __PRETTY_FUNCTION__)) |
942 | "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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 942, __PRETTY_FUNCTION__)); |
943 | |
944 | // The trip count of the loop and the base pointer of the addrec SCEV is |
945 | // guaranteed to be loop invariant, which means that it should dominate the |
946 | // header. This allows us to insert code for it in the preheader. |
947 | unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); |
948 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
949 | IRBuilder<> Builder(Preheader->getTerminator()); |
950 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
951 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
952 | |
953 | Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); |
954 | Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); |
955 | |
956 | bool Changed = false; |
957 | const SCEV *Start = Ev->getStart(); |
958 | // Handle negative strided loops. |
959 | if (NegStride) |
960 | Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSize, SE); |
961 | |
962 | // TODO: ideally we should still be able to generate memset if SCEV expander |
963 | // is taught to generate the dependencies at the latest point. |
964 | if (!isSafeToExpand(Start, *SE)) |
965 | return Changed; |
966 | |
967 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
968 | // this into a memset in the loop preheader now if we want. However, this |
969 | // would be unsafe to do if there is anything else in the loop that may read |
970 | // or write to the aliased location. Check for any overlap by generating the |
971 | // base pointer and checking the region. |
972 | Value *BasePtr = |
973 | Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); |
974 | |
975 | // From here on out, conservatively report to the pass manager that we've |
976 | // changed the IR, even if we later clean up these added instructions. There |
977 | // may be structural differences e.g. in the order of use lists not accounted |
978 | // for in just a textual dump of the IR. This is written as a variable, even |
979 | // though statically all the places this dominates could be replaced with |
980 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
981 | // the return value will read this comment, and leave them alone. |
982 | Changed = true; |
983 | |
984 | if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
985 | StoreSize, *AA, Stores)) |
986 | return Changed; |
987 | |
988 | if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) |
989 | return Changed; |
990 | |
991 | // Okay, everything looks good, insert the memset. |
992 | |
993 | const SCEV *NumBytesS = |
994 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
995 | |
996 | // TODO: ideally we should still be able to generate memset if SCEV expander |
997 | // is taught to generate the dependencies at the latest point. |
998 | if (!isSafeToExpand(NumBytesS, *SE)) |
999 | return Changed; |
1000 | |
1001 | Value *NumBytes = |
1002 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
1003 | |
1004 | CallInst *NewCall; |
1005 | if (SplatValue) { |
1006 | NewCall = Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, |
1007 | MaybeAlign(StoreAlignment)); |
1008 | } else { |
1009 | // Everything is emitted in default address space |
1010 | Type *Int8PtrTy = DestInt8PtrTy; |
1011 | |
1012 | Module *M = TheStore->getModule(); |
1013 | StringRef FuncName = "memset_pattern16"; |
1014 | FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(), |
1015 | Int8PtrTy, Int8PtrTy, IntIdxTy); |
1016 | inferLibFuncAttributes(M, FuncName, *TLI); |
1017 | |
1018 | // Otherwise we should form a memset_pattern16. PatternValue is known to be |
1019 | // an constant array of 16-bytes. Plop the value into a mergable global. |
1020 | GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, |
1021 | GlobalValue::PrivateLinkage, |
1022 | PatternValue, ".memset_pattern"); |
1023 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
1024 | GV->setAlignment(Align(16)); |
1025 | Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); |
1026 | NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); |
1027 | } |
1028 | NewCall->setDebugLoc(TheStore->getDebugLoc()); |
1029 | |
1030 | if (MSSAU) { |
1031 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1032 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1033 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1034 | } |
1035 | |
1036 | 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) |
1037 | << " 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) |
1038 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"; } } while (false); |
1039 | |
1040 | ORE.emit([&]() { |
1041 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStridedStore", |
1042 | NewCall->getDebugLoc(), Preheader) |
1043 | << "Transformed loop-strided store into a call to " |
1044 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1045 | << "() function"; |
1046 | }); |
1047 | |
1048 | // Okay, the memset has been formed. Zap the original store and anything that |
1049 | // feeds into it. |
1050 | for (auto *I : Stores) { |
1051 | if (MSSAU) |
1052 | MSSAU->removeMemoryAccess(I, true); |
1053 | deleteDeadInstruction(I); |
1054 | } |
1055 | if (MSSAU && VerifyMemorySSA) |
1056 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1057 | ++NumMemSet; |
1058 | ExpCleaner.markResultUsed(); |
1059 | return true; |
1060 | } |
1061 | |
1062 | /// If the stored value is a strided load in the same loop with the same stride |
1063 | /// this may be transformable into a memcpy. This kicks in for stuff like |
1064 | /// for (i) A[i] = B[i]; |
1065 | bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, |
1066 | const SCEV *BECount) { |
1067 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1067, __PRETTY_FUNCTION__)); |
1068 | |
1069 | Value *StorePtr = SI->getPointerOperand(); |
1070 | const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
1071 | APInt Stride = getStoreStride(StoreEv); |
1072 | unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); |
1073 | bool NegStride = StoreSize == -Stride; |
1074 | |
1075 | // The store must be feeding a non-volatile load. |
1076 | LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); |
1077 | 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-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1077, __PRETTY_FUNCTION__)); |
1078 | |
1079 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
1080 | // loop, which indicates a strided load. If we have something else, it's a |
1081 | // random load we can't handle. |
1082 | const SCEVAddRecExpr *LoadEv = |
1083 | cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); |
1084 | |
1085 | // The trip count of the loop and the base pointer of the addrec SCEV is |
1086 | // guaranteed to be loop invariant, which means that it should dominate the |
1087 | // header. This allows us to insert code for it in the preheader. |
1088 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
1089 | IRBuilder<> Builder(Preheader->getTerminator()); |
1090 | SCEVExpander Expander(*SE, *DL, "loop-idiom"); |
1091 | |
1092 | SCEVExpanderCleaner ExpCleaner(Expander, *DT); |
1093 | |
1094 | bool Changed = false; |
1095 | const SCEV *StrStart = StoreEv->getStart(); |
1096 | unsigned StrAS = SI->getPointerAddressSpace(); |
1097 | Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); |
1098 | |
1099 | // Handle negative strided loops. |
1100 | if (NegStride) |
1101 | StrStart = getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSize, SE); |
1102 | |
1103 | // Okay, we have a strided store "p[i]" of a loaded value. We can turn |
1104 | // this into a memcpy in the loop preheader now if we want. However, this |
1105 | // would be unsafe to do if there is anything else in the loop that may read |
1106 | // or write the memory region we're storing to. This includes the load that |
1107 | // feeds the stores. Check for an alias by generating the base address and |
1108 | // checking everything. |
1109 | Value *StoreBasePtr = Expander.expandCodeFor( |
1110 | StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); |
1111 | |
1112 | // From here on out, conservatively report to the pass manager that we've |
1113 | // changed the IR, even if we later clean up these added instructions. There |
1114 | // may be structural differences e.g. in the order of use lists not accounted |
1115 | // for in just a textual dump of the IR. This is written as a variable, even |
1116 | // though statically all the places this dominates could be replaced with |
1117 | // 'true', with the hope that anyone trying to be clever / "more precise" with |
1118 | // the return value will read this comment, and leave them alone. |
1119 | Changed = true; |
1120 | |
1121 | SmallPtrSet<Instruction *, 1> Stores; |
1122 | Stores.insert(SI); |
1123 | if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, |
1124 | StoreSize, *AA, Stores)) |
1125 | return Changed; |
1126 | |
1127 | const SCEV *LdStart = LoadEv->getStart(); |
1128 | unsigned LdAS = LI->getPointerAddressSpace(); |
1129 | |
1130 | // Handle negative strided loops. |
1131 | if (NegStride) |
1132 | LdStart = getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSize, SE); |
1133 | |
1134 | // For a memcpy, we have to make sure that the input array is not being |
1135 | // mutated by the loop. |
1136 | Value *LoadBasePtr = Expander.expandCodeFor( |
1137 | LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); |
1138 | |
1139 | if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, |
1140 | StoreSize, *AA, Stores)) |
1141 | return Changed; |
1142 | |
1143 | if (avoidLIRForMultiBlockLoop()) |
1144 | return Changed; |
1145 | |
1146 | // Okay, everything is safe, we can transform this! |
1147 | |
1148 | const SCEV *NumBytesS = |
1149 | getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE); |
1150 | |
1151 | Value *NumBytes = |
1152 | Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); |
1153 | |
1154 | CallInst *NewCall = nullptr; |
1155 | // Check whether to generate an unordered atomic memcpy: |
1156 | // If the load or store are atomic, then they must necessarily be unordered |
1157 | // by previous checks. |
1158 | if (!SI->isAtomic() && !LI->isAtomic()) |
1159 | NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr, |
1160 | LI->getAlign(), NumBytes); |
1161 | else { |
1162 | // We cannot allow unaligned ops for unordered load/store, so reject |
1163 | // anything where the alignment isn't at least the element size. |
1164 | const Align StoreAlign = SI->getAlign(); |
1165 | const Align LoadAlign = LI->getAlign(); |
1166 | if (StoreAlign < StoreSize || LoadAlign < StoreSize) |
1167 | return Changed; |
1168 | |
1169 | // If the element.atomic memcpy is not lowered into explicit |
1170 | // loads/stores later, then it will be lowered into an element-size |
1171 | // specific lib call. If the lib call doesn't exist for our store size, then |
1172 | // we shouldn't generate the memcpy. |
1173 | if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) |
1174 | return Changed; |
1175 | |
1176 | // Create the call. |
1177 | // Note that unordered atomic loads/stores are *required* by the spec to |
1178 | // have an alignment but non-atomic loads/stores may not. |
1179 | NewCall = Builder.CreateElementUnorderedAtomicMemCpy( |
1180 | StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, |
1181 | StoreSize); |
1182 | } |
1183 | NewCall->setDebugLoc(SI->getDebugLoc()); |
1184 | |
1185 | if (MSSAU) { |
1186 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( |
1187 | NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); |
1188 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); |
1189 | } |
1190 | |
1191 | 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) |
1192 | << " 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) |
1193 | << " 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) |
1194 | << "\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); |
1195 | |
1196 | ORE.emit([&]() { |
1197 | return OptimizationRemark(DEBUG_TYPE"loop-idiom", "ProcessLoopStoreOfLoopLoad", |
1198 | NewCall->getDebugLoc(), Preheader) |
1199 | << "Formed a call to " |
1200 | << ore::NV("NewFunction", NewCall->getCalledFunction()) |
1201 | << "() function"; |
1202 | }); |
1203 | |
1204 | // Okay, the memcpy has been formed. Zap the original store and anything that |
1205 | // feeds into it. |
1206 | if (MSSAU) |
1207 | MSSAU->removeMemoryAccess(SI, true); |
1208 | deleteDeadInstruction(SI); |
1209 | if (MSSAU && VerifyMemorySSA) |
1210 | MSSAU->getMemorySSA()->verifyMemorySSA(); |
1211 | ++NumMemCpy; |
1212 | ExpCleaner.markResultUsed(); |
1213 | return true; |
1214 | } |
1215 | |
1216 | // When compiling for codesize we avoid idiom recognition for a multi-block loop |
1217 | // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. |
1218 | // |
1219 | bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, |
1220 | bool IsLoopMemset) { |
1221 | if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { |
1222 | if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) { |
1223 | 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) |
1224 | << " : 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) |
1225 | << " 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); |
1226 | return true; |
1227 | } |
1228 | } |
1229 | |
1230 | return false; |
1231 | } |
1232 | |
1233 | bool LoopIdiomRecognize::runOnNoncountableLoop() { |
1234 | 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) |
1235 | << 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) |
1236 | << "] 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) |
1237 | << 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); |
1238 | |
1239 | return recognizePopcount() || recognizeAndInsertFFS() || |
1240 | recognizeShiftUntilBitTest(); |
1241 | } |
1242 | |
1243 | /// Check if the given conditional branch is based on the comparison between |
1244 | /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is |
1245 | /// true), the control yields to the loop entry. If the branch matches the |
1246 | /// behavior, the variable involved in the comparison is returned. This function |
1247 | /// will be called to see if the precondition and postcondition of the loop are |
1248 | /// in desirable form. |
1249 | static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, |
1250 | bool JmpOnZero = false) { |
1251 | if (!BI || !BI->isConditional()) |
1252 | return nullptr; |
1253 | |
1254 | ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); |
1255 | if (!Cond) |
1256 | return nullptr; |
1257 | |
1258 | ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1)); |
1259 | if (!CmpZero || !CmpZero->isZero()) |
1260 | return nullptr; |
1261 | |
1262 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
1263 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
1264 | if (JmpOnZero) |
1265 | std::swap(TrueSucc, FalseSucc); |
1266 | |
1267 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
1268 | if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || |
1269 | (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) |
1270 | return Cond->getOperand(0); |
1271 | |
1272 | return nullptr; |
1273 | } |
1274 | |
1275 | // Check if the recurrence variable `VarX` is in the right form to create |
1276 | // the idiom. Returns the value coerced to a PHINode if so. |
1277 | static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, |
1278 | BasicBlock *LoopEntry) { |
1279 | auto *PhiX = dyn_cast<PHINode>(VarX); |
1280 | if (PhiX && PhiX->getParent() == LoopEntry && |
1281 | (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) |
1282 | return PhiX; |
1283 | return nullptr; |
1284 | } |
1285 | |
1286 | /// Return true iff the idiom is detected in the loop. |
1287 | /// |
1288 | /// Additionally: |
1289 | /// 1) \p CntInst is set to the instruction counting the population bit. |
1290 | /// 2) \p CntPhi is set to the corresponding phi node. |
1291 | /// 3) \p Var is set to the value whose population bits are being counted. |
1292 | /// |
1293 | /// The core idiom we are trying to detect is: |
1294 | /// \code |
1295 | /// if (x0 != 0) |
1296 | /// goto loop-exit // the precondition of the loop |
1297 | /// cnt0 = init-val; |
1298 | /// do { |
1299 | /// x1 = phi (x0, x2); |
1300 | /// cnt1 = phi(cnt0, cnt2); |
1301 | /// |
1302 | /// cnt2 = cnt1 + 1; |
1303 | /// ... |
1304 | /// x2 = x1 & (x1 - 1); |
1305 | /// ... |
1306 | /// } while(x != 0); |
1307 | /// |
1308 | /// loop-exit: |
1309 | /// \endcode |
1310 | static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, |
1311 | Instruction *&CntInst, PHINode *&CntPhi, |
1312 | Value *&Var) { |
1313 | // step 1: Check to see if the look-back branch match this pattern: |
1314 | // "if (a!=0) goto loop-entry". |
1315 | BasicBlock *LoopEntry; |
1316 | Instruction *DefX2, *CountInst; |
1317 | Value *VarX1, *VarX0; |
1318 | PHINode *PhiX, *CountPhi; |
1319 | |
1320 | DefX2 = CountInst = nullptr; |
1321 | VarX1 = VarX0 = nullptr; |
1322 | PhiX = CountPhi = nullptr; |
1323 | LoopEntry = *(CurLoop->block_begin()); |
1324 | |
1325 | // step 1: Check if the loop-back branch is in desirable form. |
1326 | { |
1327 | if (Value *T = matchCondition( |
1328 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1329 | DefX2 = dyn_cast<Instruction>(T); |
1330 | else |
1331 | return false; |
1332 | } |
1333 | |
1334 | // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" |
1335 | { |
1336 | if (!DefX2 || DefX2->getOpcode() != Instruction::And) |
1337 | return false; |
1338 | |
1339 | BinaryOperator *SubOneOp; |
1340 | |
1341 | if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0)))) |
1342 | VarX1 = DefX2->getOperand(1); |
1343 | else { |
1344 | VarX1 = DefX2->getOperand(0); |
1345 | SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1)); |
1346 | } |
1347 | if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) |
1348 | return false; |
1349 | |
1350 | ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1)); |
1351 | if (!Dec || |
1352 | !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || |
1353 | (SubOneOp->getOpcode() == Instruction::Add && |
1354 | Dec->isMinusOne()))) { |
1355 | return false; |
1356 | } |
1357 | } |
1358 | |
1359 | // step 3: Check the recurrence of variable X |
1360 | PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); |
1361 | if (!PhiX) |
1362 | return false; |
1363 | |
1364 | // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 |
1365 | { |
1366 | CountInst = nullptr; |
1367 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
1368 | IterE = LoopEntry->end(); |
1369 | Iter != IterE; Iter++) { |
1370 | Instruction *Inst = &*Iter; |
1371 | if (Inst->getOpcode() != Instruction::Add) |
1372 | continue; |
1373 | |
1374 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
1375 | if (!Inc || !Inc->isOne()) |
1376 | continue; |
1377 | |
1378 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
1379 | if (!Phi) |
1380 | continue; |
1381 | |
1382 | // Check if the result of the instruction is live of the loop. |
1383 | bool LiveOutLoop = false; |
1384 | for (User *U : Inst->users()) { |
1385 | if ((cast<Instruction>(U))->getParent() != LoopEntry) { |
1386 | LiveOutLoop = true; |
1387 | break; |
1388 | } |
1389 | } |
1390 | |
1391 | if (LiveOutLoop) { |
1392 | CountInst = Inst; |
1393 | CountPhi = Phi; |
1394 | break; |
1395 | } |
1396 | } |
1397 | |
1398 | if (!CountInst) |
1399 | return false; |
1400 | } |
1401 | |
1402 | // step 5: check if the precondition is in this form: |
1403 | // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" |
1404 | { |
1405 | auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1406 | Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); |
1407 | if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) |
1408 | return false; |
1409 | |
1410 | CntInst = CountInst; |
1411 | CntPhi = CountPhi; |
1412 | Var = T; |
1413 | } |
1414 | |
1415 | return true; |
1416 | } |
1417 | |
1418 | /// Return true if the idiom is detected in the loop. |
1419 | /// |
1420 | /// Additionally: |
1421 | /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) |
1422 | /// or nullptr if there is no such. |
1423 | /// 2) \p CntPhi is set to the corresponding phi node |
1424 | /// or nullptr if there is no such. |
1425 | /// 3) \p Var is set to the value whose CTLZ could be used. |
1426 | /// 4) \p DefX is set to the instruction calculating Loop exit condition. |
1427 | /// |
1428 | /// The core idiom we are trying to detect is: |
1429 | /// \code |
1430 | /// if (x0 == 0) |
1431 | /// goto loop-exit // the precondition of the loop |
1432 | /// cnt0 = init-val; |
1433 | /// do { |
1434 | /// x = phi (x0, x.next); //PhiX |
1435 | /// cnt = phi(cnt0, cnt.next); |
1436 | /// |
1437 | /// cnt.next = cnt + 1; |
1438 | /// ... |
1439 | /// x.next = x >> 1; // DefX |
1440 | /// ... |
1441 | /// } while(x.next != 0); |
1442 | /// |
1443 | /// loop-exit: |
1444 | /// \endcode |
1445 | static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, |
1446 | Intrinsic::ID &IntrinID, Value *&InitX, |
1447 | Instruction *&CntInst, PHINode *&CntPhi, |
1448 | Instruction *&DefX) { |
1449 | BasicBlock *LoopEntry; |
1450 | Value *VarX = nullptr; |
1451 | |
1452 | DefX = nullptr; |
1453 | CntInst = nullptr; |
1454 | CntPhi = nullptr; |
1455 | LoopEntry = *(CurLoop->block_begin()); |
1456 | |
1457 | // step 1: Check if the loop-back branch is in desirable form. |
1458 | if (Value *T = matchCondition( |
1459 | dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry)) |
1460 | DefX = dyn_cast<Instruction>(T); |
1461 | else |
1462 | return false; |
1463 | |
1464 | // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" |
1465 | if (!DefX || !DefX->isShift()) |
1466 | return false; |
1467 | IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : |
1468 | Intrinsic::ctlz; |
1469 | ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)); |
1470 | if (!Shft || !Shft->isOne()) |
1471 | return false; |
1472 | VarX = DefX->getOperand(0); |
1473 | |
1474 | // step 3: Check the recurrence of variable X |
1475 | PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); |
1476 | if (!PhiX) |
1477 | return false; |
1478 | |
1479 | InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); |
1480 | |
1481 | // Make sure the initial value can't be negative otherwise the ashr in the |
1482 | // loop might never reach zero which would make the loop infinite. |
1483 | if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) |
1484 | return false; |
1485 | |
1486 | // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 |
1487 | // or cnt.next = cnt + -1. |
1488 | // TODO: We can skip the step. If loop trip count is known (CTLZ), |
1489 | // then all uses of "cnt.next" could be optimized to the trip count |
1490 | // plus "cnt0". Currently it is not optimized. |
1491 | // This step could be used to detect POPCNT instruction: |
1492 | // cnt.next = cnt + (x.next & 1) |
1493 | for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(), |
1494 | IterE = LoopEntry->end(); |
1495 | Iter != IterE; Iter++) { |
1496 | Instruction *Inst = &*Iter; |
1497 | if (Inst->getOpcode() != Instruction::Add) |
1498 | continue; |
1499 | |
1500 | ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1)); |
1501 | if (!Inc || (!Inc->isOne() && !Inc->isMinusOne())) |
1502 | continue; |
1503 | |
1504 | PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry); |
1505 | if (!Phi) |
1506 | continue; |
1507 | |
1508 | CntInst = Inst; |
1509 | CntPhi = Phi; |
1510 | break; |
1511 | } |
1512 | if (!CntInst) |
1513 | return false; |
1514 | |
1515 | return true; |
1516 | } |
1517 | |
1518 | /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop |
1519 | /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new |
1520 | /// trip count returns true; otherwise, returns false. |
1521 | bool LoopIdiomRecognize::recognizeAndInsertFFS() { |
1522 | // Give up if the loop has multiple blocks or multiple backedges. |
1523 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1524 | return false; |
1525 | |
1526 | Intrinsic::ID IntrinID; |
1527 | Value *InitX; |
1528 | Instruction *DefX = nullptr; |
1529 | PHINode *CntPhi = nullptr; |
1530 | Instruction *CntInst = nullptr; |
1531 | // Help decide if transformation is profitable. For ShiftUntilZero idiom, |
1532 | // this is always 6. |
1533 | size_t IdiomCanonicalSize = 6; |
1534 | |
1535 | if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, |
1536 | CntInst, CntPhi, DefX)) |
1537 | return false; |
1538 | |
1539 | bool IsCntPhiUsedOutsideLoop = false; |
1540 | for (User *U : CntPhi->users()) |
1541 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1542 | IsCntPhiUsedOutsideLoop = true; |
1543 | break; |
1544 | } |
1545 | bool IsCntInstUsedOutsideLoop = false; |
1546 | for (User *U : CntInst->users()) |
1547 | if (!CurLoop->contains(cast<Instruction>(U))) { |
1548 | IsCntInstUsedOutsideLoop = true; |
1549 | break; |
1550 | } |
1551 | // If both CntInst and CntPhi are used outside the loop the profitability |
1552 | // is questionable. |
1553 | if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) |
1554 | return false; |
1555 | |
1556 | // For some CPUs result of CTLZ(X) intrinsic is undefined |
1557 | // when X is 0. If we can not guarantee X != 0, we need to check this |
1558 | // when expand. |
1559 | bool ZeroCheck = false; |
1560 | // It is safe to assume Preheader exist as it was checked in |
1561 | // parent function RunOnLoop. |
1562 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1563 | |
1564 | // If we are using the count instruction outside the loop, make sure we |
1565 | // have a zero check as a precondition. Without the check the loop would run |
1566 | // one iteration for before any check of the input value. This means 0 and 1 |
1567 | // would have identical behavior in the original loop and thus |
1568 | if (!IsCntPhiUsedOutsideLoop) { |
1569 | auto *PreCondBB = PH->getSinglePredecessor(); |
1570 | if (!PreCondBB) |
1571 | return false; |
1572 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1573 | if (!PreCondBI) |
1574 | return false; |
1575 | if (matchCondition(PreCondBI, PH) != InitX) |
1576 | return false; |
1577 | ZeroCheck = true; |
1578 | } |
1579 | |
1580 | // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always |
1581 | // profitable if we delete the loop. |
1582 | |
1583 | // the loop has only 6 instructions: |
1584 | // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] |
1585 | // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] |
1586 | // %shr = ashr %n.addr.0, 1 |
1587 | // %tobool = icmp eq %shr, 0 |
1588 | // %inc = add nsw %i.0, 1 |
1589 | // br i1 %tobool |
1590 | |
1591 | const Value *Args[] = {InitX, |
1592 | ConstantInt::getBool(InitX->getContext(), ZeroCheck)}; |
1593 | |
1594 | // @llvm.dbg doesn't count as they have no semantic effect. |
1595 | auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); |
1596 | uint32_t HeaderSize = |
1597 | std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); |
1598 | |
1599 | IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); |
1600 | InstructionCost Cost = |
1601 | TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); |
1602 | if (HeaderSize != IdiomCanonicalSize && |
1603 | Cost > TargetTransformInfo::TCC_Basic) |
1604 | return false; |
1605 | |
1606 | transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, |
1607 | DefX->getDebugLoc(), ZeroCheck, |
1608 | IsCntPhiUsedOutsideLoop); |
1609 | return true; |
1610 | } |
1611 | |
1612 | /// Recognizes a population count idiom in a non-countable loop. |
1613 | /// |
1614 | /// If detected, transforms the relevant code to issue the popcount intrinsic |
1615 | /// function call, and returns true; otherwise, returns false. |
1616 | bool LoopIdiomRecognize::recognizePopcount() { |
1617 | if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) |
1618 | return false; |
1619 | |
1620 | // Counting population are usually conducted by few arithmetic instructions. |
1621 | // Such instructions can be easily "absorbed" by vacant slots in a |
1622 | // non-compact loop. Therefore, recognizing popcount idiom only makes sense |
1623 | // in a compact loop. |
1624 | |
1625 | // Give up if the loop has multiple blocks or multiple backedges. |
1626 | if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) |
1627 | return false; |
1628 | |
1629 | BasicBlock *LoopBody = *(CurLoop->block_begin()); |
1630 | if (LoopBody->size() >= 20) { |
1631 | // The loop is too big, bail out. |
1632 | return false; |
1633 | } |
1634 | |
1635 | // It should have a preheader containing nothing but an unconditional branch. |
1636 | BasicBlock *PH = CurLoop->getLoopPreheader(); |
1637 | if (!PH || &PH->front() != PH->getTerminator()) |
1638 | return false; |
1639 | auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator()); |
1640 | if (!EntryBI || EntryBI->isConditional()) |
1641 | return false; |
1642 | |
1643 | // It should have a precondition block where the generated popcount intrinsic |
1644 | // function can be inserted. |
1645 | auto *PreCondBB = PH->getSinglePredecessor(); |
1646 | if (!PreCondBB) |
1647 | return false; |
1648 | auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator()); |
1649 | if (!PreCondBI || PreCondBI->isUnconditional()) |
1650 | return false; |
1651 | |
1652 | Instruction *CntInst; |
1653 | PHINode *CntPhi; |
1654 | Value *Val; |
1655 | if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) |
1656 | return false; |
1657 | |
1658 | transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); |
1659 | return true; |
1660 | } |
1661 | |
1662 | static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
1663 | const DebugLoc &DL) { |
1664 | Value *Ops[] = {Val}; |
1665 | Type *Tys[] = {Val->getType()}; |
1666 | |
1667 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
1668 | Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); |
1669 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
1670 | CI->setDebugLoc(DL); |
1671 | |
1672 | return CI; |
1673 | } |
1674 | |
1675 | static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, |
1676 | const DebugLoc &DL, bool ZeroCheck, |
1677 | Intrinsic::ID IID) { |
1678 | Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)}; |
1679 | Type *Tys[] = {Val->getType()}; |
1680 | |
1681 | Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); |
1682 | Function *Func = Intrinsic::getDeclaration(M, IID, Tys); |
1683 | CallInst *CI = IRBuilder.CreateCall(Func, Ops); |
1684 | CI->setDebugLoc(DL); |
1685 | |
1686 | return CI; |
1687 | } |
1688 | |
1689 | /// Transform the following loop (Using CTLZ, CTTZ is similar): |
1690 | /// loop: |
1691 | /// CntPhi = PHI [Cnt0, CntInst] |
1692 | /// PhiX = PHI [InitX, DefX] |
1693 | /// CntInst = CntPhi + 1 |
1694 | /// DefX = PhiX >> 1 |
1695 | /// LOOP_BODY |
1696 | /// Br: loop if (DefX != 0) |
1697 | /// Use(CntPhi) or Use(CntInst) |
1698 | /// |
1699 | /// Into: |
1700 | /// If CntPhi used outside the loop: |
1701 | /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) |
1702 | /// Count = CountPrev + 1 |
1703 | /// else |
1704 | /// Count = BitWidth(InitX) - CTLZ(InitX) |
1705 | /// loop: |
1706 | /// CntPhi = PHI [Cnt0, CntInst] |
1707 | /// PhiX = PHI [InitX, DefX] |
1708 | /// PhiCount = PHI [Count, Dec] |
1709 | /// CntInst = CntPhi + 1 |
1710 | /// DefX = PhiX >> 1 |
1711 | /// Dec = PhiCount - 1 |
1712 | /// LOOP_BODY |
1713 | /// Br: loop if (Dec != 0) |
1714 | /// Use(CountPrev + Cnt0) // Use(CntPhi) |
1715 | /// or |
1716 | /// Use(Count + Cnt0) // Use(CntInst) |
1717 | /// |
1718 | /// If LOOP_BODY is empty the loop will be deleted. |
1719 | /// If CntInst and DefX are not used in LOOP_BODY they will be removed. |
1720 | void LoopIdiomRecognize::transformLoopToCountable( |
1721 | Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, |
1722 | PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, |
1723 | bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { |
1724 | BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator()); |
1725 | |
1726 | // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block |
1727 | IRBuilder<> Builder(PreheaderBr); |
1728 | Builder.SetCurrentDebugLocation(DL); |
1729 | |
1730 | // If there are no uses of CntPhi crate: |
1731 | // Count = BitWidth - CTLZ(InitX); |
1732 | // NewCount = Count; |
1733 | // If there are uses of CntPhi create: |
1734 | // NewCount = BitWidth - CTLZ(InitX >> 1); |
1735 | // Count = NewCount + 1; |
1736 | Value *InitXNext; |
1737 | if (IsCntPhiUsedOutsideLoop) { |
1738 | if (DefX->getOpcode() == Instruction::AShr) |
1739 | InitXNext = Builder.CreateAShr(InitX, 1); |
1740 | else if (DefX->getOpcode() == Instruction::LShr) |
1741 | InitXNext = Builder.CreateLShr(InitX, 1); |
1742 | else if (DefX->getOpcode() == Instruction::Shl) // cttz |
1743 | InitXNext = Builder.CreateShl(InitX, 1); |
1744 | else |
1745 | llvm_unreachable("Unexpected opcode!")::llvm::llvm_unreachable_internal("Unexpected opcode!", "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1745); |
1746 | } else |
1747 | InitXNext = InitX; |
1748 | Value *Count = |
1749 | createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); |
1750 | Type *CountTy = Count->getType(); |
1751 | Count = Builder.CreateSub( |
1752 | ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count); |
1753 | Value *NewCount = Count; |
1754 | if (IsCntPhiUsedOutsideLoop) |
1755 | Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1)); |
1756 | |
1757 | NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType()); |
1758 | |
1759 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); |
1760 | if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) { |
1761 | // If the counter was being incremented in the loop, add NewCount to the |
1762 | // counter's initial value, but only if the initial value is not zero. |
1763 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
1764 | if (!InitConst || !InitConst->isZero()) |
1765 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
1766 | } else { |
1767 | // If the count was being decremented in the loop, subtract NewCount from |
1768 | // the counter's initial value. |
1769 | NewCount = Builder.CreateSub(CntInitVal, NewCount); |
1770 | } |
1771 | |
1772 | // Step 2: Insert new IV and loop condition: |
1773 | // loop: |
1774 | // ... |
1775 | // PhiCount = PHI [Count, Dec] |
1776 | // ... |
1777 | // Dec = PhiCount - 1 |
1778 | // ... |
1779 | // Br: loop if (Dec != 0) |
1780 | BasicBlock *Body = *(CurLoop->block_begin()); |
1781 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
1782 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
1783 | |
1784 | PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front()); |
1785 | |
1786 | Builder.SetInsertPoint(LbCond); |
1787 | Instruction *TcDec = cast<Instruction>(Builder.CreateSub( |
1788 | TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true)); |
1789 | |
1790 | TcPhi->addIncoming(Count, Preheader); |
1791 | TcPhi->addIncoming(TcDec, Body); |
1792 | |
1793 | CmpInst::Predicate Pred = |
1794 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
1795 | LbCond->setPredicate(Pred); |
1796 | LbCond->setOperand(0, TcDec); |
1797 | LbCond->setOperand(1, ConstantInt::get(CountTy, 0)); |
1798 | |
1799 | // Step 3: All the references to the original counter outside |
1800 | // the loop are replaced with the NewCount |
1801 | if (IsCntPhiUsedOutsideLoop) |
1802 | CntPhi->replaceUsesOutsideBlock(NewCount, Body); |
1803 | else |
1804 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
1805 | |
1806 | // step 4: Forget the "non-computable" trip-count SCEV associated with the |
1807 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
1808 | SE->forgetLoop(CurLoop); |
1809 | } |
1810 | |
1811 | void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, |
1812 | Instruction *CntInst, |
1813 | PHINode *CntPhi, Value *Var) { |
1814 | BasicBlock *PreHead = CurLoop->getLoopPreheader(); |
1815 | auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator()); |
1816 | const DebugLoc &DL = CntInst->getDebugLoc(); |
1817 | |
1818 | // Assuming before transformation, the loop is following: |
1819 | // if (x) // the precondition |
1820 | // do { cnt++; x &= x - 1; } while(x); |
1821 | |
1822 | // Step 1: Insert the ctpop instruction at the end of the precondition block |
1823 | IRBuilder<> Builder(PreCondBr); |
1824 | Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; |
1825 | { |
1826 | PopCnt = createPopcntIntrinsic(Builder, Var, DL); |
1827 | NewCount = PopCntZext = |
1828 | Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType())); |
1829 | |
1830 | if (NewCount != PopCnt) |
1831 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
1832 | |
1833 | // TripCnt is exactly the number of iterations the loop has |
1834 | TripCnt = NewCount; |
1835 | |
1836 | // If the population counter's initial value is not zero, insert Add Inst. |
1837 | Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); |
1838 | ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal); |
1839 | if (!InitConst || !InitConst->isZero()) { |
1840 | NewCount = Builder.CreateAdd(NewCount, CntInitVal); |
1841 | (cast<Instruction>(NewCount))->setDebugLoc(DL); |
1842 | } |
1843 | } |
1844 | |
1845 | // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to |
1846 | // "if (NewCount == 0) loop-exit". Without this change, the intrinsic |
1847 | // function would be partial dead code, and downstream passes will drag |
1848 | // it back from the precondition block to the preheader. |
1849 | { |
1850 | ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition()); |
1851 | |
1852 | Value *Opnd0 = PopCntZext; |
1853 | Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); |
1854 | if (PreCond->getOperand(0) != Var) |
1855 | std::swap(Opnd0, Opnd1); |
1856 | |
1857 | ICmpInst *NewPreCond = cast<ICmpInst>( |
1858 | Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); |
1859 | PreCondBr->setCondition(NewPreCond); |
1860 | |
1861 | RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); |
1862 | } |
1863 | |
1864 | // Step 3: Note that the population count is exactly the trip count of the |
1865 | // loop in question, which enable us to convert the loop from noncountable |
1866 | // loop into a countable one. The benefit is twofold: |
1867 | // |
1868 | // - If the loop only counts population, the entire loop becomes dead after |
1869 | // the transformation. It is a lot easier to prove a countable loop dead |
1870 | // than to prove a noncountable one. (In some C dialects, an infinite loop |
1871 | // isn't dead even if it computes nothing useful. In general, DCE needs |
1872 | // to prove a noncountable loop finite before safely delete it.) |
1873 | // |
1874 | // - If the loop also performs something else, it remains alive. |
1875 | // Since it is transformed to countable form, it can be aggressively |
1876 | // optimized by some optimizations which are in general not applicable |
1877 | // to a noncountable loop. |
1878 | // |
1879 | // After this step, this loop (conceptually) would look like following: |
1880 | // newcnt = __builtin_ctpop(x); |
1881 | // t = newcnt; |
1882 | // if (x) |
1883 | // do { cnt++; x &= x-1; t--) } while (t > 0); |
1884 | BasicBlock *Body = *(CurLoop->block_begin()); |
1885 | { |
1886 | auto *LbBr = cast<BranchInst>(Body->getTerminator()); |
1887 | ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition()); |
1888 | Type *Ty = TripCnt->getType(); |
1889 | |
1890 | PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); |
1891 | |
1892 | Builder.SetInsertPoint(LbCond); |
1893 | Instruction *TcDec = cast<Instruction>( |
1894 | Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), |
1895 | "tcdec", false, true)); |
1896 | |
1897 | TcPhi->addIncoming(TripCnt, PreHead); |
1898 | TcPhi->addIncoming(TcDec, Body); |
1899 | |
1900 | CmpInst::Predicate Pred = |
1901 | (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; |
1902 | LbCond->setPredicate(Pred); |
1903 | LbCond->setOperand(0, TcDec); |
1904 | LbCond->setOperand(1, ConstantInt::get(Ty, 0)); |
1905 | } |
1906 | |
1907 | // Step 4: All the references to the original population counter outside |
1908 | // the loop are replaced with the NewCount -- the value returned from |
1909 | // __builtin_ctpop(). |
1910 | CntInst->replaceUsesOutsideBlock(NewCount, Body); |
1911 | |
1912 | // step 5: Forget the "non-computable" trip-count SCEV associated with the |
1913 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
1914 | SE->forgetLoop(CurLoop); |
1915 | } |
1916 | |
1917 | /// Match loop-invariant value. |
1918 | template <typename SubPattern_t> struct match_LoopInvariant { |
1919 | SubPattern_t SubPattern; |
1920 | const Loop *L; |
1921 | |
1922 | match_LoopInvariant(const SubPattern_t &SP, const Loop *L) |
1923 | : SubPattern(SP), L(L) {} |
1924 | |
1925 | template <typename ITy> bool match(ITy *V) { |
1926 | return L->isLoopInvariant(V) && SubPattern.match(V); |
1927 | } |
1928 | }; |
1929 | |
1930 | /// Matches if the value is loop-invariant. |
1931 | template <typename Ty> |
1932 | inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) { |
1933 | return match_LoopInvariant<Ty>(M, L); |
1934 | } |
1935 | |
1936 | /// Return true if the idiom is detected in the loop. |
1937 | /// |
1938 | /// The core idiom we are trying to detect is: |
1939 | /// \code |
1940 | /// entry: |
1941 | /// <...> |
1942 | /// %bitmask = shl i32 1, %bitpos |
1943 | /// br label %loop |
1944 | /// |
1945 | /// loop: |
1946 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
1947 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
1948 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
1949 | /// %x.next = shl i32 %x.curr, 1 |
1950 | /// <...> |
1951 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
1952 | /// |
1953 | /// end: |
1954 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
1955 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
1956 | /// <...> |
1957 | /// \endcode |
1958 | static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX, |
1959 | Value *&BitMask, Value *&BitPos, |
1960 | Value *&CurrX, Instruction *&NextX) { |
1961 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-bittest idiom detection.\n" ; } } while (false) |
1962 | " Performing shift-until-bittest idiom detection.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Performing shift-until-bittest idiom detection.\n" ; } } while (false); |
1963 | |
1964 | // Give up if the loop has multiple blocks or multiple backedges. |
1965 | if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { |
1966 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad block/backedge count.\n" ; } } while (false); |
1967 | return false; |
1968 | } |
1969 | |
1970 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
1971 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
1972 | assert(LoopPreheaderBB && "There is always a loop preheader.")((LoopPreheaderBB && "There is always a loop preheader." ) ? static_cast<void> (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 1972, __PRETTY_FUNCTION__)); |
1973 | |
1974 | using namespace PatternMatch; |
1975 | |
1976 | // Step 1: Check if the loop backedge is in desirable form. |
1977 | |
1978 | ICmpInst::Predicate Pred; |
1979 | Value *CmpLHS, *CmpRHS; |
1980 | BasicBlock *TrueBB, *FalseBB; |
1981 | if (!match(LoopHeaderBB->getTerminator(), |
1982 | m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)), |
1983 | m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) { |
1984 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge structure.\n" ; } } while (false); |
1985 | return false; |
1986 | } |
1987 | |
1988 | // Step 2: Check if the backedge's condition is in desirable form. |
1989 | |
1990 | auto MatchVariableBitMask = [&]() { |
1991 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
1992 | match(CmpLHS, |
1993 | m_c_And(m_Value(CurrX), |
1994 | m_CombineAnd( |
1995 | m_Value(BitMask), |
1996 | m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)), |
1997 | CurLoop)))); |
1998 | }; |
1999 | auto MatchConstantBitMask = [&]() { |
2000 | return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && |
2001 | match(CmpLHS, m_And(m_Value(CurrX), |
2002 | m_CombineAnd(m_Value(BitMask), m_Power2()))) && |
2003 | (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask))); |
2004 | }; |
2005 | auto MatchDecomposableConstantBitMask = [&]() { |
2006 | APInt Mask; |
2007 | return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) && |
2008 | ICmpInst::isEquality(Pred) && Mask.isPowerOf2() && |
2009 | (BitMask = ConstantInt::get(CurrX->getType(), Mask)) && |
2010 | (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2())); |
2011 | }; |
2012 | |
2013 | if (!MatchVariableBitMask() && !MatchConstantBitMask() && |
2014 | !MatchDecomposableConstantBitMask()) { |
2015 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge comparison.\n" ; } } while (false); |
2016 | return false; |
2017 | } |
2018 | |
2019 | // Step 3: Check if the recurrence is in desirable form. |
2020 | auto *CurrXPN = dyn_cast<PHINode>(CurrX); |
2021 | if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) { |
2022 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Not an expected PHI node.\n" ; } } while (false); |
2023 | return false; |
2024 | } |
2025 | |
2026 | BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB); |
2027 | NextX = |
2028 | dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB)); |
2029 | |
2030 | if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) { |
2031 | // FIXME: support right-shift? |
2032 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad recurrence.\n" ; } } while (false); |
2033 | return false; |
2034 | } |
2035 | |
2036 | // Step 4: Check if the backedge's destinations are in desirable form. |
2037 | |
2038 | assert(ICmpInst::isEquality(Pred) &&((ICmpInst::isEquality(Pred) && "Should only get equality predicates here." ) ? static_cast<void> (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 2039, __PRETTY_FUNCTION__)) |
2039 | "Should only get equality predicates here.")((ICmpInst::isEquality(Pred) && "Should only get equality predicates here." ) ? static_cast<void> (0) : __assert_fail ("ICmpInst::isEquality(Pred) && \"Should only get equality predicates here.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 2039, __PRETTY_FUNCTION__)); |
2040 | |
2041 | // cmp-br is commutative, so canonicalize to a single variant. |
2042 | if (Pred != ICmpInst::Predicate::ICMP_EQ) { |
2043 | Pred = ICmpInst::getInversePredicate(Pred); |
2044 | std::swap(TrueBB, FalseBB); |
2045 | } |
2046 | |
2047 | // We expect to exit loop when comparison yields false, |
2048 | // so when it yields true we should branch back to loop header. |
2049 | if (TrueBB != LoopHeaderBB) { |
2050 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Bad backedge flow.\n" ; } } while (false); |
2051 | return false; |
2052 | } |
2053 | |
2054 | // Okay, idiom checks out. |
2055 | return true; |
2056 | } |
2057 | |
2058 | /// Look for the following loop: |
2059 | /// \code |
2060 | /// entry: |
2061 | /// <...> |
2062 | /// %bitmask = shl i32 1, %bitpos |
2063 | /// br label %loop |
2064 | /// |
2065 | /// loop: |
2066 | /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] |
2067 | /// %x.curr.bitmasked = and i32 %x.curr, %bitmask |
2068 | /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 |
2069 | /// %x.next = shl i32 %x.curr, 1 |
2070 | /// <...> |
2071 | /// br i1 %x.curr.isbitunset, label %loop, label %end |
2072 | /// |
2073 | /// end: |
2074 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
2075 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
2076 | /// <...> |
2077 | /// \endcode |
2078 | /// |
2079 | /// And transform it into: |
2080 | /// \code |
2081 | /// entry: |
2082 | /// %bitmask = shl i32 1, %bitpos |
2083 | /// %lowbitmask = add i32 %bitmask, -1 |
2084 | /// %mask = or i32 %lowbitmask, %bitmask |
2085 | /// %x.masked = and i32 %x, %mask |
2086 | /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked, |
2087 | /// i1 true) |
2088 | /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros |
2089 | /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1 |
2090 | /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos |
2091 | /// %tripcount = add i32 %backedgetakencount, 1 |
2092 | /// %x.curr = shl i32 %x, %backedgetakencount |
2093 | /// %x.next = shl i32 %x, %tripcount |
2094 | /// br label %loop |
2095 | /// |
2096 | /// loop: |
2097 | /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ] |
2098 | /// %loop.iv.next = add nuw i32 %loop.iv, 1 |
2099 | /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount |
2100 | /// <...> |
2101 | /// br i1 %loop.ivcheck, label %end, label %loop |
2102 | /// |
2103 | /// end: |
2104 | /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> |
2105 | /// %x.next.res = phi i32 [ %x.next, %loop ] <...> |
2106 | /// <...> |
2107 | /// \endcode |
2108 | bool LoopIdiomRecognize::recognizeShiftUntilBitTest() { |
2109 | bool MadeChange = false; |
2110 | |
2111 | Value *X, *BitMask, *BitPos, *XCurr; |
2112 | Instruction *XNext; |
2113 | if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr, |
2114 | XNext)) { |
2115 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detection failed.\n" ; } } while (false) |
2116 | " shift-until-bittest idiom detection failed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detection failed.\n" ; } } while (false); |
2117 | return MadeChange; |
2118 | } |
2119 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom detected!\n" ; } } while (false); |
2120 | |
2121 | // Ok, it is the idiom we were looking for, we *could* transform this loop, |
2122 | // but is it profitable to transform? |
2123 | |
2124 | BasicBlock *LoopHeaderBB = CurLoop->getHeader(); |
2125 | BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); |
2126 | assert(LoopPreheaderBB && "There is always a loop preheader.")((LoopPreheaderBB && "There is always a loop preheader." ) ? static_cast<void> (0) : __assert_fail ("LoopPreheaderBB && \"There is always a loop preheader.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 2126, __PRETTY_FUNCTION__)); |
2127 | |
2128 | BasicBlock *SuccessorBB = CurLoop->getExitBlock(); |
2129 | assert(LoopPreheaderBB && "There is only a single successor.")((LoopPreheaderBB && "There is only a single successor." ) ? static_cast<void> (0) : __assert_fail ("LoopPreheaderBB && \"There is only a single successor.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp" , 2129, __PRETTY_FUNCTION__)); |
2130 | |
2131 | IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); |
2132 | Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc()); |
2133 | |
2134 | Intrinsic::ID IntrID = Intrinsic::ctlz; |
2135 | Type *Ty = X->getType(); |
2136 | unsigned Bitwidth = Ty->getScalarSizeInBits(); |
2137 | |
2138 | TargetTransformInfo::TargetCostKind CostKind = |
2139 | TargetTransformInfo::TCK_SizeAndLatency; |
2140 | |
2141 | // The rewrite is considered to be unprofitable iff and only iff the |
2142 | // intrinsic/shift we'll use are not cheap. Note that we are okay with *just* |
2143 | // making the loop countable, even if nothing else changes. |
2144 | IntrinsicCostAttributes Attrs( |
2145 | IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getTrue()}); |
2146 | InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); |
2147 | if (Cost > TargetTransformInfo::TCC_Basic) { |
2148 | LLVM_DEBUG(dbgs() << DEBUG_TYPEdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false) |
2149 | " Intrinsic is too costly, not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Intrinsic is too costly, not beneficial\n" ; } } while (false); |
2150 | return MadeChange; |
2151 | } |
2152 | if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) > |
2153 | TargetTransformInfo::TCC_Basic) { |
2154 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " Shift is too costly, not beneficial\n" ; } } while (false); |
2155 | return MadeChange; |
2156 | } |
2157 | |
2158 | // Ok, transform appears worthwhile. |
2159 | MadeChange = true; |
2160 | |
2161 | // Step 1: Compute the loop trip count. |
2162 | |
2163 | Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty), |
2164 | BitPos->getName() + ".lowbitmask"); |
2165 | Value *Mask = |
2166 | Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask"); |
2167 | Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked"); |
2168 | CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic( |
2169 | IntrID, Ty, {XMasked, /*is_zero_undef=*/Builder.getTrue()}, |
2170 | /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros"); |
2171 | Value *XMaskedNumActiveBits = Builder.CreateSub( |
2172 | ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros, |
2173 | XMasked->getName() + ".numactivebits", /*HasNUW=*/true, |
2174 | /*HasNSW=*/Bitwidth != 2); |
2175 | Value *XMaskedLeadingOnePos = |
2176 | Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty), |
2177 | XMasked->getName() + ".leadingonepos", /*HasNUW=*/false, |
2178 | /*HasNSW=*/Bitwidth > 2); |
2179 | |
2180 | Value *LoopBackedgeTakenCount = Builder.CreateSub( |
2181 | BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount", |
2182 | /*HasNUW=*/true, /*HasNSW=*/true); |
2183 | // We know loop's backedge-taken count, but what's loop's trip count? |
2184 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
2185 | Value *LoopTripCount = |
2186 | Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), |
2187 | CurLoop->getName() + ".tripcount", /*HasNUW=*/true, |
2188 | /*HasNSW=*/Bitwidth != 2); |
2189 | |
2190 | // Step 2: Compute the recurrence's final value without a loop. |
2191 | |
2192 | // NewX is always safe to compute, because `LoopBackedgeTakenCount` |
2193 | // will always be smaller than `bitwidth(X)`, i.e. we never get poison. |
2194 | Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount); |
2195 | NewX->takeName(XCurr); |
2196 | if (auto *I = dyn_cast<Instruction>(NewX)) |
2197 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
2198 | |
2199 | Value *NewXNext; |
2200 | // Rewriting XNext is more complicated, however, because `X << LoopTripCount` |
2201 | // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen |
2202 | // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know |
2203 | // that isn't the case, we'll need to emit an alternative, safe IR. |
2204 | if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() || |
2205 | PatternMatch::match( |
2206 | BitPos, PatternMatch::m_SpecificInt_ICMP( |
2207 | ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(), |
2208 | Ty->getScalarSizeInBits() - 1)))) |
2209 | NewXNext = Builder.CreateShl(X, LoopTripCount); |
2210 | else { |
2211 | // Otherwise, just additionally shift by one. It's the smallest solution, |
2212 | // alternatively, we could check that NewX is INT_MIN (or BitPos is ) |
2213 | // and select 0 instead. |
2214 | NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1)); |
2215 | } |
2216 | |
2217 | NewXNext->takeName(XNext); |
2218 | if (auto *I = dyn_cast<Instruction>(NewXNext)) |
2219 | I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); |
2220 | |
2221 | // Step 3: Adjust the successor basic block to recieve the computed |
2222 | // recurrence's final value instead of the recurrence itself. |
2223 | |
2224 | XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB); |
2225 | XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB); |
2226 | |
2227 | // Step 4: Rewrite the loop into a countable form, with canonical IV. |
2228 | |
2229 | // The new canonical induction variable. |
2230 | Builder.SetInsertPoint(&LoopHeaderBB->front()); |
2231 | auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); |
2232 | |
2233 | // The induction itself. |
2234 | // Note that while NUW is always safe, while NSW is only for bitwidths != 2. |
2235 | Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); |
2236 | auto *IVNext = |
2237 | Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next", |
2238 | /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); |
2239 | |
2240 | // The loop trip count check. |
2241 | auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount, |
2242 | CurLoop->getName() + ".ivcheck"); |
2243 | Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB); |
2244 | LoopHeaderBB->getTerminator()->eraseFromParent(); |
2245 | |
2246 | // Populate the IV PHI. |
2247 | IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); |
2248 | IV->addIncoming(IVNext, LoopHeaderBB); |
2249 | |
2250 | // Step 5: Forget the "non-computable" trip-count SCEV associated with the |
2251 | // loop. The loop would otherwise not be deleted even if it becomes empty. |
2252 | |
2253 | SE->forgetLoop(CurLoop); |
2254 | |
2255 | // Other passes will take care of actually deleting the loop if possible. |
2256 | |
2257 | LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-idiom")) { dbgs() << "loop-idiom" " shift-until-bittest idiom optimized!\n" ; } } while (false); |
2258 | |
2259 | ++NumShiftUntilBitTest; |
2260 | return MadeChange; |
2261 | } |