File: | lib/Transforms/Scalar/LICM.cpp |
Warning: | line 1092, column 7 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This pass performs loop invariant code motion, attempting to remove as much | |||
11 | // code from the body of a loop as possible. It does this by either hoisting | |||
12 | // code into the preheader block, or by sinking code to the exit blocks if it is | |||
13 | // safe. This pass also promotes must-aliased memory locations in the loop to | |||
14 | // live in registers, thus hoisting and sinking "invariant" loads and stores. | |||
15 | // | |||
16 | // This pass uses alias analysis for two purposes: | |||
17 | // | |||
18 | // 1. Moving loop invariant loads and calls out of loops. If we can determine | |||
19 | // that a load or call inside of a loop never aliases anything stored to, | |||
20 | // we can hoist it or sink it like any other instruction. | |||
21 | // 2. Scalar Promotion of Memory - If there is a store instruction inside of | |||
22 | // the loop, we try to move the store to happen AFTER the loop instead of | |||
23 | // inside of the loop. This can only happen if a few conditions are true: | |||
24 | // A. The pointer stored through is loop invariant | |||
25 | // B. There are no stores or loads in the loop which _may_ alias the | |||
26 | // pointer. There are no calls in the loop which mod/ref the pointer. | |||
27 | // If these conditions are true, we can promote the loads and stores in the | |||
28 | // loop of the pointer to use a temporary alloca'd variable. We then use | |||
29 | // the SSAUpdater to construct the appropriate SSA form for the value. | |||
30 | // | |||
31 | //===----------------------------------------------------------------------===// | |||
32 | ||||
33 | #include "llvm/Transforms/Scalar/LICM.h" | |||
34 | #include "llvm/ADT/Statistic.h" | |||
35 | #include "llvm/Analysis/AliasAnalysis.h" | |||
36 | #include "llvm/Analysis/AliasSetTracker.h" | |||
37 | #include "llvm/Analysis/BasicAliasAnalysis.h" | |||
38 | #include "llvm/Analysis/CaptureTracking.h" | |||
39 | #include "llvm/Analysis/ConstantFolding.h" | |||
40 | #include "llvm/Analysis/GlobalsModRef.h" | |||
41 | #include "llvm/Analysis/Loads.h" | |||
42 | #include "llvm/Analysis/LoopInfo.h" | |||
43 | #include "llvm/Analysis/LoopPass.h" | |||
44 | #include "llvm/Analysis/MemoryBuiltins.h" | |||
45 | #include "llvm/Analysis/MemorySSA.h" | |||
46 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
47 | #include "llvm/Analysis/ScalarEvolution.h" | |||
48 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" | |||
49 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
50 | #include "llvm/Transforms/Utils/Local.h" | |||
51 | #include "llvm/Analysis/ValueTracking.h" | |||
52 | #include "llvm/IR/CFG.h" | |||
53 | #include "llvm/IR/Constants.h" | |||
54 | #include "llvm/IR/DataLayout.h" | |||
55 | #include "llvm/IR/DerivedTypes.h" | |||
56 | #include "llvm/IR/Dominators.h" | |||
57 | #include "llvm/IR/Instructions.h" | |||
58 | #include "llvm/IR/IntrinsicInst.h" | |||
59 | #include "llvm/IR/LLVMContext.h" | |||
60 | #include "llvm/IR/Metadata.h" | |||
61 | #include "llvm/IR/PredIteratorCache.h" | |||
62 | #include "llvm/Support/CommandLine.h" | |||
63 | #include "llvm/Support/Debug.h" | |||
64 | #include "llvm/Support/raw_ostream.h" | |||
65 | #include "llvm/Transforms/Scalar.h" | |||
66 | #include "llvm/Transforms/Scalar/LoopPassManager.h" | |||
67 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
68 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
69 | #include "llvm/Transforms/Utils/SSAUpdater.h" | |||
70 | #include <algorithm> | |||
71 | #include <utility> | |||
72 | using namespace llvm; | |||
73 | ||||
74 | #define DEBUG_TYPE"licm" "licm" | |||
75 | ||||
76 | STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop" , {0}, {false}}; | |||
77 | STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop" , {0}, {false}}; | |||
78 | STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads" , "Number of load insts hoisted or sunk", {0}, {false}}; | |||
79 | STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls" , "Number of call insts hoisted or sunk", {0}, {false}}; | |||
80 | STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers" , {0}, {false}}; | |||
81 | ||||
82 | /// Memory promotion is enabled by default. | |||
83 | static cl::opt<bool> | |||
84 | DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false), | |||
85 | cl::desc("Disable memory promotion in LICM pass")); | |||
86 | ||||
87 | static cl::opt<uint32_t> MaxNumUsesTraversed( | |||
88 | "licm-max-num-uses-traversed", cl::Hidden, cl::init(8), | |||
89 | cl::desc("Max num uses visited for identifying load " | |||
90 | "invariance in loop using invariant start (default = 8)")); | |||
91 | ||||
92 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI); | |||
93 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
94 | const LoopSafetyInfo *SafetyInfo, | |||
95 | TargetTransformInfo *TTI, bool &FreeInLoop); | |||
96 | static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | |||
97 | const LoopSafetyInfo *SafetyInfo, | |||
98 | OptimizationRemarkEmitter *ORE); | |||
99 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | |||
100 | const Loop *CurLoop, LoopSafetyInfo *SafetyInfo, | |||
101 | OptimizationRemarkEmitter *ORE, bool FreeInLoop); | |||
102 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | |||
103 | const DominatorTree *DT, | |||
104 | const Loop *CurLoop, | |||
105 | const LoopSafetyInfo *SafetyInfo, | |||
106 | OptimizationRemarkEmitter *ORE, | |||
107 | const Instruction *CtxI = nullptr); | |||
108 | static bool pointerInvalidatedByLoop(Value *V, uint64_t Size, | |||
109 | const AAMDNodes &AAInfo, | |||
110 | AliasSetTracker *CurAST); | |||
111 | static Instruction * | |||
112 | CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN, | |||
113 | const LoopInfo *LI, | |||
114 | const LoopSafetyInfo *SafetyInfo); | |||
115 | ||||
116 | namespace { | |||
117 | struct LoopInvariantCodeMotion { | |||
118 | bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT, | |||
119 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | |||
120 | ScalarEvolution *SE, MemorySSA *MSSA, | |||
121 | OptimizationRemarkEmitter *ORE, bool DeleteAST); | |||
122 | ||||
123 | DenseMap<Loop *, AliasSetTracker *> &getLoopToAliasSetMap() { | |||
124 | return LoopToAliasSetMap; | |||
125 | } | |||
126 | ||||
127 | private: | |||
128 | DenseMap<Loop *, AliasSetTracker *> LoopToAliasSetMap; | |||
129 | ||||
130 | AliasSetTracker *collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | |||
131 | AliasAnalysis *AA); | |||
132 | }; | |||
133 | ||||
134 | struct LegacyLICMPass : public LoopPass { | |||
135 | static char ID; // Pass identification, replacement for typeid | |||
136 | LegacyLICMPass() : LoopPass(ID) { | |||
137 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); | |||
138 | } | |||
139 | ||||
140 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { | |||
141 | if (skipLoop(L)) { | |||
142 | // If we have run LICM on a previous loop but now we are skipping | |||
143 | // (because we've hit the opt-bisect limit), we need to clear the | |||
144 | // loop alias information. | |||
145 | for (auto <AS : LICM.getLoopToAliasSetMap()) | |||
146 | delete LTAS.second; | |||
147 | LICM.getLoopToAliasSetMap().clear(); | |||
148 | return false; | |||
149 | } | |||
150 | ||||
151 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | |||
152 | MemorySSA *MSSA = EnableMSSALoopDependency | |||
153 | ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA()) | |||
154 | : nullptr; | |||
155 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis | |||
156 | // pass. Function analyses need to be preserved across loop transformations | |||
157 | // but ORE cannot be preserved (see comment before the pass definition). | |||
158 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); | |||
159 | return LICM.runOnLoop(L, | |||
160 | &getAnalysis<AAResultsWrapperPass>().getAAResults(), | |||
161 | &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), | |||
162 | &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), | |||
163 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(), | |||
164 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( | |||
165 | *L->getHeader()->getParent()), | |||
166 | SE ? &SE->getSE() : nullptr, MSSA, &ORE, false); | |||
167 | } | |||
168 | ||||
169 | /// This transformation requires natural loop information & requires that | |||
170 | /// loop preheaders be inserted into the CFG... | |||
171 | /// | |||
172 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
173 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
174 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
175 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
176 | if (EnableMSSALoopDependency) | |||
177 | AU.addRequired<MemorySSAWrapperPass>(); | |||
178 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
179 | getLoopAnalysisUsage(AU); | |||
180 | } | |||
181 | ||||
182 | using llvm::Pass::doFinalization; | |||
183 | ||||
184 | bool doFinalization() override { | |||
185 | assert(LICM.getLoopToAliasSetMap().empty() &&(static_cast <bool> (LICM.getLoopToAliasSetMap().empty( ) && "Didn't free loop alias sets") ? void (0) : __assert_fail ("LICM.getLoopToAliasSetMap().empty() && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 186, __extension__ __PRETTY_FUNCTION__)) | |||
186 | "Didn't free loop alias sets")(static_cast <bool> (LICM.getLoopToAliasSetMap().empty( ) && "Didn't free loop alias sets") ? void (0) : __assert_fail ("LICM.getLoopToAliasSetMap().empty() && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 186, __extension__ __PRETTY_FUNCTION__)); | |||
187 | return false; | |||
188 | } | |||
189 | ||||
190 | private: | |||
191 | LoopInvariantCodeMotion LICM; | |||
192 | ||||
193 | /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. | |||
194 | void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, | |||
195 | Loop *L) override; | |||
196 | ||||
197 | /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias | |||
198 | /// set. | |||
199 | void deleteAnalysisValue(Value *V, Loop *L) override; | |||
200 | ||||
201 | /// Simple Analysis hook. Delete loop L from alias set map. | |||
202 | void deleteAnalysisLoop(Loop *L) override; | |||
203 | }; | |||
204 | } // namespace | |||
205 | ||||
206 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, | |||
207 | LoopStandardAnalysisResults &AR, LPMUpdater &) { | |||
208 | const auto &FAM = | |||
209 | AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager(); | |||
210 | Function *F = L.getHeader()->getParent(); | |||
211 | ||||
212 | auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F); | |||
213 | // FIXME: This should probably be optional rather than required. | |||
214 | if (!ORE) | |||
| ||||
215 | report_fatal_error("LICM: OptimizationRemarkEmitterAnalysis not " | |||
216 | "cached at a higher level"); | |||
217 | ||||
218 | LoopInvariantCodeMotion LICM; | |||
219 | if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, &AR.TLI, &AR.TTI, &AR.SE, | |||
220 | AR.MSSA, ORE, true)) | |||
221 | return PreservedAnalyses::all(); | |||
222 | ||||
223 | auto PA = getLoopPassPreservedAnalyses(); | |||
224 | ||||
225 | PA.preserve<DominatorTreeAnalysis>(); | |||
226 | PA.preserve<LoopAnalysis>(); | |||
227 | ||||
228 | return PA; | |||
229 | } | |||
230 | ||||
231 | char LegacyLICMPass::ID = 0; | |||
232 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | |||
233 | false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | |||
234 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | |||
235 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
236 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
237 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | |||
238 | INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | |||
239 | false)PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | |||
240 | ||||
241 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } | |||
242 | ||||
243 | /// Hoist expressions out of the specified loop. Note, alias info for inner | |||
244 | /// loop is not preserved so it is not a good idea to run LICM multiple | |||
245 | /// times on one loop. | |||
246 | /// We should delete AST for inner loops in the new pass manager to avoid | |||
247 | /// memory leak. | |||
248 | /// | |||
249 | bool LoopInvariantCodeMotion::runOnLoop( | |||
250 | Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT, | |||
251 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, ScalarEvolution *SE, | |||
252 | MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, bool DeleteAST) { | |||
253 | bool Changed = false; | |||
254 | ||||
255 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")(static_cast <bool> (L->isLCSSAForm(*DT) && "Loop is not in LCSSA form." ) ? void (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop is not in LCSSA form.\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 255, __extension__ __PRETTY_FUNCTION__)); | |||
256 | ||||
257 | AliasSetTracker *CurAST = collectAliasInfoForLoop(L, LI, AA); | |||
258 | ||||
259 | // Get the preheader block to move instructions into... | |||
260 | BasicBlock *Preheader = L->getLoopPreheader(); | |||
261 | ||||
262 | // Compute loop safety information. | |||
263 | LoopSafetyInfo SafetyInfo; | |||
264 | computeLoopSafetyInfo(&SafetyInfo, L); | |||
265 | ||||
266 | // We want to visit all of the instructions in this loop... that are not parts | |||
267 | // of our subloops (they have already had their invariants hoisted out of | |||
268 | // their loop, into this loop, so there is no need to process the BODIES of | |||
269 | // the subloops). | |||
270 | // | |||
271 | // Traverse the body of the loop in depth first order on the dominator tree so | |||
272 | // that we are guaranteed to see definitions before we see uses. This allows | |||
273 | // us to sink instructions in one pass, without iteration. After sinking | |||
274 | // instructions, we perform another pass to hoist them out of the loop. | |||
275 | // | |||
276 | if (L->hasDedicatedExits()) | |||
277 | Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, TTI, L, | |||
278 | CurAST, &SafetyInfo, ORE); | |||
279 | if (Preheader) | |||
280 | Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L, | |||
281 | CurAST, &SafetyInfo, ORE); | |||
282 | ||||
283 | // Now that all loop invariants have been removed from the loop, promote any | |||
284 | // memory references to scalars that we can. | |||
285 | // Don't sink stores from loops without dedicated block exits. Exits | |||
286 | // containing indirect branches are not transformed by loop simplify, | |||
287 | // make sure we catch that. An additional load may be generated in the | |||
288 | // preheader for SSA updater, so also avoid sinking when no preheader | |||
289 | // is available. | |||
290 | if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { | |||
291 | // Figure out the loop exits and their insertion points | |||
292 | SmallVector<BasicBlock *, 8> ExitBlocks; | |||
293 | L->getUniqueExitBlocks(ExitBlocks); | |||
294 | ||||
295 | // We can't insert into a catchswitch. | |||
296 | bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) { | |||
297 | return isa<CatchSwitchInst>(Exit->getTerminator()); | |||
298 | }); | |||
299 | ||||
300 | if (!HasCatchSwitch) { | |||
301 | SmallVector<Instruction *, 8> InsertPts; | |||
302 | InsertPts.reserve(ExitBlocks.size()); | |||
303 | for (BasicBlock *ExitBlock : ExitBlocks) | |||
304 | InsertPts.push_back(&*ExitBlock->getFirstInsertionPt()); | |||
305 | ||||
306 | PredIteratorCache PIC; | |||
307 | ||||
308 | bool Promoted = false; | |||
309 | ||||
310 | // Loop over all of the alias sets in the tracker object. | |||
311 | for (AliasSet &AS : *CurAST) { | |||
312 | // We can promote this alias set if it has a store, if it is a "Must" | |||
313 | // alias set, if the pointer is loop invariant, and if we are not | |||
314 | // eliminating any volatile loads or stores. | |||
315 | if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || | |||
316 | AS.isVolatile() || !L->isLoopInvariant(AS.begin()->getValue())) | |||
317 | continue; | |||
318 | ||||
319 | assert((static_cast <bool> (!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? void (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 321, __extension__ __PRETTY_FUNCTION__)) | |||
320 | !AS.empty() &&(static_cast <bool> (!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? void (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 321, __extension__ __PRETTY_FUNCTION__)) | |||
321 | "Must alias set should have at least one pointer element in it!")(static_cast <bool> (!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? void (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 321, __extension__ __PRETTY_FUNCTION__)); | |||
322 | ||||
323 | SmallSetVector<Value *, 8> PointerMustAliases; | |||
324 | for (const auto &ASI : AS) | |||
325 | PointerMustAliases.insert(ASI.getValue()); | |||
326 | ||||
327 | Promoted |= promoteLoopAccessesToScalars(PointerMustAliases, ExitBlocks, | |||
328 | InsertPts, PIC, LI, DT, TLI, L, | |||
329 | CurAST, &SafetyInfo, ORE); | |||
330 | } | |||
331 | ||||
332 | // Once we have promoted values across the loop body we have to | |||
333 | // recursively reform LCSSA as any nested loop may now have values defined | |||
334 | // within the loop used in the outer loop. | |||
335 | // FIXME: This is really heavy handed. It would be a bit better to use an | |||
336 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed | |||
337 | // it as it went. | |||
338 | if (Promoted) | |||
339 | formLCSSARecursively(*L, *DT, LI, SE); | |||
340 | ||||
341 | Changed |= Promoted; | |||
342 | } | |||
343 | } | |||
344 | ||||
345 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is | |||
346 | // specifically moving instructions across the loop boundary and so it is | |||
347 | // especially in need of sanity checking here. | |||
348 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")(static_cast <bool> (L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!" ) ? void (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 348, __extension__ __PRETTY_FUNCTION__)); | |||
349 | assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&(static_cast <bool> ((!L->getParentLoop() || L->getParentLoop ()->isLCSSAForm(*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? void (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 350, __extension__ __PRETTY_FUNCTION__)) | |||
350 | "Parent loop not left in LCSSA form after LICM!")(static_cast <bool> ((!L->getParentLoop() || L->getParentLoop ()->isLCSSAForm(*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? void (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 350, __extension__ __PRETTY_FUNCTION__)); | |||
351 | ||||
352 | // If this loop is nested inside of another one, save the alias information | |||
353 | // for when we process the outer loop. | |||
354 | if (L->getParentLoop() && !DeleteAST) | |||
355 | LoopToAliasSetMap[L] = CurAST; | |||
356 | else | |||
357 | delete CurAST; | |||
358 | ||||
359 | if (Changed && SE) | |||
360 | SE->forgetLoopDispositions(L); | |||
361 | return Changed; | |||
362 | } | |||
363 | ||||
364 | /// Walk the specified region of the CFG (defined by all blocks dominated by | |||
365 | /// the specified block, and that are in the current loop) in reverse depth | |||
366 | /// first order w.r.t the DominatorTree. This allows us to visit uses before | |||
367 | /// definitions, allowing us to sink a loop body in one pass without iteration. | |||
368 | /// | |||
369 | bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI, | |||
370 | DominatorTree *DT, TargetLibraryInfo *TLI, | |||
371 | TargetTransformInfo *TTI, Loop *CurLoop, | |||
372 | AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo, | |||
373 | OptimizationRemarkEmitter *ORE) { | |||
374 | ||||
375 | // Verify inputs. | |||
376 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 378, __extension__ __PRETTY_FUNCTION__)) | |||
377 | CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 378, __extension__ __PRETTY_FUNCTION__)) | |||
378 | "Unexpected input to sinkRegion")(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 378, __extension__ __PRETTY_FUNCTION__)); | |||
379 | ||||
380 | // We want to visit children before parents. We will enque all the parents | |||
381 | // before their children in the worklist and process the worklist in reverse | |||
382 | // order. | |||
383 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | |||
384 | ||||
385 | bool Changed = false; | |||
386 | for (DomTreeNode *DTN : reverse(Worklist)) { | |||
387 | BasicBlock *BB = DTN->getBlock(); | |||
388 | // Only need to process the contents of this block if it is not part of a | |||
389 | // subloop (which would already have been processed). | |||
390 | if (inSubLoop(BB, CurLoop, LI)) | |||
391 | continue; | |||
392 | ||||
393 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { | |||
394 | Instruction &I = *--II; | |||
395 | ||||
396 | // If the instruction is dead, we would try to sink it because it isn't | |||
397 | // used in the loop, instead, just delete it. | |||
398 | if (isInstructionTriviallyDead(&I, TLI)) { | |||
399 | LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM deleting dead inst: " << I << '\n'; } } while (false); | |||
400 | salvageDebugInfo(I); | |||
401 | ++II; | |||
402 | CurAST->deleteValue(&I); | |||
403 | I.eraseFromParent(); | |||
404 | Changed = true; | |||
405 | continue; | |||
406 | } | |||
407 | ||||
408 | // Check to see if we can sink this instruction to the exit blocks | |||
409 | // of the loop. We can do this if the all users of the instruction are | |||
410 | // outside of the loop. In this case, it doesn't even matter if the | |||
411 | // operands of the instruction are loop invariant. | |||
412 | // | |||
413 | bool FreeInLoop = false; | |||
414 | if (isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) && | |||
415 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo, ORE)) { | |||
416 | if (sink(I, LI, DT, CurLoop, SafetyInfo, ORE, FreeInLoop)) { | |||
417 | if (!FreeInLoop) { | |||
418 | ++II; | |||
419 | CurAST->deleteValue(&I); | |||
420 | I.eraseFromParent(); | |||
421 | } | |||
422 | Changed = true; | |||
423 | } | |||
424 | } | |||
425 | } | |||
426 | } | |||
427 | return Changed; | |||
428 | } | |||
429 | ||||
430 | /// Walk the specified region of the CFG (defined by all blocks dominated by | |||
431 | /// the specified block, and that are in the current loop) in depth first | |||
432 | /// order w.r.t the DominatorTree. This allows us to visit definitions before | |||
433 | /// uses, allowing us to hoist a loop body in one pass without iteration. | |||
434 | /// | |||
435 | bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI, | |||
436 | DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop, | |||
437 | AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo, | |||
438 | OptimizationRemarkEmitter *ORE) { | |||
439 | // Verify inputs. | |||
440 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 442, __extension__ __PRETTY_FUNCTION__)) | |||
441 | CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 442, __extension__ __PRETTY_FUNCTION__)) | |||
442 | "Unexpected input to hoistRegion")(static_cast <bool> (N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion") ? void (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 442, __extension__ __PRETTY_FUNCTION__)); | |||
443 | ||||
444 | // We want to visit parents before children. We will enque all the parents | |||
445 | // before their children in the worklist and process the worklist in order. | |||
446 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | |||
447 | ||||
448 | bool Changed = false; | |||
449 | for (DomTreeNode *DTN : Worklist) { | |||
450 | BasicBlock *BB = DTN->getBlock(); | |||
451 | // Only need to process the contents of this block if it is not part of a | |||
452 | // subloop (which would already have been processed). | |||
453 | if (inSubLoop(BB, CurLoop, LI)) | |||
454 | continue; | |||
455 | ||||
456 | // Keep track of whether the prefix of instructions visited so far are such | |||
457 | // that the next instruction visited is guaranteed to execute if the loop | |||
458 | // is entered. | |||
459 | bool IsMustExecute = CurLoop->getHeader() == BB; | |||
460 | ||||
461 | for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) { | |||
462 | Instruction &I = *II++; | |||
463 | // Try constant folding this instruction. If all the operands are | |||
464 | // constants, it is technically hoistable, but it would be better to | |||
465 | // just fold it. | |||
466 | if (Constant *C = ConstantFoldInstruction( | |||
467 | &I, I.getModule()->getDataLayout(), TLI)) { | |||
468 | LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false) | |||
469 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false); | |||
470 | CurAST->copyValue(&I, C); | |||
471 | I.replaceAllUsesWith(C); | |||
472 | if (isInstructionTriviallyDead(&I, TLI)) { | |||
473 | CurAST->deleteValue(&I); | |||
474 | I.eraseFromParent(); | |||
475 | } | |||
476 | Changed = true; | |||
477 | continue; | |||
478 | } | |||
479 | ||||
480 | // Try hoisting the instruction out to the preheader. We can only do | |||
481 | // this if all of the operands of the instruction are loop invariant and | |||
482 | // if it is safe to hoist the instruction. | |||
483 | // | |||
484 | if (CurLoop->hasLoopInvariantOperands(&I) && | |||
485 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo, ORE) && | |||
486 | (IsMustExecute || | |||
487 | isSafeToExecuteUnconditionally( | |||
488 | I, DT, CurLoop, SafetyInfo, ORE, | |||
489 | CurLoop->getLoopPreheader()->getTerminator()))) { | |||
490 | Changed |= hoist(I, DT, CurLoop, SafetyInfo, ORE); | |||
491 | continue; | |||
492 | } | |||
493 | ||||
494 | // Attempt to remove floating point division out of the loop by | |||
495 | // converting it to a reciprocal multiplication. | |||
496 | if (I.getOpcode() == Instruction::FDiv && | |||
497 | CurLoop->isLoopInvariant(I.getOperand(1)) && | |||
498 | I.hasAllowReciprocal()) { | |||
499 | auto Divisor = I.getOperand(1); | |||
500 | auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0); | |||
501 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor); | |||
502 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); | |||
503 | ReciprocalDivisor->insertBefore(&I); | |||
504 | ||||
505 | auto Product = | |||
506 | BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor); | |||
507 | Product->setFastMathFlags(I.getFastMathFlags()); | |||
508 | Product->insertAfter(&I); | |||
509 | I.replaceAllUsesWith(Product); | |||
510 | I.eraseFromParent(); | |||
511 | ||||
512 | hoist(*ReciprocalDivisor, DT, CurLoop, SafetyInfo, ORE); | |||
513 | Changed = true; | |||
514 | continue; | |||
515 | } | |||
516 | ||||
517 | if (IsMustExecute) | |||
518 | IsMustExecute = isGuaranteedToTransferExecutionToSuccessor(&I); | |||
519 | } | |||
520 | } | |||
521 | ||||
522 | return Changed; | |||
523 | } | |||
524 | ||||
525 | // Return true if LI is invariant within scope of the loop. LI is invariant if | |||
526 | // CurLoop is dominated by an invariant.start representing the same memory | |||
527 | // location and size as the memory location LI loads from, and also the | |||
528 | // invariant.start has no uses. | |||
529 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, | |||
530 | Loop *CurLoop) { | |||
531 | Value *Addr = LI->getOperand(0); | |||
532 | const DataLayout &DL = LI->getModule()->getDataLayout(); | |||
533 | const uint32_t LocSizeInBits = DL.getTypeSizeInBits( | |||
534 | cast<PointerType>(Addr->getType())->getElementType()); | |||
535 | ||||
536 | // if the type is i8 addrspace(x)*, we know this is the type of | |||
537 | // llvm.invariant.start operand | |||
538 | auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()), | |||
539 | LI->getPointerAddressSpace()); | |||
540 | unsigned BitcastsVisited = 0; | |||
541 | // Look through bitcasts until we reach the i8* type (this is invariant.start | |||
542 | // operand type). | |||
543 | while (Addr->getType() != PtrInt8Ty) { | |||
544 | auto *BC = dyn_cast<BitCastInst>(Addr); | |||
545 | // Avoid traversing high number of bitcast uses. | |||
546 | if (++BitcastsVisited > MaxNumUsesTraversed || !BC) | |||
547 | return false; | |||
548 | Addr = BC->getOperand(0); | |||
549 | } | |||
550 | ||||
551 | unsigned UsesVisited = 0; | |||
552 | // Traverse all uses of the load operand value, to see if invariant.start is | |||
553 | // one of the uses, and whether it dominates the load instruction. | |||
554 | for (auto *U : Addr->users()) { | |||
555 | // Avoid traversing for Load operand with high number of users. | |||
556 | if (++UsesVisited > MaxNumUsesTraversed) | |||
557 | return false; | |||
558 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | |||
559 | // If there are escaping uses of invariant.start instruction, the load maybe | |||
560 | // non-invariant. | |||
561 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || | |||
562 | !II->use_empty()) | |||
563 | continue; | |||
564 | unsigned InvariantSizeInBits = | |||
565 | cast<ConstantInt>(II->getArgOperand(0))->getSExtValue() * 8; | |||
566 | // Confirm the invariant.start location size contains the load operand size | |||
567 | // in bits. Also, the invariant.start should dominate the load, and we | |||
568 | // should not hoist the load out of a loop that contains this dominating | |||
569 | // invariant.start. | |||
570 | if (LocSizeInBits <= InvariantSizeInBits && | |||
571 | DT->properlyDominates(II->getParent(), CurLoop->getHeader())) | |||
572 | return true; | |||
573 | } | |||
574 | ||||
575 | return false; | |||
576 | } | |||
577 | ||||
578 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, | |||
579 | Loop *CurLoop, AliasSetTracker *CurAST, | |||
580 | LoopSafetyInfo *SafetyInfo, | |||
581 | OptimizationRemarkEmitter *ORE) { | |||
582 | // SafetyInfo is nullptr if we are checking for sinking from preheader to | |||
583 | // loop body. | |||
584 | const bool SinkingToLoopBody = !SafetyInfo; | |||
585 | // Loads have extra constraints we have to verify before we can hoist them. | |||
586 | if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { | |||
587 | if (!LI->isUnordered()) | |||
588 | return false; // Don't sink/hoist volatile or ordered atomic loads! | |||
589 | ||||
590 | // Loads from constant memory are always safe to move, even if they end up | |||
591 | // in the same alias set as something that ends up being modified. | |||
592 | if (AA->pointsToConstantMemory(LI->getOperand(0))) | |||
593 | return true; | |||
594 | if (LI->getMetadata(LLVMContext::MD_invariant_load)) | |||
595 | return true; | |||
596 | ||||
597 | if (LI->isAtomic() && SinkingToLoopBody) | |||
598 | return false; // Don't sink unordered atomic loads to loop body. | |||
599 | ||||
600 | // This checks for an invariant.start dominating the load. | |||
601 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) | |||
602 | return true; | |||
603 | ||||
604 | // Don't hoist loads which have may-aliased stores in loop. | |||
605 | uint64_t Size = 0; | |||
606 | if (LI->getType()->isSized()) | |||
607 | Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType()); | |||
608 | ||||
609 | AAMDNodes AAInfo; | |||
610 | LI->getAAMetadata(AAInfo); | |||
611 | ||||
612 | bool Invalidated = | |||
613 | pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST); | |||
614 | // Check loop-invariant address because this may also be a sinkable load | |||
615 | // whose address is not necessarily loop-invariant. | |||
616 | if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand())) | |||
617 | ORE->emit([&]() { | |||
618 | return OptimizationRemarkMissed( | |||
619 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI) | |||
620 | << "failed to move load with loop-invariant address " | |||
621 | "because the loop may invalidate its value"; | |||
622 | }); | |||
623 | ||||
624 | return !Invalidated; | |||
625 | } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { | |||
626 | // Don't sink or hoist dbg info; it's legal, but not useful. | |||
627 | if (isa<DbgInfoIntrinsic>(I)) | |||
628 | return false; | |||
629 | ||||
630 | // Don't sink calls which can throw. | |||
631 | if (CI->mayThrow()) | |||
632 | return false; | |||
633 | ||||
634 | // Handle simple cases by querying alias analysis. | |||
635 | FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI); | |||
636 | if (Behavior == FMRB_DoesNotAccessMemory) | |||
637 | return true; | |||
638 | if (AliasAnalysis::onlyReadsMemory(Behavior)) { | |||
639 | // A readonly argmemonly function only reads from memory pointed to by | |||
640 | // it's arguments with arbitrary offsets. If we can prove there are no | |||
641 | // writes to this memory in the loop, we can hoist or sink. | |||
642 | if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) { | |||
643 | for (Value *Op : CI->arg_operands()) | |||
644 | if (Op->getType()->isPointerTy() && | |||
645 | pointerInvalidatedByLoop(Op, MemoryLocation::UnknownSize, | |||
646 | AAMDNodes(), CurAST)) | |||
647 | return false; | |||
648 | return true; | |||
649 | } | |||
650 | // If this call only reads from memory and there are no writes to memory | |||
651 | // in the loop, we can hoist or sink the call as appropriate. | |||
652 | bool FoundMod = false; | |||
653 | for (AliasSet &AS : *CurAST) { | |||
654 | if (!AS.isForwardingAliasSet() && AS.isMod()) { | |||
655 | FoundMod = true; | |||
656 | break; | |||
657 | } | |||
658 | } | |||
659 | if (!FoundMod) | |||
660 | return true; | |||
661 | } | |||
662 | ||||
663 | // FIXME: This should use mod/ref information to see if we can hoist or | |||
664 | // sink the call. | |||
665 | ||||
666 | return false; | |||
667 | } | |||
668 | ||||
669 | // Only these instructions are hoistable/sinkable. | |||
670 | if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) && | |||
671 | !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) && | |||
672 | !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) && | |||
673 | !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) && | |||
674 | !isa<InsertValueInst>(I)) | |||
675 | return false; | |||
676 | ||||
677 | // If we are checking for sinking from preheader to loop body it will be | |||
678 | // always safe as there is no speculative execution. | |||
679 | if (SinkingToLoopBody) | |||
680 | return true; | |||
681 | ||||
682 | // TODO: Plumb the context instruction through to make hoisting and sinking | |||
683 | // more powerful. Hoisting of loads already works due to the special casing | |||
684 | // above. | |||
685 | return isSafeToExecuteUnconditionally(I, DT, CurLoop, SafetyInfo, nullptr); | |||
686 | } | |||
687 | ||||
688 | /// Returns true if a PHINode is a trivially replaceable with an | |||
689 | /// Instruction. | |||
690 | /// This is true when all incoming values are that instruction. | |||
691 | /// This pattern occurs most often with LCSSA PHI nodes. | |||
692 | /// | |||
693 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { | |||
694 | for (const Value *IncValue : PN.incoming_values()) | |||
695 | if (IncValue != &I) | |||
696 | return false; | |||
697 | ||||
698 | return true; | |||
699 | } | |||
700 | ||||
701 | /// Return true if the instruction is free in the loop. | |||
702 | static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
703 | const TargetTransformInfo *TTI) { | |||
704 | ||||
705 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
706 | if (TTI->getUserCost(GEP) != TargetTransformInfo::TCC_Free) | |||
707 | return false; | |||
708 | // For a GEP, we cannot simply use getUserCost because currently it | |||
709 | // optimistically assume that a GEP will fold into addressing mode | |||
710 | // regardless of its users. | |||
711 | const BasicBlock *BB = GEP->getParent(); | |||
712 | for (const User *U : GEP->users()) { | |||
713 | const Instruction *UI = cast<Instruction>(U); | |||
714 | if (CurLoop->contains(UI) && | |||
715 | (BB != UI->getParent() || | |||
716 | (!isa<StoreInst>(UI) && !isa<LoadInst>(UI)))) | |||
717 | return false; | |||
718 | } | |||
719 | return true; | |||
720 | } else | |||
721 | return TTI->getUserCost(&I) == TargetTransformInfo::TCC_Free; | |||
722 | } | |||
723 | ||||
724 | /// Return true if the only users of this instruction are outside of | |||
725 | /// the loop. If this is true, we can sink the instruction to the exit | |||
726 | /// blocks of the loop. | |||
727 | /// | |||
728 | /// We also return true if the instruction could be folded away in lowering. | |||
729 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). | |||
730 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
731 | const LoopSafetyInfo *SafetyInfo, | |||
732 | TargetTransformInfo *TTI, bool &FreeInLoop) { | |||
733 | const auto &BlockColors = SafetyInfo->BlockColors; | |||
734 | bool IsFree = isFreeInLoop(I, CurLoop, TTI); | |||
735 | for (const User *U : I.users()) { | |||
736 | const Instruction *UI = cast<Instruction>(U); | |||
737 | if (const PHINode *PN = dyn_cast<PHINode>(UI)) { | |||
738 | const BasicBlock *BB = PN->getParent(); | |||
739 | // We cannot sink uses in catchswitches. | |||
740 | if (isa<CatchSwitchInst>(BB->getTerminator())) | |||
741 | return false; | |||
742 | ||||
743 | // We need to sink a callsite to a unique funclet. Avoid sinking if the | |||
744 | // phi use is too muddled. | |||
745 | if (isa<CallInst>(I)) | |||
746 | if (!BlockColors.empty() && | |||
747 | BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1) | |||
748 | return false; | |||
749 | } | |||
750 | ||||
751 | if (CurLoop->contains(UI)) { | |||
752 | if (IsFree) { | |||
753 | FreeInLoop = true; | |||
754 | continue; | |||
755 | } | |||
756 | return false; | |||
757 | } | |||
758 | } | |||
759 | return true; | |||
760 | } | |||
761 | ||||
762 | static Instruction * | |||
763 | CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN, | |||
764 | const LoopInfo *LI, | |||
765 | const LoopSafetyInfo *SafetyInfo) { | |||
766 | Instruction *New; | |||
767 | if (auto *CI = dyn_cast<CallInst>(&I)) { | |||
768 | const auto &BlockColors = SafetyInfo->BlockColors; | |||
769 | ||||
770 | // Sinking call-sites need to be handled differently from other | |||
771 | // instructions. The cloned call-site needs a funclet bundle operand | |||
772 | // appropriate for it's location in the CFG. | |||
773 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
774 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); | |||
775 | BundleIdx != BundleEnd; ++BundleIdx) { | |||
776 | OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx); | |||
777 | if (Bundle.getTagID() == LLVMContext::OB_funclet) | |||
778 | continue; | |||
779 | ||||
780 | OpBundles.emplace_back(Bundle); | |||
781 | } | |||
782 | ||||
783 | if (!BlockColors.empty()) { | |||
784 | const ColorVector &CV = BlockColors.find(&ExitBlock)->second; | |||
785 | assert(CV.size() == 1 && "non-unique color for exit block!")(static_cast <bool> (CV.size() == 1 && "non-unique color for exit block!" ) ? void (0) : __assert_fail ("CV.size() == 1 && \"non-unique color for exit block!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 785, __extension__ __PRETTY_FUNCTION__)); | |||
786 | BasicBlock *BBColor = CV.front(); | |||
787 | Instruction *EHPad = BBColor->getFirstNonPHI(); | |||
788 | if (EHPad->isEHPad()) | |||
789 | OpBundles.emplace_back("funclet", EHPad); | |||
790 | } | |||
791 | ||||
792 | New = CallInst::Create(CI, OpBundles); | |||
793 | } else { | |||
794 | New = I.clone(); | |||
795 | } | |||
796 | ||||
797 | ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New); | |||
798 | if (!I.getName().empty()) | |||
799 | New->setName(I.getName() + ".le"); | |||
800 | ||||
801 | // Build LCSSA PHI nodes for any in-loop operands. Note that this is | |||
802 | // particularly cheap because we can rip off the PHI node that we're | |||
803 | // replacing for the number and blocks of the predecessors. | |||
804 | // OPT: If this shows up in a profile, we can instead finish sinking all | |||
805 | // invariant instructions, and then walk their operands to re-establish | |||
806 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when | |||
807 | // sinking bottom-up. | |||
808 | for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE; | |||
809 | ++OI) | |||
810 | if (Instruction *OInst = dyn_cast<Instruction>(*OI)) | |||
811 | if (Loop *OLoop = LI->getLoopFor(OInst->getParent())) | |||
812 | if (!OLoop->contains(&PN)) { | |||
813 | PHINode *OpPN = | |||
814 | PHINode::Create(OInst->getType(), PN.getNumIncomingValues(), | |||
815 | OInst->getName() + ".lcssa", &ExitBlock.front()); | |||
816 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) | |||
817 | OpPN->addIncoming(OInst, PN.getIncomingBlock(i)); | |||
818 | *OI = OpPN; | |||
819 | } | |||
820 | return New; | |||
821 | } | |||
822 | ||||
823 | static Instruction *sinkThroughTriviallyReplaceablePHI( | |||
824 | PHINode *TPN, Instruction *I, LoopInfo *LI, | |||
825 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, | |||
826 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop) { | |||
827 | assert(isTriviallyReplaceablePHI(*TPN, *I) &&(static_cast <bool> (isTriviallyReplaceablePHI(*TPN, *I ) && "Expect only trivially replaceable PHI") ? void ( 0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 828, __extension__ __PRETTY_FUNCTION__)) | |||
828 | "Expect only trivially replaceable PHI")(static_cast <bool> (isTriviallyReplaceablePHI(*TPN, *I ) && "Expect only trivially replaceable PHI") ? void ( 0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 828, __extension__ __PRETTY_FUNCTION__)); | |||
829 | BasicBlock *ExitBlock = TPN->getParent(); | |||
830 | Instruction *New; | |||
831 | auto It = SunkCopies.find(ExitBlock); | |||
832 | if (It != SunkCopies.end()) | |||
833 | New = It->second; | |||
834 | else | |||
835 | New = SunkCopies[ExitBlock] = | |||
836 | CloneInstructionInExitBlock(*I, *ExitBlock, *TPN, LI, SafetyInfo); | |||
837 | return New; | |||
838 | } | |||
839 | ||||
840 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { | |||
841 | BasicBlock *BB = PN->getParent(); | |||
842 | if (!BB->canSplitPredecessors()) | |||
843 | return false; | |||
844 | // It's not impossible to split EHPad blocks, but if BlockColors already exist | |||
845 | // it require updating BlockColors for all offspring blocks accordingly. By | |||
846 | // skipping such corner case, we can make updating BlockColors after splitting | |||
847 | // predecessor fairly simple. | |||
848 | if (!SafetyInfo->BlockColors.empty() && BB->getFirstNonPHI()->isEHPad()) | |||
849 | return false; | |||
850 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | |||
851 | BasicBlock *BBPred = *PI; | |||
852 | if (isa<IndirectBrInst>(BBPred->getTerminator())) | |||
853 | return false; | |||
854 | } | |||
855 | return true; | |||
856 | } | |||
857 | ||||
858 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, | |||
859 | LoopInfo *LI, const Loop *CurLoop, | |||
860 | LoopSafetyInfo *SafetyInfo) { | |||
861 | #ifndef NDEBUG | |||
862 | SmallVector<BasicBlock *, 32> ExitBlocks; | |||
863 | CurLoop->getUniqueExitBlocks(ExitBlocks); | |||
864 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | |||
865 | ExitBlocks.end()); | |||
866 | #endif | |||
867 | BasicBlock *ExitBB = PN->getParent(); | |||
868 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")(static_cast <bool> (ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.") ? void (0) : __assert_fail ("ExitBlockSet.count(ExitBB) && \"Expect the PHI is in an exit block.\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 868, __extension__ __PRETTY_FUNCTION__)); | |||
869 | ||||
870 | // Split predecessors of the loop exit to make instructions in the loop are | |||
871 | // exposed to exit blocks through trivially replaceable PHIs while keeping the | |||
872 | // loop in the canonical form where each predecessor of each exit block should | |||
873 | // be contained within the loop. For example, this will convert the loop below | |||
874 | // from | |||
875 | // | |||
876 | // LB1: | |||
877 | // %v1 = | |||
878 | // br %LE, %LB2 | |||
879 | // LB2: | |||
880 | // %v2 = | |||
881 | // br %LE, %LB1 | |||
882 | // LE: | |||
883 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable | |||
884 | // | |||
885 | // to | |||
886 | // | |||
887 | // LB1: | |||
888 | // %v1 = | |||
889 | // br %LE.split, %LB2 | |||
890 | // LB2: | |||
891 | // %v2 = | |||
892 | // br %LE.split2, %LB1 | |||
893 | // LE.split: | |||
894 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable | |||
895 | // br %LE | |||
896 | // LE.split2: | |||
897 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable | |||
898 | // br %LE | |||
899 | // LE: | |||
900 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] | |||
901 | // | |||
902 | auto &BlockColors = SafetyInfo->BlockColors; | |||
903 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB)); | |||
904 | while (!PredBBs.empty()) { | |||
905 | BasicBlock *PredBB = *PredBBs.begin(); | |||
906 | assert(CurLoop->contains(PredBB) &&(static_cast <bool> (CurLoop->contains(PredBB) && "Expect all predecessors are in the loop") ? void (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 907, __extension__ __PRETTY_FUNCTION__)) | |||
907 | "Expect all predecessors are in the loop")(static_cast <bool> (CurLoop->contains(PredBB) && "Expect all predecessors are in the loop") ? void (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 907, __extension__ __PRETTY_FUNCTION__)); | |||
908 | if (PN->getBasicBlockIndex(PredBB) >= 0) { | |||
909 | BasicBlock *NewPred = SplitBlockPredecessors( | |||
910 | ExitBB, PredBB, ".split.loop.exit", DT, LI, true); | |||
911 | // Since we do not allow splitting EH-block with BlockColors in | |||
912 | // canSplitPredecessors(), we can simply assign predecessor's color to | |||
913 | // the new block. | |||
914 | if (!BlockColors.empty()) { | |||
915 | // Grab a reference to the ColorVector to be inserted before getting the | |||
916 | // reference to the vector we are copying because inserting the new | |||
917 | // element in BlockColors might cause the map to be reallocated. | |||
918 | ColorVector &ColorsForNewBlock = BlockColors[NewPred]; | |||
919 | ColorVector &ColorsForOldBlock = BlockColors[PredBB]; | |||
920 | ColorsForNewBlock = ColorsForOldBlock; | |||
921 | } | |||
922 | } | |||
923 | PredBBs.remove(PredBB); | |||
924 | } | |||
925 | } | |||
926 | ||||
927 | /// When an instruction is found to only be used outside of the loop, this | |||
928 | /// function moves it to the exit blocks and patches up SSA form as needed. | |||
929 | /// This method is guaranteed to remove the original instruction from its | |||
930 | /// position, and may either delete it or move it to outside of the loop. | |||
931 | /// | |||
932 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | |||
933 | const Loop *CurLoop, LoopSafetyInfo *SafetyInfo, | |||
934 | OptimizationRemarkEmitter *ORE, bool FreeInLoop) { | |||
935 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM sinking instruction: " << I << "\n"; } } while (false); | |||
936 | ORE->emit([&]() { | |||
937 | return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I) | |||
938 | << "sinking " << ore::NV("Inst", &I); | |||
939 | }); | |||
940 | bool Changed = false; | |||
941 | if (isa<LoadInst>(I)) | |||
942 | ++NumMovedLoads; | |||
943 | else if (isa<CallInst>(I)) | |||
944 | ++NumMovedCalls; | |||
945 | ++NumSunk; | |||
946 | ||||
947 | // Iterate over users to be ready for actual sinking. Replace users via | |||
948 | // unrechable blocks with undef and make all user PHIs trivially replcable. | |||
949 | SmallPtrSet<Instruction *, 8> VisitedUsers; | |||
950 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { | |||
951 | auto *User = cast<Instruction>(*UI); | |||
952 | Use &U = UI.getUse(); | |||
953 | ++UI; | |||
954 | ||||
955 | if (VisitedUsers.count(User) || CurLoop->contains(User)) | |||
956 | continue; | |||
957 | ||||
958 | if (!DT->isReachableFromEntry(User->getParent())) { | |||
959 | U = UndefValue::get(I.getType()); | |||
960 | Changed = true; | |||
961 | continue; | |||
962 | } | |||
963 | ||||
964 | // The user must be a PHI node. | |||
965 | PHINode *PN = cast<PHINode>(User); | |||
966 | ||||
967 | // Surprisingly, instructions can be used outside of loops without any | |||
968 | // exits. This can only happen in PHI nodes if the incoming block is | |||
969 | // unreachable. | |||
970 | BasicBlock *BB = PN->getIncomingBlock(U); | |||
971 | if (!DT->isReachableFromEntry(BB)) { | |||
972 | U = UndefValue::get(I.getType()); | |||
973 | Changed = true; | |||
974 | continue; | |||
975 | } | |||
976 | ||||
977 | VisitedUsers.insert(PN); | |||
978 | if (isTriviallyReplaceablePHI(*PN, I)) | |||
979 | continue; | |||
980 | ||||
981 | if (!canSplitPredecessors(PN, SafetyInfo)) | |||
982 | return Changed; | |||
983 | ||||
984 | // Split predecessors of the PHI so that we can make users trivially | |||
985 | // replaceable. | |||
986 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo); | |||
987 | ||||
988 | // Should rebuild the iterators, as they may be invalidated by | |||
989 | // splitPredecessorsOfLoopExit(). | |||
990 | UI = I.user_begin(); | |||
991 | UE = I.user_end(); | |||
992 | } | |||
993 | ||||
994 | if (VisitedUsers.empty()) | |||
995 | return Changed; | |||
996 | ||||
997 | #ifndef NDEBUG | |||
998 | SmallVector<BasicBlock *, 32> ExitBlocks; | |||
999 | CurLoop->getUniqueExitBlocks(ExitBlocks); | |||
1000 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | |||
1001 | ExitBlocks.end()); | |||
1002 | #endif | |||
1003 | ||||
1004 | // Clones of this instruction. Don't create more than one per exit block! | |||
1005 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; | |||
1006 | ||||
1007 | // If this instruction is only used outside of the loop, then all users are | |||
1008 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of | |||
1009 | // the instruction. | |||
1010 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); | |||
1011 | for (auto *UI : Users) { | |||
1012 | auto *User = cast<Instruction>(UI); | |||
1013 | ||||
1014 | if (CurLoop->contains(User)) | |||
1015 | continue; | |||
1016 | ||||
1017 | PHINode *PN = cast<PHINode>(User); | |||
1018 | assert(ExitBlockSet.count(PN->getParent()) &&(static_cast <bool> (ExitBlockSet.count(PN->getParent ()) && "The LCSSA PHI is not in an exit block!") ? void (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1019, __extension__ __PRETTY_FUNCTION__)) | |||
1019 | "The LCSSA PHI is not in an exit block!")(static_cast <bool> (ExitBlockSet.count(PN->getParent ()) && "The LCSSA PHI is not in an exit block!") ? void (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1019, __extension__ __PRETTY_FUNCTION__)); | |||
1020 | // The PHI must be trivially replaceable. | |||
1021 | Instruction *New = sinkThroughTriviallyReplaceablePHI(PN, &I, LI, SunkCopies, | |||
1022 | SafetyInfo, CurLoop); | |||
1023 | PN->replaceAllUsesWith(New); | |||
1024 | PN->eraseFromParent(); | |||
1025 | Changed = true; | |||
1026 | } | |||
1027 | return Changed; | |||
1028 | } | |||
1029 | ||||
1030 | /// When an instruction is found to only use loop invariant operands that | |||
1031 | /// is safe to hoist, this instruction is called to do the dirty work. | |||
1032 | /// | |||
1033 | static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | |||
1034 | const LoopSafetyInfo *SafetyInfo, | |||
1035 | OptimizationRemarkEmitter *ORE) { | |||
1036 | auto *Preheader = CurLoop->getLoopPreheader(); | |||
1037 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " << Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Preheader ->getName() << ": " << I << "\n"; } } while (false) | |||
1038 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Preheader ->getName() << ": " << I << "\n"; } } while (false); | |||
1039 | ORE->emit([&]() { | |||
1040 | return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting " | |||
1041 | << ore::NV("Inst", &I); | |||
1042 | }); | |||
1043 | ||||
1044 | // Metadata can be dependent on conditions we are hoisting above. | |||
1045 | // Conservatively strip all metadata on the instruction unless we were | |||
1046 | // guaranteed to execute I if we entered the loop, in which case the metadata | |||
1047 | // is valid in the loop preheader. | |||
1048 | if (I.hasMetadataOtherThanDebugLoc() && | |||
1049 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning | |||
1050 | // time in isGuaranteedToExecute if we don't actually have anything to | |||
1051 | // drop. It is a compile time optimization, not required for correctness. | |||
1052 | !isGuaranteedToExecute(I, DT, CurLoop, SafetyInfo)) | |||
1053 | I.dropUnknownNonDebugMetadata(); | |||
1054 | ||||
1055 | // Move the new node to the Preheader, before its terminator. | |||
1056 | I.moveBefore(Preheader->getTerminator()); | |||
1057 | ||||
1058 | // Do not retain debug locations when we are moving instructions to different | |||
1059 | // basic blocks, because we want to avoid jumpy line tables. Calls, however, | |||
1060 | // need to retain their debug locs because they may be inlined. | |||
1061 | // FIXME: How do we retain source locations without causing poor debugging | |||
1062 | // behavior? | |||
1063 | if (!isa<CallInst>(I)) | |||
1064 | I.setDebugLoc(DebugLoc()); | |||
1065 | ||||
1066 | if (isa<LoadInst>(I)) | |||
1067 | ++NumMovedLoads; | |||
1068 | else if (isa<CallInst>(I)) | |||
1069 | ++NumMovedCalls; | |||
1070 | ++NumHoisted; | |||
1071 | return true; | |||
1072 | } | |||
1073 | ||||
1074 | /// Only sink or hoist an instruction if it is not a trapping instruction, | |||
1075 | /// or if the instruction is known not to trap when moved to the preheader. | |||
1076 | /// or if it is a trapping instruction and is guaranteed to execute. | |||
1077 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | |||
1078 | const DominatorTree *DT, | |||
1079 | const Loop *CurLoop, | |||
1080 | const LoopSafetyInfo *SafetyInfo, | |||
1081 | OptimizationRemarkEmitter *ORE, | |||
1082 | const Instruction *CtxI) { | |||
1083 | if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT)) | |||
1084 | return true; | |||
1085 | ||||
1086 | bool GuaranteedToExecute = | |||
1087 | isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo); | |||
1088 | ||||
1089 | if (!GuaranteedToExecute) { | |||
1090 | auto *LI = dyn_cast<LoadInst>(&Inst); | |||
1091 | if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand())) | |||
1092 | ORE->emit([&]() { | |||
| ||||
1093 | return OptimizationRemarkMissed( | |||
1094 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI) | |||
1095 | << "failed to hoist load with loop-invariant address " | |||
1096 | "because load is conditionally executed"; | |||
1097 | }); | |||
1098 | } | |||
1099 | ||||
1100 | return GuaranteedToExecute; | |||
1101 | } | |||
1102 | ||||
1103 | namespace { | |||
1104 | class LoopPromoter : public LoadAndStorePromoter { | |||
1105 | Value *SomePtr; // Designated pointer to store to. | |||
1106 | const SmallSetVector<Value *, 8> &PointerMustAliases; | |||
1107 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; | |||
1108 | SmallVectorImpl<Instruction *> &LoopInsertPts; | |||
1109 | PredIteratorCache &PredCache; | |||
1110 | AliasSetTracker &AST; | |||
1111 | LoopInfo &LI; | |||
1112 | DebugLoc DL; | |||
1113 | int Alignment; | |||
1114 | bool UnorderedAtomic; | |||
1115 | AAMDNodes AATags; | |||
1116 | ||||
1117 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { | |||
1118 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
1119 | if (Loop *L = LI.getLoopFor(I->getParent())) | |||
1120 | if (!L->contains(BB)) { | |||
1121 | // We need to create an LCSSA PHI node for the incoming value and | |||
1122 | // store that. | |||
1123 | PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB), | |||
1124 | I->getName() + ".lcssa", &BB->front()); | |||
1125 | for (BasicBlock *Pred : PredCache.get(BB)) | |||
1126 | PN->addIncoming(I, Pred); | |||
1127 | return PN; | |||
1128 | } | |||
1129 | return V; | |||
1130 | } | |||
1131 | ||||
1132 | public: | |||
1133 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, | |||
1134 | const SmallSetVector<Value *, 8> &PMA, | |||
1135 | SmallVectorImpl<BasicBlock *> &LEB, | |||
1136 | SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC, | |||
1137 | AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment, | |||
1138 | bool UnorderedAtomic, const AAMDNodes &AATags) | |||
1139 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), | |||
1140 | LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast), | |||
1141 | LI(li), DL(std::move(dl)), Alignment(alignment), | |||
1142 | UnorderedAtomic(UnorderedAtomic), AATags(AATags) {} | |||
1143 | ||||
1144 | bool isInstInList(Instruction *I, | |||
1145 | const SmallVectorImpl<Instruction *> &) const override { | |||
1146 | Value *Ptr; | |||
1147 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
1148 | Ptr = LI->getOperand(0); | |||
1149 | else | |||
1150 | Ptr = cast<StoreInst>(I)->getPointerOperand(); | |||
1151 | return PointerMustAliases.count(Ptr); | |||
1152 | } | |||
1153 | ||||
1154 | void doExtraRewritesBeforeFinalDeletion() const override { | |||
1155 | // Insert stores after in the loop exit blocks. Each exit block gets a | |||
1156 | // store of the live-out values that feed them. Since we've already told | |||
1157 | // the SSA updater about the defs in the loop and the preheader | |||
1158 | // definition, it is all set and we can start using it. | |||
1159 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { | |||
1160 | BasicBlock *ExitBlock = LoopExitBlocks[i]; | |||
1161 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); | |||
1162 | LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); | |||
1163 | Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); | |||
1164 | Instruction *InsertPos = LoopInsertPts[i]; | |||
1165 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); | |||
1166 | if (UnorderedAtomic) | |||
1167 | NewSI->setOrdering(AtomicOrdering::Unordered); | |||
1168 | NewSI->setAlignment(Alignment); | |||
1169 | NewSI->setDebugLoc(DL); | |||
1170 | if (AATags) | |||
1171 | NewSI->setAAMetadata(AATags); | |||
1172 | } | |||
1173 | } | |||
1174 | ||||
1175 | void replaceLoadWithValue(LoadInst *LI, Value *V) const override { | |||
1176 | // Update alias analysis. | |||
1177 | AST.copyValue(LI, V); | |||
1178 | } | |||
1179 | void instructionDeleted(Instruction *I) const override { AST.deleteValue(I); } | |||
1180 | }; | |||
1181 | ||||
1182 | ||||
1183 | /// Return true iff we can prove that a caller of this function can not inspect | |||
1184 | /// the contents of the provided object in a well defined program. | |||
1185 | bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) { | |||
1186 | if (isa<AllocaInst>(Object)) | |||
1187 | // Since the alloca goes out of scope, we know the caller can't retain a | |||
1188 | // reference to it and be well defined. Thus, we don't need to check for | |||
1189 | // capture. | |||
1190 | return true; | |||
1191 | ||||
1192 | // For all other objects we need to know that the caller can't possibly | |||
1193 | // have gotten a reference to the object. There are two components of | |||
1194 | // that: | |||
1195 | // 1) Object can't be escaped by this function. This is what | |||
1196 | // PointerMayBeCaptured checks. | |||
1197 | // 2) Object can't have been captured at definition site. For this, we | |||
1198 | // need to know the return value is noalias. At the moment, we use a | |||
1199 | // weaker condition and handle only AllocLikeFunctions (which are | |||
1200 | // known to be noalias). TODO | |||
1201 | return isAllocLikeFn(Object, TLI) && | |||
1202 | !PointerMayBeCaptured(Object, true, true); | |||
1203 | } | |||
1204 | ||||
1205 | } // namespace | |||
1206 | ||||
1207 | /// Try to promote memory values to scalars by sinking stores out of the | |||
1208 | /// loop and moving loads to before the loop. We do this by looping over | |||
1209 | /// the stores in the loop, looking for stores to Must pointers which are | |||
1210 | /// loop invariant. | |||
1211 | /// | |||
1212 | bool llvm::promoteLoopAccessesToScalars( | |||
1213 | const SmallSetVector<Value *, 8> &PointerMustAliases, | |||
1214 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | |||
1215 | SmallVectorImpl<Instruction *> &InsertPts, PredIteratorCache &PIC, | |||
1216 | LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, | |||
1217 | Loop *CurLoop, AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo, | |||
1218 | OptimizationRemarkEmitter *ORE) { | |||
1219 | // Verify inputs. | |||
1220 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&(static_cast <bool> (LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? void (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1222, __extension__ __PRETTY_FUNCTION__)) | |||
1221 | CurAST != nullptr && SafetyInfo != nullptr &&(static_cast <bool> (LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? void (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1222, __extension__ __PRETTY_FUNCTION__)) | |||
1222 | "Unexpected Input to promoteLoopAccessesToScalars")(static_cast <bool> (LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? void (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1222, __extension__ __PRETTY_FUNCTION__)); | |||
1223 | ||||
1224 | Value *SomePtr = *PointerMustAliases.begin(); | |||
1225 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); | |||
1226 | ||||
1227 | // It is not safe to promote a load/store from the loop if the load/store is | |||
1228 | // conditional. For example, turning: | |||
1229 | // | |||
1230 | // for () { if (c) *P += 1; } | |||
1231 | // | |||
1232 | // into: | |||
1233 | // | |||
1234 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; | |||
1235 | // | |||
1236 | // is not safe, because *P may only be valid to access if 'c' is true. | |||
1237 | // | |||
1238 | // The safety property divides into two parts: | |||
1239 | // p1) The memory may not be dereferenceable on entry to the loop. In this | |||
1240 | // case, we can't insert the required load in the preheader. | |||
1241 | // p2) The memory model does not allow us to insert a store along any dynamic | |||
1242 | // path which did not originally have one. | |||
1243 | // | |||
1244 | // If at least one store is guaranteed to execute, both properties are | |||
1245 | // satisfied, and promotion is legal. | |||
1246 | // | |||
1247 | // This, however, is not a necessary condition. Even if no store/load is | |||
1248 | // guaranteed to execute, we can still establish these properties. | |||
1249 | // We can establish (p1) by proving that hoisting the load into the preheader | |||
1250 | // is safe (i.e. proving dereferenceability on all paths through the loop). We | |||
1251 | // can use any access within the alias set to prove dereferenceability, | |||
1252 | // since they're all must alias. | |||
1253 | // | |||
1254 | // There are two ways establish (p2): | |||
1255 | // a) Prove the location is thread-local. In this case the memory model | |||
1256 | // requirement does not apply, and stores are safe to insert. | |||
1257 | // b) Prove a store dominates every exit block. In this case, if an exit | |||
1258 | // blocks is reached, the original dynamic path would have taken us through | |||
1259 | // the store, so inserting a store into the exit block is safe. Note that this | |||
1260 | // is different from the store being guaranteed to execute. For instance, | |||
1261 | // if an exception is thrown on the first iteration of the loop, the original | |||
1262 | // store is never executed, but the exit blocks are not executed either. | |||
1263 | ||||
1264 | bool DereferenceableInPH = false; | |||
1265 | bool SafeToInsertStore = false; | |||
1266 | ||||
1267 | SmallVector<Instruction *, 64> LoopUses; | |||
1268 | ||||
1269 | // We start with an alignment of one and try to find instructions that allow | |||
1270 | // us to prove better alignment. | |||
1271 | unsigned Alignment = 1; | |||
1272 | // Keep track of which types of access we see | |||
1273 | bool SawUnorderedAtomic = false; | |||
1274 | bool SawNotAtomic = false; | |||
1275 | AAMDNodes AATags; | |||
1276 | ||||
1277 | const DataLayout &MDL = Preheader->getModule()->getDataLayout(); | |||
1278 | ||||
1279 | bool IsKnownThreadLocalObject = false; | |||
1280 | if (SafetyInfo->MayThrow) { | |||
1281 | // If a loop can throw, we have to insert a store along each unwind edge. | |||
1282 | // That said, we can't actually make the unwind edge explicit. Therefore, | |||
1283 | // we have to prove that the store is dead along the unwind edge. We do | |||
1284 | // this by proving that the caller can't have a reference to the object | |||
1285 | // after return and thus can't possibly load from the object. | |||
1286 | Value *Object = GetUnderlyingObject(SomePtr, MDL); | |||
1287 | if (!isKnownNonEscaping(Object, TLI)) | |||
1288 | return false; | |||
1289 | // Subtlety: Alloca's aren't visible to callers, but *are* potentially | |||
1290 | // visible to other threads if captured and used during their lifetimes. | |||
1291 | IsKnownThreadLocalObject = !isa<AllocaInst>(Object); | |||
1292 | } | |||
1293 | ||||
1294 | // Check that all of the pointers in the alias set have the same type. We | |||
1295 | // cannot (yet) promote a memory location that is loaded and stored in | |||
1296 | // different sizes. While we are at it, collect alignment and AA info. | |||
1297 | for (Value *ASIV : PointerMustAliases) { | |||
1298 | // Check that all of the pointers in the alias set have the same type. We | |||
1299 | // cannot (yet) promote a memory location that is loaded and stored in | |||
1300 | // different sizes. | |||
1301 | if (SomePtr->getType() != ASIV->getType()) | |||
1302 | return false; | |||
1303 | ||||
1304 | for (User *U : ASIV->users()) { | |||
1305 | // Ignore instructions that are outside the loop. | |||
1306 | Instruction *UI = dyn_cast<Instruction>(U); | |||
1307 | if (!UI || !CurLoop->contains(UI)) | |||
1308 | continue; | |||
1309 | ||||
1310 | // If there is an non-load/store instruction in the loop, we can't promote | |||
1311 | // it. | |||
1312 | if (LoadInst *Load = dyn_cast<LoadInst>(UI)) { | |||
1313 | assert(!Load->isVolatile() && "AST broken")(static_cast <bool> (!Load->isVolatile() && "AST broken" ) ? void (0) : __assert_fail ("!Load->isVolatile() && \"AST broken\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1313, __extension__ __PRETTY_FUNCTION__)); | |||
1314 | if (!Load->isUnordered()) | |||
1315 | return false; | |||
1316 | ||||
1317 | SawUnorderedAtomic |= Load->isAtomic(); | |||
1318 | SawNotAtomic |= !Load->isAtomic(); | |||
1319 | ||||
1320 | if (!DereferenceableInPH) | |||
1321 | DereferenceableInPH = isSafeToExecuteUnconditionally( | |||
1322 | *Load, DT, CurLoop, SafetyInfo, ORE, Preheader->getTerminator()); | |||
1323 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) { | |||
1324 | // Stores *of* the pointer are not interesting, only stores *to* the | |||
1325 | // pointer. | |||
1326 | if (UI->getOperand(1) != ASIV) | |||
1327 | continue; | |||
1328 | assert(!Store->isVolatile() && "AST broken")(static_cast <bool> (!Store->isVolatile() && "AST broken") ? void (0) : __assert_fail ("!Store->isVolatile() && \"AST broken\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1328, __extension__ __PRETTY_FUNCTION__)); | |||
1329 | if (!Store->isUnordered()) | |||
1330 | return false; | |||
1331 | ||||
1332 | SawUnorderedAtomic |= Store->isAtomic(); | |||
1333 | SawNotAtomic |= !Store->isAtomic(); | |||
1334 | ||||
1335 | // If the store is guaranteed to execute, both properties are satisfied. | |||
1336 | // We may want to check if a store is guaranteed to execute even if we | |||
1337 | // already know that promotion is safe, since it may have higher | |||
1338 | // alignment than any other guaranteed stores, in which case we can | |||
1339 | // raise the alignment on the promoted store. | |||
1340 | unsigned InstAlignment = Store->getAlignment(); | |||
1341 | if (!InstAlignment) | |||
1342 | InstAlignment = | |||
1343 | MDL.getABITypeAlignment(Store->getValueOperand()->getType()); | |||
1344 | ||||
1345 | if (!DereferenceableInPH || !SafeToInsertStore || | |||
1346 | (InstAlignment > Alignment)) { | |||
1347 | if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) { | |||
1348 | DereferenceableInPH = true; | |||
1349 | SafeToInsertStore = true; | |||
1350 | Alignment = std::max(Alignment, InstAlignment); | |||
1351 | } | |||
1352 | } | |||
1353 | ||||
1354 | // If a store dominates all exit blocks, it is safe to sink. | |||
1355 | // As explained above, if an exit block was executed, a dominating | |||
1356 | // store must have been executed at least once, so we are not | |||
1357 | // introducing stores on paths that did not have them. | |||
1358 | // Note that this only looks at explicit exit blocks. If we ever | |||
1359 | // start sinking stores into unwind edges (see above), this will break. | |||
1360 | if (!SafeToInsertStore) | |||
1361 | SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) { | |||
1362 | return DT->dominates(Store->getParent(), Exit); | |||
1363 | }); | |||
1364 | ||||
1365 | // If the store is not guaranteed to execute, we may still get | |||
1366 | // deref info through it. | |||
1367 | if (!DereferenceableInPH) { | |||
1368 | DereferenceableInPH = isDereferenceableAndAlignedPointer( | |||
1369 | Store->getPointerOperand(), Store->getAlignment(), MDL, | |||
1370 | Preheader->getTerminator(), DT); | |||
1371 | } | |||
1372 | } else | |||
1373 | return false; // Not a load or store. | |||
1374 | ||||
1375 | // Merge the AA tags. | |||
1376 | if (LoopUses.empty()) { | |||
1377 | // On the first load/store, just take its AA tags. | |||
1378 | UI->getAAMetadata(AATags); | |||
1379 | } else if (AATags) { | |||
1380 | UI->getAAMetadata(AATags, /* Merge = */ true); | |||
1381 | } | |||
1382 | ||||
1383 | LoopUses.push_back(UI); | |||
1384 | } | |||
1385 | } | |||
1386 | ||||
1387 | // If we found both an unordered atomic instruction and a non-atomic memory | |||
1388 | // access, bail. We can't blindly promote non-atomic to atomic since we | |||
1389 | // might not be able to lower the result. We can't downgrade since that | |||
1390 | // would violate memory model. Also, align 0 is an error for atomics. | |||
1391 | if (SawUnorderedAtomic && SawNotAtomic) | |||
1392 | return false; | |||
1393 | ||||
1394 | // If we couldn't prove we can hoist the load, bail. | |||
1395 | if (!DereferenceableInPH) | |||
1396 | return false; | |||
1397 | ||||
1398 | // We know we can hoist the load, but don't have a guaranteed store. | |||
1399 | // Check whether the location is thread-local. If it is, then we can insert | |||
1400 | // stores along paths which originally didn't have them without violating the | |||
1401 | // memory model. | |||
1402 | if (!SafeToInsertStore) { | |||
1403 | if (IsKnownThreadLocalObject) | |||
1404 | SafeToInsertStore = true; | |||
1405 | else { | |||
1406 | Value *Object = GetUnderlyingObject(SomePtr, MDL); | |||
1407 | SafeToInsertStore = | |||
1408 | (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) && | |||
1409 | !PointerMayBeCaptured(Object, true, true); | |||
1410 | } | |||
1411 | } | |||
1412 | ||||
1413 | // If we've still failed to prove we can sink the store, give up. | |||
1414 | if (!SafeToInsertStore) | |||
1415 | return false; | |||
1416 | ||||
1417 | // Otherwise, this is safe to promote, lets do it! | |||
1418 | LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr << '\n'; } } while (false) | |||
1419 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr << '\n'; } } while (false); | |||
1420 | ORE->emit([&]() { | |||
1421 | return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar", | |||
1422 | LoopUses[0]) | |||
1423 | << "Moving accesses to memory location out of the loop"; | |||
1424 | }); | |||
1425 | ++NumPromoted; | |||
1426 | ||||
1427 | // Grab a debug location for the inserted loads/stores; given that the | |||
1428 | // inserted loads/stores have little relation to the original loads/stores, | |||
1429 | // this code just arbitrarily picks a location from one, since any debug | |||
1430 | // location is better than none. | |||
1431 | DebugLoc DL = LoopUses[0]->getDebugLoc(); | |||
1432 | ||||
1433 | // We use the SSAUpdater interface to insert phi nodes as required. | |||
1434 | SmallVector<PHINode *, 16> NewPHIs; | |||
1435 | SSAUpdater SSA(&NewPHIs); | |||
1436 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, | |||
1437 | InsertPts, PIC, *CurAST, *LI, DL, Alignment, | |||
1438 | SawUnorderedAtomic, AATags); | |||
1439 | ||||
1440 | // Set up the preheader to have a definition of the value. It is the live-out | |||
1441 | // value from the preheader that uses in the loop will use. | |||
1442 | LoadInst *PreheaderLoad = new LoadInst( | |||
1443 | SomePtr, SomePtr->getName() + ".promoted", Preheader->getTerminator()); | |||
1444 | if (SawUnorderedAtomic) | |||
1445 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); | |||
1446 | PreheaderLoad->setAlignment(Alignment); | |||
1447 | PreheaderLoad->setDebugLoc(DL); | |||
1448 | if (AATags) | |||
1449 | PreheaderLoad->setAAMetadata(AATags); | |||
1450 | SSA.AddAvailableValue(Preheader, PreheaderLoad); | |||
1451 | ||||
1452 | // Rewrite all the loads in the loop and remember all the definitions from | |||
1453 | // stores in the loop. | |||
1454 | Promoter.run(LoopUses); | |||
1455 | ||||
1456 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. | |||
1457 | if (PreheaderLoad->use_empty()) | |||
1458 | PreheaderLoad->eraseFromParent(); | |||
1459 | ||||
1460 | return true; | |||
1461 | } | |||
1462 | ||||
1463 | /// Returns an owning pointer to an alias set which incorporates aliasing info | |||
1464 | /// from L and all subloops of L. | |||
1465 | /// FIXME: In new pass manager, there is no helper function to handle loop | |||
1466 | /// analysis such as cloneBasicBlockAnalysis, so the AST needs to be recomputed | |||
1467 | /// from scratch for every loop. Hook up with the helper functions when | |||
1468 | /// available in the new pass manager to avoid redundant computation. | |||
1469 | AliasSetTracker * | |||
1470 | LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | |||
1471 | AliasAnalysis *AA) { | |||
1472 | AliasSetTracker *CurAST = nullptr; | |||
1473 | SmallVector<Loop *, 4> RecomputeLoops; | |||
1474 | for (Loop *InnerL : L->getSubLoops()) { | |||
1475 | auto MapI = LoopToAliasSetMap.find(InnerL); | |||
1476 | // If the AST for this inner loop is missing it may have been merged into | |||
1477 | // some other loop's AST and then that loop unrolled, and so we need to | |||
1478 | // recompute it. | |||
1479 | if (MapI == LoopToAliasSetMap.end()) { | |||
1480 | RecomputeLoops.push_back(InnerL); | |||
1481 | continue; | |||
1482 | } | |||
1483 | AliasSetTracker *InnerAST = MapI->second; | |||
1484 | ||||
1485 | if (CurAST != nullptr) { | |||
1486 | // What if InnerLoop was modified by other passes ? | |||
1487 | CurAST->add(*InnerAST); | |||
1488 | ||||
1489 | // Once we've incorporated the inner loop's AST into ours, we don't need | |||
1490 | // the subloop's anymore. | |||
1491 | delete InnerAST; | |||
1492 | } else { | |||
1493 | CurAST = InnerAST; | |||
1494 | } | |||
1495 | LoopToAliasSetMap.erase(MapI); | |||
1496 | } | |||
1497 | if (CurAST == nullptr) | |||
1498 | CurAST = new AliasSetTracker(*AA); | |||
1499 | ||||
1500 | auto mergeLoop = [&](Loop *L) { | |||
1501 | // Loop over the body of this loop, looking for calls, invokes, and stores. | |||
1502 | for (BasicBlock *BB : L->blocks()) | |||
1503 | CurAST->add(*BB); // Incorporate the specified basic block | |||
1504 | }; | |||
1505 | ||||
1506 | // Add everything from the sub loops that are no longer directly available. | |||
1507 | for (Loop *InnerL : RecomputeLoops) | |||
1508 | mergeLoop(InnerL); | |||
1509 | ||||
1510 | // And merge in this loop. | |||
1511 | mergeLoop(L); | |||
1512 | ||||
1513 | return CurAST; | |||
1514 | } | |||
1515 | ||||
1516 | /// Simple analysis hook. Clone alias set info. | |||
1517 | /// | |||
1518 | void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, | |||
1519 | Loop *L) { | |||
1520 | AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L); | |||
1521 | if (!AST) | |||
1522 | return; | |||
1523 | ||||
1524 | AST->copyValue(From, To); | |||
1525 | } | |||
1526 | ||||
1527 | /// Simple Analysis hook. Delete value V from alias set | |||
1528 | /// | |||
1529 | void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) { | |||
1530 | AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L); | |||
1531 | if (!AST) | |||
1532 | return; | |||
1533 | ||||
1534 | AST->deleteValue(V); | |||
1535 | } | |||
1536 | ||||
1537 | /// Simple Analysis hook. Delete value L from alias set map. | |||
1538 | /// | |||
1539 | void LegacyLICMPass::deleteAnalysisLoop(Loop *L) { | |||
1540 | AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L); | |||
1541 | if (!AST) | |||
1542 | return; | |||
1543 | ||||
1544 | delete AST; | |||
1545 | LICM.getLoopToAliasSetMap().erase(L); | |||
1546 | } | |||
1547 | ||||
1548 | /// Return true if the body of this loop may store into the memory | |||
1549 | /// location pointed to by V. | |||
1550 | /// | |||
1551 | static bool pointerInvalidatedByLoop(Value *V, uint64_t Size, | |||
1552 | const AAMDNodes &AAInfo, | |||
1553 | AliasSetTracker *CurAST) { | |||
1554 | // Check to see if any of the basic blocks in CurLoop invalidate *V. | |||
1555 | return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod(); | |||
1556 | } | |||
1557 | ||||
1558 | /// Little predicate that returns true if the specified basic block is in | |||
1559 | /// a subloop of the current one, not the current one itself. | |||
1560 | /// | |||
1561 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { | |||
1562 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")(static_cast <bool> (CurLoop->contains(BB) && "Only valid if BB is IN the loop") ? void (0) : __assert_fail ("CurLoop->contains(BB) && \"Only valid if BB is IN the loop\"" , "/build/llvm-toolchain-snapshot-7~svn338205/lib/Transforms/Scalar/LICM.cpp" , 1562, __extension__ __PRETTY_FUNCTION__)); | |||
1563 | return LI->getLoopFor(BB) != CurLoop; | |||
1564 | } |