File: | lib/Transforms/Scalar/LICM.cpp |
Warning: | line 1226, column 30 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// | |||
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 performs loop invariant code motion, attempting to remove as much | |||
10 | // code from the body of a loop as possible. It does this by either hoisting | |||
11 | // code into the preheader block, or by sinking code to the exit blocks if it is | |||
12 | // safe. This pass also promotes must-aliased memory locations in the loop to | |||
13 | // live in registers, thus hoisting and sinking "invariant" loads and stores. | |||
14 | // | |||
15 | // This pass uses alias analysis for two purposes: | |||
16 | // | |||
17 | // 1. Moving loop invariant loads and calls out of loops. If we can determine | |||
18 | // that a load or call inside of a loop never aliases anything stored to, | |||
19 | // we can hoist it or sink it like any other instruction. | |||
20 | // 2. Scalar Promotion of Memory - If there is a store instruction inside of | |||
21 | // the loop, we try to move the store to happen AFTER the loop instead of | |||
22 | // inside of the loop. This can only happen if a few conditions are true: | |||
23 | // A. The pointer stored through is loop invariant | |||
24 | // B. There are no stores or loads in the loop which _may_ alias the | |||
25 | // pointer. There are no calls in the loop which mod/ref the pointer. | |||
26 | // If these conditions are true, we can promote the loads and stores in the | |||
27 | // loop of the pointer to use a temporary alloca'd variable. We then use | |||
28 | // the SSAUpdater to construct the appropriate SSA form for the value. | |||
29 | // | |||
30 | //===----------------------------------------------------------------------===// | |||
31 | ||||
32 | #include "llvm/Transforms/Scalar/LICM.h" | |||
33 | #include "llvm/ADT/SetOperations.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/GuardUtils.h" | |||
42 | #include "llvm/Analysis/Loads.h" | |||
43 | #include "llvm/Analysis/LoopInfo.h" | |||
44 | #include "llvm/Analysis/LoopIterator.h" | |||
45 | #include "llvm/Analysis/LoopPass.h" | |||
46 | #include "llvm/Analysis/MemoryBuiltins.h" | |||
47 | #include "llvm/Analysis/MemorySSA.h" | |||
48 | #include "llvm/Analysis/MemorySSAUpdater.h" | |||
49 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
50 | #include "llvm/Analysis/ScalarEvolution.h" | |||
51 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" | |||
52 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
53 | #include "llvm/Analysis/ValueTracking.h" | |||
54 | #include "llvm/IR/CFG.h" | |||
55 | #include "llvm/IR/Constants.h" | |||
56 | #include "llvm/IR/DataLayout.h" | |||
57 | #include "llvm/IR/DebugInfoMetadata.h" | |||
58 | #include "llvm/IR/DerivedTypes.h" | |||
59 | #include "llvm/IR/Dominators.h" | |||
60 | #include "llvm/IR/Instructions.h" | |||
61 | #include "llvm/IR/IntrinsicInst.h" | |||
62 | #include "llvm/IR/LLVMContext.h" | |||
63 | #include "llvm/IR/Metadata.h" | |||
64 | #include "llvm/IR/PatternMatch.h" | |||
65 | #include "llvm/IR/PredIteratorCache.h" | |||
66 | #include "llvm/Support/CommandLine.h" | |||
67 | #include "llvm/Support/Debug.h" | |||
68 | #include "llvm/Support/raw_ostream.h" | |||
69 | #include "llvm/Transforms/Scalar.h" | |||
70 | #include "llvm/Transforms/Scalar/LoopPassManager.h" | |||
71 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
72 | #include "llvm/Transforms/Utils/Local.h" | |||
73 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
74 | #include "llvm/Transforms/Utils/SSAUpdater.h" | |||
75 | #include <algorithm> | |||
76 | #include <utility> | |||
77 | using namespace llvm; | |||
78 | ||||
79 | #define DEBUG_TYPE"licm" "licm" | |||
80 | ||||
81 | STATISTIC(NumCreatedBlocks, "Number of blocks created")static llvm::Statistic NumCreatedBlocks = {"licm", "NumCreatedBlocks" , "Number of blocks created", {0}, {false}}; | |||
82 | STATISTIC(NumClonedBranches, "Number of branches cloned")static llvm::Statistic NumClonedBranches = {"licm", "NumClonedBranches" , "Number of branches cloned", {0}, {false}}; | |||
83 | STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop" , {0}, {false}}; | |||
84 | STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop" , {0}, {false}}; | |||
85 | STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads" , "Number of load insts hoisted or sunk", {0}, {false}}; | |||
86 | STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls" , "Number of call insts hoisted or sunk", {0}, {false}}; | |||
87 | STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers" , {0}, {false}}; | |||
88 | ||||
89 | /// Memory promotion is enabled by default. | |||
90 | static cl::opt<bool> | |||
91 | DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false), | |||
92 | cl::desc("Disable memory promotion in LICM pass")); | |||
93 | ||||
94 | static cl::opt<bool> ControlFlowHoisting( | |||
95 | "licm-control-flow-hoisting", cl::Hidden, cl::init(false), | |||
96 | cl::desc("Enable control flow (and PHI) hoisting in LICM")); | |||
97 | ||||
98 | static cl::opt<uint32_t> MaxNumUsesTraversed( | |||
99 | "licm-max-num-uses-traversed", cl::Hidden, cl::init(8), | |||
100 | cl::desc("Max num uses visited for identifying load " | |||
101 | "invariance in loop using invariant start (default = 8)")); | |||
102 | ||||
103 | // Default value of zero implies we use the regular alias set tracker mechanism | |||
104 | // instead of the cross product using AA to identify aliasing of the memory | |||
105 | // location we are interested in. | |||
106 | static cl::opt<int> | |||
107 | LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0), | |||
108 | cl::desc("How many instruction to cross product using AA")); | |||
109 | ||||
110 | // Experimental option to allow imprecision in LICM in pathological cases, in | |||
111 | // exchange for faster compile. This is to be removed if MemorySSA starts to | |||
112 | // address the same issue. This flag applies only when LICM uses MemorySSA | |||
113 | // instead on AliasSetTracker. LICM calls MemorySSAWalker's | |||
114 | // getClobberingMemoryAccess, up to the value of the Cap, getting perfect | |||
115 | // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess, | |||
116 | // which may not be precise, since optimizeUses is capped. The result is | |||
117 | // correct, but we may not get as "far up" as possible to get which access is | |||
118 | // clobbering the one queried. | |||
119 | cl::opt<unsigned> llvm::SetLicmMssaOptCap( | |||
120 | "licm-mssa-optimization-cap", cl::init(100), cl::Hidden, | |||
121 | cl::desc("Enable imprecision in LICM in pathological cases, in exchange " | |||
122 | "for faster compile. Caps the MemorySSA clobbering calls.")); | |||
123 | ||||
124 | // Experimentally, memory promotion carries less importance than sinking and | |||
125 | // hoisting. Limit when we do promotion when using MemorySSA, in order to save | |||
126 | // compile time. | |||
127 | cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap( | |||
128 | "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden, | |||
129 | cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no " | |||
130 | "effect. When MSSA in LICM is enabled, then this is the maximum " | |||
131 | "number of accesses allowed to be present in a loop in order to " | |||
132 | "enable memory promotion.")); | |||
133 | ||||
134 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI); | |||
135 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
136 | const LoopSafetyInfo *SafetyInfo, | |||
137 | TargetTransformInfo *TTI, bool &FreeInLoop); | |||
138 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | |||
139 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | |||
140 | MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE); | |||
141 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | |||
142 | const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo, | |||
143 | MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE); | |||
144 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | |||
145 | const DominatorTree *DT, | |||
146 | const Loop *CurLoop, | |||
147 | const LoopSafetyInfo *SafetyInfo, | |||
148 | OptimizationRemarkEmitter *ORE, | |||
149 | const Instruction *CtxI = nullptr); | |||
150 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | |||
151 | AliasSetTracker *CurAST, Loop *CurLoop, | |||
152 | AliasAnalysis *AA); | |||
153 | static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | |||
154 | Loop *CurLoop, | |||
155 | SinkAndHoistLICMFlags &Flags); | |||
156 | static Instruction *CloneInstructionInExitBlock( | |||
157 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | |||
158 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU); | |||
159 | ||||
160 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | |||
161 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU); | |||
162 | ||||
163 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | |||
164 | ICFLoopSafetyInfo &SafetyInfo, | |||
165 | MemorySSAUpdater *MSSAU); | |||
166 | ||||
167 | namespace { | |||
168 | struct LoopInvariantCodeMotion { | |||
169 | using ASTrackerMapTy = DenseMap<Loop *, std::unique_ptr<AliasSetTracker>>; | |||
170 | bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT, | |||
171 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | |||
172 | ScalarEvolution *SE, MemorySSA *MSSA, | |||
173 | OptimizationRemarkEmitter *ORE, bool DeleteAST); | |||
174 | ||||
175 | ASTrackerMapTy &getLoopToAliasSetMap() { return LoopToAliasSetMap; } | |||
176 | LoopInvariantCodeMotion(unsigned LicmMssaOptCap, | |||
177 | unsigned LicmMssaNoAccForPromotionCap) | |||
178 | : LicmMssaOptCap(LicmMssaOptCap), | |||
179 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {} | |||
180 | ||||
181 | private: | |||
182 | ASTrackerMapTy LoopToAliasSetMap; | |||
183 | unsigned LicmMssaOptCap; | |||
184 | unsigned LicmMssaNoAccForPromotionCap; | |||
185 | ||||
186 | std::unique_ptr<AliasSetTracker> | |||
187 | collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AliasAnalysis *AA); | |||
188 | std::unique_ptr<AliasSetTracker> | |||
189 | collectAliasInfoForLoopWithMSSA(Loop *L, AliasAnalysis *AA, | |||
190 | MemorySSAUpdater *MSSAU); | |||
191 | }; | |||
192 | ||||
193 | struct LegacyLICMPass : public LoopPass { | |||
194 | static char ID; // Pass identification, replacement for typeid | |||
195 | LegacyLICMPass( | |||
196 | unsigned LicmMssaOptCap = SetLicmMssaOptCap, | |||
197 | unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap) | |||
198 | : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) { | |||
199 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); | |||
200 | } | |||
201 | ||||
202 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { | |||
203 | if (skipLoop(L)) { | |||
204 | // If we have run LICM on a previous loop but now we are skipping | |||
205 | // (because we've hit the opt-bisect limit), we need to clear the | |||
206 | // loop alias information. | |||
207 | LICM.getLoopToAliasSetMap().clear(); | |||
208 | return false; | |||
209 | } | |||
210 | ||||
211 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | |||
212 | MemorySSA *MSSA = EnableMSSALoopDependency | |||
213 | ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA()) | |||
214 | : nullptr; | |||
215 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis | |||
216 | // pass. Function analyses need to be preserved across loop transformations | |||
217 | // but ORE cannot be preserved (see comment before the pass definition). | |||
218 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); | |||
219 | return LICM.runOnLoop(L, | |||
220 | &getAnalysis<AAResultsWrapperPass>().getAAResults(), | |||
221 | &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), | |||
222 | &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), | |||
223 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(), | |||
224 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( | |||
225 | *L->getHeader()->getParent()), | |||
226 | SE ? &SE->getSE() : nullptr, MSSA, &ORE, false); | |||
227 | } | |||
228 | ||||
229 | /// This transformation requires natural loop information & requires that | |||
230 | /// loop preheaders be inserted into the CFG... | |||
231 | /// | |||
232 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
233 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
234 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
235 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
236 | if (EnableMSSALoopDependency) { | |||
237 | AU.addRequired<MemorySSAWrapperPass>(); | |||
238 | AU.addPreserved<MemorySSAWrapperPass>(); | |||
239 | } | |||
240 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
241 | getLoopAnalysisUsage(AU); | |||
242 | } | |||
243 | ||||
244 | using llvm::Pass::doFinalization; | |||
245 | ||||
246 | bool doFinalization() override { | |||
247 | auto &AliasSetMap = LICM.getLoopToAliasSetMap(); | |||
248 | // All loops in the AliasSetMap should be cleaned up already. The only case | |||
249 | // where we fail to do so is if an outer loop gets deleted before LICM | |||
250 | // visits it. | |||
251 | assert(all_of(AliasSetMap,((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy ::value_type &KV) { return !KV.first->getParentLoop(); }) && "Didn't free loop alias sets") ? static_cast< void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 255, __PRETTY_FUNCTION__)) | |||
252 | [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) {((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy ::value_type &KV) { return !KV.first->getParentLoop(); }) && "Didn't free loop alias sets") ? static_cast< void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 255, __PRETTY_FUNCTION__)) | |||
253 | return !KV.first->getParentLoop();((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy ::value_type &KV) { return !KV.first->getParentLoop(); }) && "Didn't free loop alias sets") ? static_cast< void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 255, __PRETTY_FUNCTION__)) | |||
254 | }) &&((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy ::value_type &KV) { return !KV.first->getParentLoop(); }) && "Didn't free loop alias sets") ? static_cast< void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 255, __PRETTY_FUNCTION__)) | |||
255 | "Didn't free loop alias sets")((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy ::value_type &KV) { return !KV.first->getParentLoop(); }) && "Didn't free loop alias sets") ? static_cast< void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 255, __PRETTY_FUNCTION__)); | |||
256 | AliasSetMap.clear(); | |||
257 | return false; | |||
258 | } | |||
259 | ||||
260 | private: | |||
261 | LoopInvariantCodeMotion LICM; | |||
262 | ||||
263 | /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. | |||
264 | void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, | |||
265 | Loop *L) override; | |||
266 | ||||
267 | /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias | |||
268 | /// set. | |||
269 | void deleteAnalysisValue(Value *V, Loop *L) override; | |||
270 | ||||
271 | /// Simple Analysis hook. Delete loop L from alias set map. | |||
272 | void deleteAnalysisLoop(Loop *L) override; | |||
273 | }; | |||
274 | } // namespace | |||
275 | ||||
276 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, | |||
277 | LoopStandardAnalysisResults &AR, LPMUpdater &) { | |||
278 | const auto &FAM = | |||
279 | AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager(); | |||
280 | Function *F = L.getHeader()->getParent(); | |||
281 | ||||
282 | auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F); | |||
283 | // FIXME: This should probably be optional rather than required. | |||
284 | if (!ORE) | |||
285 | report_fatal_error("LICM: OptimizationRemarkEmitterAnalysis not " | |||
286 | "cached at a higher level"); | |||
287 | ||||
288 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | |||
289 | if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, &AR.TLI, &AR.TTI, &AR.SE, | |||
290 | AR.MSSA, ORE, true)) | |||
291 | return PreservedAnalyses::all(); | |||
292 | ||||
293 | auto PA = getLoopPassPreservedAnalyses(); | |||
294 | ||||
295 | PA.preserve<DominatorTreeAnalysis>(); | |||
296 | PA.preserve<LoopAnalysis>(); | |||
297 | ||||
298 | return PA; | |||
299 | } | |||
300 | ||||
301 | char LegacyLICMPass::ID = 0; | |||
302 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | |||
303 | false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | |||
304 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | |||
305 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
306 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
307 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | |||
308 | 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)); } | |||
309 | 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)); } | |||
310 | ||||
311 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } | |||
312 | Pass *llvm::createLICMPass(unsigned LicmMssaOptCap, | |||
313 | unsigned LicmMssaNoAccForPromotionCap) { | |||
314 | return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | |||
315 | } | |||
316 | ||||
317 | /// Hoist expressions out of the specified loop. Note, alias info for inner | |||
318 | /// loop is not preserved so it is not a good idea to run LICM multiple | |||
319 | /// times on one loop. | |||
320 | /// We should delete AST for inner loops in the new pass manager to avoid | |||
321 | /// memory leak. | |||
322 | /// | |||
323 | bool LoopInvariantCodeMotion::runOnLoop( | |||
324 | Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT, | |||
325 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, ScalarEvolution *SE, | |||
326 | MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, bool DeleteAST) { | |||
327 | bool Changed = false; | |||
328 | ||||
329 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")((L->isLCSSAForm(*DT) && "Loop is not in LCSSA form." ) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop is not in LCSSA form.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 329, __PRETTY_FUNCTION__)); | |||
330 | ||||
331 | std::unique_ptr<AliasSetTracker> CurAST; | |||
332 | std::unique_ptr<MemorySSAUpdater> MSSAU; | |||
333 | bool NoOfMemAccTooLarge = false; | |||
334 | unsigned LicmMssaOptCounter = 0; | |||
335 | ||||
336 | if (!MSSA) { | |||
337 | LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Using Alias Set Tracker.\n" ; } } while (false); | |||
338 | CurAST = collectAliasInfoForLoop(L, LI, AA); | |||
339 | } else { | |||
340 | LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Using MemorySSA.\n"; } } while (false); | |||
341 | MSSAU = make_unique<MemorySSAUpdater>(MSSA); | |||
342 | ||||
343 | unsigned AccessCapCount = 0; | |||
344 | for (auto *BB : L->getBlocks()) { | |||
345 | if (auto *Accesses = MSSA->getBlockAccesses(BB)) { | |||
346 | for (const auto &MA : *Accesses) { | |||
347 | (void)MA; | |||
348 | AccessCapCount++; | |||
349 | if (AccessCapCount > LicmMssaNoAccForPromotionCap) { | |||
350 | NoOfMemAccTooLarge = true; | |||
351 | break; | |||
352 | } | |||
353 | } | |||
354 | } | |||
355 | if (NoOfMemAccTooLarge) | |||
356 | break; | |||
357 | } | |||
358 | } | |||
359 | ||||
360 | // Get the preheader block to move instructions into... | |||
361 | BasicBlock *Preheader = L->getLoopPreheader(); | |||
362 | ||||
363 | // Compute loop safety information. | |||
364 | ICFLoopSafetyInfo SafetyInfo(DT); | |||
365 | SafetyInfo.computeLoopSafetyInfo(L); | |||
366 | ||||
367 | // We want to visit all of the instructions in this loop... that are not parts | |||
368 | // of our subloops (they have already had their invariants hoisted out of | |||
369 | // their loop, into this loop, so there is no need to process the BODIES of | |||
370 | // the subloops). | |||
371 | // | |||
372 | // Traverse the body of the loop in depth first order on the dominator tree so | |||
373 | // that we are guaranteed to see definitions before we see uses. This allows | |||
374 | // us to sink instructions in one pass, without iteration. After sinking | |||
375 | // instructions, we perform another pass to hoist them out of the loop. | |||
376 | SinkAndHoistLICMFlags Flags = {NoOfMemAccTooLarge, LicmMssaOptCounter, | |||
377 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap}; | |||
378 | if (L->hasDedicatedExits()) | |||
379 | Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, TTI, L, | |||
380 | CurAST.get(), MSSAU.get(), &SafetyInfo, Flags, ORE); | |||
381 | if (Preheader) | |||
382 | Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L, | |||
383 | CurAST.get(), MSSAU.get(), &SafetyInfo, Flags, ORE); | |||
384 | ||||
385 | // Now that all loop invariants have been removed from the loop, promote any | |||
386 | // memory references to scalars that we can. | |||
387 | // Don't sink stores from loops without dedicated block exits. Exits | |||
388 | // containing indirect branches are not transformed by loop simplify, | |||
389 | // make sure we catch that. An additional load may be generated in the | |||
390 | // preheader for SSA updater, so also avoid sinking when no preheader | |||
391 | // is available. | |||
392 | if (!DisablePromotion && Preheader && L->hasDedicatedExits() && | |||
393 | !NoOfMemAccTooLarge) { | |||
394 | // Figure out the loop exits and their insertion points | |||
395 | SmallVector<BasicBlock *, 8> ExitBlocks; | |||
396 | L->getUniqueExitBlocks(ExitBlocks); | |||
397 | ||||
398 | // We can't insert into a catchswitch. | |||
399 | bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) { | |||
400 | return isa<CatchSwitchInst>(Exit->getTerminator()); | |||
401 | }); | |||
402 | ||||
403 | if (!HasCatchSwitch) { | |||
404 | SmallVector<Instruction *, 8> InsertPts; | |||
405 | SmallVector<MemoryAccess *, 8> MSSAInsertPts; | |||
406 | InsertPts.reserve(ExitBlocks.size()); | |||
407 | if (MSSAU) | |||
408 | MSSAInsertPts.reserve(ExitBlocks.size()); | |||
409 | for (BasicBlock *ExitBlock : ExitBlocks) { | |||
410 | InsertPts.push_back(&*ExitBlock->getFirstInsertionPt()); | |||
411 | if (MSSAU) | |||
412 | MSSAInsertPts.push_back(nullptr); | |||
413 | } | |||
414 | ||||
415 | PredIteratorCache PIC; | |||
416 | ||||
417 | bool Promoted = false; | |||
418 | ||||
419 | // Build an AST using MSSA. | |||
420 | if (!CurAST.get()) | |||
421 | CurAST = collectAliasInfoForLoopWithMSSA(L, AA, MSSAU.get()); | |||
422 | ||||
423 | // Loop over all of the alias sets in the tracker object. | |||
424 | for (AliasSet &AS : *CurAST) { | |||
425 | // We can promote this alias set if it has a store, if it is a "Must" | |||
426 | // alias set, if the pointer is loop invariant, and if we are not | |||
427 | // eliminating any volatile loads or stores. | |||
428 | if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || | |||
429 | !L->isLoopInvariant(AS.begin()->getValue())) | |||
430 | continue; | |||
431 | ||||
432 | assert(((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 434, __PRETTY_FUNCTION__)) | |||
433 | !AS.empty() &&((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 434, __PRETTY_FUNCTION__)) | |||
434 | "Must alias set should have at least one pointer element in it!")((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 434, __PRETTY_FUNCTION__)); | |||
435 | ||||
436 | SmallSetVector<Value *, 8> PointerMustAliases; | |||
437 | for (const auto &ASI : AS) | |||
438 | PointerMustAliases.insert(ASI.getValue()); | |||
439 | ||||
440 | Promoted |= promoteLoopAccessesToScalars( | |||
441 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI, | |||
442 | DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE); | |||
443 | } | |||
444 | ||||
445 | // Once we have promoted values across the loop body we have to | |||
446 | // recursively reform LCSSA as any nested loop may now have values defined | |||
447 | // within the loop used in the outer loop. | |||
448 | // FIXME: This is really heavy handed. It would be a bit better to use an | |||
449 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed | |||
450 | // it as it went. | |||
451 | if (Promoted) | |||
452 | formLCSSARecursively(*L, *DT, LI, SE); | |||
453 | ||||
454 | Changed |= Promoted; | |||
455 | } | |||
456 | } | |||
457 | ||||
458 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is | |||
459 | // specifically moving instructions across the loop boundary and so it is | |||
460 | // especially in need of sanity checking here. | |||
461 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")((L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 461, __PRETTY_FUNCTION__)); | |||
462 | assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&(((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm (*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 463, __PRETTY_FUNCTION__)) | |||
463 | "Parent loop not left in LCSSA form after LICM!")(((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm (*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 463, __PRETTY_FUNCTION__)); | |||
464 | ||||
465 | // If this loop is nested inside of another one, save the alias information | |||
466 | // for when we process the outer loop. | |||
467 | if (!MSSAU.get() && CurAST.get() && L->getParentLoop() && !DeleteAST) | |||
468 | LoopToAliasSetMap[L] = std::move(CurAST); | |||
469 | ||||
470 | if (MSSAU.get() && VerifyMemorySSA) | |||
471 | MSSAU->getMemorySSA()->verifyMemorySSA(); | |||
472 | ||||
473 | if (Changed && SE) | |||
474 | SE->forgetLoopDispositions(L); | |||
475 | return Changed; | |||
476 | } | |||
477 | ||||
478 | /// Walk the specified region of the CFG (defined by all blocks dominated by | |||
479 | /// the specified block, and that are in the current loop) in reverse depth | |||
480 | /// first order w.r.t the DominatorTree. This allows us to visit uses before | |||
481 | /// definitions, allowing us to sink a loop body in one pass without iteration. | |||
482 | /// | |||
483 | bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI, | |||
484 | DominatorTree *DT, TargetLibraryInfo *TLI, | |||
485 | TargetTransformInfo *TTI, Loop *CurLoop, | |||
486 | AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | |||
487 | ICFLoopSafetyInfo *SafetyInfo, | |||
488 | SinkAndHoistLICMFlags &Flags, | |||
489 | OptimizationRemarkEmitter *ORE) { | |||
490 | ||||
491 | // Verify inputs. | |||
492 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 494, __PRETTY_FUNCTION__)) | |||
493 | CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 494, __PRETTY_FUNCTION__)) | |||
494 | "Unexpected input to sinkRegion.")((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 494, __PRETTY_FUNCTION__)); | |||
495 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 496, __PRETTY_FUNCTION__)) | |||
496 | "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 496, __PRETTY_FUNCTION__)); | |||
497 | ||||
498 | // We want to visit children before parents. We will enque all the parents | |||
499 | // before their children in the worklist and process the worklist in reverse | |||
500 | // order. | |||
501 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | |||
502 | ||||
503 | bool Changed = false; | |||
504 | for (DomTreeNode *DTN : reverse(Worklist)) { | |||
505 | BasicBlock *BB = DTN->getBlock(); | |||
506 | // Only need to process the contents of this block if it is not part of a | |||
507 | // subloop (which would already have been processed). | |||
508 | if (inSubLoop(BB, CurLoop, LI)) | |||
509 | continue; | |||
510 | ||||
511 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { | |||
512 | Instruction &I = *--II; | |||
513 | ||||
514 | // If the instruction is dead, we would try to sink it because it isn't | |||
515 | // used in the loop, instead, just delete it. | |||
516 | if (isInstructionTriviallyDead(&I, TLI)) { | |||
517 | 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); | |||
518 | salvageDebugInfo(I); | |||
519 | ++II; | |||
520 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | |||
521 | Changed = true; | |||
522 | continue; | |||
523 | } | |||
524 | ||||
525 | // Check to see if we can sink this instruction to the exit blocks | |||
526 | // of the loop. We can do this if the all users of the instruction are | |||
527 | // outside of the loop. In this case, it doesn't even matter if the | |||
528 | // operands of the instruction are loop invariant. | |||
529 | // | |||
530 | bool FreeInLoop = false; | |||
531 | if (isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) && | |||
532 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | |||
533 | ORE) && | |||
534 | !I.mayHaveSideEffects()) { | |||
535 | if (sink(I, LI, DT, CurLoop, SafetyInfo, MSSAU, ORE)) { | |||
536 | if (!FreeInLoop) { | |||
537 | ++II; | |||
538 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | |||
539 | } | |||
540 | Changed = true; | |||
541 | } | |||
542 | } | |||
543 | } | |||
544 | } | |||
545 | if (MSSAU && VerifyMemorySSA) | |||
546 | MSSAU->getMemorySSA()->verifyMemorySSA(); | |||
547 | return Changed; | |||
548 | } | |||
549 | ||||
550 | namespace { | |||
551 | // This is a helper class for hoistRegion to make it able to hoist control flow | |||
552 | // in order to be able to hoist phis. The way this works is that we initially | |||
553 | // start hoisting to the loop preheader, and when we see a loop invariant branch | |||
554 | // we make note of this. When we then come to hoist an instruction that's | |||
555 | // conditional on such a branch we duplicate the branch and the relevant control | |||
556 | // flow, then hoist the instruction into the block corresponding to its original | |||
557 | // block in the duplicated control flow. | |||
558 | class ControlFlowHoister { | |||
559 | private: | |||
560 | // Information about the loop we are hoisting from | |||
561 | LoopInfo *LI; | |||
562 | DominatorTree *DT; | |||
563 | Loop *CurLoop; | |||
564 | MemorySSAUpdater *MSSAU; | |||
565 | ||||
566 | // A map of blocks in the loop to the block their instructions will be hoisted | |||
567 | // to. | |||
568 | DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap; | |||
569 | ||||
570 | // The branches that we can hoist, mapped to the block that marks a | |||
571 | // convergence point of their control flow. | |||
572 | DenseMap<BranchInst *, BasicBlock *> HoistableBranches; | |||
573 | ||||
574 | public: | |||
575 | ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop, | |||
576 | MemorySSAUpdater *MSSAU) | |||
577 | : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {} | |||
578 | ||||
579 | void registerPossiblyHoistableBranch(BranchInst *BI) { | |||
580 | // We can only hoist conditional branches with loop invariant operands. | |||
581 | if (!ControlFlowHoisting || !BI->isConditional() || | |||
582 | !CurLoop->hasLoopInvariantOperands(BI)) | |||
583 | return; | |||
584 | ||||
585 | // The branch destinations need to be in the loop, and we don't gain | |||
586 | // anything by duplicating conditional branches with duplicate successors, | |||
587 | // as it's essentially the same as an unconditional branch. | |||
588 | BasicBlock *TrueDest = BI->getSuccessor(0); | |||
589 | BasicBlock *FalseDest = BI->getSuccessor(1); | |||
590 | if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) || | |||
591 | TrueDest == FalseDest) | |||
592 | return; | |||
593 | ||||
594 | // We can hoist BI if one branch destination is the successor of the other, | |||
595 | // or both have common successor which we check by seeing if the | |||
596 | // intersection of their successors is non-empty. | |||
597 | // TODO: This could be expanded to allowing branches where both ends | |||
598 | // eventually converge to a single block. | |||
599 | SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc; | |||
600 | TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest)); | |||
601 | FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest)); | |||
602 | BasicBlock *CommonSucc = nullptr; | |||
603 | if (TrueDestSucc.count(FalseDest)) { | |||
604 | CommonSucc = FalseDest; | |||
605 | } else if (FalseDestSucc.count(TrueDest)) { | |||
606 | CommonSucc = TrueDest; | |||
607 | } else { | |||
608 | set_intersect(TrueDestSucc, FalseDestSucc); | |||
609 | // If there's one common successor use that. | |||
610 | if (TrueDestSucc.size() == 1) | |||
611 | CommonSucc = *TrueDestSucc.begin(); | |||
612 | // If there's more than one pick whichever appears first in the block list | |||
613 | // (we can't use the value returned by TrueDestSucc.begin() as it's | |||
614 | // unpredicatable which element gets returned). | |||
615 | else if (!TrueDestSucc.empty()) { | |||
616 | Function *F = TrueDest->getParent(); | |||
617 | auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); }; | |||
618 | auto It = std::find_if(F->begin(), F->end(), IsSucc); | |||
619 | assert(It != F->end() && "Could not find successor in function")((It != F->end() && "Could not find successor in function" ) ? static_cast<void> (0) : __assert_fail ("It != F->end() && \"Could not find successor in function\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 619, __PRETTY_FUNCTION__)); | |||
620 | CommonSucc = &*It; | |||
621 | } | |||
622 | } | |||
623 | // The common successor has to be dominated by the branch, as otherwise | |||
624 | // there will be some other path to the successor that will not be | |||
625 | // controlled by this branch so any phi we hoist would be controlled by the | |||
626 | // wrong condition. This also takes care of avoiding hoisting of loop back | |||
627 | // edges. | |||
628 | // TODO: In some cases this could be relaxed if the successor is dominated | |||
629 | // by another block that's been hoisted and we can guarantee that the | |||
630 | // control flow has been replicated exactly. | |||
631 | if (CommonSucc && DT->dominates(BI, CommonSucc)) | |||
632 | HoistableBranches[BI] = CommonSucc; | |||
633 | } | |||
634 | ||||
635 | bool canHoistPHI(PHINode *PN) { | |||
636 | // The phi must have loop invariant operands. | |||
637 | if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN)) | |||
638 | return false; | |||
639 | // We can hoist phis if the block they are in is the target of hoistable | |||
640 | // branches which cover all of the predecessors of the block. | |||
641 | SmallPtrSet<BasicBlock *, 8> PredecessorBlocks; | |||
642 | BasicBlock *BB = PN->getParent(); | |||
643 | for (BasicBlock *PredBB : predecessors(BB)) | |||
644 | PredecessorBlocks.insert(PredBB); | |||
645 | // If we have less predecessor blocks than predecessors then the phi will | |||
646 | // have more than one incoming value for the same block which we can't | |||
647 | // handle. | |||
648 | // TODO: This could be handled be erasing some of the duplicate incoming | |||
649 | // values. | |||
650 | if (PredecessorBlocks.size() != pred_size(BB)) | |||
651 | return false; | |||
652 | for (auto &Pair : HoistableBranches) { | |||
653 | if (Pair.second == BB) { | |||
654 | // Which blocks are predecessors via this branch depends on if the | |||
655 | // branch is triangle-like or diamond-like. | |||
656 | if (Pair.first->getSuccessor(0) == BB) { | |||
657 | PredecessorBlocks.erase(Pair.first->getParent()); | |||
658 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | |||
659 | } else if (Pair.first->getSuccessor(1) == BB) { | |||
660 | PredecessorBlocks.erase(Pair.first->getParent()); | |||
661 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | |||
662 | } else { | |||
663 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | |||
664 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | |||
665 | } | |||
666 | } | |||
667 | } | |||
668 | // PredecessorBlocks will now be empty if for every predecessor of BB we | |||
669 | // found a hoistable branch source. | |||
670 | return PredecessorBlocks.empty(); | |||
671 | } | |||
672 | ||||
673 | BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) { | |||
674 | if (!ControlFlowHoisting) | |||
675 | return CurLoop->getLoopPreheader(); | |||
676 | // If BB has already been hoisted, return that | |||
677 | if (HoistDestinationMap.count(BB)) | |||
678 | return HoistDestinationMap[BB]; | |||
679 | ||||
680 | // Check if this block is conditional based on a pending branch | |||
681 | auto HasBBAsSuccessor = | |||
682 | [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) { | |||
683 | return BB != Pair.second && (Pair.first->getSuccessor(0) == BB || | |||
684 | Pair.first->getSuccessor(1) == BB); | |||
685 | }; | |||
686 | auto It = std::find_if(HoistableBranches.begin(), HoistableBranches.end(), | |||
687 | HasBBAsSuccessor); | |||
688 | ||||
689 | // If not involved in a pending branch, hoist to preheader | |||
690 | BasicBlock *InitialPreheader = CurLoop->getLoopPreheader(); | |||
691 | if (It == HoistableBranches.end()) { | |||
692 | LLVM_DEBUG(dbgs() << "LICM using " << InitialPreheader->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getName() << " as hoist destination for " << BB->getName() << "\n"; } } while (false) | |||
693 | << " as hoist destination for " << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getName() << " as hoist destination for " << BB->getName() << "\n"; } } while (false) | |||
694 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getName() << " as hoist destination for " << BB->getName() << "\n"; } } while (false); | |||
695 | HoistDestinationMap[BB] = InitialPreheader; | |||
696 | return InitialPreheader; | |||
697 | } | |||
698 | BranchInst *BI = It->first; | |||
699 | assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 701, __PRETTY_FUNCTION__)) | |||
700 | HoistableBranches.end() &&((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 701, __PRETTY_FUNCTION__)) | |||
701 | "BB is expected to be the target of at most one branch")((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 701, __PRETTY_FUNCTION__)); | |||
702 | ||||
703 | LLVMContext &C = BB->getContext(); | |||
704 | BasicBlock *TrueDest = BI->getSuccessor(0); | |||
705 | BasicBlock *FalseDest = BI->getSuccessor(1); | |||
706 | BasicBlock *CommonSucc = HoistableBranches[BI]; | |||
707 | BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent()); | |||
708 | ||||
709 | // Create hoisted versions of blocks that currently don't have them | |||
710 | auto CreateHoistedBlock = [&](BasicBlock *Orig) { | |||
711 | if (HoistDestinationMap.count(Orig)) | |||
712 | return HoistDestinationMap[Orig]; | |||
713 | BasicBlock *New = | |||
714 | BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent()); | |||
715 | HoistDestinationMap[Orig] = New; | |||
716 | DT->addNewBlock(New, HoistTarget); | |||
717 | if (CurLoop->getParentLoop()) | |||
718 | CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI); | |||
719 | ++NumCreatedBlocks; | |||
720 | LLVM_DEBUG(dbgs() << "LICM created " << New->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false) | |||
721 | << " as hoist destination for " << Orig->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false) | |||
722 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false); | |||
723 | return New; | |||
724 | }; | |||
725 | BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest); | |||
726 | BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest); | |||
727 | BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc); | |||
728 | ||||
729 | // Link up these blocks with branches. | |||
730 | if (!HoistCommonSucc->getTerminator()) { | |||
731 | // The new common successor we've generated will branch to whatever that | |||
732 | // hoist target branched to. | |||
733 | BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor(); | |||
734 | assert(TargetSucc && "Expected hoist target to have a single successor")((TargetSucc && "Expected hoist target to have a single successor" ) ? static_cast<void> (0) : __assert_fail ("TargetSucc && \"Expected hoist target to have a single successor\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 734, __PRETTY_FUNCTION__)); | |||
735 | HoistCommonSucc->moveBefore(TargetSucc); | |||
736 | BranchInst::Create(TargetSucc, HoistCommonSucc); | |||
737 | } | |||
738 | if (!HoistTrueDest->getTerminator()) { | |||
739 | HoistTrueDest->moveBefore(HoistCommonSucc); | |||
740 | BranchInst::Create(HoistCommonSucc, HoistTrueDest); | |||
741 | } | |||
742 | if (!HoistFalseDest->getTerminator()) { | |||
743 | HoistFalseDest->moveBefore(HoistCommonSucc); | |||
744 | BranchInst::Create(HoistCommonSucc, HoistFalseDest); | |||
745 | } | |||
746 | ||||
747 | // If BI is being cloned to what was originally the preheader then | |||
748 | // HoistCommonSucc will now be the new preheader. | |||
749 | if (HoistTarget == InitialPreheader) { | |||
750 | // Phis in the loop header now need to use the new preheader. | |||
751 | InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc); | |||
752 | if (MSSAU) | |||
753 | MSSAU->wireOldPredecessorsToNewImmediatePredecessor( | |||
754 | HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget}); | |||
755 | // The new preheader dominates the loop header. | |||
756 | DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc); | |||
757 | DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader()); | |||
758 | DT->changeImmediateDominator(HeaderNode, PreheaderNode); | |||
759 | // The preheader hoist destination is now the new preheader, with the | |||
760 | // exception of the hoist destination of this branch. | |||
761 | for (auto &Pair : HoistDestinationMap) | |||
762 | if (Pair.second == InitialPreheader && Pair.first != BI->getParent()) | |||
763 | Pair.second = HoistCommonSucc; | |||
764 | } | |||
765 | ||||
766 | // Now finally clone BI. | |||
767 | ReplaceInstWithInst( | |||
768 | HoistTarget->getTerminator(), | |||
769 | BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition())); | |||
770 | ++NumClonedBranches; | |||
771 | ||||
772 | assert(CurLoop->getLoopPreheader() &&((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 773, __PRETTY_FUNCTION__)) | |||
773 | "Hoisting blocks should not have destroyed preheader")((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 773, __PRETTY_FUNCTION__)); | |||
774 | return HoistDestinationMap[BB]; | |||
775 | } | |||
776 | }; | |||
777 | } // namespace | |||
778 | ||||
779 | /// Walk the specified region of the CFG (defined by all blocks dominated by | |||
780 | /// the specified block, and that are in the current loop) in depth first | |||
781 | /// order w.r.t the DominatorTree. This allows us to visit definitions before | |||
782 | /// uses, allowing us to hoist a loop body in one pass without iteration. | |||
783 | /// | |||
784 | bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI, | |||
785 | DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop, | |||
786 | AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | |||
787 | ICFLoopSafetyInfo *SafetyInfo, | |||
788 | SinkAndHoistLICMFlags &Flags, | |||
789 | OptimizationRemarkEmitter *ORE) { | |||
790 | // Verify inputs. | |||
791 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 793, __PRETTY_FUNCTION__)) | |||
792 | CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 793, __PRETTY_FUNCTION__)) | |||
793 | "Unexpected input to hoistRegion.")((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 793, __PRETTY_FUNCTION__)); | |||
794 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 795, __PRETTY_FUNCTION__)) | |||
795 | "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 795, __PRETTY_FUNCTION__)); | |||
796 | ||||
797 | ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU); | |||
798 | ||||
799 | // Keep track of instructions that have been hoisted, as they may need to be | |||
800 | // re-hoisted if they end up not dominating all of their uses. | |||
801 | SmallVector<Instruction *, 16> HoistedInstructions; | |||
802 | ||||
803 | // For PHI hoisting to work we need to hoist blocks before their successors. | |||
804 | // We can do this by iterating through the blocks in the loop in reverse | |||
805 | // post-order. | |||
806 | LoopBlocksRPO Worklist(CurLoop); | |||
807 | Worklist.perform(LI); | |||
808 | bool Changed = false; | |||
809 | for (BasicBlock *BB : Worklist) { | |||
810 | // Only need to process the contents of this block if it is not part of a | |||
811 | // subloop (which would already have been processed). | |||
812 | if (inSubLoop(BB, CurLoop, LI)) | |||
813 | continue; | |||
814 | ||||
815 | for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) { | |||
816 | Instruction &I = *II++; | |||
817 | // Try constant folding this instruction. If all the operands are | |||
818 | // constants, it is technically hoistable, but it would be better to | |||
819 | // just fold it. | |||
820 | if (Constant *C = ConstantFoldInstruction( | |||
821 | &I, I.getModule()->getDataLayout(), TLI)) { | |||
822 | LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false) | |||
823 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false); | |||
824 | if (CurAST) | |||
825 | CurAST->copyValue(&I, C); | |||
826 | // FIXME MSSA: Such replacements may make accesses unoptimized (D51960). | |||
827 | I.replaceAllUsesWith(C); | |||
828 | if (isInstructionTriviallyDead(&I, TLI)) | |||
829 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | |||
830 | Changed = true; | |||
831 | continue; | |||
832 | } | |||
833 | ||||
834 | // Try hoisting the instruction out to the preheader. We can only do | |||
835 | // this if all of the operands of the instruction are loop invariant and | |||
836 | // if it is safe to hoist the instruction. | |||
837 | // TODO: It may be safe to hoist if we are hoisting to a conditional block | |||
838 | // and we have accurately duplicated the control flow from the loop header | |||
839 | // to that block. | |||
840 | if (CurLoop->hasLoopInvariantOperands(&I) && | |||
841 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | |||
842 | ORE) && | |||
843 | isSafeToExecuteUnconditionally( | |||
844 | I, DT, CurLoop, SafetyInfo, ORE, | |||
845 | CurLoop->getLoopPreheader()->getTerminator())) { | |||
846 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | |||
847 | MSSAU, ORE); | |||
848 | HoistedInstructions.push_back(&I); | |||
849 | Changed = true; | |||
850 | continue; | |||
851 | } | |||
852 | ||||
853 | // Attempt to remove floating point division out of the loop by | |||
854 | // converting it to a reciprocal multiplication. | |||
855 | if (I.getOpcode() == Instruction::FDiv && | |||
856 | CurLoop->isLoopInvariant(I.getOperand(1)) && | |||
857 | I.hasAllowReciprocal()) { | |||
858 | auto Divisor = I.getOperand(1); | |||
859 | auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0); | |||
860 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor); | |||
861 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); | |||
862 | SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent()); | |||
863 | ReciprocalDivisor->insertBefore(&I); | |||
864 | ||||
865 | auto Product = | |||
866 | BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor); | |||
867 | Product->setFastMathFlags(I.getFastMathFlags()); | |||
868 | SafetyInfo->insertInstructionTo(Product, I.getParent()); | |||
869 | Product->insertAfter(&I); | |||
870 | I.replaceAllUsesWith(Product); | |||
871 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | |||
872 | ||||
873 | hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), | |||
874 | SafetyInfo, MSSAU, ORE); | |||
875 | HoistedInstructions.push_back(ReciprocalDivisor); | |||
876 | Changed = true; | |||
877 | continue; | |||
878 | } | |||
879 | ||||
880 | auto IsInvariantStart = [&](Instruction &I) { | |||
881 | using namespace PatternMatch; | |||
882 | return I.use_empty() && | |||
883 | match(&I, m_Intrinsic<Intrinsic::invariant_start>()); | |||
884 | }; | |||
885 | auto MustExecuteWithoutWritesBefore = [&](Instruction &I) { | |||
886 | return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) && | |||
887 | SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop); | |||
888 | }; | |||
889 | if ((IsInvariantStart(I) || isGuard(&I)) && | |||
890 | CurLoop->hasLoopInvariantOperands(&I) && | |||
891 | MustExecuteWithoutWritesBefore(I)) { | |||
892 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | |||
893 | MSSAU, ORE); | |||
894 | HoistedInstructions.push_back(&I); | |||
895 | Changed = true; | |||
896 | continue; | |||
897 | } | |||
898 | ||||
899 | if (PHINode *PN = dyn_cast<PHINode>(&I)) { | |||
900 | if (CFH.canHoistPHI(PN)) { | |||
901 | // Redirect incoming blocks first to ensure that we create hoisted | |||
902 | // versions of those blocks before we hoist the phi. | |||
903 | for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i) | |||
904 | PN->setIncomingBlock( | |||
905 | i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i))); | |||
906 | hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | |||
907 | MSSAU, ORE); | |||
908 | assert(DT->dominates(PN, BB) && "Conditional PHIs not expected")((DT->dominates(PN, BB) && "Conditional PHIs not expected" ) ? static_cast<void> (0) : __assert_fail ("DT->dominates(PN, BB) && \"Conditional PHIs not expected\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 908, __PRETTY_FUNCTION__)); | |||
909 | Changed = true; | |||
910 | continue; | |||
911 | } | |||
912 | } | |||
913 | ||||
914 | // Remember possibly hoistable branches so we can actually hoist them | |||
915 | // later if needed. | |||
916 | if (BranchInst *BI = dyn_cast<BranchInst>(&I)) | |||
917 | CFH.registerPossiblyHoistableBranch(BI); | |||
918 | } | |||
919 | } | |||
920 | ||||
921 | // If we hoisted instructions to a conditional block they may not dominate | |||
922 | // their uses that weren't hoisted (such as phis where some operands are not | |||
923 | // loop invariant). If so make them unconditional by moving them to their | |||
924 | // immediate dominator. We iterate through the instructions in reverse order | |||
925 | // which ensures that when we rehoist an instruction we rehoist its operands, | |||
926 | // and also keep track of where in the block we are rehoisting to to make sure | |||
927 | // that we rehoist instructions before the instructions that use them. | |||
928 | Instruction *HoistPoint = nullptr; | |||
929 | if (ControlFlowHoisting) { | |||
930 | for (Instruction *I : reverse(HoistedInstructions)) { | |||
931 | if (!llvm::all_of(I->uses(), | |||
932 | [&](Use &U) { return DT->dominates(I, U); })) { | |||
933 | BasicBlock *Dominator = | |||
934 | DT->getNode(I->getParent())->getIDom()->getBlock(); | |||
935 | if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) { | |||
936 | if (HoistPoint) | |||
937 | assert(DT->dominates(Dominator, HoistPoint->getParent()) &&((DT->dominates(Dominator, HoistPoint->getParent()) && "New hoist point expected to dominate old hoist point") ? static_cast <void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 938, __PRETTY_FUNCTION__)) | |||
938 | "New hoist point expected to dominate old hoist point")((DT->dominates(Dominator, HoistPoint->getParent()) && "New hoist point expected to dominate old hoist point") ? static_cast <void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 938, __PRETTY_FUNCTION__)); | |||
939 | HoistPoint = Dominator->getTerminator(); | |||
940 | } | |||
941 | LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getName() << ": " << *I << "\n"; } } while (false) | |||
942 | << HoistPoint->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getName() << ": " << *I << "\n"; } } while (false) | |||
943 | << ": " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getName() << ": " << *I << "\n"; } } while (false); | |||
944 | moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU); | |||
945 | HoistPoint = I; | |||
946 | Changed = true; | |||
947 | } | |||
948 | } | |||
949 | } | |||
950 | if (MSSAU && VerifyMemorySSA) | |||
951 | MSSAU->getMemorySSA()->verifyMemorySSA(); | |||
952 | ||||
953 | // Now that we've finished hoisting make sure that LI and DT are still | |||
954 | // valid. | |||
955 | #ifndef NDEBUG | |||
956 | if (Changed) { | |||
957 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&((DT->verify(DominatorTree::VerificationLevel::Fast) && "Dominator tree verification failed") ? static_cast<void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 958, __PRETTY_FUNCTION__)) | |||
958 | "Dominator tree verification failed")((DT->verify(DominatorTree::VerificationLevel::Fast) && "Dominator tree verification failed") ? static_cast<void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 958, __PRETTY_FUNCTION__)); | |||
959 | LI->verify(*DT); | |||
960 | } | |||
961 | #endif | |||
962 | ||||
963 | return Changed; | |||
964 | } | |||
965 | ||||
966 | // Return true if LI is invariant within scope of the loop. LI is invariant if | |||
967 | // CurLoop is dominated by an invariant.start representing the same memory | |||
968 | // location and size as the memory location LI loads from, and also the | |||
969 | // invariant.start has no uses. | |||
970 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, | |||
971 | Loop *CurLoop) { | |||
972 | Value *Addr = LI->getOperand(0); | |||
973 | const DataLayout &DL = LI->getModule()->getDataLayout(); | |||
974 | const uint32_t LocSizeInBits = DL.getTypeSizeInBits( | |||
975 | cast<PointerType>(Addr->getType())->getElementType()); | |||
976 | ||||
977 | // if the type is i8 addrspace(x)*, we know this is the type of | |||
978 | // llvm.invariant.start operand | |||
979 | auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()), | |||
980 | LI->getPointerAddressSpace()); | |||
981 | unsigned BitcastsVisited = 0; | |||
982 | // Look through bitcasts until we reach the i8* type (this is invariant.start | |||
983 | // operand type). | |||
984 | while (Addr->getType() != PtrInt8Ty) { | |||
985 | auto *BC = dyn_cast<BitCastInst>(Addr); | |||
986 | // Avoid traversing high number of bitcast uses. | |||
987 | if (++BitcastsVisited > MaxNumUsesTraversed || !BC) | |||
988 | return false; | |||
989 | Addr = BC->getOperand(0); | |||
990 | } | |||
991 | ||||
992 | unsigned UsesVisited = 0; | |||
993 | // Traverse all uses of the load operand value, to see if invariant.start is | |||
994 | // one of the uses, and whether it dominates the load instruction. | |||
995 | for (auto *U : Addr->users()) { | |||
996 | // Avoid traversing for Load operand with high number of users. | |||
997 | if (++UsesVisited > MaxNumUsesTraversed) | |||
998 | return false; | |||
999 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | |||
1000 | // If there are escaping uses of invariant.start instruction, the load maybe | |||
1001 | // non-invariant. | |||
1002 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || | |||
1003 | !II->use_empty()) | |||
1004 | continue; | |||
1005 | unsigned InvariantSizeInBits = | |||
1006 | cast<ConstantInt>(II->getArgOperand(0))->getSExtValue() * 8; | |||
1007 | // Confirm the invariant.start location size contains the load operand size | |||
1008 | // in bits. Also, the invariant.start should dominate the load, and we | |||
1009 | // should not hoist the load out of a loop that contains this dominating | |||
1010 | // invariant.start. | |||
1011 | if (LocSizeInBits <= InvariantSizeInBits && | |||
1012 | DT->properlyDominates(II->getParent(), CurLoop->getHeader())) | |||
1013 | return true; | |||
1014 | } | |||
1015 | ||||
1016 | return false; | |||
1017 | } | |||
1018 | ||||
1019 | namespace { | |||
1020 | /// Return true if-and-only-if we know how to (mechanically) both hoist and | |||
1021 | /// sink a given instruction out of a loop. Does not address legality | |||
1022 | /// concerns such as aliasing or speculation safety. | |||
1023 | bool isHoistableAndSinkableInst(Instruction &I) { | |||
1024 | // Only these instructions are hoistable/sinkable. | |||
1025 | return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) || | |||
1026 | isa<FenceInst>(I) || isa<BinaryOperator>(I) || isa<CastInst>(I) || | |||
1027 | isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || | |||
1028 | isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | |||
1029 | isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) || | |||
1030 | isa<InsertValueInst>(I)); | |||
1031 | } | |||
1032 | /// Return true if all of the alias sets within this AST are known not to | |||
1033 | /// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop. | |||
1034 | bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU, | |||
1035 | const Loop *L) { | |||
1036 | if (CurAST) { | |||
1037 | for (AliasSet &AS : *CurAST) { | |||
1038 | if (!AS.isForwardingAliasSet() && AS.isMod()) { | |||
1039 | return false; | |||
1040 | } | |||
1041 | } | |||
1042 | return true; | |||
1043 | } else { /*MSSAU*/ | |||
1044 | for (auto *BB : L->getBlocks()) | |||
1045 | if (MSSAU->getMemorySSA()->getBlockDefs(BB)) | |||
1046 | return false; | |||
1047 | return true; | |||
1048 | } | |||
1049 | } | |||
1050 | ||||
1051 | /// Return true if I is the only Instruction with a MemoryAccess in L. | |||
1052 | bool isOnlyMemoryAccess(const Instruction *I, const Loop *L, | |||
1053 | const MemorySSAUpdater *MSSAU) { | |||
1054 | for (auto *BB : L->getBlocks()) | |||
1055 | if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) { | |||
1056 | int NotAPhi = 0; | |||
1057 | for (const auto &Acc : *Accs) { | |||
1058 | if (isa<MemoryPhi>(&Acc)) | |||
1059 | continue; | |||
1060 | const auto *MUD = cast<MemoryUseOrDef>(&Acc); | |||
1061 | if (MUD->getMemoryInst() != I || NotAPhi++ == 1) | |||
1062 | return false; | |||
1063 | } | |||
1064 | } | |||
1065 | return true; | |||
1066 | } | |||
1067 | } | |||
1068 | ||||
1069 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, | |||
1070 | Loop *CurLoop, AliasSetTracker *CurAST, | |||
1071 | MemorySSAUpdater *MSSAU, | |||
1072 | bool TargetExecutesOncePerLoop, | |||
1073 | SinkAndHoistLICMFlags *Flags, | |||
1074 | OptimizationRemarkEmitter *ORE) { | |||
1075 | // If we don't understand the instruction, bail early. | |||
1076 | if (!isHoistableAndSinkableInst(I)) | |||
| ||||
1077 | return false; | |||
1078 | ||||
1079 | MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr; | |||
1080 | if (MSSA) | |||
1081 | assert(Flags != nullptr && "Flags cannot be null.")((Flags != nullptr && "Flags cannot be null.") ? static_cast <void> (0) : __assert_fail ("Flags != nullptr && \"Flags cannot be null.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1081, __PRETTY_FUNCTION__)); | |||
1082 | ||||
1083 | // Loads have extra constraints we have to verify before we can hoist them. | |||
1084 | if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { | |||
1085 | if (!LI->isUnordered()) | |||
1086 | return false; // Don't sink/hoist volatile or ordered atomic loads! | |||
1087 | ||||
1088 | // Loads from constant memory are always safe to move, even if they end up | |||
1089 | // in the same alias set as something that ends up being modified. | |||
1090 | if (AA->pointsToConstantMemory(LI->getOperand(0))) | |||
1091 | return true; | |||
1092 | if (LI->getMetadata(LLVMContext::MD_invariant_load)) | |||
1093 | return true; | |||
1094 | ||||
1095 | if (LI->isAtomic() && !TargetExecutesOncePerLoop) | |||
1096 | return false; // Don't risk duplicating unordered loads | |||
1097 | ||||
1098 | // This checks for an invariant.start dominating the load. | |||
1099 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) | |||
1100 | return true; | |||
1101 | ||||
1102 | bool Invalidated; | |||
1103 | if (CurAST) | |||
1104 | Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST, | |||
1105 | CurLoop, AA); | |||
1106 | else | |||
1107 | Invalidated = pointerInvalidatedByLoopWithMSSA( | |||
1108 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, *Flags); | |||
1109 | // Check loop-invariant address because this may also be a sinkable load | |||
1110 | // whose address is not necessarily loop-invariant. | |||
1111 | if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand())) | |||
1112 | ORE->emit([&]() { | |||
1113 | return OptimizationRemarkMissed( | |||
1114 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI) | |||
1115 | << "failed to move load with loop-invariant address " | |||
1116 | "because the loop may invalidate its value"; | |||
1117 | }); | |||
1118 | ||||
1119 | return !Invalidated; | |||
1120 | } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { | |||
1121 | // Don't sink or hoist dbg info; it's legal, but not useful. | |||
1122 | if (isa<DbgInfoIntrinsic>(I)) | |||
1123 | return false; | |||
1124 | ||||
1125 | // Don't sink calls which can throw. | |||
1126 | if (CI->mayThrow()) | |||
1127 | return false; | |||
1128 | ||||
1129 | using namespace PatternMatch; | |||
1130 | if (match(CI, m_Intrinsic<Intrinsic::assume>())) | |||
1131 | // Assumes don't actually alias anything or throw | |||
1132 | return true; | |||
1133 | ||||
1134 | // Handle simple cases by querying alias analysis. | |||
1135 | FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI); | |||
1136 | if (Behavior == FMRB_DoesNotAccessMemory) | |||
1137 | return true; | |||
1138 | if (AliasAnalysis::onlyReadsMemory(Behavior)) { | |||
1139 | // A readonly argmemonly function only reads from memory pointed to by | |||
1140 | // it's arguments with arbitrary offsets. If we can prove there are no | |||
1141 | // writes to this memory in the loop, we can hoist or sink. | |||
1142 | if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) { | |||
1143 | // TODO: expand to writeable arguments | |||
1144 | for (Value *Op : CI->arg_operands()) | |||
1145 | if (Op->getType()->isPointerTy()) { | |||
1146 | bool Invalidated; | |||
1147 | if (CurAST) | |||
1148 | Invalidated = pointerInvalidatedByLoop( | |||
1149 | MemoryLocation(Op, LocationSize::unknown(), AAMDNodes()), | |||
1150 | CurAST, CurLoop, AA); | |||
1151 | else | |||
1152 | Invalidated = pointerInvalidatedByLoopWithMSSA( | |||
1153 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, | |||
1154 | *Flags); | |||
1155 | if (Invalidated) | |||
1156 | return false; | |||
1157 | } | |||
1158 | return true; | |||
1159 | } | |||
1160 | ||||
1161 | // If this call only reads from memory and there are no writes to memory | |||
1162 | // in the loop, we can hoist or sink the call as appropriate. | |||
1163 | if (isReadOnly(CurAST, MSSAU, CurLoop)) | |||
1164 | return true; | |||
1165 | } | |||
1166 | ||||
1167 | // FIXME: This should use mod/ref information to see if we can hoist or | |||
1168 | // sink the call. | |||
1169 | ||||
1170 | return false; | |||
1171 | } else if (auto *FI = dyn_cast<FenceInst>(&I)) { | |||
1172 | // Fences alias (most) everything to provide ordering. For the moment, | |||
1173 | // just give up if there are any other memory operations in the loop. | |||
1174 | if (CurAST) { | |||
1175 | auto Begin = CurAST->begin(); | |||
1176 | assert(Begin != CurAST->end() && "must contain FI")((Begin != CurAST->end() && "must contain FI") ? static_cast <void> (0) : __assert_fail ("Begin != CurAST->end() && \"must contain FI\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1176, __PRETTY_FUNCTION__)); | |||
1177 | if (std::next(Begin) != CurAST->end()) | |||
1178 | // constant memory for instance, TODO: handle better | |||
1179 | return false; | |||
1180 | auto *UniqueI = Begin->getUniqueInstruction(); | |||
1181 | if (!UniqueI) | |||
1182 | // other memory op, give up | |||
1183 | return false; | |||
1184 | (void)FI; // suppress unused variable warning | |||
1185 | assert(UniqueI == FI && "AS must contain FI")((UniqueI == FI && "AS must contain FI") ? static_cast <void> (0) : __assert_fail ("UniqueI == FI && \"AS must contain FI\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1185, __PRETTY_FUNCTION__)); | |||
1186 | return true; | |||
1187 | } else // MSSAU | |||
1188 | return isOnlyMemoryAccess(FI, CurLoop, MSSAU); | |||
1189 | } else if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
1190 | if (!SI->isUnordered()) | |||
1191 | return false; // Don't sink/hoist volatile or ordered atomic store! | |||
1192 | ||||
1193 | // We can only hoist a store that we can prove writes a value which is not | |||
1194 | // read or overwritten within the loop. For those cases, we fallback to | |||
1195 | // load store promotion instead. TODO: We can extend this to cases where | |||
1196 | // there is exactly one write to the location and that write dominates an | |||
1197 | // arbitrary number of reads in the loop. | |||
1198 | if (CurAST) { | |||
1199 | auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI)); | |||
1200 | ||||
1201 | if (AS.isRef() || !AS.isMustAlias()) | |||
1202 | // Quick exit test, handled by the full path below as well. | |||
1203 | return false; | |||
1204 | auto *UniqueI = AS.getUniqueInstruction(); | |||
1205 | if (!UniqueI) | |||
1206 | // other memory op, give up | |||
1207 | return false; | |||
1208 | assert(UniqueI == SI && "AS must contain SI")((UniqueI == SI && "AS must contain SI") ? static_cast <void> (0) : __assert_fail ("UniqueI == SI && \"AS must contain SI\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1208, __PRETTY_FUNCTION__)); | |||
1209 | return true; | |||
1210 | } else { // MSSAU | |||
1211 | if (isOnlyMemoryAccess(SI, CurLoop, MSSAU)) | |||
1212 | return true; | |||
1213 | // If there are more accesses than the Promotion cap, give up, we're not | |||
1214 | // walking a list that long. | |||
1215 | if (Flags->NoOfMemAccTooLarge) | |||
1216 | return false; | |||
1217 | // Check store only if there's still "quota" to check clobber. | |||
1218 | if (Flags->LicmMssaOptCounter >= Flags->LicmMssaOptCap) | |||
1219 | return false; | |||
1220 | // If there are interfering Uses (i.e. their defining access is in the | |||
1221 | // loop), or ordered loads (stored as Defs!), don't move this store. | |||
1222 | // Could do better here, but this is conservatively correct. | |||
1223 | // TODO: Cache set of Uses on the first walk in runOnLoop, update when | |||
1224 | // moving accesses. Can also extend to dominating uses. | |||
1225 | for (auto *BB : CurLoop->getBlocks()) | |||
1226 | if (auto *Accesses = MSSA->getBlockAccesses(BB)) { | |||
| ||||
1227 | for (const auto &MA : *Accesses) | |||
1228 | if (const auto *MU = dyn_cast<MemoryUse>(&MA)) { | |||
1229 | auto *MD = MU->getDefiningAccess(); | |||
1230 | if (!MSSA->isLiveOnEntryDef(MD) && | |||
1231 | CurLoop->contains(MD->getBlock())) | |||
1232 | return false; | |||
1233 | } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) | |||
1234 | if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) { | |||
1235 | (void)LI; // Silence warning. | |||
1236 | assert(!LI->isUnordered() && "Expected unordered load")((!LI->isUnordered() && "Expected unordered load") ? static_cast<void> (0) : __assert_fail ("!LI->isUnordered() && \"Expected unordered load\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1236, __PRETTY_FUNCTION__)); | |||
1237 | return false; | |||
1238 | } | |||
1239 | } | |||
1240 | ||||
1241 | auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI); | |||
1242 | Flags->LicmMssaOptCounter++; | |||
1243 | // If there are no clobbering Defs in the loop, store is safe to hoist. | |||
1244 | return MSSA->isLiveOnEntryDef(Source) || | |||
1245 | !CurLoop->contains(Source->getBlock()); | |||
1246 | } | |||
1247 | } | |||
1248 | ||||
1249 | assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((!I.mayReadOrWriteMemory() && "unhandled aliasing") ? static_cast<void> (0) : __assert_fail ("!I.mayReadOrWriteMemory() && \"unhandled aliasing\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1249, __PRETTY_FUNCTION__)); | |||
1250 | ||||
1251 | // We've established mechanical ability and aliasing, it's up to the caller | |||
1252 | // to check fault safety | |||
1253 | return true; | |||
1254 | } | |||
1255 | ||||
1256 | /// Returns true if a PHINode is a trivially replaceable with an | |||
1257 | /// Instruction. | |||
1258 | /// This is true when all incoming values are that instruction. | |||
1259 | /// This pattern occurs most often with LCSSA PHI nodes. | |||
1260 | /// | |||
1261 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { | |||
1262 | for (const Value *IncValue : PN.incoming_values()) | |||
1263 | if (IncValue != &I) | |||
1264 | return false; | |||
1265 | ||||
1266 | return true; | |||
1267 | } | |||
1268 | ||||
1269 | /// Return true if the instruction is free in the loop. | |||
1270 | static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
1271 | const TargetTransformInfo *TTI) { | |||
1272 | ||||
1273 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
1274 | if (TTI->getUserCost(GEP) != TargetTransformInfo::TCC_Free) | |||
1275 | return false; | |||
1276 | // For a GEP, we cannot simply use getUserCost because currently it | |||
1277 | // optimistically assume that a GEP will fold into addressing mode | |||
1278 | // regardless of its users. | |||
1279 | const BasicBlock *BB = GEP->getParent(); | |||
1280 | for (const User *U : GEP->users()) { | |||
1281 | const Instruction *UI = cast<Instruction>(U); | |||
1282 | if (CurLoop->contains(UI) && | |||
1283 | (BB != UI->getParent() || | |||
1284 | (!isa<StoreInst>(UI) && !isa<LoadInst>(UI)))) | |||
1285 | return false; | |||
1286 | } | |||
1287 | return true; | |||
1288 | } else | |||
1289 | return TTI->getUserCost(&I) == TargetTransformInfo::TCC_Free; | |||
1290 | } | |||
1291 | ||||
1292 | /// Return true if the only users of this instruction are outside of | |||
1293 | /// the loop. If this is true, we can sink the instruction to the exit | |||
1294 | /// blocks of the loop. | |||
1295 | /// | |||
1296 | /// We also return true if the instruction could be folded away in lowering. | |||
1297 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). | |||
1298 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | |||
1299 | const LoopSafetyInfo *SafetyInfo, | |||
1300 | TargetTransformInfo *TTI, bool &FreeInLoop) { | |||
1301 | const auto &BlockColors = SafetyInfo->getBlockColors(); | |||
1302 | bool IsFree = isFreeInLoop(I, CurLoop, TTI); | |||
1303 | for (const User *U : I.users()) { | |||
1304 | const Instruction *UI = cast<Instruction>(U); | |||
1305 | if (const PHINode *PN = dyn_cast<PHINode>(UI)) { | |||
1306 | const BasicBlock *BB = PN->getParent(); | |||
1307 | // We cannot sink uses in catchswitches. | |||
1308 | if (isa<CatchSwitchInst>(BB->getTerminator())) | |||
1309 | return false; | |||
1310 | ||||
1311 | // We need to sink a callsite to a unique funclet. Avoid sinking if the | |||
1312 | // phi use is too muddled. | |||
1313 | if (isa<CallInst>(I)) | |||
1314 | if (!BlockColors.empty() && | |||
1315 | BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1) | |||
1316 | return false; | |||
1317 | } | |||
1318 | ||||
1319 | if (CurLoop->contains(UI)) { | |||
1320 | if (IsFree) { | |||
1321 | FreeInLoop = true; | |||
1322 | continue; | |||
1323 | } | |||
1324 | return false; | |||
1325 | } | |||
1326 | } | |||
1327 | return true; | |||
1328 | } | |||
1329 | ||||
1330 | static Instruction *CloneInstructionInExitBlock( | |||
1331 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | |||
1332 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) { | |||
1333 | Instruction *New; | |||
1334 | if (auto *CI = dyn_cast<CallInst>(&I)) { | |||
1335 | const auto &BlockColors = SafetyInfo->getBlockColors(); | |||
1336 | ||||
1337 | // Sinking call-sites need to be handled differently from other | |||
1338 | // instructions. The cloned call-site needs a funclet bundle operand | |||
1339 | // appropriate for its location in the CFG. | |||
1340 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
1341 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); | |||
1342 | BundleIdx != BundleEnd; ++BundleIdx) { | |||
1343 | OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx); | |||
1344 | if (Bundle.getTagID() == LLVMContext::OB_funclet) | |||
1345 | continue; | |||
1346 | ||||
1347 | OpBundles.emplace_back(Bundle); | |||
1348 | } | |||
1349 | ||||
1350 | if (!BlockColors.empty()) { | |||
1351 | const ColorVector &CV = BlockColors.find(&ExitBlock)->second; | |||
1352 | assert(CV.size() == 1 && "non-unique color for exit block!")((CV.size() == 1 && "non-unique color for exit block!" ) ? static_cast<void> (0) : __assert_fail ("CV.size() == 1 && \"non-unique color for exit block!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1352, __PRETTY_FUNCTION__)); | |||
1353 | BasicBlock *BBColor = CV.front(); | |||
1354 | Instruction *EHPad = BBColor->getFirstNonPHI(); | |||
1355 | if (EHPad->isEHPad()) | |||
1356 | OpBundles.emplace_back("funclet", EHPad); | |||
1357 | } | |||
1358 | ||||
1359 | New = CallInst::Create(CI, OpBundles); | |||
1360 | } else { | |||
1361 | New = I.clone(); | |||
1362 | } | |||
1363 | ||||
1364 | ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New); | |||
1365 | if (!I.getName().empty()) | |||
1366 | New->setName(I.getName() + ".le"); | |||
1367 | ||||
1368 | MemoryAccess *OldMemAcc; | |||
1369 | if (MSSAU && (OldMemAcc = MSSAU->getMemorySSA()->getMemoryAccess(&I))) { | |||
1370 | // Create a new MemoryAccess and let MemorySSA set its defining access. | |||
1371 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( | |||
1372 | New, nullptr, New->getParent(), MemorySSA::Beginning); | |||
1373 | if (NewMemAcc) { | |||
1374 | if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc)) | |||
1375 | MSSAU->insertDef(MemDef, /*RenameUses=*/true); | |||
1376 | else { | |||
1377 | auto *MemUse = cast<MemoryUse>(NewMemAcc); | |||
1378 | MSSAU->insertUse(MemUse); | |||
1379 | } | |||
1380 | } | |||
1381 | } | |||
1382 | ||||
1383 | // Build LCSSA PHI nodes for any in-loop operands. Note that this is | |||
1384 | // particularly cheap because we can rip off the PHI node that we're | |||
1385 | // replacing for the number and blocks of the predecessors. | |||
1386 | // OPT: If this shows up in a profile, we can instead finish sinking all | |||
1387 | // invariant instructions, and then walk their operands to re-establish | |||
1388 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when | |||
1389 | // sinking bottom-up. | |||
1390 | for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE; | |||
1391 | ++OI) | |||
1392 | if (Instruction *OInst = dyn_cast<Instruction>(*OI)) | |||
1393 | if (Loop *OLoop = LI->getLoopFor(OInst->getParent())) | |||
1394 | if (!OLoop->contains(&PN)) { | |||
1395 | PHINode *OpPN = | |||
1396 | PHINode::Create(OInst->getType(), PN.getNumIncomingValues(), | |||
1397 | OInst->getName() + ".lcssa", &ExitBlock.front()); | |||
1398 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) | |||
1399 | OpPN->addIncoming(OInst, PN.getIncomingBlock(i)); | |||
1400 | *OI = OpPN; | |||
1401 | } | |||
1402 | return New; | |||
1403 | } | |||
1404 | ||||
1405 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | |||
1406 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU) { | |||
1407 | if (AST) | |||
1408 | AST->deleteValue(&I); | |||
1409 | if (MSSAU) | |||
1410 | MSSAU->removeMemoryAccess(&I); | |||
1411 | SafetyInfo.removeInstruction(&I); | |||
1412 | I.eraseFromParent(); | |||
1413 | } | |||
1414 | ||||
1415 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | |||
1416 | ICFLoopSafetyInfo &SafetyInfo, | |||
1417 | MemorySSAUpdater *MSSAU) { | |||
1418 | SafetyInfo.removeInstruction(&I); | |||
1419 | SafetyInfo.insertInstructionTo(&I, Dest.getParent()); | |||
1420 | I.moveBefore(&Dest); | |||
1421 | if (MSSAU) | |||
1422 | if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>( | |||
1423 | MSSAU->getMemorySSA()->getMemoryAccess(&I))) | |||
1424 | MSSAU->moveToPlace(OldMemAcc, Dest.getParent(), MemorySSA::End); | |||
1425 | } | |||
1426 | ||||
1427 | static Instruction *sinkThroughTriviallyReplaceablePHI( | |||
1428 | PHINode *TPN, Instruction *I, LoopInfo *LI, | |||
1429 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, | |||
1430 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop, | |||
1431 | MemorySSAUpdater *MSSAU) { | |||
1432 | assert(isTriviallyReplaceablePHI(*TPN, *I) &&((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI" ) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1433, __PRETTY_FUNCTION__)) | |||
1433 | "Expect only trivially replaceable PHI")((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI" ) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1433, __PRETTY_FUNCTION__)); | |||
1434 | BasicBlock *ExitBlock = TPN->getParent(); | |||
1435 | Instruction *New; | |||
1436 | auto It = SunkCopies.find(ExitBlock); | |||
1437 | if (It != SunkCopies.end()) | |||
1438 | New = It->second; | |||
1439 | else | |||
1440 | New = SunkCopies[ExitBlock] = CloneInstructionInExitBlock( | |||
1441 | *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU); | |||
1442 | return New; | |||
1443 | } | |||
1444 | ||||
1445 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { | |||
1446 | BasicBlock *BB = PN->getParent(); | |||
1447 | if (!BB->canSplitPredecessors()) | |||
1448 | return false; | |||
1449 | // It's not impossible to split EHPad blocks, but if BlockColors already exist | |||
1450 | // it require updating BlockColors for all offspring blocks accordingly. By | |||
1451 | // skipping such corner case, we can make updating BlockColors after splitting | |||
1452 | // predecessor fairly simple. | |||
1453 | if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad()) | |||
1454 | return false; | |||
1455 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | |||
1456 | BasicBlock *BBPred = *PI; | |||
1457 | if (isa<IndirectBrInst>(BBPred->getTerminator())) | |||
1458 | return false; | |||
1459 | } | |||
1460 | return true; | |||
1461 | } | |||
1462 | ||||
1463 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, | |||
1464 | LoopInfo *LI, const Loop *CurLoop, | |||
1465 | LoopSafetyInfo *SafetyInfo, | |||
1466 | MemorySSAUpdater *MSSAU) { | |||
1467 | #ifndef NDEBUG | |||
1468 | SmallVector<BasicBlock *, 32> ExitBlocks; | |||
1469 | CurLoop->getUniqueExitBlocks(ExitBlocks); | |||
1470 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | |||
1471 | ExitBlocks.end()); | |||
1472 | #endif | |||
1473 | BasicBlock *ExitBB = PN->getParent(); | |||
1474 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")((ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block." ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(ExitBB) && \"Expect the PHI is in an exit block.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1474, __PRETTY_FUNCTION__)); | |||
1475 | ||||
1476 | // Split predecessors of the loop exit to make instructions in the loop are | |||
1477 | // exposed to exit blocks through trivially replaceable PHIs while keeping the | |||
1478 | // loop in the canonical form where each predecessor of each exit block should | |||
1479 | // be contained within the loop. For example, this will convert the loop below | |||
1480 | // from | |||
1481 | // | |||
1482 | // LB1: | |||
1483 | // %v1 = | |||
1484 | // br %LE, %LB2 | |||
1485 | // LB2: | |||
1486 | // %v2 = | |||
1487 | // br %LE, %LB1 | |||
1488 | // LE: | |||
1489 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable | |||
1490 | // | |||
1491 | // to | |||
1492 | // | |||
1493 | // LB1: | |||
1494 | // %v1 = | |||
1495 | // br %LE.split, %LB2 | |||
1496 | // LB2: | |||
1497 | // %v2 = | |||
1498 | // br %LE.split2, %LB1 | |||
1499 | // LE.split: | |||
1500 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable | |||
1501 | // br %LE | |||
1502 | // LE.split2: | |||
1503 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable | |||
1504 | // br %LE | |||
1505 | // LE: | |||
1506 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] | |||
1507 | // | |||
1508 | const auto &BlockColors = SafetyInfo->getBlockColors(); | |||
1509 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB)); | |||
1510 | while (!PredBBs.empty()) { | |||
1511 | BasicBlock *PredBB = *PredBBs.begin(); | |||
1512 | assert(CurLoop->contains(PredBB) &&((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1513, __PRETTY_FUNCTION__)) | |||
1513 | "Expect all predecessors are in the loop")((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1513, __PRETTY_FUNCTION__)); | |||
1514 | if (PN->getBasicBlockIndex(PredBB) >= 0) { | |||
1515 | BasicBlock *NewPred = SplitBlockPredecessors( | |||
1516 | ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true); | |||
1517 | // Since we do not allow splitting EH-block with BlockColors in | |||
1518 | // canSplitPredecessors(), we can simply assign predecessor's color to | |||
1519 | // the new block. | |||
1520 | if (!BlockColors.empty()) | |||
1521 | // Grab a reference to the ColorVector to be inserted before getting the | |||
1522 | // reference to the vector we are copying because inserting the new | |||
1523 | // element in BlockColors might cause the map to be reallocated. | |||
1524 | SafetyInfo->copyColors(NewPred, PredBB); | |||
1525 | } | |||
1526 | PredBBs.remove(PredBB); | |||
1527 | } | |||
1528 | } | |||
1529 | ||||
1530 | /// When an instruction is found to only be used outside of the loop, this | |||
1531 | /// function moves it to the exit blocks and patches up SSA form as needed. | |||
1532 | /// This method is guaranteed to remove the original instruction from its | |||
1533 | /// position, and may either delete it or move it to outside of the loop. | |||
1534 | /// | |||
1535 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | |||
1536 | const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo, | |||
1537 | MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE) { | |||
1538 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM sinking instruction: " << I << "\n"; } } while (false); | |||
1539 | ORE->emit([&]() { | |||
1540 | return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I) | |||
1541 | << "sinking " << ore::NV("Inst", &I); | |||
1542 | }); | |||
1543 | bool Changed = false; | |||
1544 | if (isa<LoadInst>(I)) | |||
1545 | ++NumMovedLoads; | |||
1546 | else if (isa<CallInst>(I)) | |||
1547 | ++NumMovedCalls; | |||
1548 | ++NumSunk; | |||
1549 | ||||
1550 | // Iterate over users to be ready for actual sinking. Replace users via | |||
1551 | // unreachable blocks with undef and make all user PHIs trivially replaceable. | |||
1552 | SmallPtrSet<Instruction *, 8> VisitedUsers; | |||
1553 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { | |||
1554 | auto *User = cast<Instruction>(*UI); | |||
1555 | Use &U = UI.getUse(); | |||
1556 | ++UI; | |||
1557 | ||||
1558 | if (VisitedUsers.count(User) || CurLoop->contains(User)) | |||
1559 | continue; | |||
1560 | ||||
1561 | if (!DT->isReachableFromEntry(User->getParent())) { | |||
1562 | U = UndefValue::get(I.getType()); | |||
1563 | Changed = true; | |||
1564 | continue; | |||
1565 | } | |||
1566 | ||||
1567 | // The user must be a PHI node. | |||
1568 | PHINode *PN = cast<PHINode>(User); | |||
1569 | ||||
1570 | // Surprisingly, instructions can be used outside of loops without any | |||
1571 | // exits. This can only happen in PHI nodes if the incoming block is | |||
1572 | // unreachable. | |||
1573 | BasicBlock *BB = PN->getIncomingBlock(U); | |||
1574 | if (!DT->isReachableFromEntry(BB)) { | |||
1575 | U = UndefValue::get(I.getType()); | |||
1576 | Changed = true; | |||
1577 | continue; | |||
1578 | } | |||
1579 | ||||
1580 | VisitedUsers.insert(PN); | |||
1581 | if (isTriviallyReplaceablePHI(*PN, I)) | |||
1582 | continue; | |||
1583 | ||||
1584 | if (!canSplitPredecessors(PN, SafetyInfo)) | |||
1585 | return Changed; | |||
1586 | ||||
1587 | // Split predecessors of the PHI so that we can make users trivially | |||
1588 | // replaceable. | |||
1589 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU); | |||
1590 | ||||
1591 | // Should rebuild the iterators, as they may be invalidated by | |||
1592 | // splitPredecessorsOfLoopExit(). | |||
1593 | UI = I.user_begin(); | |||
1594 | UE = I.user_end(); | |||
1595 | } | |||
1596 | ||||
1597 | if (VisitedUsers.empty()) | |||
1598 | return Changed; | |||
1599 | ||||
1600 | #ifndef NDEBUG | |||
1601 | SmallVector<BasicBlock *, 32> ExitBlocks; | |||
1602 | CurLoop->getUniqueExitBlocks(ExitBlocks); | |||
1603 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | |||
1604 | ExitBlocks.end()); | |||
1605 | #endif | |||
1606 | ||||
1607 | // Clones of this instruction. Don't create more than one per exit block! | |||
1608 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; | |||
1609 | ||||
1610 | // If this instruction is only used outside of the loop, then all users are | |||
1611 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of | |||
1612 | // the instruction. | |||
1613 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); | |||
1614 | for (auto *UI : Users) { | |||
1615 | auto *User = cast<Instruction>(UI); | |||
1616 | ||||
1617 | if (CurLoop->contains(User)) | |||
1618 | continue; | |||
1619 | ||||
1620 | PHINode *PN = cast<PHINode>(User); | |||
1621 | assert(ExitBlockSet.count(PN->getParent()) &&((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!" ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1622, __PRETTY_FUNCTION__)) | |||
1622 | "The LCSSA PHI is not in an exit block!")((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!" ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1622, __PRETTY_FUNCTION__)); | |||
1623 | // The PHI must be trivially replaceable. | |||
1624 | Instruction *New = sinkThroughTriviallyReplaceablePHI( | |||
1625 | PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU); | |||
1626 | PN->replaceAllUsesWith(New); | |||
1627 | eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr); | |||
1628 | Changed = true; | |||
1629 | } | |||
1630 | return Changed; | |||
1631 | } | |||
1632 | ||||
1633 | /// When an instruction is found to only use loop invariant operands that | |||
1634 | /// is safe to hoist, this instruction is called to do the dirty work. | |||
1635 | /// | |||
1636 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | |||
1637 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | |||
1638 | MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE) { | |||
1639 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getName() << ": " << Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Dest ->getName() << ": " << I << "\n"; } } while (false) | |||
1640 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Dest ->getName() << ": " << I << "\n"; } } while (false); | |||
1641 | ORE->emit([&]() { | |||
1642 | return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting " | |||
1643 | << ore::NV("Inst", &I); | |||
1644 | }); | |||
1645 | ||||
1646 | // Metadata can be dependent on conditions we are hoisting above. | |||
1647 | // Conservatively strip all metadata on the instruction unless we were | |||
1648 | // guaranteed to execute I if we entered the loop, in which case the metadata | |||
1649 | // is valid in the loop preheader. | |||
1650 | if (I.hasMetadataOtherThanDebugLoc() && | |||
1651 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning | |||
1652 | // time in isGuaranteedToExecute if we don't actually have anything to | |||
1653 | // drop. It is a compile time optimization, not required for correctness. | |||
1654 | !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop)) | |||
1655 | I.dropUnknownNonDebugMetadata(); | |||
1656 | ||||
1657 | if (isa<PHINode>(I)) | |||
1658 | // Move the new node to the end of the phi list in the destination block. | |||
1659 | moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU); | |||
1660 | else | |||
1661 | // Move the new node to the destination block, before its terminator. | |||
1662 | moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU); | |||
1663 | ||||
1664 | // Apply line 0 debug locations when we are moving instructions to different | |||
1665 | // basic blocks because we want to avoid jumpy line tables. | |||
1666 | if (const DebugLoc &DL = I.getDebugLoc()) | |||
1667 | I.setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt())); | |||
1668 | ||||
1669 | if (isa<LoadInst>(I)) | |||
1670 | ++NumMovedLoads; | |||
1671 | else if (isa<CallInst>(I)) | |||
1672 | ++NumMovedCalls; | |||
1673 | ++NumHoisted; | |||
1674 | } | |||
1675 | ||||
1676 | /// Only sink or hoist an instruction if it is not a trapping instruction, | |||
1677 | /// or if the instruction is known not to trap when moved to the preheader. | |||
1678 | /// or if it is a trapping instruction and is guaranteed to execute. | |||
1679 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | |||
1680 | const DominatorTree *DT, | |||
1681 | const Loop *CurLoop, | |||
1682 | const LoopSafetyInfo *SafetyInfo, | |||
1683 | OptimizationRemarkEmitter *ORE, | |||
1684 | const Instruction *CtxI) { | |||
1685 | if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT)) | |||
1686 | return true; | |||
1687 | ||||
1688 | bool GuaranteedToExecute = | |||
1689 | SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop); | |||
1690 | ||||
1691 | if (!GuaranteedToExecute) { | |||
1692 | auto *LI = dyn_cast<LoadInst>(&Inst); | |||
1693 | if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand())) | |||
1694 | ORE->emit([&]() { | |||
1695 | return OptimizationRemarkMissed( | |||
1696 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI) | |||
1697 | << "failed to hoist load with loop-invariant address " | |||
1698 | "because load is conditionally executed"; | |||
1699 | }); | |||
1700 | } | |||
1701 | ||||
1702 | return GuaranteedToExecute; | |||
1703 | } | |||
1704 | ||||
1705 | namespace { | |||
1706 | class LoopPromoter : public LoadAndStorePromoter { | |||
1707 | Value *SomePtr; // Designated pointer to store to. | |||
1708 | const SmallSetVector<Value *, 8> &PointerMustAliases; | |||
1709 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; | |||
1710 | SmallVectorImpl<Instruction *> &LoopInsertPts; | |||
1711 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts; | |||
1712 | PredIteratorCache &PredCache; | |||
1713 | AliasSetTracker &AST; | |||
1714 | MemorySSAUpdater *MSSAU; | |||
1715 | LoopInfo &LI; | |||
1716 | DebugLoc DL; | |||
1717 | int Alignment; | |||
1718 | bool UnorderedAtomic; | |||
1719 | AAMDNodes AATags; | |||
1720 | ICFLoopSafetyInfo &SafetyInfo; | |||
1721 | ||||
1722 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { | |||
1723 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
1724 | if (Loop *L = LI.getLoopFor(I->getParent())) | |||
1725 | if (!L->contains(BB)) { | |||
1726 | // We need to create an LCSSA PHI node for the incoming value and | |||
1727 | // store that. | |||
1728 | PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB), | |||
1729 | I->getName() + ".lcssa", &BB->front()); | |||
1730 | for (BasicBlock *Pred : PredCache.get(BB)) | |||
1731 | PN->addIncoming(I, Pred); | |||
1732 | return PN; | |||
1733 | } | |||
1734 | return V; | |||
1735 | } | |||
1736 | ||||
1737 | public: | |||
1738 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, | |||
1739 | const SmallSetVector<Value *, 8> &PMA, | |||
1740 | SmallVectorImpl<BasicBlock *> &LEB, | |||
1741 | SmallVectorImpl<Instruction *> &LIP, | |||
1742 | SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC, | |||
1743 | AliasSetTracker &ast, MemorySSAUpdater *MSSAU, LoopInfo &li, | |||
1744 | DebugLoc dl, int alignment, bool UnorderedAtomic, | |||
1745 | const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo) | |||
1746 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), | |||
1747 | LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), | |||
1748 | PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)), | |||
1749 | Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags), | |||
1750 | SafetyInfo(SafetyInfo) {} | |||
1751 | ||||
1752 | bool isInstInList(Instruction *I, | |||
1753 | const SmallVectorImpl<Instruction *> &) const override { | |||
1754 | Value *Ptr; | |||
1755 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
1756 | Ptr = LI->getOperand(0); | |||
1757 | else | |||
1758 | Ptr = cast<StoreInst>(I)->getPointerOperand(); | |||
1759 | return PointerMustAliases.count(Ptr); | |||
1760 | } | |||
1761 | ||||
1762 | void doExtraRewritesBeforeFinalDeletion() override { | |||
1763 | // Insert stores after in the loop exit blocks. Each exit block gets a | |||
1764 | // store of the live-out values that feed them. Since we've already told | |||
1765 | // the SSA updater about the defs in the loop and the preheader | |||
1766 | // definition, it is all set and we can start using it. | |||
1767 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { | |||
1768 | BasicBlock *ExitBlock = LoopExitBlocks[i]; | |||
1769 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); | |||
1770 | LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); | |||
1771 | Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); | |||
1772 | Instruction *InsertPos = LoopInsertPts[i]; | |||
1773 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); | |||
1774 | if (UnorderedAtomic) | |||
1775 | NewSI->setOrdering(AtomicOrdering::Unordered); | |||
1776 | NewSI->setAlignment(Alignment); | |||
1777 | NewSI->setDebugLoc(DL); | |||
1778 | if (AATags) | |||
1779 | NewSI->setAAMetadata(AATags); | |||
1780 | ||||
1781 | if (MSSAU) { | |||
1782 | MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i]; | |||
1783 | MemoryAccess *NewMemAcc; | |||
1784 | if (!MSSAInsertPoint) { | |||
1785 | NewMemAcc = MSSAU->createMemoryAccessInBB( | |||
1786 | NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning); | |||
1787 | } else { | |||
1788 | NewMemAcc = | |||
1789 | MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint); | |||
1790 | } | |||
1791 | MSSAInsertPts[i] = NewMemAcc; | |||
1792 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); | |||
1793 | // FIXME: true for safety, false may still be correct. | |||
1794 | } | |||
1795 | } | |||
1796 | } | |||
1797 | ||||
1798 | void replaceLoadWithValue(LoadInst *LI, Value *V) const override { | |||
1799 | // Update alias analysis. | |||
1800 | AST.copyValue(LI, V); | |||
1801 | } | |||
1802 | void instructionDeleted(Instruction *I) const override { | |||
1803 | SafetyInfo.removeInstruction(I); | |||
1804 | AST.deleteValue(I); | |||
1805 | if (MSSAU) | |||
1806 | MSSAU->removeMemoryAccess(I); | |||
1807 | } | |||
1808 | }; | |||
1809 | ||||
1810 | ||||
1811 | /// Return true iff we can prove that a caller of this function can not inspect | |||
1812 | /// the contents of the provided object in a well defined program. | |||
1813 | bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) { | |||
1814 | if (isa<AllocaInst>(Object)) | |||
1815 | // Since the alloca goes out of scope, we know the caller can't retain a | |||
1816 | // reference to it and be well defined. Thus, we don't need to check for | |||
1817 | // capture. | |||
1818 | return true; | |||
1819 | ||||
1820 | // For all other objects we need to know that the caller can't possibly | |||
1821 | // have gotten a reference to the object. There are two components of | |||
1822 | // that: | |||
1823 | // 1) Object can't be escaped by this function. This is what | |||
1824 | // PointerMayBeCaptured checks. | |||
1825 | // 2) Object can't have been captured at definition site. For this, we | |||
1826 | // need to know the return value is noalias. At the moment, we use a | |||
1827 | // weaker condition and handle only AllocLikeFunctions (which are | |||
1828 | // known to be noalias). TODO | |||
1829 | return isAllocLikeFn(Object, TLI) && | |||
1830 | !PointerMayBeCaptured(Object, true, true); | |||
1831 | } | |||
1832 | ||||
1833 | } // namespace | |||
1834 | ||||
1835 | /// Try to promote memory values to scalars by sinking stores out of the | |||
1836 | /// loop and moving loads to before the loop. We do this by looping over | |||
1837 | /// the stores in the loop, looking for stores to Must pointers which are | |||
1838 | /// loop invariant. | |||
1839 | /// | |||
1840 | bool llvm::promoteLoopAccessesToScalars( | |||
1841 | const SmallSetVector<Value *, 8> &PointerMustAliases, | |||
1842 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | |||
1843 | SmallVectorImpl<Instruction *> &InsertPts, | |||
1844 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC, | |||
1845 | LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, | |||
1846 | Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | |||
1847 | ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) { | |||
1848 | // Verify inputs. | |||
1849 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1851, __PRETTY_FUNCTION__)) | |||
1850 | CurAST != nullptr && SafetyInfo != nullptr &&((LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1851, __PRETTY_FUNCTION__)) | |||
1851 | "Unexpected Input to promoteLoopAccessesToScalars")((LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 1851, __PRETTY_FUNCTION__)); | |||
1852 | ||||
1853 | Value *SomePtr = *PointerMustAliases.begin(); | |||
1854 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); | |||
1855 | ||||
1856 | // It is not safe to promote a load/store from the loop if the load/store is | |||
1857 | // conditional. For example, turning: | |||
1858 | // | |||
1859 | // for () { if (c) *P += 1; } | |||
1860 | // | |||
1861 | // into: | |||
1862 | // | |||
1863 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; | |||
1864 | // | |||
1865 | // is not safe, because *P may only be valid to access if 'c' is true. | |||
1866 | // | |||
1867 | // The safety property divides into two parts: | |||
1868 | // p1) The memory may not be dereferenceable on entry to the loop. In this | |||
1869 | // case, we can't insert the required load in the preheader. | |||
1870 | // p2) The memory model does not allow us to insert a store along any dynamic | |||
1871 | // path which did not originally have one. | |||
1872 | // | |||
1873 | // If at least one store is guaranteed to execute, both properties are | |||
1874 | // satisfied, and promotion is legal. | |||
1875 | // | |||
1876 | // This, however, is not a necessary condition. Even if no store/load is | |||
1877 | // guaranteed to execute, we can still establish these properties. | |||
1878 | // We can establish (p1) by proving that hoisting the load into the preheader | |||
1879 | // is safe (i.e. proving dereferenceability on all paths through the loop). We | |||
1880 | // can use any access within the alias set to prove dereferenceability, | |||
1881 | // since they're all must alias. | |||
1882 | // | |||
1883 | // There are two ways establish (p2): | |||
1884 | // a) Prove the location is thread-local. In this case the memory model | |||
1885 | // requirement does not apply, and stores are safe to insert. | |||
1886 | // b) Prove a store dominates every exit block. In this case, if an exit | |||
1887 | // blocks is reached, the original dynamic path would have taken us through | |||
1888 | // the store, so inserting a store into the exit block is safe. Note that this | |||
1889 | // is different from the store being guaranteed to execute. For instance, | |||
1890 | // if an exception is thrown on the first iteration of the loop, the original | |||
1891 | // store is never executed, but the exit blocks are not executed either. | |||
1892 | ||||
1893 | bool DereferenceableInPH = false; | |||
1894 | bool SafeToInsertStore = false; | |||
1895 | ||||
1896 | SmallVector<Instruction *, 64> LoopUses; | |||
1897 | ||||
1898 | // We start with an alignment of one and try to find instructions that allow | |||
1899 | // us to prove better alignment. | |||
1900 | unsigned Alignment = 1; | |||
1901 | // Keep track of which types of access we see | |||
1902 | bool SawUnorderedAtomic = false; | |||
1903 | bool SawNotAtomic = false; | |||
1904 | AAMDNodes AATags; | |||
1905 | ||||
1906 | const DataLayout &MDL = Preheader->getModule()->getDataLayout(); | |||
1907 | ||||
1908 | bool IsKnownThreadLocalObject = false; | |||
1909 | if (SafetyInfo->anyBlockMayThrow()) { | |||
1910 | // If a loop can throw, we have to insert a store along each unwind edge. | |||
1911 | // That said, we can't actually make the unwind edge explicit. Therefore, | |||
1912 | // we have to prove that the store is dead along the unwind edge. We do | |||
1913 | // this by proving that the caller can't have a reference to the object | |||
1914 | // after return and thus can't possibly load from the object. | |||
1915 | Value *Object = GetUnderlyingObject(SomePtr, MDL); | |||
1916 | if (!isKnownNonEscaping(Object, TLI)) | |||
1917 | return false; | |||
1918 | // Subtlety: Alloca's aren't visible to callers, but *are* potentially | |||
1919 | // visible to other threads if captured and used during their lifetimes. | |||
1920 | IsKnownThreadLocalObject = !isa<AllocaInst>(Object); | |||
1921 | } | |||
1922 | ||||
1923 | // Check that all of the pointers in the alias set have the same type. We | |||
1924 | // cannot (yet) promote a memory location that is loaded and stored in | |||
1925 | // different sizes. While we are at it, collect alignment and AA info. | |||
1926 | for (Value *ASIV : PointerMustAliases) { | |||
1927 | // Check that all of the pointers in the alias set have the same type. We | |||
1928 | // cannot (yet) promote a memory location that is loaded and stored in | |||
1929 | // different sizes. | |||
1930 | if (SomePtr->getType() != ASIV->getType()) | |||
1931 | return false; | |||
1932 | ||||
1933 | for (User *U : ASIV->users()) { | |||
1934 | // Ignore instructions that are outside the loop. | |||
1935 | Instruction *UI = dyn_cast<Instruction>(U); | |||
1936 | if (!UI || !CurLoop->contains(UI)) | |||
1937 | continue; | |||
1938 | ||||
1939 | // If there is an non-load/store instruction in the loop, we can't promote | |||
1940 | // it. | |||
1941 | if (LoadInst *Load = dyn_cast<LoadInst>(UI)) { | |||
1942 | if (!Load->isUnordered()) | |||
1943 | return false; | |||
1944 | ||||
1945 | SawUnorderedAtomic |= Load->isAtomic(); | |||
1946 | SawNotAtomic |= !Load->isAtomic(); | |||
1947 | ||||
1948 | unsigned InstAlignment = Load->getAlignment(); | |||
1949 | if (!InstAlignment) | |||
1950 | InstAlignment = | |||
1951 | MDL.getABITypeAlignment(Load->getType()); | |||
1952 | ||||
1953 | // Note that proving a load safe to speculate requires proving | |||
1954 | // sufficient alignment at the target location. Proving it guaranteed | |||
1955 | // to execute does as well. Thus we can increase our guaranteed | |||
1956 | // alignment as well. | |||
1957 | if (!DereferenceableInPH || (InstAlignment > Alignment)) | |||
1958 | if (isSafeToExecuteUnconditionally(*Load, DT, CurLoop, SafetyInfo, | |||
1959 | ORE, Preheader->getTerminator())) { | |||
1960 | DereferenceableInPH = true; | |||
1961 | Alignment = std::max(Alignment, InstAlignment); | |||
1962 | } | |||
1963 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) { | |||
1964 | // Stores *of* the pointer are not interesting, only stores *to* the | |||
1965 | // pointer. | |||
1966 | if (UI->getOperand(1) != ASIV) | |||
1967 | continue; | |||
1968 | if (!Store->isUnordered()) | |||
1969 | return false; | |||
1970 | ||||
1971 | SawUnorderedAtomic |= Store->isAtomic(); | |||
1972 | SawNotAtomic |= !Store->isAtomic(); | |||
1973 | ||||
1974 | // If the store is guaranteed to execute, both properties are satisfied. | |||
1975 | // We may want to check if a store is guaranteed to execute even if we | |||
1976 | // already know that promotion is safe, since it may have higher | |||
1977 | // alignment than any other guaranteed stores, in which case we can | |||
1978 | // raise the alignment on the promoted store. | |||
1979 | unsigned InstAlignment = Store->getAlignment(); | |||
1980 | if (!InstAlignment) | |||
1981 | InstAlignment = | |||
1982 | MDL.getABITypeAlignment(Store->getValueOperand()->getType()); | |||
1983 | ||||
1984 | if (!DereferenceableInPH || !SafeToInsertStore || | |||
1985 | (InstAlignment > Alignment)) { | |||
1986 | if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) { | |||
1987 | DereferenceableInPH = true; | |||
1988 | SafeToInsertStore = true; | |||
1989 | Alignment = std::max(Alignment, InstAlignment); | |||
1990 | } | |||
1991 | } | |||
1992 | ||||
1993 | // If a store dominates all exit blocks, it is safe to sink. | |||
1994 | // As explained above, if an exit block was executed, a dominating | |||
1995 | // store must have been executed at least once, so we are not | |||
1996 | // introducing stores on paths that did not have them. | |||
1997 | // Note that this only looks at explicit exit blocks. If we ever | |||
1998 | // start sinking stores into unwind edges (see above), this will break. | |||
1999 | if (!SafeToInsertStore) | |||
2000 | SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) { | |||
2001 | return DT->dominates(Store->getParent(), Exit); | |||
2002 | }); | |||
2003 | ||||
2004 | // If the store is not guaranteed to execute, we may still get | |||
2005 | // deref info through it. | |||
2006 | if (!DereferenceableInPH) { | |||
2007 | DereferenceableInPH = isDereferenceableAndAlignedPointer( | |||
2008 | Store->getPointerOperand(), Store->getAlignment(), MDL, | |||
2009 | Preheader->getTerminator(), DT); | |||
2010 | } | |||
2011 | } else | |||
2012 | return false; // Not a load or store. | |||
2013 | ||||
2014 | // Merge the AA tags. | |||
2015 | if (LoopUses.empty()) { | |||
2016 | // On the first load/store, just take its AA tags. | |||
2017 | UI->getAAMetadata(AATags); | |||
2018 | } else if (AATags) { | |||
2019 | UI->getAAMetadata(AATags, /* Merge = */ true); | |||
2020 | } | |||
2021 | ||||
2022 | LoopUses.push_back(UI); | |||
2023 | } | |||
2024 | } | |||
2025 | ||||
2026 | // If we found both an unordered atomic instruction and a non-atomic memory | |||
2027 | // access, bail. We can't blindly promote non-atomic to atomic since we | |||
2028 | // might not be able to lower the result. We can't downgrade since that | |||
2029 | // would violate memory model. Also, align 0 is an error for atomics. | |||
2030 | if (SawUnorderedAtomic && SawNotAtomic) | |||
2031 | return false; | |||
2032 | ||||
2033 | // If we're inserting an atomic load in the preheader, we must be able to | |||
2034 | // lower it. We're only guaranteed to be able to lower naturally aligned | |||
2035 | // atomics. | |||
2036 | auto *SomePtrElemType = SomePtr->getType()->getPointerElementType(); | |||
2037 | if (SawUnorderedAtomic && | |||
2038 | Alignment < MDL.getTypeStoreSize(SomePtrElemType)) | |||
2039 | return false; | |||
2040 | ||||
2041 | // If we couldn't prove we can hoist the load, bail. | |||
2042 | if (!DereferenceableInPH) | |||
2043 | return false; | |||
2044 | ||||
2045 | // We know we can hoist the load, but don't have a guaranteed store. | |||
2046 | // Check whether the location is thread-local. If it is, then we can insert | |||
2047 | // stores along paths which originally didn't have them without violating the | |||
2048 | // memory model. | |||
2049 | if (!SafeToInsertStore) { | |||
2050 | if (IsKnownThreadLocalObject) | |||
2051 | SafeToInsertStore = true; | |||
2052 | else { | |||
2053 | Value *Object = GetUnderlyingObject(SomePtr, MDL); | |||
2054 | SafeToInsertStore = | |||
2055 | (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) && | |||
2056 | !PointerMayBeCaptured(Object, true, true); | |||
2057 | } | |||
2058 | } | |||
2059 | ||||
2060 | // If we've still failed to prove we can sink the store, give up. | |||
2061 | if (!SafeToInsertStore) | |||
2062 | return false; | |||
2063 | ||||
2064 | // Otherwise, this is safe to promote, lets do it! | |||
2065 | 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) | |||
2066 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr << '\n'; } } while (false); | |||
2067 | ORE->emit([&]() { | |||
2068 | return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar", | |||
2069 | LoopUses[0]) | |||
2070 | << "Moving accesses to memory location out of the loop"; | |||
2071 | }); | |||
2072 | ++NumPromoted; | |||
2073 | ||||
2074 | // Grab a debug location for the inserted loads/stores; given that the | |||
2075 | // inserted loads/stores have little relation to the original loads/stores, | |||
2076 | // this code just arbitrarily picks a location from one, since any debug | |||
2077 | // location is better than none. | |||
2078 | DebugLoc DL = LoopUses[0]->getDebugLoc(); | |||
2079 | ||||
2080 | // We use the SSAUpdater interface to insert phi nodes as required. | |||
2081 | SmallVector<PHINode *, 16> NewPHIs; | |||
2082 | SSAUpdater SSA(&NewPHIs); | |||
2083 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, | |||
2084 | InsertPts, MSSAInsertPts, PIC, *CurAST, MSSAU, *LI, DL, | |||
2085 | Alignment, SawUnorderedAtomic, AATags, *SafetyInfo); | |||
2086 | ||||
2087 | // Set up the preheader to have a definition of the value. It is the live-out | |||
2088 | // value from the preheader that uses in the loop will use. | |||
2089 | LoadInst *PreheaderLoad = new LoadInst( | |||
2090 | SomePtr->getType()->getPointerElementType(), SomePtr, | |||
2091 | SomePtr->getName() + ".promoted", Preheader->getTerminator()); | |||
2092 | if (SawUnorderedAtomic) | |||
2093 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); | |||
2094 | PreheaderLoad->setAlignment(Alignment); | |||
2095 | PreheaderLoad->setDebugLoc(DL); | |||
2096 | if (AATags) | |||
2097 | PreheaderLoad->setAAMetadata(AATags); | |||
2098 | SSA.AddAvailableValue(Preheader, PreheaderLoad); | |||
2099 | ||||
2100 | MemoryAccess *PreheaderLoadMemoryAccess; | |||
2101 | if (MSSAU) { | |||
2102 | PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB( | |||
2103 | PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End); | |||
2104 | MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess); | |||
2105 | MSSAU->insertUse(NewMemUse); | |||
2106 | } | |||
2107 | ||||
2108 | // Rewrite all the loads in the loop and remember all the definitions from | |||
2109 | // stores in the loop. | |||
2110 | Promoter.run(LoopUses); | |||
2111 | ||||
2112 | if (MSSAU && VerifyMemorySSA) | |||
2113 | MSSAU->getMemorySSA()->verifyMemorySSA(); | |||
2114 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. | |||
2115 | if (PreheaderLoad->use_empty()) | |||
2116 | eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU); | |||
2117 | ||||
2118 | return true; | |||
2119 | } | |||
2120 | ||||
2121 | /// Returns an owning pointer to an alias set which incorporates aliasing info | |||
2122 | /// from L and all subloops of L. | |||
2123 | /// FIXME: In new pass manager, there is no helper function to handle loop | |||
2124 | /// analysis such as cloneBasicBlockAnalysis, so the AST needs to be recomputed | |||
2125 | /// from scratch for every loop. Hook up with the helper functions when | |||
2126 | /// available in the new pass manager to avoid redundant computation. | |||
2127 | std::unique_ptr<AliasSetTracker> | |||
2128 | LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | |||
2129 | AliasAnalysis *AA) { | |||
2130 | std::unique_ptr<AliasSetTracker> CurAST; | |||
2131 | SmallVector<Loop *, 4> RecomputeLoops; | |||
2132 | for (Loop *InnerL : L->getSubLoops()) { | |||
2133 | auto MapI = LoopToAliasSetMap.find(InnerL); | |||
2134 | // If the AST for this inner loop is missing it may have been merged into | |||
2135 | // some other loop's AST and then that loop unrolled, and so we need to | |||
2136 | // recompute it. | |||
2137 | if (MapI == LoopToAliasSetMap.end()) { | |||
2138 | RecomputeLoops.push_back(InnerL); | |||
2139 | continue; | |||
2140 | } | |||
2141 | std::unique_ptr<AliasSetTracker> InnerAST = std::move(MapI->second); | |||
2142 | ||||
2143 | if (CurAST) { | |||
2144 | // What if InnerLoop was modified by other passes ? | |||
2145 | // Once we've incorporated the inner loop's AST into ours, we don't need | |||
2146 | // the subloop's anymore. | |||
2147 | CurAST->add(*InnerAST); | |||
2148 | } else { | |||
2149 | CurAST = std::move(InnerAST); | |||
2150 | } | |||
2151 | LoopToAliasSetMap.erase(MapI); | |||
2152 | } | |||
2153 | if (!CurAST) | |||
2154 | CurAST = make_unique<AliasSetTracker>(*AA); | |||
2155 | ||||
2156 | // Add everything from the sub loops that are no longer directly available. | |||
2157 | for (Loop *InnerL : RecomputeLoops) | |||
2158 | for (BasicBlock *BB : InnerL->blocks()) | |||
2159 | CurAST->add(*BB); | |||
2160 | ||||
2161 | // And merge in this loop (without anything from inner loops). | |||
2162 | for (BasicBlock *BB : L->blocks()) | |||
2163 | if (LI->getLoopFor(BB) == L) | |||
2164 | CurAST->add(*BB); | |||
2165 | ||||
2166 | return CurAST; | |||
2167 | } | |||
2168 | ||||
2169 | std::unique_ptr<AliasSetTracker> | |||
2170 | LoopInvariantCodeMotion::collectAliasInfoForLoopWithMSSA( | |||
2171 | Loop *L, AliasAnalysis *AA, MemorySSAUpdater *MSSAU) { | |||
2172 | auto *MSSA = MSSAU->getMemorySSA(); | |||
2173 | auto CurAST = make_unique<AliasSetTracker>(*AA, MSSA, L); | |||
2174 | CurAST->addAllInstructionsInLoopUsingMSSA(); | |||
2175 | return CurAST; | |||
2176 | } | |||
2177 | ||||
2178 | /// Simple analysis hook. Clone alias set info. | |||
2179 | /// | |||
2180 | void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, | |||
2181 | Loop *L) { | |||
2182 | auto ASTIt = LICM.getLoopToAliasSetMap().find(L); | |||
2183 | if (ASTIt == LICM.getLoopToAliasSetMap().end()) | |||
2184 | return; | |||
2185 | ||||
2186 | ASTIt->second->copyValue(From, To); | |||
2187 | } | |||
2188 | ||||
2189 | /// Simple Analysis hook. Delete value V from alias set | |||
2190 | /// | |||
2191 | void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) { | |||
2192 | auto ASTIt = LICM.getLoopToAliasSetMap().find(L); | |||
2193 | if (ASTIt == LICM.getLoopToAliasSetMap().end()) | |||
2194 | return; | |||
2195 | ||||
2196 | ASTIt->second->deleteValue(V); | |||
2197 | } | |||
2198 | ||||
2199 | /// Simple Analysis hook. Delete value L from alias set map. | |||
2200 | /// | |||
2201 | void LegacyLICMPass::deleteAnalysisLoop(Loop *L) { | |||
2202 | if (!LICM.getLoopToAliasSetMap().count(L)) | |||
2203 | return; | |||
2204 | ||||
2205 | LICM.getLoopToAliasSetMap().erase(L); | |||
2206 | } | |||
2207 | ||||
2208 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | |||
2209 | AliasSetTracker *CurAST, Loop *CurLoop, | |||
2210 | AliasAnalysis *AA) { | |||
2211 | // First check to see if any of the basic blocks in CurLoop invalidate *V. | |||
2212 | bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod(); | |||
2213 | ||||
2214 | if (!isInvalidatedAccordingToAST || !LICMN2Theshold) | |||
2215 | return isInvalidatedAccordingToAST; | |||
2216 | ||||
2217 | // Check with a diagnostic analysis if we can refine the information above. | |||
2218 | // This is to identify the limitations of using the AST. | |||
2219 | // The alias set mechanism used by LICM has a major weakness in that it | |||
2220 | // combines all things which may alias into a single set *before* asking | |||
2221 | // modref questions. As a result, a single readonly call within a loop will | |||
2222 | // collapse all loads and stores into a single alias set and report | |||
2223 | // invalidation if the loop contains any store. For example, readonly calls | |||
2224 | // with deopt states have this form and create a general alias set with all | |||
2225 | // loads and stores. In order to get any LICM in loops containing possible | |||
2226 | // deopt states we need a more precise invalidation of checking the mod ref | |||
2227 | // info of each instruction within the loop and LI. This has a complexity of | |||
2228 | // O(N^2), so currently, it is used only as a diagnostic tool since the | |||
2229 | // default value of LICMN2Threshold is zero. | |||
2230 | ||||
2231 | // Don't look at nested loops. | |||
2232 | if (CurLoop->begin() != CurLoop->end()) | |||
2233 | return true; | |||
2234 | ||||
2235 | int N = 0; | |||
2236 | for (BasicBlock *BB : CurLoop->getBlocks()) | |||
2237 | for (Instruction &I : *BB) { | |||
2238 | if (N >= LICMN2Theshold) { | |||
2239 | LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Alasing N2 threshold exhausted for " << *(MemLoc.Ptr) << "\n"; } } while (false) | |||
2240 | << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Alasing N2 threshold exhausted for " << *(MemLoc.Ptr) << "\n"; } } while (false); | |||
2241 | return true; | |||
2242 | } | |||
2243 | N++; | |||
2244 | auto Res = AA->getModRefInfo(&I, MemLoc); | |||
2245 | if (isModSet(Res)) { | |||
2246 | LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing failed on " << I << " for " << *(MemLoc.Ptr) << "\n"; } } while (false ) | |||
2247 | << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing failed on " << I << " for " << *(MemLoc.Ptr) << "\n"; } } while (false ); | |||
2248 | return true; | |||
2249 | } | |||
2250 | } | |||
2251 | LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing okay for " << *(MemLoc .Ptr) << "\n"; } } while (false); | |||
2252 | return false; | |||
2253 | } | |||
2254 | ||||
2255 | static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | |||
2256 | Loop *CurLoop, | |||
2257 | SinkAndHoistLICMFlags &Flags) { | |||
2258 | MemoryAccess *Source; | |||
2259 | // See declaration of SetLicmMssaOptCap for usage details. | |||
2260 | if (Flags.LicmMssaOptCounter >= Flags.LicmMssaOptCap) | |||
2261 | Source = MU->getDefiningAccess(); | |||
2262 | else { | |||
2263 | Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU); | |||
2264 | Flags.LicmMssaOptCounter++; | |||
2265 | } | |||
2266 | return !MSSA->isLiveOnEntryDef(Source) && | |||
2267 | CurLoop->contains(Source->getBlock()); | |||
2268 | } | |||
2269 | ||||
2270 | /// Little predicate that returns true if the specified basic block is in | |||
2271 | /// a subloop of the current one, not the current one itself. | |||
2272 | /// | |||
2273 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { | |||
2274 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")((CurLoop->contains(BB) && "Only valid if BB is IN the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(BB) && \"Only valid if BB is IN the loop\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Scalar/LICM.cpp" , 2274, __PRETTY_FUNCTION__)); | |||
2275 | return LI->getLoopFor(BB) != CurLoop; | |||
2276 | } |