1 | //===- Inliner.cpp - Code common to all inliners --------------------------===// |
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 file implements the mechanics required to implement inlining without |
10 | // missing any calls and updating the call graph. The decisions of which calls |
11 | // are profitable to inline are implemented elsewhere. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/Transforms/IPO/Inliner.h" |
16 | #include "llvm/ADT/PriorityWorklist.h" |
17 | #include "llvm/ADT/STLExtras.h" |
18 | #include "llvm/ADT/ScopeExit.h" |
19 | #include "llvm/ADT/SetVector.h" |
20 | #include "llvm/ADT/SmallPtrSet.h" |
21 | #include "llvm/ADT/SmallVector.h" |
22 | #include "llvm/ADT/Statistic.h" |
23 | #include "llvm/ADT/StringExtras.h" |
24 | #include "llvm/ADT/StringRef.h" |
25 | #include "llvm/Analysis/AssumptionCache.h" |
26 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
27 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
28 | #include "llvm/Analysis/CGSCCPassManager.h" |
29 | #include "llvm/Analysis/EphemeralValuesCache.h" |
30 | #include "llvm/Analysis/InlineAdvisor.h" |
31 | #include "llvm/Analysis/InlineCost.h" |
32 | #include "llvm/Analysis/LazyCallGraph.h" |
33 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
34 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
35 | #include "llvm/Analysis/ReplayInlineAdvisor.h" |
36 | #include "llvm/Analysis/Utils/ImportedFunctionsInliningStatistics.h" |
37 | #include "llvm/IR/Attributes.h" |
38 | #include "llvm/IR/BasicBlock.h" |
39 | #include "llvm/IR/DebugLoc.h" |
40 | #include "llvm/IR/DerivedTypes.h" |
41 | #include "llvm/IR/DiagnosticInfo.h" |
42 | #include "llvm/IR/Function.h" |
43 | #include "llvm/IR/InstIterator.h" |
44 | #include "llvm/IR/Instruction.h" |
45 | #include "llvm/IR/Instructions.h" |
46 | #include "llvm/IR/IntrinsicInst.h" |
47 | #include "llvm/IR/Metadata.h" |
48 | #include "llvm/IR/Module.h" |
49 | #include "llvm/IR/PassManager.h" |
50 | #include "llvm/IR/Value.h" |
51 | #include "llvm/Pass.h" |
52 | #include "llvm/Support/Casting.h" |
53 | #include "llvm/Support/CommandLine.h" |
54 | #include "llvm/Support/Debug.h" |
55 | #include "llvm/Support/raw_ostream.h" |
56 | #include "llvm/Transforms/Utils/CallPromotionUtils.h" |
57 | #include "llvm/Transforms/Utils/Cloning.h" |
58 | #include "llvm/Transforms/Utils/Local.h" |
59 | #include "llvm/Transforms/Utils/ModuleUtils.h" |
60 | #include <algorithm> |
61 | #include <cassert> |
62 | #include <utility> |
63 | |
64 | using namespace llvm; |
65 | |
66 | #define DEBUG_TYPE "inline" |
67 | |
68 | STATISTIC(NumInlined, "Number of functions inlined" ); |
69 | STATISTIC(NumDeleted, "Number of functions deleted because all callers found" ); |
70 | |
71 | static cl::opt<int> IntraSCCCostMultiplier( |
72 | "intra-scc-cost-multiplier" , cl::init(Val: 2), cl::Hidden, |
73 | cl::desc( |
74 | "Cost multiplier to multiply onto inlined call sites where the " |
75 | "new call was previously an intra-SCC call (not relevant when the " |
76 | "original call was already intra-SCC). This can accumulate over " |
77 | "multiple inlinings (e.g. if a call site already had a cost " |
78 | "multiplier and one of its inlined calls was also subject to " |
79 | "this, the inlined call would have the original multiplier " |
80 | "multiplied by intra-scc-cost-multiplier). This is to prevent tons of " |
81 | "inlining through a child SCC which can cause terrible compile times" )); |
82 | |
83 | /// A flag for test, so we can print the content of the advisor when running it |
84 | /// as part of the default (e.g. -O3) pipeline. |
85 | static cl::opt<bool> KeepAdvisorForPrinting("keep-inline-advisor-for-printing" , |
86 | cl::init(Val: false), cl::Hidden); |
87 | |
88 | /// Allows printing the contents of the advisor after each SCC inliner pass. |
89 | static cl::opt<bool> |
90 | EnablePostSCCAdvisorPrinting("enable-scc-inline-advisor-printing" , |
91 | cl::init(Val: false), cl::Hidden); |
92 | |
93 | |
94 | static cl::opt<std::string> CGSCCInlineReplayFile( |
95 | "cgscc-inline-replay" , cl::init(Val: "" ), cl::value_desc("filename" ), |
96 | cl::desc( |
97 | "Optimization remarks file containing inline remarks to be replayed " |
98 | "by cgscc inlining." ), |
99 | cl::Hidden); |
100 | |
101 | static cl::opt<ReplayInlinerSettings::Scope> CGSCCInlineReplayScope( |
102 | "cgscc-inline-replay-scope" , |
103 | cl::init(Val: ReplayInlinerSettings::Scope::Function), |
104 | cl::values(clEnumValN(ReplayInlinerSettings::Scope::Function, "Function" , |
105 | "Replay on functions that have remarks associated " |
106 | "with them (default)" ), |
107 | clEnumValN(ReplayInlinerSettings::Scope::Module, "Module" , |
108 | "Replay on the entire module" )), |
109 | cl::desc("Whether inline replay should be applied to the entire " |
110 | "Module or just the Functions (default) that are present as " |
111 | "callers in remarks during cgscc inlining." ), |
112 | cl::Hidden); |
113 | |
114 | static cl::opt<ReplayInlinerSettings::Fallback> CGSCCInlineReplayFallback( |
115 | "cgscc-inline-replay-fallback" , |
116 | cl::init(Val: ReplayInlinerSettings::Fallback::Original), |
117 | cl::values( |
118 | clEnumValN( |
119 | ReplayInlinerSettings::Fallback::Original, "Original" , |
120 | "All decisions not in replay send to original advisor (default)" ), |
121 | clEnumValN(ReplayInlinerSettings::Fallback::AlwaysInline, |
122 | "AlwaysInline" , "All decisions not in replay are inlined" ), |
123 | clEnumValN(ReplayInlinerSettings::Fallback::NeverInline, "NeverInline" , |
124 | "All decisions not in replay are not inlined" )), |
125 | cl::desc( |
126 | "How cgscc inline replay treats sites that don't come from the replay. " |
127 | "Original: defers to original advisor, AlwaysInline: inline all sites " |
128 | "not in replay, NeverInline: inline no sites not in replay" ), |
129 | cl::Hidden); |
130 | |
131 | static cl::opt<CallSiteFormat::Format> CGSCCInlineReplayFormat( |
132 | "cgscc-inline-replay-format" , |
133 | cl::init(Val: CallSiteFormat::Format::LineColumnDiscriminator), |
134 | cl::values( |
135 | clEnumValN(CallSiteFormat::Format::Line, "Line" , "<Line Number>" ), |
136 | clEnumValN(CallSiteFormat::Format::LineColumn, "LineColumn" , |
137 | "<Line Number>:<Column Number>" ), |
138 | clEnumValN(CallSiteFormat::Format::LineDiscriminator, |
139 | "LineDiscriminator" , "<Line Number>.<Discriminator>" ), |
140 | clEnumValN(CallSiteFormat::Format::LineColumnDiscriminator, |
141 | "LineColumnDiscriminator" , |
142 | "<Line Number>:<Column Number>.<Discriminator> (default)" )), |
143 | cl::desc("How cgscc inline replay file is formatted" ), cl::Hidden); |
144 | |
145 | /// Return true if the specified inline history ID |
146 | /// indicates an inline history that includes the specified function. |
147 | static bool inlineHistoryIncludes( |
148 | Function *F, int InlineHistoryID, |
149 | const SmallVectorImpl<std::pair<Function *, int>> &InlineHistory) { |
150 | while (InlineHistoryID != -1) { |
151 | assert(unsigned(InlineHistoryID) < InlineHistory.size() && |
152 | "Invalid inline history ID" ); |
153 | if (InlineHistory[InlineHistoryID].first == F) |
154 | return true; |
155 | InlineHistoryID = InlineHistory[InlineHistoryID].second; |
156 | } |
157 | return false; |
158 | } |
159 | |
160 | InlineAdvisor & |
161 | InlinerPass::getAdvisor(const ModuleAnalysisManagerCGSCCProxy::Result &MAM, |
162 | FunctionAnalysisManager &FAM, Module &M) { |
163 | if (OwnedAdvisor) |
164 | return *OwnedAdvisor; |
165 | |
166 | auto *IAA = MAM.getCachedResult<InlineAdvisorAnalysis>(IR&: M); |
167 | if (!IAA) { |
168 | // It should still be possible to run the inliner as a stand-alone SCC pass, |
169 | // for test scenarios. In that case, we default to the |
170 | // DefaultInlineAdvisor, which doesn't need to keep state between SCC pass |
171 | // runs. It also uses just the default InlineParams. |
172 | // In this case, we need to use the provided FAM, which is valid for the |
173 | // duration of the inliner pass, and thus the lifetime of the owned advisor. |
174 | // The one we would get from the MAM can be invalidated as a result of the |
175 | // inliner's activity. |
176 | OwnedAdvisor = std::make_unique<DefaultInlineAdvisor>( |
177 | args&: M, args&: FAM, args: getInlineParams(), |
178 | args: InlineContext{.LTOPhase: LTOPhase, .Pass: InlinePass::CGSCCInliner}); |
179 | |
180 | if (!CGSCCInlineReplayFile.empty()) |
181 | OwnedAdvisor = getReplayInlineAdvisor( |
182 | M, FAM, Context&: M.getContext(), OriginalAdvisor: std::move(OwnedAdvisor), |
183 | ReplaySettings: ReplayInlinerSettings{.ReplayFile: CGSCCInlineReplayFile, |
184 | .ReplayScope: CGSCCInlineReplayScope, |
185 | .ReplayFallback: CGSCCInlineReplayFallback, |
186 | .ReplayFormat: {.OutputFormat: CGSCCInlineReplayFormat}}, |
187 | /*EmitRemarks=*/true, |
188 | IC: InlineContext{.LTOPhase: LTOPhase, .Pass: InlinePass::ReplayCGSCCInliner}); |
189 | |
190 | return *OwnedAdvisor; |
191 | } |
192 | assert(IAA->getAdvisor() && |
193 | "Expected a present InlineAdvisorAnalysis also have an " |
194 | "InlineAdvisor initialized" ); |
195 | return *IAA->getAdvisor(); |
196 | } |
197 | |
198 | void makeFunctionBodyUnreachable(Function &F) { |
199 | F.dropAllReferences(); |
200 | for (BasicBlock &BB : make_early_inc_range(Range&: F)) |
201 | BB.eraseFromParent(); |
202 | BasicBlock *BB = BasicBlock::Create(Context&: F.getContext(), Name: "" , Parent: &F); |
203 | new UnreachableInst(F.getContext(), BB); |
204 | } |
205 | |
206 | PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC, |
207 | CGSCCAnalysisManager &AM, LazyCallGraph &CG, |
208 | CGSCCUpdateResult &UR) { |
209 | const auto &MAMProxy = |
210 | AM.getResult<ModuleAnalysisManagerCGSCCProxy>(IR&: InitialC, ExtraArgs&: CG); |
211 | bool Changed = false; |
212 | |
213 | assert(InitialC.size() > 0 && "Cannot handle an empty SCC!" ); |
214 | Module &M = *InitialC.begin()->getFunction().getParent(); |
215 | ProfileSummaryInfo *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: M); |
216 | |
217 | FunctionAnalysisManager &FAM = |
218 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(IR&: InitialC, ExtraArgs&: CG) |
219 | .getManager(); |
220 | |
221 | InlineAdvisor &Advisor = getAdvisor(MAM: MAMProxy, FAM, M); |
222 | Advisor.onPassEntry(SCC: &InitialC); |
223 | |
224 | // We use a single common worklist for calls across the entire SCC. We |
225 | // process these in-order and append new calls introduced during inlining to |
226 | // the end. The PriorityInlineOrder is optional here, in which the smaller |
227 | // callee would have a higher priority to inline. |
228 | // |
229 | // Note that this particular order of processing is actually critical to |
230 | // avoid very bad behaviors. Consider *highly connected* call graphs where |
231 | // each function contains a small amount of code and a couple of calls to |
232 | // other functions. Because the LLVM inliner is fundamentally a bottom-up |
233 | // inliner, it can handle gracefully the fact that these all appear to be |
234 | // reasonable inlining candidates as it will flatten things until they become |
235 | // too big to inline, and then move on and flatten another batch. |
236 | // |
237 | // However, when processing call edges *within* an SCC we cannot rely on this |
238 | // bottom-up behavior. As a consequence, with heavily connected *SCCs* of |
239 | // functions we can end up incrementally inlining N calls into each of |
240 | // N functions because each incremental inlining decision looks good and we |
241 | // don't have a topological ordering to prevent explosions. |
242 | // |
243 | // To compensate for this, we don't process transitive edges made immediate |
244 | // by inlining until we've done one pass of inlining across the entire SCC. |
245 | // Large, highly connected SCCs still lead to some amount of code bloat in |
246 | // this model, but it is uniformly spread across all the functions in the SCC |
247 | // and eventually they all become too large to inline, rather than |
248 | // incrementally maknig a single function grow in a super linear fashion. |
249 | SmallVector<std::pair<CallBase *, int>, 16> Calls; |
250 | |
251 | // Populate the initial list of calls in this SCC. |
252 | for (auto &N : InitialC) { |
253 | auto &ORE = |
254 | FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: N.getFunction()); |
255 | // We want to generally process call sites top-down in order for |
256 | // simplifications stemming from replacing the call with the returned value |
257 | // after inlining to be visible to subsequent inlining decisions. |
258 | // FIXME: Using instructions sequence is a really bad way to do this. |
259 | // Instead we should do an actual RPO walk of the function body. |
260 | for (Instruction &I : instructions(F&: N.getFunction())) |
261 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
262 | if (Function *Callee = CB->getCalledFunction()) { |
263 | if (!Callee->isDeclaration()) |
264 | Calls.push_back(Elt: {CB, -1}); |
265 | else if (!isa<IntrinsicInst>(Val: I)) { |
266 | using namespace ore; |
267 | setInlineRemark(CB&: *CB, Message: "unavailable definition" ); |
268 | ORE.emit(RemarkBuilder: [&]() { |
269 | return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition" , &I) |
270 | << NV("Callee" , Callee) << " will not be inlined into " |
271 | << NV("Caller" , CB->getCaller()) |
272 | << " because its definition is unavailable" |
273 | << setIsVerbose(); |
274 | }); |
275 | } |
276 | } |
277 | } |
278 | |
279 | // Capture updatable variable for the current SCC. |
280 | auto *C = &InitialC; |
281 | |
282 | auto AdvisorOnExit = make_scope_exit(F: [&] { Advisor.onPassExit(SCC: C); }); |
283 | |
284 | if (Calls.empty()) |
285 | return PreservedAnalyses::all(); |
286 | |
287 | // When inlining a callee produces new call sites, we want to keep track of |
288 | // the fact that they were inlined from the callee. This allows us to avoid |
289 | // infinite inlining in some obscure cases. To represent this, we use an |
290 | // index into the InlineHistory vector. |
291 | SmallVector<std::pair<Function *, int>, 16> InlineHistory; |
292 | |
293 | // Track a set vector of inlined callees so that we can augment the caller |
294 | // with all of their edges in the call graph before pruning out the ones that |
295 | // got simplified away. |
296 | SmallSetVector<Function *, 4> InlinedCallees; |
297 | |
298 | // Track the dead functions to delete once finished with inlining calls. We |
299 | // defer deleting these to make it easier to handle the call graph updates. |
300 | SmallVector<Function *, 4> DeadFunctions; |
301 | |
302 | // Track potentially dead non-local functions with comdats to see if they can |
303 | // be deleted as a batch after inlining. |
304 | SmallVector<Function *, 4> DeadFunctionsInComdats; |
305 | |
306 | // Loop forward over all of the calls. Note that we cannot cache the size as |
307 | // inlining can introduce new calls that need to be processed. |
308 | for (int I = 0; I < (int)Calls.size(); ++I) { |
309 | // We expect the calls to typically be batched with sequences of calls that |
310 | // have the same caller, so we first set up some shared infrastructure for |
311 | // this caller. We also do any pruning we can at this layer on the caller |
312 | // alone. |
313 | Function &F = *Calls[I].first->getCaller(); |
314 | LazyCallGraph::Node &N = *CG.lookup(F); |
315 | if (CG.lookupSCC(N) != C) |
316 | continue; |
317 | |
318 | LLVM_DEBUG(dbgs() << "Inlining calls in: " << F.getName() << "\n" |
319 | << " Function size: " << F.getInstructionCount() |
320 | << "\n" ); |
321 | |
322 | auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { |
323 | return FAM.getResult<AssumptionAnalysis>(IR&: F); |
324 | }; |
325 | |
326 | // Now process as many calls as we have within this caller in the sequence. |
327 | // We bail out as soon as the caller has to change so we can update the |
328 | // call graph and prepare the context of that new caller. |
329 | bool DidInline = false; |
330 | for (; I < (int)Calls.size() && Calls[I].first->getCaller() == &F; ++I) { |
331 | auto &P = Calls[I]; |
332 | CallBase *CB = P.first; |
333 | const int InlineHistoryID = P.second; |
334 | Function &Callee = *CB->getCalledFunction(); |
335 | |
336 | if (InlineHistoryID != -1 && |
337 | inlineHistoryIncludes(F: &Callee, InlineHistoryID, InlineHistory)) { |
338 | LLVM_DEBUG(dbgs() << "Skipping inlining due to history: " << F.getName() |
339 | << " -> " << Callee.getName() << "\n" ); |
340 | setInlineRemark(CB&: *CB, Message: "recursive" ); |
341 | // Set noinline so that we don't forget this decision across CGSCC |
342 | // iterations. |
343 | CB->setIsNoInline(); |
344 | continue; |
345 | } |
346 | |
347 | // Check if this inlining may repeat breaking an SCC apart that has |
348 | // already been split once before. In that case, inlining here may |
349 | // trigger infinite inlining, much like is prevented within the inliner |
350 | // itself by the InlineHistory above, but spread across CGSCC iterations |
351 | // and thus hidden from the full inline history. |
352 | LazyCallGraph::SCC *CalleeSCC = CG.lookupSCC(N&: *CG.lookup(F: Callee)); |
353 | if (CalleeSCC == C && UR.InlinedInternalEdges.count(V: {&N, C})) { |
354 | LLVM_DEBUG(dbgs() << "Skipping inlining internal SCC edge from a node " |
355 | "previously split out of this SCC by inlining: " |
356 | << F.getName() << " -> " << Callee.getName() << "\n" ); |
357 | setInlineRemark(CB&: *CB, Message: "recursive SCC split" ); |
358 | continue; |
359 | } |
360 | |
361 | std::unique_ptr<InlineAdvice> Advice = |
362 | Advisor.getAdvice(CB&: *CB, MandatoryOnly: OnlyMandatory); |
363 | |
364 | // Check whether we want to inline this callsite. |
365 | if (!Advice) |
366 | continue; |
367 | |
368 | if (!Advice->isInliningRecommended()) { |
369 | Advice->recordUnattemptedInlining(); |
370 | continue; |
371 | } |
372 | |
373 | int CBCostMult = |
374 | getStringFnAttrAsInt( |
375 | CB&: *CB, AttrKind: InlineConstants::FunctionInlineCostMultiplierAttributeName) |
376 | .value_or(u: 1); |
377 | |
378 | // Setup the data structure used to plumb customization into the |
379 | // `InlineFunction` routine. |
380 | InlineFunctionInfo IFI( |
381 | GetAssumptionCache, PSI, |
382 | &FAM.getResult<BlockFrequencyAnalysis>(IR&: *(CB->getCaller())), |
383 | &FAM.getResult<BlockFrequencyAnalysis>(IR&: Callee)); |
384 | |
385 | InlineResult IR = InlineFunction( |
386 | CB&: *CB, IFI, /*MergeAttributes=*/true, |
387 | CalleeAAR: &FAM.getResult<AAManager>(IR&: *CB->getCaller()), InsertLifetime: true, ForwardVarArgsTo: nullptr, |
388 | ORE: &FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: *CB->getCaller())); |
389 | if (!IR.isSuccess()) { |
390 | Advice->recordUnsuccessfulInlining(Result: IR); |
391 | continue; |
392 | } |
393 | // TODO: Shouldn't we be invalidating all analyses on F here? |
394 | // The caller was modified, so invalidate Ephemeral Values. |
395 | FAM.getResult<EphemeralValuesAnalysis>(IR&: F).clear(); |
396 | |
397 | DidInline = true; |
398 | InlinedCallees.insert(X: &Callee); |
399 | ++NumInlined; |
400 | |
401 | LLVM_DEBUG(dbgs() << " Size after inlining: " |
402 | << F.getInstructionCount() << "\n" ); |
403 | |
404 | // Add any new callsites to defined functions to the worklist. |
405 | if (!IFI.InlinedCallSites.empty()) { |
406 | int NewHistoryID = InlineHistory.size(); |
407 | InlineHistory.push_back(Elt: {&Callee, InlineHistoryID}); |
408 | |
409 | for (CallBase *ICB : reverse(C&: IFI.InlinedCallSites)) { |
410 | Function *NewCallee = ICB->getCalledFunction(); |
411 | assert(!(NewCallee && NewCallee->isIntrinsic()) && |
412 | "Intrinsic calls should not be tracked." ); |
413 | if (!NewCallee) { |
414 | // Try to promote an indirect (virtual) call without waiting for |
415 | // the post-inline cleanup and the next DevirtSCCRepeatedPass |
416 | // iteration because the next iteration may not happen and we may |
417 | // miss inlining it. |
418 | if (tryPromoteCall(CB&: *ICB)) |
419 | NewCallee = ICB->getCalledFunction(); |
420 | } |
421 | if (NewCallee) { |
422 | if (!NewCallee->isDeclaration()) { |
423 | Calls.push_back(Elt: {ICB, NewHistoryID}); |
424 | // Continually inlining through an SCC can result in huge compile |
425 | // times and bloated code since we arbitrarily stop at some point |
426 | // when the inliner decides it's not profitable to inline anymore. |
427 | // We attempt to mitigate this by making these calls exponentially |
428 | // more expensive. |
429 | // This doesn't apply to calls in the same SCC since if we do |
430 | // inline through the SCC the function will end up being |
431 | // self-recursive which the inliner bails out on, and inlining |
432 | // within an SCC is necessary for performance. |
433 | if (CalleeSCC != C && |
434 | CalleeSCC == CG.lookupSCC(N&: CG.get(F&: *NewCallee))) { |
435 | Attribute NewCBCostMult = Attribute::get( |
436 | Context&: M.getContext(), |
437 | Kind: InlineConstants::FunctionInlineCostMultiplierAttributeName, |
438 | Val: itostr(X: CBCostMult * IntraSCCCostMultiplier)); |
439 | ICB->addFnAttr(Attr: NewCBCostMult); |
440 | } |
441 | } |
442 | } |
443 | } |
444 | } |
445 | |
446 | // For local functions or discardable functions without comdats, check |
447 | // whether this makes the callee trivially dead. In that case, we can drop |
448 | // the body of the function eagerly which may reduce the number of callers |
449 | // of other functions to one, changing inline cost thresholds. Non-local |
450 | // discardable functions with comdats are checked later on. |
451 | bool CalleeWasDeleted = false; |
452 | if (Callee.isDiscardableIfUnused() && Callee.hasZeroLiveUses() && |
453 | !CG.isLibFunction(F&: Callee)) { |
454 | if (Callee.hasLocalLinkage() || !Callee.hasComdat()) { |
455 | Calls.erase( |
456 | CS: std::remove_if(first: Calls.begin() + I + 1, last: Calls.end(), |
457 | pred: [&](const std::pair<CallBase *, int> &Call) { |
458 | return Call.first->getCaller() == &Callee; |
459 | }), |
460 | CE: Calls.end()); |
461 | |
462 | // Clear the body and queue the function itself for call graph |
463 | // updating when we finish inlining. |
464 | makeFunctionBodyUnreachable(F&: Callee); |
465 | assert(!is_contained(DeadFunctions, &Callee) && |
466 | "Cannot put cause a function to become dead twice!" ); |
467 | DeadFunctions.push_back(Elt: &Callee); |
468 | CalleeWasDeleted = true; |
469 | } else { |
470 | DeadFunctionsInComdats.push_back(Elt: &Callee); |
471 | } |
472 | } |
473 | if (CalleeWasDeleted) |
474 | Advice->recordInliningWithCalleeDeleted(); |
475 | else |
476 | Advice->recordInlining(); |
477 | } |
478 | |
479 | // Back the call index up by one to put us in a good position to go around |
480 | // the outer loop. |
481 | --I; |
482 | |
483 | if (!DidInline) |
484 | continue; |
485 | Changed = true; |
486 | |
487 | // At this point, since we have made changes we have at least removed |
488 | // a call instruction. However, in the process we do some incremental |
489 | // simplification of the surrounding code. This simplification can |
490 | // essentially do all of the same things as a function pass and we can |
491 | // re-use the exact same logic for updating the call graph to reflect the |
492 | // change. |
493 | |
494 | // Inside the update, we also update the FunctionAnalysisManager in the |
495 | // proxy for this particular SCC. We do this as the SCC may have changed and |
496 | // as we're going to mutate this particular function we want to make sure |
497 | // the proxy is in place to forward any invalidation events. |
498 | LazyCallGraph::SCC *OldC = C; |
499 | C = &updateCGAndAnalysisManagerForCGSCCPass(G&: CG, C&: *C, N, AM, UR, FAM); |
500 | LLVM_DEBUG(dbgs() << "Updated inlining SCC: " << *C << "\n" ); |
501 | |
502 | // If this causes an SCC to split apart into multiple smaller SCCs, there |
503 | // is a subtle risk we need to prepare for. Other transformations may |
504 | // expose an "infinite inlining" opportunity later, and because of the SCC |
505 | // mutation, we will revisit this function and potentially re-inline. If we |
506 | // do, and that re-inlining also has the potentially to mutate the SCC |
507 | // structure, the infinite inlining problem can manifest through infinite |
508 | // SCC splits and merges. To avoid this, we capture the originating caller |
509 | // node and the SCC containing the call edge. This is a slight over |
510 | // approximation of the possible inlining decisions that must be avoided, |
511 | // but is relatively efficient to store. We use C != OldC to know when |
512 | // a new SCC is generated and the original SCC may be generated via merge |
513 | // in later iterations. |
514 | // |
515 | // It is also possible that even if no new SCC is generated |
516 | // (i.e., C == OldC), the original SCC could be split and then merged |
517 | // into the same one as itself. and the original SCC will be added into |
518 | // UR.CWorklist again, we want to catch such cases too. |
519 | // |
520 | // FIXME: This seems like a very heavyweight way of retaining the inline |
521 | // history, we should look for a more efficient way of tracking it. |
522 | if ((C != OldC || UR.CWorklist.count(key: OldC)) && |
523 | llvm::any_of(Range&: InlinedCallees, P: [&](Function *Callee) { |
524 | return CG.lookupSCC(N&: *CG.lookup(F: *Callee)) == OldC; |
525 | })) { |
526 | LLVM_DEBUG(dbgs() << "Inlined an internal call edge and split an SCC, " |
527 | "retaining this to avoid infinite inlining.\n" ); |
528 | UR.InlinedInternalEdges.insert(V: {&N, OldC}); |
529 | } |
530 | InlinedCallees.clear(); |
531 | |
532 | // Invalidate analyses for this function now so that we don't have to |
533 | // invalidate analyses for all functions in this SCC later. |
534 | FAM.invalidate(IR&: F, PA: PreservedAnalyses::none()); |
535 | } |
536 | |
537 | // We must ensure that we only delete functions with comdats if every function |
538 | // in the comdat is going to be deleted. |
539 | if (!DeadFunctionsInComdats.empty()) { |
540 | filterDeadComdatFunctions(DeadComdatFunctions&: DeadFunctionsInComdats); |
541 | for (auto *Callee : DeadFunctionsInComdats) |
542 | makeFunctionBodyUnreachable(F&: *Callee); |
543 | DeadFunctions.append(RHS: DeadFunctionsInComdats); |
544 | } |
545 | |
546 | // Now that we've finished inlining all of the calls across this SCC, delete |
547 | // all of the trivially dead functions, updating the call graph and the CGSCC |
548 | // pass manager in the process. |
549 | // |
550 | // Note that this walks a pointer set which has non-deterministic order but |
551 | // that is OK as all we do is delete things and add pointers to unordered |
552 | // sets. |
553 | for (Function *DeadF : DeadFunctions) { |
554 | CG.markDeadFunction(F&: *DeadF); |
555 | // Get the necessary information out of the call graph and nuke the |
556 | // function there. Also, clear out any cached analyses. |
557 | auto &DeadC = *CG.lookupSCC(N&: *CG.lookup(F: *DeadF)); |
558 | FAM.clear(IR&: *DeadF, Name: DeadF->getName()); |
559 | AM.clear(IR&: DeadC, Name: DeadC.getName()); |
560 | |
561 | // Mark the relevant parts of the call graph as invalid so we don't visit |
562 | // them. |
563 | UR.InvalidatedSCCs.insert(Ptr: &DeadC); |
564 | |
565 | UR.DeadFunctions.push_back(Elt: DeadF); |
566 | |
567 | ++NumDeleted; |
568 | } |
569 | |
570 | if (!Changed) |
571 | return PreservedAnalyses::all(); |
572 | |
573 | PreservedAnalyses PA; |
574 | // Even if we change the IR, we update the core CGSCC data structures and so |
575 | // can preserve the proxy to the function analysis manager. |
576 | PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); |
577 | // We have already invalidated all analyses on modified functions. |
578 | PA.preserveSet<AllAnalysesOn<Function>>(); |
579 | return PA; |
580 | } |
581 | |
582 | ModuleInlinerWrapperPass::ModuleInlinerWrapperPass(InlineParams Params, |
583 | bool MandatoryFirst, |
584 | InlineContext IC, |
585 | InliningAdvisorMode Mode, |
586 | unsigned MaxDevirtIterations) |
587 | : Params(Params), IC(IC), Mode(Mode), |
588 | MaxDevirtIterations(MaxDevirtIterations) { |
589 | // Run the inliner first. The theory is that we are walking bottom-up and so |
590 | // the callees have already been fully optimized, and we want to inline them |
591 | // into the callers so that our optimizations can reflect that. |
592 | // For PreLinkThinLTO pass, we disable hot-caller heuristic for sample PGO |
593 | // because it makes profile annotation in the backend inaccurate. |
594 | if (MandatoryFirst) { |
595 | PM.addPass(Pass: InlinerPass(/*OnlyMandatory*/ true)); |
596 | if (EnablePostSCCAdvisorPrinting) |
597 | PM.addPass(Pass: InlineAdvisorAnalysisPrinterPass(dbgs())); |
598 | } |
599 | PM.addPass(Pass: InlinerPass()); |
600 | if (EnablePostSCCAdvisorPrinting) |
601 | PM.addPass(Pass: InlineAdvisorAnalysisPrinterPass(dbgs())); |
602 | } |
603 | |
604 | PreservedAnalyses ModuleInlinerWrapperPass::run(Module &M, |
605 | ModuleAnalysisManager &MAM) { |
606 | auto &IAA = MAM.getResult<InlineAdvisorAnalysis>(IR&: M); |
607 | if (!IAA.tryCreate(Params, Mode, |
608 | ReplaySettings: {.ReplayFile: CGSCCInlineReplayFile, |
609 | .ReplayScope: CGSCCInlineReplayScope, |
610 | .ReplayFallback: CGSCCInlineReplayFallback, |
611 | .ReplayFormat: {.OutputFormat: CGSCCInlineReplayFormat}}, |
612 | IC)) { |
613 | M.getContext().emitError( |
614 | ErrorStr: "Could not setup Inlining Advisor for the requested " |
615 | "mode and/or options" ); |
616 | return PreservedAnalyses::all(); |
617 | } |
618 | |
619 | // We wrap the CGSCC pipeline in a devirtualization repeater. This will try |
620 | // to detect when we devirtualize indirect calls and iterate the SCC passes |
621 | // in that case to try and catch knock-on inlining or function attrs |
622 | // opportunities. Then we add it to the module pipeline by walking the SCCs |
623 | // in postorder (or bottom-up). |
624 | // If MaxDevirtIterations is 0, we just don't use the devirtualization |
625 | // wrapper. |
626 | if (MaxDevirtIterations == 0) |
627 | MPM.addPass(Pass: createModuleToPostOrderCGSCCPassAdaptor(Pass: std::move(PM))); |
628 | else |
629 | MPM.addPass(Pass: createModuleToPostOrderCGSCCPassAdaptor( |
630 | Pass: createDevirtSCCRepeatedPass(Pass: std::move(PM), MaxIterations: MaxDevirtIterations))); |
631 | |
632 | MPM.addPass(Pass: std::move(AfterCGMPM)); |
633 | MPM.run(IR&: M, AM&: MAM); |
634 | |
635 | // Discard the InlineAdvisor, a subsequent inlining session should construct |
636 | // its own. |
637 | auto PA = PreservedAnalyses::all(); |
638 | if (!KeepAdvisorForPrinting) |
639 | PA.abandon<InlineAdvisorAnalysis>(); |
640 | return PA; |
641 | } |
642 | |
643 | void InlinerPass::printPipeline( |
644 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
645 | static_cast<PassInfoMixin<InlinerPass> *>(this)->printPipeline( |
646 | OS, MapClassName2PassName); |
647 | if (OnlyMandatory) |
648 | OS << "<only-mandatory>" ; |
649 | } |
650 | |
651 | void ModuleInlinerWrapperPass::printPipeline( |
652 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
653 | // Print some info about passes added to the wrapper. This is however |
654 | // incomplete as InlineAdvisorAnalysis part isn't included (which also depends |
655 | // on Params and Mode). |
656 | if (!MPM.isEmpty()) { |
657 | MPM.printPipeline(OS, MapClassName2PassName); |
658 | OS << ','; |
659 | } |
660 | OS << "cgscc(" ; |
661 | if (MaxDevirtIterations != 0) |
662 | OS << "devirt<" << MaxDevirtIterations << ">(" ; |
663 | PM.printPipeline(OS, MapClassName2PassName); |
664 | if (MaxDevirtIterations != 0) |
665 | OS << ')'; |
666 | OS << ')'; |
667 | } |
668 | |