1 //===- InlineAdvisor.cpp - analysis pass implementation -------------------===//
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 InlineAdvisorAnalysis and DefaultInlineAdvisor, and
10 // related types.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Analysis/InlineAdvisor.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/InlineCost.h"
17 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
18 #include "llvm/Analysis/ProfileSummaryInfo.h"
19 #include "llvm/Analysis/ReplayInlineAdvisor.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/DebugInfoMetadata.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/raw_ostream.h"
26
27 using namespace llvm;
28 #define DEBUG_TYPE "inline"
29
30 // This weirdly named statistic tracks the number of times that, when attempting
31 // to inline a function A into B, we analyze the callers of B in order to see
32 // if those would be more profitable and blocked inline steps.
33 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
34
35 /// Flag to add inline messages as callsite attributes 'inline-remark'.
36 static cl::opt<bool>
37 InlineRemarkAttribute("inline-remark-attribute", cl::init(false),
38 cl::Hidden,
39 cl::desc("Enable adding inline-remark attribute to"
40 " callsites processed by inliner but decided"
41 " to be not inlined"));
42
43 // An integer used to limit the cost of inline deferral. The default negative
44 // number tells shouldBeDeferred to only take the secondary cost into account.
45 static cl::opt<int>
46 InlineDeferralScale("inline-deferral-scale",
47 cl::desc("Scale to limit the cost of inline deferral"),
48 cl::init(2), cl::Hidden);
49
50 extern cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats;
51
recordUnsuccessfulInliningImpl(const InlineResult & Result)52 void DefaultInlineAdvice::recordUnsuccessfulInliningImpl(
53 const InlineResult &Result) {
54 using namespace ore;
55 llvm::setInlineRemark(*OriginalCB, std::string(Result.getFailureReason()) +
56 "; " + inlineCostStr(*OIC));
57 ORE.emit([&]() {
58 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, Block)
59 << NV("Callee", Callee) << " will not be inlined into "
60 << NV("Caller", Caller) << ": "
61 << NV("Reason", Result.getFailureReason());
62 });
63 }
64
recordInliningWithCalleeDeletedImpl()65 void DefaultInlineAdvice::recordInliningWithCalleeDeletedImpl() {
66 if (EmitRemarks)
67 emitInlinedInto(ORE, DLoc, Block, *Callee, *Caller, *OIC);
68 }
69
recordInliningImpl()70 void DefaultInlineAdvice::recordInliningImpl() {
71 if (EmitRemarks)
72 emitInlinedInto(ORE, DLoc, Block, *Callee, *Caller, *OIC);
73 }
74
getDefaultInlineAdvice(CallBase & CB,FunctionAnalysisManager & FAM,const InlineParams & Params)75 llvm::Optional<llvm::InlineCost> static getDefaultInlineAdvice(
76 CallBase &CB, FunctionAnalysisManager &FAM, const InlineParams &Params) {
77 Function &Caller = *CB.getCaller();
78 ProfileSummaryInfo *PSI =
79 FAM.getResult<ModuleAnalysisManagerFunctionProxy>(Caller)
80 .getCachedResult<ProfileSummaryAnalysis>(
81 *CB.getParent()->getParent()->getParent());
82
83 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(Caller);
84 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
85 return FAM.getResult<AssumptionAnalysis>(F);
86 };
87 auto GetBFI = [&](Function &F) -> BlockFrequencyInfo & {
88 return FAM.getResult<BlockFrequencyAnalysis>(F);
89 };
90 auto GetTLI = [&](Function &F) -> const TargetLibraryInfo & {
91 return FAM.getResult<TargetLibraryAnalysis>(F);
92 };
93
94 auto GetInlineCost = [&](CallBase &CB) {
95 Function &Callee = *CB.getCalledFunction();
96 auto &CalleeTTI = FAM.getResult<TargetIRAnalysis>(Callee);
97 bool RemarksEnabled =
98 Callee.getContext().getDiagHandlerPtr()->isMissedOptRemarkEnabled(
99 DEBUG_TYPE);
100 return getInlineCost(CB, Params, CalleeTTI, GetAssumptionCache, GetTLI,
101 GetBFI, PSI, RemarksEnabled ? &ORE : nullptr);
102 };
103 return llvm::shouldInline(CB, GetInlineCost, ORE,
104 Params.EnableDeferral.getValueOr(false));
105 }
106
107 std::unique_ptr<InlineAdvice>
getAdviceImpl(CallBase & CB)108 DefaultInlineAdvisor::getAdviceImpl(CallBase &CB) {
109 auto OIC = getDefaultInlineAdvice(CB, FAM, Params);
110 return std::make_unique<DefaultInlineAdvice>(
111 this, CB, OIC,
112 FAM.getResult<OptimizationRemarkEmitterAnalysis>(*CB.getCaller()));
113 }
114
InlineAdvice(InlineAdvisor * Advisor,CallBase & CB,OptimizationRemarkEmitter & ORE,bool IsInliningRecommended)115 InlineAdvice::InlineAdvice(InlineAdvisor *Advisor, CallBase &CB,
116 OptimizationRemarkEmitter &ORE,
117 bool IsInliningRecommended)
118 : Advisor(Advisor), Caller(CB.getCaller()), Callee(CB.getCalledFunction()),
119 DLoc(CB.getDebugLoc()), Block(CB.getParent()), ORE(ORE),
120 IsInliningRecommended(IsInliningRecommended) {}
121
markFunctionAsDeleted(Function * F)122 void InlineAdvisor::markFunctionAsDeleted(Function *F) {
123 assert((!DeletedFunctions.count(F)) &&
124 "Cannot put cause a function to become dead twice!");
125 DeletedFunctions.insert(F);
126 }
127
freeDeletedFunctions()128 void InlineAdvisor::freeDeletedFunctions() {
129 for (auto *F : DeletedFunctions)
130 delete F;
131 DeletedFunctions.clear();
132 }
133
recordInlineStatsIfNeeded()134 void InlineAdvice::recordInlineStatsIfNeeded() {
135 if (Advisor->ImportedFunctionsStats)
136 Advisor->ImportedFunctionsStats->recordInline(*Caller, *Callee);
137 }
138
recordInlining()139 void InlineAdvice::recordInlining() {
140 markRecorded();
141 recordInlineStatsIfNeeded();
142 recordInliningImpl();
143 }
144
recordInliningWithCalleeDeleted()145 void InlineAdvice::recordInliningWithCalleeDeleted() {
146 markRecorded();
147 recordInlineStatsIfNeeded();
148 Advisor->markFunctionAsDeleted(Callee);
149 recordInliningWithCalleeDeletedImpl();
150 }
151
152 AnalysisKey InlineAdvisorAnalysis::Key;
153
tryCreate(InlineParams Params,InliningAdvisorMode Mode,StringRef ReplayFile)154 bool InlineAdvisorAnalysis::Result::tryCreate(InlineParams Params,
155 InliningAdvisorMode Mode,
156 StringRef ReplayFile) {
157 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
158 switch (Mode) {
159 case InliningAdvisorMode::Default:
160 Advisor.reset(new DefaultInlineAdvisor(M, FAM, Params));
161 // Restrict replay to default advisor, ML advisors are stateful so
162 // replay will need augmentations to interleave with them correctly.
163 if (!ReplayFile.empty()) {
164 Advisor = std::make_unique<ReplayInlineAdvisor>(
165 M, FAM, M.getContext(), std::move(Advisor), ReplayFile,
166 /* EmitRemarks =*/true);
167 }
168 break;
169 case InliningAdvisorMode::Development:
170 #ifdef LLVM_HAVE_TF_API
171 Advisor =
172 llvm::getDevelopmentModeAdvisor(M, MAM, [&FAM, Params](CallBase &CB) {
173 auto OIC = getDefaultInlineAdvice(CB, FAM, Params);
174 return OIC.hasValue();
175 });
176 #endif
177 break;
178 case InliningAdvisorMode::Release:
179 #ifdef LLVM_HAVE_TF_AOT
180 Advisor = llvm::getReleaseModeAdvisor(M, MAM);
181 #endif
182 break;
183 }
184
185 return !!Advisor;
186 }
187
188 /// Return true if inlining of CB can block the caller from being
189 /// inlined which is proved to be more beneficial. \p IC is the
190 /// estimated inline cost associated with callsite \p CB.
191 /// \p TotalSecondaryCost will be set to the estimated cost of inlining the
192 /// caller if \p CB is suppressed for inlining.
193 static bool
shouldBeDeferred(Function * Caller,InlineCost IC,int & TotalSecondaryCost,function_ref<InlineCost (CallBase & CB)> GetInlineCost)194 shouldBeDeferred(Function *Caller, InlineCost IC, int &TotalSecondaryCost,
195 function_ref<InlineCost(CallBase &CB)> GetInlineCost) {
196 // For now we only handle local or inline functions.
197 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage())
198 return false;
199 // If the cost of inlining CB is non-positive, it is not going to prevent the
200 // caller from being inlined into its callers and hence we don't need to
201 // defer.
202 if (IC.getCost() <= 0)
203 return false;
204 // Try to detect the case where the current inlining candidate caller (call
205 // it B) is a static or linkonce-ODR function and is an inlining candidate
206 // elsewhere, and the current candidate callee (call it C) is large enough
207 // that inlining it into B would make B too big to inline later. In these
208 // circumstances it may be best not to inline C into B, but to inline B into
209 // its callers.
210 //
211 // This only applies to static and linkonce-ODR functions because those are
212 // expected to be available for inlining in the translation units where they
213 // are used. Thus we will always have the opportunity to make local inlining
214 // decisions. Importantly the linkonce-ODR linkage covers inline functions
215 // and templates in C++.
216 //
217 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
218 // the internal implementation of the inline cost metrics rather than
219 // treating them as truly abstract units etc.
220 TotalSecondaryCost = 0;
221 // The candidate cost to be imposed upon the current function.
222 int CandidateCost = IC.getCost() - 1;
223 // If the caller has local linkage and can be inlined to all its callers, we
224 // can apply a huge negative bonus to TotalSecondaryCost.
225 bool ApplyLastCallBonus = Caller->hasLocalLinkage() && !Caller->hasOneUse();
226 // This bool tracks what happens if we DO inline C into B.
227 bool InliningPreventsSomeOuterInline = false;
228 unsigned NumCallerUsers = 0;
229 for (User *U : Caller->users()) {
230 CallBase *CS2 = dyn_cast<CallBase>(U);
231
232 // If this isn't a call to Caller (it could be some other sort
233 // of reference) skip it. Such references will prevent the caller
234 // from being removed.
235 if (!CS2 || CS2->getCalledFunction() != Caller) {
236 ApplyLastCallBonus = false;
237 continue;
238 }
239
240 InlineCost IC2 = GetInlineCost(*CS2);
241 ++NumCallerCallersAnalyzed;
242 if (!IC2) {
243 ApplyLastCallBonus = false;
244 continue;
245 }
246 if (IC2.isAlways())
247 continue;
248
249 // See if inlining of the original callsite would erase the cost delta of
250 // this callsite. We subtract off the penalty for the call instruction,
251 // which we would be deleting.
252 if (IC2.getCostDelta() <= CandidateCost) {
253 InliningPreventsSomeOuterInline = true;
254 TotalSecondaryCost += IC2.getCost();
255 NumCallerUsers++;
256 }
257 }
258
259 if (!InliningPreventsSomeOuterInline)
260 return false;
261
262 // If all outer calls to Caller would get inlined, the cost for the last
263 // one is set very low by getInlineCost, in anticipation that Caller will
264 // be removed entirely. We did not account for this above unless there
265 // is only one caller of Caller.
266 if (ApplyLastCallBonus)
267 TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus;
268
269 // If InlineDeferralScale is negative, then ignore the cost of primary
270 // inlining -- IC.getCost() multiplied by the number of callers to Caller.
271 if (InlineDeferralScale < 0)
272 return TotalSecondaryCost < IC.getCost();
273
274 int TotalCost = TotalSecondaryCost + IC.getCost() * NumCallerUsers;
275 int Allowance = IC.getCost() * InlineDeferralScale;
276 return TotalCost < Allowance;
277 }
278
279 namespace llvm {
operator <<(raw_ostream & R,const ore::NV & Arg)280 static raw_ostream &operator<<(raw_ostream &R, const ore::NV &Arg) {
281 return R << Arg.Val;
282 }
283
284 template <class RemarkT>
operator <<(RemarkT && R,const InlineCost & IC)285 RemarkT &operator<<(RemarkT &&R, const InlineCost &IC) {
286 using namespace ore;
287 if (IC.isAlways()) {
288 R << "(cost=always)";
289 } else if (IC.isNever()) {
290 R << "(cost=never)";
291 } else {
292 R << "(cost=" << ore::NV("Cost", IC.getCost())
293 << ", threshold=" << ore::NV("Threshold", IC.getThreshold()) << ")";
294 }
295 if (const char *Reason = IC.getReason())
296 R << ": " << ore::NV("Reason", Reason);
297 return R;
298 }
299 } // namespace llvm
300
inlineCostStr(const InlineCost & IC)301 std::string llvm::inlineCostStr(const InlineCost &IC) {
302 std::string Buffer;
303 raw_string_ostream Remark(Buffer);
304 Remark << IC;
305 return Remark.str();
306 }
307
setInlineRemark(CallBase & CB,StringRef Message)308 void llvm::setInlineRemark(CallBase &CB, StringRef Message) {
309 if (!InlineRemarkAttribute)
310 return;
311
312 Attribute Attr = Attribute::get(CB.getContext(), "inline-remark", Message);
313 CB.addAttribute(AttributeList::FunctionIndex, Attr);
314 }
315
316 /// Return the cost only if the inliner should attempt to inline at the given
317 /// CallSite. If we return the cost, we will emit an optimisation remark later
318 /// using that cost, so we won't do so from this function. Return None if
319 /// inlining should not be attempted.
320 Optional<InlineCost>
shouldInline(CallBase & CB,function_ref<InlineCost (CallBase & CB)> GetInlineCost,OptimizationRemarkEmitter & ORE,bool EnableDeferral)321 llvm::shouldInline(CallBase &CB,
322 function_ref<InlineCost(CallBase &CB)> GetInlineCost,
323 OptimizationRemarkEmitter &ORE, bool EnableDeferral) {
324 using namespace ore;
325
326 InlineCost IC = GetInlineCost(CB);
327 Instruction *Call = &CB;
328 Function *Callee = CB.getCalledFunction();
329 Function *Caller = CB.getCaller();
330
331 if (IC.isAlways()) {
332 LLVM_DEBUG(dbgs() << " Inlining " << inlineCostStr(IC)
333 << ", Call: " << CB << "\n");
334 return IC;
335 }
336
337 if (!IC) {
338 LLVM_DEBUG(dbgs() << " NOT Inlining " << inlineCostStr(IC)
339 << ", Call: " << CB << "\n");
340 if (IC.isNever()) {
341 ORE.emit([&]() {
342 return OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call)
343 << NV("Callee", Callee) << " not inlined into "
344 << NV("Caller", Caller) << " because it should never be inlined "
345 << IC;
346 });
347 } else {
348 ORE.emit([&]() {
349 return OptimizationRemarkMissed(DEBUG_TYPE, "TooCostly", Call)
350 << NV("Callee", Callee) << " not inlined into "
351 << NV("Caller", Caller) << " because too costly to inline "
352 << IC;
353 });
354 }
355 setInlineRemark(CB, inlineCostStr(IC));
356 return None;
357 }
358
359 int TotalSecondaryCost = 0;
360 if (EnableDeferral &&
361 shouldBeDeferred(Caller, IC, TotalSecondaryCost, GetInlineCost)) {
362 LLVM_DEBUG(dbgs() << " NOT Inlining: " << CB
363 << " Cost = " << IC.getCost()
364 << ", outer Cost = " << TotalSecondaryCost << '\n');
365 ORE.emit([&]() {
366 return OptimizationRemarkMissed(DEBUG_TYPE, "IncreaseCostInOtherContexts",
367 Call)
368 << "Not inlining. Cost of inlining " << NV("Callee", Callee)
369 << " increases the cost of inlining " << NV("Caller", Caller)
370 << " in other contexts";
371 });
372 setInlineRemark(CB, "deferred");
373 // IC does not bool() to false, so get an InlineCost that will.
374 // This will not be inspected to make an error message.
375 return None;
376 }
377
378 LLVM_DEBUG(dbgs() << " Inlining " << inlineCostStr(IC) << ", Call: " << CB
379 << '\n');
380 return IC;
381 }
382
getCallSiteLocation(DebugLoc DLoc)383 std::string llvm::getCallSiteLocation(DebugLoc DLoc) {
384 std::string Buffer;
385 raw_string_ostream CallSiteLoc(Buffer);
386 bool First = true;
387 for (DILocation *DIL = DLoc.get(); DIL; DIL = DIL->getInlinedAt()) {
388 if (!First)
389 CallSiteLoc << " @ ";
390 // Note that negative line offset is actually possible, but we use
391 // unsigned int to match line offset representation in remarks so
392 // it's directly consumable by relay advisor.
393 uint32_t Offset =
394 DIL->getLine() - DIL->getScope()->getSubprogram()->getLine();
395 uint32_t Discriminator = DIL->getBaseDiscriminator();
396 StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName();
397 if (Name.empty())
398 Name = DIL->getScope()->getSubprogram()->getName();
399 CallSiteLoc << Name.str() << ":" << llvm::utostr(Offset) << ":"
400 << llvm::utostr(DIL->getColumn());
401 if (Discriminator)
402 CallSiteLoc << "." << llvm::utostr(Discriminator);
403 First = false;
404 }
405
406 return CallSiteLoc.str();
407 }
408
addLocationToRemarks(OptimizationRemark & Remark,DebugLoc DLoc)409 void llvm::addLocationToRemarks(OptimizationRemark &Remark, DebugLoc DLoc) {
410 if (!DLoc.get()) {
411 return;
412 }
413
414 bool First = true;
415 Remark << " at callsite ";
416 for (DILocation *DIL = DLoc.get(); DIL; DIL = DIL->getInlinedAt()) {
417 if (!First)
418 Remark << " @ ";
419 unsigned int Offset = DIL->getLine();
420 Offset -= DIL->getScope()->getSubprogram()->getLine();
421 unsigned int Discriminator = DIL->getBaseDiscriminator();
422 StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName();
423 if (Name.empty())
424 Name = DIL->getScope()->getSubprogram()->getName();
425 Remark << Name << ":" << ore::NV("Line", Offset) << ":"
426 << ore::NV("Column", DIL->getColumn());
427 if (Discriminator)
428 Remark << "." << ore::NV("Disc", Discriminator);
429 First = false;
430 }
431
432 Remark << ";";
433 }
434
emitInlinedInto(OptimizationRemarkEmitter & ORE,DebugLoc DLoc,const BasicBlock * Block,const Function & Callee,const Function & Caller,const InlineCost & IC,bool ForProfileContext,const char * PassName)435 void llvm::emitInlinedInto(OptimizationRemarkEmitter &ORE, DebugLoc DLoc,
436 const BasicBlock *Block, const Function &Callee,
437 const Function &Caller, const InlineCost &IC,
438 bool ForProfileContext, const char *PassName) {
439 ORE.emit([&]() {
440 bool AlwaysInline = IC.isAlways();
441 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined";
442 OptimizationRemark Remark(PassName ? PassName : DEBUG_TYPE, RemarkName,
443 DLoc, Block);
444 Remark << ore::NV("Callee", &Callee) << " inlined into ";
445 Remark << ore::NV("Caller", &Caller);
446 if (ForProfileContext)
447 Remark << " to match profiling context";
448 Remark << " with " << IC;
449 addLocationToRemarks(Remark, DLoc);
450 return Remark;
451 });
452 }
453
InlineAdvisor(Module & M,FunctionAnalysisManager & FAM)454 InlineAdvisor::InlineAdvisor(Module &M, FunctionAnalysisManager &FAM)
455 : M(M), FAM(FAM) {
456 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No) {
457 ImportedFunctionsStats =
458 std::make_unique<ImportedFunctionsInliningStatistics>();
459 ImportedFunctionsStats->setModuleInfo(M);
460 }
461 }
462
~InlineAdvisor()463 InlineAdvisor::~InlineAdvisor() {
464 if (ImportedFunctionsStats) {
465 assert(InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No);
466 ImportedFunctionsStats->dump(InlinerFunctionImportStats ==
467 InlinerFunctionImportStatsOpts::Verbose);
468 }
469
470 freeDeletedFunctions();
471 }
472
getMandatoryAdvice(CallBase & CB,bool Advice)473 std::unique_ptr<InlineAdvice> InlineAdvisor::getMandatoryAdvice(CallBase &CB,
474 bool Advice) {
475 return std::make_unique<InlineAdvice>(this, CB, getCallerORE(CB), Advice);
476 }
477
478 InlineAdvisor::MandatoryInliningKind
getMandatoryKind(CallBase & CB,FunctionAnalysisManager & FAM,OptimizationRemarkEmitter & ORE)479 InlineAdvisor::getMandatoryKind(CallBase &CB, FunctionAnalysisManager &FAM,
480 OptimizationRemarkEmitter &ORE) {
481 auto &Callee = *CB.getCalledFunction();
482
483 auto GetTLI = [&](Function &F) -> const TargetLibraryInfo & {
484 return FAM.getResult<TargetLibraryAnalysis>(F);
485 };
486
487 auto &TIR = FAM.getResult<TargetIRAnalysis>(Callee);
488
489 auto TrivialDecision =
490 llvm::getAttributeBasedInliningDecision(CB, &Callee, TIR, GetTLI);
491
492 if (TrivialDecision.hasValue()) {
493 if (TrivialDecision->isSuccess())
494 return MandatoryInliningKind::Always;
495 else
496 return MandatoryInliningKind::Never;
497 }
498 return MandatoryInliningKind::NotMandatory;
499 }
500
getAdvice(CallBase & CB,bool MandatoryOnly)501 std::unique_ptr<InlineAdvice> InlineAdvisor::getAdvice(CallBase &CB,
502 bool MandatoryOnly) {
503 if (!MandatoryOnly)
504 return getAdviceImpl(CB);
505 bool Advice = CB.getCaller() != CB.getCalledFunction() &&
506 MandatoryInliningKind::Always ==
507 getMandatoryKind(CB, FAM, getCallerORE(CB));
508 return getMandatoryAdvice(CB, Advice);
509 }
510
getCallerORE(CallBase & CB)511 OptimizationRemarkEmitter &InlineAdvisor::getCallerORE(CallBase &CB) {
512 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*CB.getCaller());
513 }
514