1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
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 Function import based on summaries.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/Transforms/IPO/FunctionImport.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/Bitcode/BitcodeReader.h"
23 #include "llvm/IR/AutoUpgrade.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalObject.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/ModuleSummaryIndex.h"
33 #include "llvm/IRReader/IRReader.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/Linker/IRMover.h"
36 #include "llvm/Object/ModuleSymbolTable.h"
37 #include "llvm/Object/SymbolicFile.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/Error.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/FileSystem.h"
45 #include "llvm/Support/SourceMgr.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Transforms/IPO/Internalize.h"
48 #include "llvm/Transforms/Utils/Cloning.h"
49 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
50 #include "llvm/Transforms/Utils/ValueMapper.h"
51 #include <cassert>
52 #include <memory>
53 #include <set>
54 #include <string>
55 #include <system_error>
56 #include <tuple>
57 #include <utility>
58
59 using namespace llvm;
60
61 #define DEBUG_TYPE "function-import"
62
63 STATISTIC(NumImportedFunctionsThinLink,
64 "Number of functions thin link decided to import");
65 STATISTIC(NumImportedHotFunctionsThinLink,
66 "Number of hot functions thin link decided to import");
67 STATISTIC(NumImportedCriticalFunctionsThinLink,
68 "Number of critical functions thin link decided to import");
69 STATISTIC(NumImportedGlobalVarsThinLink,
70 "Number of global variables thin link decided to import");
71 STATISTIC(NumImportedFunctions, "Number of functions imported in backend");
72 STATISTIC(NumImportedGlobalVars,
73 "Number of global variables imported in backend");
74 STATISTIC(NumImportedModules, "Number of modules imported from");
75 STATISTIC(NumDeadSymbols, "Number of dead stripped symbols in index");
76 STATISTIC(NumLiveSymbols, "Number of live symbols in index");
77
78 /// Limit on instruction count of imported functions.
79 static cl::opt<unsigned> ImportInstrLimit(
80 "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
81 cl::desc("Only import functions with less than N instructions"));
82
83 static cl::opt<int> ImportCutoff(
84 "import-cutoff", cl::init(-1), cl::Hidden, cl::value_desc("N"),
85 cl::desc("Only import first N functions if N>=0 (default -1)"));
86
87 static cl::opt<float>
88 ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7),
89 cl::Hidden, cl::value_desc("x"),
90 cl::desc("As we import functions, multiply the "
91 "`import-instr-limit` threshold by this factor "
92 "before processing newly imported functions"));
93
94 static cl::opt<float> ImportHotInstrFactor(
95 "import-hot-evolution-factor", cl::init(1.0), cl::Hidden,
96 cl::value_desc("x"),
97 cl::desc("As we import functions called from hot callsite, multiply the "
98 "`import-instr-limit` threshold by this factor "
99 "before processing newly imported functions"));
100
101 static cl::opt<float> ImportHotMultiplier(
102 "import-hot-multiplier", cl::init(10.0), cl::Hidden, cl::value_desc("x"),
103 cl::desc("Multiply the `import-instr-limit` threshold for hot callsites"));
104
105 static cl::opt<float> ImportCriticalMultiplier(
106 "import-critical-multiplier", cl::init(100.0), cl::Hidden,
107 cl::value_desc("x"),
108 cl::desc(
109 "Multiply the `import-instr-limit` threshold for critical callsites"));
110
111 // FIXME: This multiplier was not really tuned up.
112 static cl::opt<float> ImportColdMultiplier(
113 "import-cold-multiplier", cl::init(0), cl::Hidden, cl::value_desc("N"),
114 cl::desc("Multiply the `import-instr-limit` threshold for cold callsites"));
115
116 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden,
117 cl::desc("Print imported functions"));
118
119 static cl::opt<bool> PrintImportFailures(
120 "print-import-failures", cl::init(false), cl::Hidden,
121 cl::desc("Print information for functions rejected for importing"));
122
123 static cl::opt<bool> ComputeDead("compute-dead", cl::init(true), cl::Hidden,
124 cl::desc("Compute dead symbols"));
125
126 static cl::opt<bool> EnableImportMetadata(
127 "enable-import-metadata", cl::init(
128 #if !defined(NDEBUG)
129 true /*Enabled with asserts.*/
130 #else
131 false
132 #endif
133 ),
134 cl::Hidden, cl::desc("Enable import metadata like 'thinlto_src_module'"));
135
136 /// Summary file to use for function importing when using -function-import from
137 /// the command line.
138 static cl::opt<std::string>
139 SummaryFile("summary-file",
140 cl::desc("The summary file to use for function importing."));
141
142 /// Used when testing importing from distributed indexes via opt
143 // -function-import.
144 static cl::opt<bool>
145 ImportAllIndex("import-all-index",
146 cl::desc("Import all external functions in index."));
147
148 // Load lazily a module from \p FileName in \p Context.
loadFile(const std::string & FileName,LLVMContext & Context)149 static std::unique_ptr<Module> loadFile(const std::string &FileName,
150 LLVMContext &Context) {
151 SMDiagnostic Err;
152 LLVM_DEBUG(dbgs() << "Loading '" << FileName << "'\n");
153 // Metadata isn't loaded until functions are imported, to minimize
154 // the memory overhead.
155 std::unique_ptr<Module> Result =
156 getLazyIRFileModule(FileName, Err, Context,
157 /* ShouldLazyLoadMetadata = */ true);
158 if (!Result) {
159 Err.print("function-import", errs());
160 report_fatal_error("Abort");
161 }
162
163 return Result;
164 }
165
166 /// Given a list of possible callee implementation for a call site, select one
167 /// that fits the \p Threshold.
168 ///
169 /// FIXME: select "best" instead of first that fits. But what is "best"?
170 /// - The smallest: more likely to be inlined.
171 /// - The one with the least outgoing edges (already well optimized).
172 /// - One from a module already being imported from in order to reduce the
173 /// number of source modules parsed/linked.
174 /// - One that has PGO data attached.
175 /// - [insert you fancy metric here]
176 static const GlobalValueSummary *
selectCallee(const ModuleSummaryIndex & Index,ArrayRef<std::unique_ptr<GlobalValueSummary>> CalleeSummaryList,unsigned Threshold,StringRef CallerModulePath,FunctionImporter::ImportFailureReason & Reason,GlobalValue::GUID GUID)177 selectCallee(const ModuleSummaryIndex &Index,
178 ArrayRef<std::unique_ptr<GlobalValueSummary>> CalleeSummaryList,
179 unsigned Threshold, StringRef CallerModulePath,
180 FunctionImporter::ImportFailureReason &Reason,
181 GlobalValue::GUID GUID) {
182 Reason = FunctionImporter::ImportFailureReason::None;
183 auto It = llvm::find_if(
184 CalleeSummaryList,
185 [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
186 auto *GVSummary = SummaryPtr.get();
187 if (!Index.isGlobalValueLive(GVSummary)) {
188 Reason = FunctionImporter::ImportFailureReason::NotLive;
189 return false;
190 }
191
192 // For SamplePGO, in computeImportForFunction the OriginalId
193 // may have been used to locate the callee summary list (See
194 // comment there).
195 // The mapping from OriginalId to GUID may return a GUID
196 // that corresponds to a static variable. Filter it out here.
197 // This can happen when
198 // 1) There is a call to a library function which is not defined
199 // in the index.
200 // 2) There is a static variable with the OriginalGUID identical
201 // to the GUID of the library function in 1);
202 // When this happens, the logic for SamplePGO kicks in and
203 // the static variable in 2) will be found, which needs to be
204 // filtered out.
205 if (GVSummary->getSummaryKind() == GlobalValueSummary::GlobalVarKind) {
206 Reason = FunctionImporter::ImportFailureReason::GlobalVar;
207 return false;
208 }
209 if (GlobalValue::isInterposableLinkage(GVSummary->linkage())) {
210 Reason = FunctionImporter::ImportFailureReason::InterposableLinkage;
211 // There is no point in importing these, we can't inline them
212 return false;
213 }
214
215 auto *Summary = cast<FunctionSummary>(GVSummary->getBaseObject());
216
217 // If this is a local function, make sure we import the copy
218 // in the caller's module. The only time a local function can
219 // share an entry in the index is if there is a local with the same name
220 // in another module that had the same source file name (in a different
221 // directory), where each was compiled in their own directory so there
222 // was not distinguishing path.
223 // However, do the import from another module if there is only one
224 // entry in the list - in that case this must be a reference due
225 // to indirect call profile data, since a function pointer can point to
226 // a local in another module.
227 if (GlobalValue::isLocalLinkage(Summary->linkage()) &&
228 CalleeSummaryList.size() > 1 &&
229 Summary->modulePath() != CallerModulePath) {
230 Reason =
231 FunctionImporter::ImportFailureReason::LocalLinkageNotInModule;
232 return false;
233 }
234
235 if ((Summary->instCount() > Threshold) &&
236 !Summary->fflags().AlwaysInline) {
237 Reason = FunctionImporter::ImportFailureReason::TooLarge;
238 return false;
239 }
240
241 // Skip if it isn't legal to import (e.g. may reference unpromotable
242 // locals).
243 if (Summary->notEligibleToImport()) {
244 Reason = FunctionImporter::ImportFailureReason::NotEligible;
245 return false;
246 }
247
248 // Don't bother importing if we can't inline it anyway.
249 if (Summary->fflags().NoInline) {
250 Reason = FunctionImporter::ImportFailureReason::NoInline;
251 return false;
252 }
253
254 return true;
255 });
256 if (It == CalleeSummaryList.end())
257 return nullptr;
258
259 return cast<GlobalValueSummary>(It->get());
260 }
261
262 namespace {
263
264 using EdgeInfo = std::tuple<const FunctionSummary *, unsigned /* Threshold */,
265 GlobalValue::GUID>;
266
267 } // anonymous namespace
268
269 static ValueInfo
updateValueInfoForIndirectCalls(const ModuleSummaryIndex & Index,ValueInfo VI)270 updateValueInfoForIndirectCalls(const ModuleSummaryIndex &Index, ValueInfo VI) {
271 if (!VI.getSummaryList().empty())
272 return VI;
273 // For SamplePGO, the indirect call targets for local functions will
274 // have its original name annotated in profile. We try to find the
275 // corresponding PGOFuncName as the GUID.
276 // FIXME: Consider updating the edges in the graph after building
277 // it, rather than needing to perform this mapping on each walk.
278 auto GUID = Index.getGUIDFromOriginalID(VI.getGUID());
279 if (GUID == 0)
280 return ValueInfo();
281 return Index.getValueInfo(GUID);
282 }
283
computeImportForReferencedGlobals(const FunctionSummary & Summary,const ModuleSummaryIndex & Index,const GVSummaryMapTy & DefinedGVSummaries,FunctionImporter::ImportMapTy & ImportList,StringMap<FunctionImporter::ExportSetTy> * ExportLists)284 static void computeImportForReferencedGlobals(
285 const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
286 const GVSummaryMapTy &DefinedGVSummaries,
287 FunctionImporter::ImportMapTy &ImportList,
288 StringMap<FunctionImporter::ExportSetTy> *ExportLists) {
289 for (auto &VI : Summary.refs()) {
290 if (DefinedGVSummaries.count(VI.getGUID())) {
291 LLVM_DEBUG(
292 dbgs() << "Ref ignored! Target already in destination module.\n");
293 continue;
294 }
295
296 LLVM_DEBUG(dbgs() << " ref -> " << VI << "\n");
297
298 // If this is a local variable, make sure we import the copy
299 // in the caller's module. The only time a local variable can
300 // share an entry in the index is if there is a local with the same name
301 // in another module that had the same source file name (in a different
302 // directory), where each was compiled in their own directory so there
303 // was not distinguishing path.
304 auto LocalNotInModule = [&](const GlobalValueSummary *RefSummary) -> bool {
305 return GlobalValue::isLocalLinkage(RefSummary->linkage()) &&
306 RefSummary->modulePath() != Summary.modulePath();
307 };
308
309 auto MarkExported = [&](const ValueInfo &VI, const GlobalValueSummary *S) {
310 if (ExportLists)
311 (*ExportLists)[S->modulePath()].insert(VI);
312 };
313
314 for (auto &RefSummary : VI.getSummaryList())
315 if (isa<GlobalVarSummary>(RefSummary.get()) &&
316 Index.canImportGlobalVar(RefSummary.get(), /* AnalyzeRefs */ true) &&
317 !LocalNotInModule(RefSummary.get())) {
318 auto ILI = ImportList[RefSummary->modulePath()].insert(VI.getGUID());
319 // Only update stat if we haven't already imported this variable.
320 if (ILI.second)
321 NumImportedGlobalVarsThinLink++;
322 MarkExported(VI, RefSummary.get());
323 // Promote referenced functions and variables. We don't promote
324 // objects referenced by writeonly variable initializer, because
325 // we convert such variables initializers to "zeroinitializer".
326 // See processGlobalForThinLTO.
327 if (!Index.isWriteOnly(cast<GlobalVarSummary>(RefSummary.get())))
328 for (const auto &VI : RefSummary->refs())
329 for (const auto &RefFn : VI.getSummaryList())
330 MarkExported(VI, RefFn.get());
331 break;
332 }
333 }
334 }
335
336 static const char *
getFailureName(FunctionImporter::ImportFailureReason Reason)337 getFailureName(FunctionImporter::ImportFailureReason Reason) {
338 switch (Reason) {
339 case FunctionImporter::ImportFailureReason::None:
340 return "None";
341 case FunctionImporter::ImportFailureReason::GlobalVar:
342 return "GlobalVar";
343 case FunctionImporter::ImportFailureReason::NotLive:
344 return "NotLive";
345 case FunctionImporter::ImportFailureReason::TooLarge:
346 return "TooLarge";
347 case FunctionImporter::ImportFailureReason::InterposableLinkage:
348 return "InterposableLinkage";
349 case FunctionImporter::ImportFailureReason::LocalLinkageNotInModule:
350 return "LocalLinkageNotInModule";
351 case FunctionImporter::ImportFailureReason::NotEligible:
352 return "NotEligible";
353 case FunctionImporter::ImportFailureReason::NoInline:
354 return "NoInline";
355 }
356 llvm_unreachable("invalid reason");
357 }
358
359 /// Compute the list of functions to import for a given caller. Mark these
360 /// imported functions and the symbols they reference in their source module as
361 /// exported from their source module.
computeImportForFunction(const FunctionSummary & Summary,const ModuleSummaryIndex & Index,const unsigned Threshold,const GVSummaryMapTy & DefinedGVSummaries,SmallVectorImpl<EdgeInfo> & Worklist,FunctionImporter::ImportMapTy & ImportList,StringMap<FunctionImporter::ExportSetTy> * ExportLists,FunctionImporter::ImportThresholdsTy & ImportThresholds)362 static void computeImportForFunction(
363 const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
364 const unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries,
365 SmallVectorImpl<EdgeInfo> &Worklist,
366 FunctionImporter::ImportMapTy &ImportList,
367 StringMap<FunctionImporter::ExportSetTy> *ExportLists,
368 FunctionImporter::ImportThresholdsTy &ImportThresholds) {
369 computeImportForReferencedGlobals(Summary, Index, DefinedGVSummaries,
370 ImportList, ExportLists);
371 static int ImportCount = 0;
372 for (auto &Edge : Summary.calls()) {
373 ValueInfo VI = Edge.first;
374 LLVM_DEBUG(dbgs() << " edge -> " << VI << " Threshold:" << Threshold
375 << "\n");
376
377 if (ImportCutoff >= 0 && ImportCount >= ImportCutoff) {
378 LLVM_DEBUG(dbgs() << "ignored! import-cutoff value of " << ImportCutoff
379 << " reached.\n");
380 continue;
381 }
382
383 VI = updateValueInfoForIndirectCalls(Index, VI);
384 if (!VI)
385 continue;
386
387 if (DefinedGVSummaries.count(VI.getGUID())) {
388 LLVM_DEBUG(dbgs() << "ignored! Target already in destination module.\n");
389 continue;
390 }
391
392 auto GetBonusMultiplier = [](CalleeInfo::HotnessType Hotness) -> float {
393 if (Hotness == CalleeInfo::HotnessType::Hot)
394 return ImportHotMultiplier;
395 if (Hotness == CalleeInfo::HotnessType::Cold)
396 return ImportColdMultiplier;
397 if (Hotness == CalleeInfo::HotnessType::Critical)
398 return ImportCriticalMultiplier;
399 return 1.0;
400 };
401
402 const auto NewThreshold =
403 Threshold * GetBonusMultiplier(Edge.second.getHotness());
404
405 auto IT = ImportThresholds.insert(std::make_pair(
406 VI.getGUID(), std::make_tuple(NewThreshold, nullptr, nullptr)));
407 bool PreviouslyVisited = !IT.second;
408 auto &ProcessedThreshold = std::get<0>(IT.first->second);
409 auto &CalleeSummary = std::get<1>(IT.first->second);
410 auto &FailureInfo = std::get<2>(IT.first->second);
411
412 bool IsHotCallsite =
413 Edge.second.getHotness() == CalleeInfo::HotnessType::Hot;
414 bool IsCriticalCallsite =
415 Edge.second.getHotness() == CalleeInfo::HotnessType::Critical;
416
417 const FunctionSummary *ResolvedCalleeSummary = nullptr;
418 if (CalleeSummary) {
419 assert(PreviouslyVisited);
420 // Since the traversal of the call graph is DFS, we can revisit a function
421 // a second time with a higher threshold. In this case, it is added back
422 // to the worklist with the new threshold (so that its own callee chains
423 // can be considered with the higher threshold).
424 if (NewThreshold <= ProcessedThreshold) {
425 LLVM_DEBUG(
426 dbgs() << "ignored! Target was already imported with Threshold "
427 << ProcessedThreshold << "\n");
428 continue;
429 }
430 // Update with new larger threshold.
431 ProcessedThreshold = NewThreshold;
432 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
433 } else {
434 // If we already rejected importing a callee at the same or higher
435 // threshold, don't waste time calling selectCallee.
436 if (PreviouslyVisited && NewThreshold <= ProcessedThreshold) {
437 LLVM_DEBUG(
438 dbgs() << "ignored! Target was already rejected with Threshold "
439 << ProcessedThreshold << "\n");
440 if (PrintImportFailures) {
441 assert(FailureInfo &&
442 "Expected FailureInfo for previously rejected candidate");
443 FailureInfo->Attempts++;
444 }
445 continue;
446 }
447
448 FunctionImporter::ImportFailureReason Reason;
449 CalleeSummary = selectCallee(Index, VI.getSummaryList(), NewThreshold,
450 Summary.modulePath(), Reason, VI.getGUID());
451 if (!CalleeSummary) {
452 // Update with new larger threshold if this was a retry (otherwise
453 // we would have already inserted with NewThreshold above). Also
454 // update failure info if requested.
455 if (PreviouslyVisited) {
456 ProcessedThreshold = NewThreshold;
457 if (PrintImportFailures) {
458 assert(FailureInfo &&
459 "Expected FailureInfo for previously rejected candidate");
460 FailureInfo->Reason = Reason;
461 FailureInfo->Attempts++;
462 FailureInfo->MaxHotness =
463 std::max(FailureInfo->MaxHotness, Edge.second.getHotness());
464 }
465 } else if (PrintImportFailures) {
466 assert(!FailureInfo &&
467 "Expected no FailureInfo for newly rejected candidate");
468 FailureInfo = std::make_unique<FunctionImporter::ImportFailureInfo>(
469 VI, Edge.second.getHotness(), Reason, 1);
470 }
471 LLVM_DEBUG(
472 dbgs() << "ignored! No qualifying callee with summary found.\n");
473 continue;
474 }
475
476 // "Resolve" the summary
477 CalleeSummary = CalleeSummary->getBaseObject();
478 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
479
480 assert((ResolvedCalleeSummary->fflags().AlwaysInline ||
481 (ResolvedCalleeSummary->instCount() <= NewThreshold)) &&
482 "selectCallee() didn't honor the threshold");
483
484 auto ExportModulePath = ResolvedCalleeSummary->modulePath();
485 auto ILI = ImportList[ExportModulePath].insert(VI.getGUID());
486 // We previously decided to import this GUID definition if it was already
487 // inserted in the set of imports from the exporting module.
488 bool PreviouslyImported = !ILI.second;
489 if (!PreviouslyImported) {
490 NumImportedFunctionsThinLink++;
491 if (IsHotCallsite)
492 NumImportedHotFunctionsThinLink++;
493 if (IsCriticalCallsite)
494 NumImportedCriticalFunctionsThinLink++;
495 }
496
497 // Make exports in the source module.
498 if (ExportLists) {
499 auto &ExportList = (*ExportLists)[ExportModulePath];
500 ExportList.insert(VI);
501 if (!PreviouslyImported) {
502 // This is the first time this function was exported from its source
503 // module, so mark all functions and globals it references as exported
504 // to the outside if they are defined in the same source module.
505 // For efficiency, we unconditionally add all the referenced GUIDs
506 // to the ExportList for this module, and will prune out any not
507 // defined in the module later in a single pass.
508 for (auto &Edge : ResolvedCalleeSummary->calls())
509 ExportList.insert(Edge.first);
510
511 for (auto &Ref : ResolvedCalleeSummary->refs())
512 ExportList.insert(Ref);
513 }
514 }
515 }
516
517 auto GetAdjustedThreshold = [](unsigned Threshold, bool IsHotCallsite) {
518 // Adjust the threshold for next level of imported functions.
519 // The threshold is different for hot callsites because we can then
520 // inline chains of hot calls.
521 if (IsHotCallsite)
522 return Threshold * ImportHotInstrFactor;
523 return Threshold * ImportInstrFactor;
524 };
525
526 const auto AdjThreshold = GetAdjustedThreshold(Threshold, IsHotCallsite);
527
528 ImportCount++;
529
530 // Insert the newly imported function to the worklist.
531 Worklist.emplace_back(ResolvedCalleeSummary, AdjThreshold, VI.getGUID());
532 }
533 }
534
535 /// Given the list of globals defined in a module, compute the list of imports
536 /// as well as the list of "exports", i.e. the list of symbols referenced from
537 /// another module (that may require promotion).
ComputeImportForModule(const GVSummaryMapTy & DefinedGVSummaries,const ModuleSummaryIndex & Index,StringRef ModName,FunctionImporter::ImportMapTy & ImportList,StringMap<FunctionImporter::ExportSetTy> * ExportLists=nullptr)538 static void ComputeImportForModule(
539 const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
540 StringRef ModName, FunctionImporter::ImportMapTy &ImportList,
541 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
542 // Worklist contains the list of function imported in this module, for which
543 // we will analyse the callees and may import further down the callgraph.
544 SmallVector<EdgeInfo, 128> Worklist;
545 FunctionImporter::ImportThresholdsTy ImportThresholds;
546
547 // Populate the worklist with the import for the functions in the current
548 // module
549 for (auto &GVSummary : DefinedGVSummaries) {
550 #ifndef NDEBUG
551 // FIXME: Change the GVSummaryMapTy to hold ValueInfo instead of GUID
552 // so this map look up (and possibly others) can be avoided.
553 auto VI = Index.getValueInfo(GVSummary.first);
554 #endif
555 if (!Index.isGlobalValueLive(GVSummary.second)) {
556 LLVM_DEBUG(dbgs() << "Ignores Dead GUID: " << VI << "\n");
557 continue;
558 }
559 auto *FuncSummary =
560 dyn_cast<FunctionSummary>(GVSummary.second->getBaseObject());
561 if (!FuncSummary)
562 // Skip import for global variables
563 continue;
564 LLVM_DEBUG(dbgs() << "Initialize import for " << VI << "\n");
565 computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
566 DefinedGVSummaries, Worklist, ImportList,
567 ExportLists, ImportThresholds);
568 }
569
570 // Process the newly imported functions and add callees to the worklist.
571 while (!Worklist.empty()) {
572 auto FuncInfo = Worklist.pop_back_val();
573 auto *Summary = std::get<0>(FuncInfo);
574 auto Threshold = std::get<1>(FuncInfo);
575
576 computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries,
577 Worklist, ImportList, ExportLists,
578 ImportThresholds);
579 }
580
581 // Print stats about functions considered but rejected for importing
582 // when requested.
583 if (PrintImportFailures) {
584 dbgs() << "Missed imports into module " << ModName << "\n";
585 for (auto &I : ImportThresholds) {
586 auto &ProcessedThreshold = std::get<0>(I.second);
587 auto &CalleeSummary = std::get<1>(I.second);
588 auto &FailureInfo = std::get<2>(I.second);
589 if (CalleeSummary)
590 continue; // We are going to import.
591 assert(FailureInfo);
592 FunctionSummary *FS = nullptr;
593 if (!FailureInfo->VI.getSummaryList().empty())
594 FS = dyn_cast<FunctionSummary>(
595 FailureInfo->VI.getSummaryList()[0]->getBaseObject());
596 dbgs() << FailureInfo->VI
597 << ": Reason = " << getFailureName(FailureInfo->Reason)
598 << ", Threshold = " << ProcessedThreshold
599 << ", Size = " << (FS ? (int)FS->instCount() : -1)
600 << ", MaxHotness = " << getHotnessName(FailureInfo->MaxHotness)
601 << ", Attempts = " << FailureInfo->Attempts << "\n";
602 }
603 }
604 }
605
606 #ifndef NDEBUG
isGlobalVarSummary(const ModuleSummaryIndex & Index,ValueInfo VI)607 static bool isGlobalVarSummary(const ModuleSummaryIndex &Index, ValueInfo VI) {
608 auto SL = VI.getSummaryList();
609 return SL.empty()
610 ? false
611 : SL[0]->getSummaryKind() == GlobalValueSummary::GlobalVarKind;
612 }
613
isGlobalVarSummary(const ModuleSummaryIndex & Index,GlobalValue::GUID G)614 static bool isGlobalVarSummary(const ModuleSummaryIndex &Index,
615 GlobalValue::GUID G) {
616 if (const auto &VI = Index.getValueInfo(G))
617 return isGlobalVarSummary(Index, VI);
618 return false;
619 }
620
621 template <class T>
numGlobalVarSummaries(const ModuleSummaryIndex & Index,T & Cont)622 static unsigned numGlobalVarSummaries(const ModuleSummaryIndex &Index,
623 T &Cont) {
624 unsigned NumGVS = 0;
625 for (auto &V : Cont)
626 if (isGlobalVarSummary(Index, V))
627 ++NumGVS;
628 return NumGVS;
629 }
630 #endif
631
632 #ifndef NDEBUG
633 static bool
checkVariableImport(const ModuleSummaryIndex & Index,StringMap<FunctionImporter::ImportMapTy> & ImportLists,StringMap<FunctionImporter::ExportSetTy> & ExportLists)634 checkVariableImport(const ModuleSummaryIndex &Index,
635 StringMap<FunctionImporter::ImportMapTy> &ImportLists,
636 StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
637
638 DenseSet<GlobalValue::GUID> FlattenedImports;
639
640 for (auto &ImportPerModule : ImportLists)
641 for (auto &ExportPerModule : ImportPerModule.second)
642 FlattenedImports.insert(ExportPerModule.second.begin(),
643 ExportPerModule.second.end());
644
645 // Checks that all GUIDs of read/writeonly vars we see in export lists
646 // are also in the import lists. Otherwise we my face linker undefs,
647 // because readonly and writeonly vars are internalized in their
648 // source modules.
649 auto IsReadOrWriteOnlyVar = [&](StringRef ModulePath, const ValueInfo &VI) {
650 auto *GVS = dyn_cast_or_null<GlobalVarSummary>(
651 Index.findSummaryInModule(VI, ModulePath));
652 return GVS && (Index.isReadOnly(GVS) || Index.isWriteOnly(GVS));
653 };
654
655 for (auto &ExportPerModule : ExportLists)
656 for (auto &VI : ExportPerModule.second)
657 if (!FlattenedImports.count(VI.getGUID()) &&
658 IsReadOrWriteOnlyVar(ExportPerModule.first(), VI))
659 return false;
660
661 return true;
662 }
663 #endif
664
665 /// Compute all the import and export for every module using the Index.
ComputeCrossModuleImport(const ModuleSummaryIndex & Index,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,StringMap<FunctionImporter::ImportMapTy> & ImportLists,StringMap<FunctionImporter::ExportSetTy> & ExportLists)666 void llvm::ComputeCrossModuleImport(
667 const ModuleSummaryIndex &Index,
668 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
669 StringMap<FunctionImporter::ImportMapTy> &ImportLists,
670 StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
671 // For each module that has function defined, compute the import/export lists.
672 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
673 auto &ImportList = ImportLists[DefinedGVSummaries.first()];
674 LLVM_DEBUG(dbgs() << "Computing import for Module '"
675 << DefinedGVSummaries.first() << "'\n");
676 ComputeImportForModule(DefinedGVSummaries.second, Index,
677 DefinedGVSummaries.first(), ImportList,
678 &ExportLists);
679 }
680
681 // When computing imports we added all GUIDs referenced by anything
682 // imported from the module to its ExportList. Now we prune each ExportList
683 // of any not defined in that module. This is more efficient than checking
684 // while computing imports because some of the summary lists may be long
685 // due to linkonce (comdat) copies.
686 for (auto &ELI : ExportLists) {
687 const auto &DefinedGVSummaries =
688 ModuleToDefinedGVSummaries.lookup(ELI.first());
689 for (auto EI = ELI.second.begin(); EI != ELI.second.end();) {
690 if (!DefinedGVSummaries.count(EI->getGUID()))
691 ELI.second.erase(EI++);
692 else
693 ++EI;
694 }
695 }
696
697 assert(checkVariableImport(Index, ImportLists, ExportLists));
698 #ifndef NDEBUG
699 LLVM_DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
700 << " modules:\n");
701 for (auto &ModuleImports : ImportLists) {
702 auto ModName = ModuleImports.first();
703 auto &Exports = ExportLists[ModName];
704 unsigned NumGVS = numGlobalVarSummaries(Index, Exports);
705 LLVM_DEBUG(dbgs() << "* Module " << ModName << " exports "
706 << Exports.size() - NumGVS << " functions and " << NumGVS
707 << " vars. Imports from " << ModuleImports.second.size()
708 << " modules.\n");
709 for (auto &Src : ModuleImports.second) {
710 auto SrcModName = Src.first();
711 unsigned NumGVSPerMod = numGlobalVarSummaries(Index, Src.second);
712 LLVM_DEBUG(dbgs() << " - " << Src.second.size() - NumGVSPerMod
713 << " functions imported from " << SrcModName << "\n");
714 LLVM_DEBUG(dbgs() << " - " << NumGVSPerMod
715 << " global vars imported from " << SrcModName << "\n");
716 }
717 }
718 #endif
719 }
720
721 #ifndef NDEBUG
dumpImportListForModule(const ModuleSummaryIndex & Index,StringRef ModulePath,FunctionImporter::ImportMapTy & ImportList)722 static void dumpImportListForModule(const ModuleSummaryIndex &Index,
723 StringRef ModulePath,
724 FunctionImporter::ImportMapTy &ImportList) {
725 LLVM_DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
726 << ImportList.size() << " modules.\n");
727 for (auto &Src : ImportList) {
728 auto SrcModName = Src.first();
729 unsigned NumGVSPerMod = numGlobalVarSummaries(Index, Src.second);
730 LLVM_DEBUG(dbgs() << " - " << Src.second.size() - NumGVSPerMod
731 << " functions imported from " << SrcModName << "\n");
732 LLVM_DEBUG(dbgs() << " - " << NumGVSPerMod << " vars imported from "
733 << SrcModName << "\n");
734 }
735 }
736 #endif
737
738 /// Compute all the imports for the given module in the Index.
ComputeCrossModuleImportForModule(StringRef ModulePath,const ModuleSummaryIndex & Index,FunctionImporter::ImportMapTy & ImportList)739 void llvm::ComputeCrossModuleImportForModule(
740 StringRef ModulePath, const ModuleSummaryIndex &Index,
741 FunctionImporter::ImportMapTy &ImportList) {
742 // Collect the list of functions this module defines.
743 // GUID -> Summary
744 GVSummaryMapTy FunctionSummaryMap;
745 Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
746
747 // Compute the import list for this module.
748 LLVM_DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
749 ComputeImportForModule(FunctionSummaryMap, Index, ModulePath, ImportList);
750
751 #ifndef NDEBUG
752 dumpImportListForModule(Index, ModulePath, ImportList);
753 #endif
754 }
755
756 // Mark all external summaries in Index for import into the given module.
757 // Used for distributed builds using a distributed index.
ComputeCrossModuleImportForModuleFromIndex(StringRef ModulePath,const ModuleSummaryIndex & Index,FunctionImporter::ImportMapTy & ImportList)758 void llvm::ComputeCrossModuleImportForModuleFromIndex(
759 StringRef ModulePath, const ModuleSummaryIndex &Index,
760 FunctionImporter::ImportMapTy &ImportList) {
761 for (auto &GlobalList : Index) {
762 // Ignore entries for undefined references.
763 if (GlobalList.second.SummaryList.empty())
764 continue;
765
766 auto GUID = GlobalList.first;
767 assert(GlobalList.second.SummaryList.size() == 1 &&
768 "Expected individual combined index to have one summary per GUID");
769 auto &Summary = GlobalList.second.SummaryList[0];
770 // Skip the summaries for the importing module. These are included to
771 // e.g. record required linkage changes.
772 if (Summary->modulePath() == ModulePath)
773 continue;
774 // Add an entry to provoke importing by thinBackend.
775 ImportList[Summary->modulePath()].insert(GUID);
776 }
777 #ifndef NDEBUG
778 dumpImportListForModule(Index, ModulePath, ImportList);
779 #endif
780 }
781
computeDeadSymbols(ModuleSummaryIndex & Index,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols,function_ref<PrevailingType (GlobalValue::GUID)> isPrevailing)782 void llvm::computeDeadSymbols(
783 ModuleSummaryIndex &Index,
784 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
785 function_ref<PrevailingType(GlobalValue::GUID)> isPrevailing) {
786 assert(!Index.withGlobalValueDeadStripping());
787 if (!ComputeDead)
788 return;
789 if (GUIDPreservedSymbols.empty())
790 // Don't do anything when nothing is live, this is friendly with tests.
791 return;
792 unsigned LiveSymbols = 0;
793 SmallVector<ValueInfo, 128> Worklist;
794 Worklist.reserve(GUIDPreservedSymbols.size() * 2);
795 for (auto GUID : GUIDPreservedSymbols) {
796 ValueInfo VI = Index.getValueInfo(GUID);
797 if (!VI)
798 continue;
799 for (auto &S : VI.getSummaryList())
800 S->setLive(true);
801 }
802
803 // Add values flagged in the index as live roots to the worklist.
804 for (const auto &Entry : Index) {
805 auto VI = Index.getValueInfo(Entry);
806 for (auto &S : Entry.second.SummaryList)
807 if (S->isLive()) {
808 LLVM_DEBUG(dbgs() << "Live root: " << VI << "\n");
809 Worklist.push_back(VI);
810 ++LiveSymbols;
811 break;
812 }
813 }
814
815 // Make value live and add it to the worklist if it was not live before.
816 auto visit = [&](ValueInfo VI, bool IsAliasee) {
817 // FIXME: If we knew which edges were created for indirect call profiles,
818 // we could skip them here. Any that are live should be reached via
819 // other edges, e.g. reference edges. Otherwise, using a profile collected
820 // on a slightly different binary might provoke preserving, importing
821 // and ultimately promoting calls to functions not linked into this
822 // binary, which increases the binary size unnecessarily. Note that
823 // if this code changes, the importer needs to change so that edges
824 // to functions marked dead are skipped.
825 VI = updateValueInfoForIndirectCalls(Index, VI);
826 if (!VI)
827 return;
828
829 if (llvm::any_of(VI.getSummaryList(),
830 [](const std::unique_ptr<llvm::GlobalValueSummary> &S) {
831 return S->isLive();
832 }))
833 return;
834
835 // We only keep live symbols that are known to be non-prevailing if any are
836 // available_externally, linkonceodr, weakodr. Those symbols are discarded
837 // later in the EliminateAvailableExternally pass and setting them to
838 // not-live could break downstreams users of liveness information (PR36483)
839 // or limit optimization opportunities.
840 if (isPrevailing(VI.getGUID()) == PrevailingType::No) {
841 bool KeepAliveLinkage = false;
842 bool Interposable = false;
843 for (auto &S : VI.getSummaryList()) {
844 if (S->linkage() == GlobalValue::AvailableExternallyLinkage ||
845 S->linkage() == GlobalValue::WeakODRLinkage ||
846 S->linkage() == GlobalValue::LinkOnceODRLinkage)
847 KeepAliveLinkage = true;
848 else if (GlobalValue::isInterposableLinkage(S->linkage()))
849 Interposable = true;
850 }
851
852 if (!IsAliasee) {
853 if (!KeepAliveLinkage)
854 return;
855
856 if (Interposable)
857 report_fatal_error(
858 "Interposable and available_externally/linkonce_odr/weak_odr "
859 "symbol");
860 }
861 }
862
863 for (auto &S : VI.getSummaryList())
864 S->setLive(true);
865 ++LiveSymbols;
866 Worklist.push_back(VI);
867 };
868
869 while (!Worklist.empty()) {
870 auto VI = Worklist.pop_back_val();
871 for (auto &Summary : VI.getSummaryList()) {
872 if (auto *AS = dyn_cast<AliasSummary>(Summary.get())) {
873 // If this is an alias, visit the aliasee VI to ensure that all copies
874 // are marked live and it is added to the worklist for further
875 // processing of its references.
876 visit(AS->getAliaseeVI(), true);
877 continue;
878 }
879
880 Summary->setLive(true);
881 for (auto Ref : Summary->refs())
882 visit(Ref, false);
883 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get()))
884 for (auto Call : FS->calls())
885 visit(Call.first, false);
886 }
887 }
888 Index.setWithGlobalValueDeadStripping();
889
890 unsigned DeadSymbols = Index.size() - LiveSymbols;
891 LLVM_DEBUG(dbgs() << LiveSymbols << " symbols Live, and " << DeadSymbols
892 << " symbols Dead \n");
893 NumDeadSymbols += DeadSymbols;
894 NumLiveSymbols += LiveSymbols;
895 }
896
897 // Compute dead symbols and propagate constants in combined index.
computeDeadSymbolsWithConstProp(ModuleSummaryIndex & Index,const DenseSet<GlobalValue::GUID> & GUIDPreservedSymbols,function_ref<PrevailingType (GlobalValue::GUID)> isPrevailing,bool ImportEnabled)898 void llvm::computeDeadSymbolsWithConstProp(
899 ModuleSummaryIndex &Index,
900 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
901 function_ref<PrevailingType(GlobalValue::GUID)> isPrevailing,
902 bool ImportEnabled) {
903 computeDeadSymbols(Index, GUIDPreservedSymbols, isPrevailing);
904 if (ImportEnabled)
905 Index.propagateAttributes(GUIDPreservedSymbols);
906 }
907
908 /// Compute the set of summaries needed for a ThinLTO backend compilation of
909 /// \p ModulePath.
gatherImportedSummariesForModule(StringRef ModulePath,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,const FunctionImporter::ImportMapTy & ImportList,std::map<std::string,GVSummaryMapTy> & ModuleToSummariesForIndex)910 void llvm::gatherImportedSummariesForModule(
911 StringRef ModulePath,
912 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
913 const FunctionImporter::ImportMapTy &ImportList,
914 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
915 // Include all summaries from the importing module.
916 ModuleToSummariesForIndex[ModulePath] =
917 ModuleToDefinedGVSummaries.lookup(ModulePath);
918 // Include summaries for imports.
919 for (auto &ILI : ImportList) {
920 auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()];
921 const auto &DefinedGVSummaries =
922 ModuleToDefinedGVSummaries.lookup(ILI.first());
923 for (auto &GI : ILI.second) {
924 const auto &DS = DefinedGVSummaries.find(GI);
925 assert(DS != DefinedGVSummaries.end() &&
926 "Expected a defined summary for imported global value");
927 SummariesForIndex[GI] = DS->second;
928 }
929 }
930 }
931
932 /// Emit the files \p ModulePath will import from into \p OutputFilename.
EmitImportsFiles(StringRef ModulePath,StringRef OutputFilename,const std::map<std::string,GVSummaryMapTy> & ModuleToSummariesForIndex)933 std::error_code llvm::EmitImportsFiles(
934 StringRef ModulePath, StringRef OutputFilename,
935 const std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
936 std::error_code EC;
937 raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::OF_None);
938 if (EC)
939 return EC;
940 for (auto &ILI : ModuleToSummariesForIndex)
941 // The ModuleToSummariesForIndex map includes an entry for the current
942 // Module (needed for writing out the index files). We don't want to
943 // include it in the imports file, however, so filter it out.
944 if (ILI.first != ModulePath)
945 ImportsOS << ILI.first << "\n";
946 return std::error_code();
947 }
948
convertToDeclaration(GlobalValue & GV)949 bool llvm::convertToDeclaration(GlobalValue &GV) {
950 LLVM_DEBUG(dbgs() << "Converting to a declaration: `" << GV.getName()
951 << "\n");
952 if (Function *F = dyn_cast<Function>(&GV)) {
953 F->deleteBody();
954 F->clearMetadata();
955 F->setComdat(nullptr);
956 } else if (GlobalVariable *V = dyn_cast<GlobalVariable>(&GV)) {
957 V->setInitializer(nullptr);
958 V->setLinkage(GlobalValue::ExternalLinkage);
959 V->clearMetadata();
960 V->setComdat(nullptr);
961 } else {
962 GlobalValue *NewGV;
963 if (GV.getValueType()->isFunctionTy())
964 NewGV =
965 Function::Create(cast<FunctionType>(GV.getValueType()),
966 GlobalValue::ExternalLinkage, GV.getAddressSpace(),
967 "", GV.getParent());
968 else
969 NewGV =
970 new GlobalVariable(*GV.getParent(), GV.getValueType(),
971 /*isConstant*/ false, GlobalValue::ExternalLinkage,
972 /*init*/ nullptr, "",
973 /*insertbefore*/ nullptr, GV.getThreadLocalMode(),
974 GV.getType()->getAddressSpace());
975 NewGV->takeName(&GV);
976 GV.replaceAllUsesWith(NewGV);
977 return false;
978 }
979 return true;
980 }
981
982 /// Fixup prevailing symbol linkages in \p TheModule based on summary analysis.
thinLTOResolvePrevailingInModule(Module & TheModule,const GVSummaryMapTy & DefinedGlobals)983 void llvm::thinLTOResolvePrevailingInModule(
984 Module &TheModule, const GVSummaryMapTy &DefinedGlobals) {
985 auto updateLinkage = [&](GlobalValue &GV) {
986 // See if the global summary analysis computed a new resolved linkage.
987 const auto &GS = DefinedGlobals.find(GV.getGUID());
988 if (GS == DefinedGlobals.end())
989 return;
990 auto NewLinkage = GS->second->linkage();
991 if (NewLinkage == GV.getLinkage())
992 return;
993 if (GlobalValue::isLocalLinkage(GV.getLinkage()) ||
994 // Don't internalize anything here, because the code below
995 // lacks necessary correctness checks. Leave this job to
996 // LLVM 'internalize' pass.
997 GlobalValue::isLocalLinkage(NewLinkage) ||
998 // In case it was dead and already converted to declaration.
999 GV.isDeclaration())
1000 return;
1001
1002 // Check for a non-prevailing def that has interposable linkage
1003 // (e.g. non-odr weak or linkonce). In that case we can't simply
1004 // convert to available_externally, since it would lose the
1005 // interposable property and possibly get inlined. Simply drop
1006 // the definition in that case.
1007 if (GlobalValue::isAvailableExternallyLinkage(NewLinkage) &&
1008 GlobalValue::isInterposableLinkage(GV.getLinkage())) {
1009 if (!convertToDeclaration(GV))
1010 // FIXME: Change this to collect replaced GVs and later erase
1011 // them from the parent module once thinLTOResolvePrevailingGUID is
1012 // changed to enable this for aliases.
1013 llvm_unreachable("Expected GV to be converted");
1014 } else {
1015 // If all copies of the original symbol had global unnamed addr and
1016 // linkonce_odr linkage, it should be an auto hide symbol. In that case
1017 // the thin link would have marked it as CanAutoHide. Add hidden visibility
1018 // to the symbol to preserve the property.
1019 if (NewLinkage == GlobalValue::WeakODRLinkage &&
1020 GS->second->canAutoHide()) {
1021 assert(GV.hasLinkOnceODRLinkage() && GV.hasGlobalUnnamedAddr());
1022 GV.setVisibility(GlobalValue::HiddenVisibility);
1023 }
1024
1025 LLVM_DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName()
1026 << "` from " << GV.getLinkage() << " to " << NewLinkage
1027 << "\n");
1028 GV.setLinkage(NewLinkage);
1029 }
1030 // Remove declarations from comdats, including available_externally
1031 // as this is a declaration for the linker, and will be dropped eventually.
1032 // It is illegal for comdats to contain declarations.
1033 auto *GO = dyn_cast_or_null<GlobalObject>(&GV);
1034 if (GO && GO->isDeclarationForLinker() && GO->hasComdat())
1035 GO->setComdat(nullptr);
1036 };
1037
1038 // Process functions and global now
1039 for (auto &GV : TheModule)
1040 updateLinkage(GV);
1041 for (auto &GV : TheModule.globals())
1042 updateLinkage(GV);
1043 for (auto &GV : TheModule.aliases())
1044 updateLinkage(GV);
1045 }
1046
1047 /// Run internalization on \p TheModule based on symmary analysis.
thinLTOInternalizeModule(Module & TheModule,const GVSummaryMapTy & DefinedGlobals)1048 void llvm::thinLTOInternalizeModule(Module &TheModule,
1049 const GVSummaryMapTy &DefinedGlobals) {
1050 // Declare a callback for the internalize pass that will ask for every
1051 // candidate GlobalValue if it can be internalized or not.
1052 auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
1053 // Lookup the linkage recorded in the summaries during global analysis.
1054 auto GS = DefinedGlobals.find(GV.getGUID());
1055 if (GS == DefinedGlobals.end()) {
1056 // Must have been promoted (possibly conservatively). Find original
1057 // name so that we can access the correct summary and see if it can
1058 // be internalized again.
1059 // FIXME: Eventually we should control promotion instead of promoting
1060 // and internalizing again.
1061 StringRef OrigName =
1062 ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
1063 std::string OrigId = GlobalValue::getGlobalIdentifier(
1064 OrigName, GlobalValue::InternalLinkage,
1065 TheModule.getSourceFileName());
1066 GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
1067 if (GS == DefinedGlobals.end()) {
1068 // Also check the original non-promoted non-globalized name. In some
1069 // cases a preempted weak value is linked in as a local copy because
1070 // it is referenced by an alias (IRLinker::linkGlobalValueProto).
1071 // In that case, since it was originally not a local value, it was
1072 // recorded in the index using the original name.
1073 // FIXME: This may not be needed once PR27866 is fixed.
1074 GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
1075 assert(GS != DefinedGlobals.end());
1076 }
1077 }
1078 return !GlobalValue::isLocalLinkage(GS->second->linkage());
1079 };
1080
1081 // FIXME: See if we can just internalize directly here via linkage changes
1082 // based on the index, rather than invoking internalizeModule.
1083 internalizeModule(TheModule, MustPreserveGV);
1084 }
1085
1086 /// Make alias a clone of its aliasee.
replaceAliasWithAliasee(Module * SrcModule,GlobalAlias * GA)1087 static Function *replaceAliasWithAliasee(Module *SrcModule, GlobalAlias *GA) {
1088 Function *Fn = cast<Function>(GA->getBaseObject());
1089
1090 ValueToValueMapTy VMap;
1091 Function *NewFn = CloneFunction(Fn, VMap);
1092 // Clone should use the original alias's linkage, visibility and name, and we
1093 // ensure all uses of alias instead use the new clone (casted if necessary).
1094 NewFn->setLinkage(GA->getLinkage());
1095 NewFn->setVisibility(GA->getVisibility());
1096 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewFn, GA->getType()));
1097 NewFn->takeName(GA);
1098 return NewFn;
1099 }
1100
1101 // Internalize values that we marked with specific attribute
1102 // in processGlobalForThinLTO.
internalizeGVsAfterImport(Module & M)1103 static void internalizeGVsAfterImport(Module &M) {
1104 for (auto &GV : M.globals())
1105 // Skip GVs which have been converted to declarations
1106 // by dropDeadSymbols.
1107 if (!GV.isDeclaration() && GV.hasAttribute("thinlto-internalize")) {
1108 GV.setLinkage(GlobalValue::InternalLinkage);
1109 GV.setVisibility(GlobalValue::DefaultVisibility);
1110 }
1111 }
1112
1113 // Automatically import functions in Module \p DestModule based on the summaries
1114 // index.
importFunctions(Module & DestModule,const FunctionImporter::ImportMapTy & ImportList)1115 Expected<bool> FunctionImporter::importFunctions(
1116 Module &DestModule, const FunctionImporter::ImportMapTy &ImportList) {
1117 LLVM_DEBUG(dbgs() << "Starting import for Module "
1118 << DestModule.getModuleIdentifier() << "\n");
1119 unsigned ImportedCount = 0, ImportedGVCount = 0;
1120
1121 IRMover Mover(DestModule);
1122 // Do the actual import of functions now, one Module at a time
1123 std::set<StringRef> ModuleNameOrderedList;
1124 for (auto &FunctionsToImportPerModule : ImportList) {
1125 ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
1126 }
1127 for (auto &Name : ModuleNameOrderedList) {
1128 // Get the module for the import
1129 const auto &FunctionsToImportPerModule = ImportList.find(Name);
1130 assert(FunctionsToImportPerModule != ImportList.end());
1131 Expected<std::unique_ptr<Module>> SrcModuleOrErr = ModuleLoader(Name);
1132 if (!SrcModuleOrErr)
1133 return SrcModuleOrErr.takeError();
1134 std::unique_ptr<Module> SrcModule = std::move(*SrcModuleOrErr);
1135 assert(&DestModule.getContext() == &SrcModule->getContext() &&
1136 "Context mismatch");
1137
1138 // If modules were created with lazy metadata loading, materialize it
1139 // now, before linking it (otherwise this will be a noop).
1140 if (Error Err = SrcModule->materializeMetadata())
1141 return std::move(Err);
1142
1143 auto &ImportGUIDs = FunctionsToImportPerModule->second;
1144 // Find the globals to import
1145 SetVector<GlobalValue *> GlobalsToImport;
1146 for (Function &F : *SrcModule) {
1147 if (!F.hasName())
1148 continue;
1149 auto GUID = F.getGUID();
1150 auto Import = ImportGUIDs.count(GUID);
1151 LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function "
1152 << GUID << " " << F.getName() << " from "
1153 << SrcModule->getSourceFileName() << "\n");
1154 if (Import) {
1155 if (Error Err = F.materialize())
1156 return std::move(Err);
1157 if (EnableImportMetadata) {
1158 // Add 'thinlto_src_module' metadata for statistics and debugging.
1159 F.setMetadata(
1160 "thinlto_src_module",
1161 MDNode::get(DestModule.getContext(),
1162 {MDString::get(DestModule.getContext(),
1163 SrcModule->getSourceFileName())}));
1164 }
1165 GlobalsToImport.insert(&F);
1166 }
1167 }
1168 for (GlobalVariable &GV : SrcModule->globals()) {
1169 if (!GV.hasName())
1170 continue;
1171 auto GUID = GV.getGUID();
1172 auto Import = ImportGUIDs.count(GUID);
1173 LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global "
1174 << GUID << " " << GV.getName() << " from "
1175 << SrcModule->getSourceFileName() << "\n");
1176 if (Import) {
1177 if (Error Err = GV.materialize())
1178 return std::move(Err);
1179 ImportedGVCount += GlobalsToImport.insert(&GV);
1180 }
1181 }
1182 for (GlobalAlias &GA : SrcModule->aliases()) {
1183 if (!GA.hasName())
1184 continue;
1185 auto GUID = GA.getGUID();
1186 auto Import = ImportGUIDs.count(GUID);
1187 LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias "
1188 << GUID << " " << GA.getName() << " from "
1189 << SrcModule->getSourceFileName() << "\n");
1190 if (Import) {
1191 if (Error Err = GA.materialize())
1192 return std::move(Err);
1193 // Import alias as a copy of its aliasee.
1194 GlobalObject *Base = GA.getBaseObject();
1195 if (Error Err = Base->materialize())
1196 return std::move(Err);
1197 auto *Fn = replaceAliasWithAliasee(SrcModule.get(), &GA);
1198 LLVM_DEBUG(dbgs() << "Is importing aliasee fn " << Base->getGUID()
1199 << " " << Base->getName() << " from "
1200 << SrcModule->getSourceFileName() << "\n");
1201 if (EnableImportMetadata) {
1202 // Add 'thinlto_src_module' metadata for statistics and debugging.
1203 Fn->setMetadata(
1204 "thinlto_src_module",
1205 MDNode::get(DestModule.getContext(),
1206 {MDString::get(DestModule.getContext(),
1207 SrcModule->getSourceFileName())}));
1208 }
1209 GlobalsToImport.insert(Fn);
1210 }
1211 }
1212
1213 // Upgrade debug info after we're done materializing all the globals and we
1214 // have loaded all the required metadata!
1215 UpgradeDebugInfo(*SrcModule);
1216
1217 // Link in the specified functions.
1218 if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport))
1219 return true;
1220
1221 if (PrintImports) {
1222 for (const auto *GV : GlobalsToImport)
1223 dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
1224 << " from " << SrcModule->getSourceFileName() << "\n";
1225 }
1226
1227 if (Mover.move(std::move(SrcModule), GlobalsToImport.getArrayRef(),
1228 [](GlobalValue &, IRMover::ValueAdder) {},
1229 /*IsPerformingImport=*/true))
1230 report_fatal_error("Function Import: link error");
1231
1232 ImportedCount += GlobalsToImport.size();
1233 NumImportedModules++;
1234 }
1235
1236 internalizeGVsAfterImport(DestModule);
1237
1238 NumImportedFunctions += (ImportedCount - ImportedGVCount);
1239 NumImportedGlobalVars += ImportedGVCount;
1240
1241 LLVM_DEBUG(dbgs() << "Imported " << ImportedCount - ImportedGVCount
1242 << " functions for Module "
1243 << DestModule.getModuleIdentifier() << "\n");
1244 LLVM_DEBUG(dbgs() << "Imported " << ImportedGVCount
1245 << " global variables for Module "
1246 << DestModule.getModuleIdentifier() << "\n");
1247 return ImportedCount;
1248 }
1249
doImportingForModule(Module & M)1250 static bool doImportingForModule(Module &M) {
1251 if (SummaryFile.empty())
1252 report_fatal_error("error: -function-import requires -summary-file\n");
1253 Expected<std::unique_ptr<ModuleSummaryIndex>> IndexPtrOrErr =
1254 getModuleSummaryIndexForFile(SummaryFile);
1255 if (!IndexPtrOrErr) {
1256 logAllUnhandledErrors(IndexPtrOrErr.takeError(), errs(),
1257 "Error loading file '" + SummaryFile + "': ");
1258 return false;
1259 }
1260 std::unique_ptr<ModuleSummaryIndex> Index = std::move(*IndexPtrOrErr);
1261
1262 // First step is collecting the import list.
1263 FunctionImporter::ImportMapTy ImportList;
1264 // If requested, simply import all functions in the index. This is used
1265 // when testing distributed backend handling via the opt tool, when
1266 // we have distributed indexes containing exactly the summaries to import.
1267 if (ImportAllIndex)
1268 ComputeCrossModuleImportForModuleFromIndex(M.getModuleIdentifier(), *Index,
1269 ImportList);
1270 else
1271 ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
1272 ImportList);
1273
1274 // Conservatively mark all internal values as promoted. This interface is
1275 // only used when doing importing via the function importing pass. The pass
1276 // is only enabled when testing importing via the 'opt' tool, which does
1277 // not do the ThinLink that would normally determine what values to promote.
1278 for (auto &I : *Index) {
1279 for (auto &S : I.second.SummaryList) {
1280 if (GlobalValue::isLocalLinkage(S->linkage()))
1281 S->setLinkage(GlobalValue::ExternalLinkage);
1282 }
1283 }
1284
1285 // Next we need to promote to global scope and rename any local values that
1286 // are potentially exported to other modules.
1287 if (renameModuleForThinLTO(M, *Index, nullptr)) {
1288 errs() << "Error renaming module\n";
1289 return false;
1290 }
1291
1292 // Perform the import now.
1293 auto ModuleLoader = [&M](StringRef Identifier) {
1294 return loadFile(Identifier, M.getContext());
1295 };
1296 FunctionImporter Importer(*Index, ModuleLoader);
1297 Expected<bool> Result = Importer.importFunctions(M, ImportList);
1298
1299 // FIXME: Probably need to propagate Errors through the pass manager.
1300 if (!Result) {
1301 logAllUnhandledErrors(Result.takeError(), errs(),
1302 "Error importing module: ");
1303 return false;
1304 }
1305
1306 return *Result;
1307 }
1308
1309 namespace {
1310
1311 /// Pass that performs cross-module function import provided a summary file.
1312 class FunctionImportLegacyPass : public ModulePass {
1313 public:
1314 /// Pass identification, replacement for typeid
1315 static char ID;
1316
FunctionImportLegacyPass()1317 explicit FunctionImportLegacyPass() : ModulePass(ID) {}
1318
1319 /// Specify pass name for debug output
getPassName() const1320 StringRef getPassName() const override { return "Function Importing"; }
1321
runOnModule(Module & M)1322 bool runOnModule(Module &M) override {
1323 if (skipModule(M))
1324 return false;
1325
1326 return doImportingForModule(M);
1327 }
1328 };
1329
1330 } // end anonymous namespace
1331
run(Module & M,ModuleAnalysisManager & AM)1332 PreservedAnalyses FunctionImportPass::run(Module &M,
1333 ModuleAnalysisManager &AM) {
1334 if (!doImportingForModule(M))
1335 return PreservedAnalyses::all();
1336
1337 return PreservedAnalyses::none();
1338 }
1339
1340 char FunctionImportLegacyPass::ID = 0;
1341 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import",
1342 "Summary Based Function Import", false, false)
1343
1344 namespace llvm {
1345
createFunctionImportPass()1346 Pass *createFunctionImportPass() {
1347 return new FunctionImportLegacyPass();
1348 }
1349
1350 } // end namespace llvm
1351