1//===-LTOBackend.cpp - LLVM Link Time Optimizer Backend -------------------===//
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 "backend" phase of LTO, i.e. it performs
10// optimization and code generation on a loaded module. It is generally used
11// internally by the LTO class but can also be used independently, for example
12// to implement a standalone ThinLTO backend.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/LTO/LTOBackend.h"
17#include "llvm/Analysis/AliasAnalysis.h"
18#include "llvm/Analysis/CGSCCPassManager.h"
19#include "llvm/Analysis/ModuleSummaryAnalysis.h"
20#include "llvm/Analysis/RuntimeLibcallInfo.h"
21#include "llvm/Analysis/TargetLibraryInfo.h"
22#include "llvm/Bitcode/BitcodeReader.h"
23#include "llvm/Bitcode/BitcodeWriter.h"
24#include "llvm/CGData/CodeGenData.h"
25#include "llvm/IR/LLVMRemarkStreamer.h"
26#include "llvm/IR/LegacyPassManager.h"
27#include "llvm/IR/PassManager.h"
28#include "llvm/IR/Verifier.h"
29#include "llvm/LTO/LTO.h"
30#include "llvm/MC/TargetRegistry.h"
31#include "llvm/Object/ModuleSymbolTable.h"
32#include "llvm/Passes/PassBuilder.h"
33#include "llvm/Passes/StandardInstrumentations.h"
34#include "llvm/Plugins/PassPlugin.h"
35#include "llvm/Support/Error.h"
36#include "llvm/Support/FileSystem.h"
37#include "llvm/Support/MemoryBuffer.h"
38#include "llvm/Support/Path.h"
39#include "llvm/Support/ThreadPool.h"
40#include "llvm/Support/ToolOutputFile.h"
41#include "llvm/Support/VirtualFileSystem.h"
42#include "llvm/Support/raw_ostream.h"
43#include "llvm/Target/TargetMachine.h"
44#include "llvm/TargetParser/SubtargetFeature.h"
45#include "llvm/Transforms/IPO/WholeProgramDevirt.h"
46#include "llvm/Transforms/Utils/FunctionImportUtils.h"
47#include "llvm/Transforms/Utils/SplitModule.h"
48#include <optional>
49
50using namespace llvm;
51using namespace lto;
52
53#define DEBUG_TYPE "lto-backend"
54
55enum class LTOBitcodeEmbedding {
56 DoNotEmbed = 0,
57 EmbedOptimized = 1,
58 EmbedPostMergePreOptimized = 2
59};
60
61static cl::opt<LTOBitcodeEmbedding> EmbedBitcode(
62 "lto-embed-bitcode", cl::init(Val: LTOBitcodeEmbedding::DoNotEmbed),
63 cl::values(clEnumValN(LTOBitcodeEmbedding::DoNotEmbed, "none",
64 "Do not embed"),
65 clEnumValN(LTOBitcodeEmbedding::EmbedOptimized, "optimized",
66 "Embed after all optimization passes"),
67 clEnumValN(LTOBitcodeEmbedding::EmbedPostMergePreOptimized,
68 "post-merge-pre-opt",
69 "Embed post merge, but before optimizations")),
70 cl::desc("Embed LLVM bitcode in object files produced by LTO"));
71
72static cl::opt<bool> ThinLTOAssumeMerged(
73 "thinlto-assume-merged", cl::init(Val: false),
74 cl::desc("Assume the input has already undergone ThinLTO function "
75 "importing and the other pre-optimization pipeline changes."));
76
77static cl::list<std::string>
78 SaveModulesList("filter-save-modules", cl::value_desc("module names"),
79 cl::desc("Only save bitcode for module whose name without "
80 "path matches this for -save-temps options"),
81 cl::CommaSeparated, cl::Hidden);
82
83namespace llvm {
84extern cl::opt<bool> NoPGOWarnMismatch;
85}
86
87[[noreturn]] static void reportOpenError(StringRef Path, Twine Msg) {
88 errs() << "failed to open " << Path << ": " << Msg << '\n';
89 errs().flush();
90 exit(status: 1);
91}
92
93Error Config::addSaveTemps(std::string OutputFileName, bool UseInputModulePath,
94 const DenseSet<StringRef> &SaveTempsArgs) {
95 ShouldDiscardValueNames = false;
96
97 std::error_code EC;
98 if (SaveTempsArgs.empty() || SaveTempsArgs.contains(V: "resolution")) {
99 ResolutionFile =
100 std::make_unique<raw_fd_ostream>(args: OutputFileName + "resolution.txt", args&: EC,
101 args: sys::fs::OpenFlags::OF_TextWithCRLF);
102 if (EC) {
103 ResolutionFile.reset();
104 return errorCodeToError(EC);
105 }
106 }
107
108 auto setHook = [&](std::string PathSuffix, ModuleHookFn &Hook) {
109 // Keep track of the hook provided by the linker, which also needs to run.
110 ModuleHookFn LinkerHook = Hook;
111 Hook = [=, SaveModNames = llvm::SmallVector<std::string, 1>(
112 SaveModulesList.begin(), SaveModulesList.end())](
113 unsigned Task, const Module &M) {
114 // If SaveModulesList is not empty, only do save-temps if the module's
115 // filename (without path) matches a name in the list.
116 if (!SaveModNames.empty() &&
117 !llvm::is_contained(
118 Range: SaveModNames,
119 Element: std::string(llvm::sys::path::filename(path: M.getName()))))
120 return false;
121
122 // If the linker's hook returned false, we need to pass that result
123 // through.
124 if (LinkerHook && !LinkerHook(Task, M))
125 return false;
126
127 std::string PathPrefix;
128 // If this is the combined module (not a ThinLTO backend compile) or the
129 // user hasn't requested using the input module's path, emit to a file
130 // named from the provided OutputFileName with the Task ID appended.
131 if (M.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath) {
132 PathPrefix = OutputFileName;
133 if (Task != (unsigned)-1)
134 PathPrefix += utostr(X: Task) + ".";
135 } else
136 PathPrefix = M.getModuleIdentifier() + ".";
137 std::string Path = PathPrefix + PathSuffix + ".bc";
138 std::error_code EC;
139 raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::OF_None);
140 // Because -save-temps is a debugging feature, we report the error
141 // directly and exit.
142 if (EC)
143 reportOpenError(Path, Msg: EC.message());
144 WriteBitcodeToFile(M, Out&: OS, /*ShouldPreserveUseListOrder=*/false);
145 return true;
146 };
147 };
148
149 auto SaveCombinedIndex =
150 [=](const ModuleSummaryIndex &Index,
151 const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols) {
152 std::string Path = OutputFileName + "index.bc";
153 std::error_code EC;
154 raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::OF_None);
155 // Because -save-temps is a debugging feature, we report the error
156 // directly and exit.
157 if (EC)
158 reportOpenError(Path, Msg: EC.message());
159 writeIndexToFile(Index, Out&: OS);
160
161 Path = OutputFileName + "index.dot";
162 raw_fd_ostream OSDot(Path, EC, sys::fs::OpenFlags::OF_Text);
163 if (EC)
164 reportOpenError(Path, Msg: EC.message());
165 Index.exportToDot(OS&: OSDot, GUIDPreservedSymbols);
166 return true;
167 };
168
169 if (SaveTempsArgs.empty()) {
170 setHook("0.preopt", PreOptModuleHook);
171 setHook("1.promote", PostPromoteModuleHook);
172 setHook("2.internalize", PostInternalizeModuleHook);
173 setHook("3.import", PostImportModuleHook);
174 setHook("4.opt", PostOptModuleHook);
175 setHook("5.precodegen", PreCodeGenModuleHook);
176 CombinedIndexHook = SaveCombinedIndex;
177 } else {
178 if (SaveTempsArgs.contains(V: "preopt"))
179 setHook("0.preopt", PreOptModuleHook);
180 if (SaveTempsArgs.contains(V: "promote"))
181 setHook("1.promote", PostPromoteModuleHook);
182 if (SaveTempsArgs.contains(V: "internalize"))
183 setHook("2.internalize", PostInternalizeModuleHook);
184 if (SaveTempsArgs.contains(V: "import"))
185 setHook("3.import", PostImportModuleHook);
186 if (SaveTempsArgs.contains(V: "opt"))
187 setHook("4.opt", PostOptModuleHook);
188 if (SaveTempsArgs.contains(V: "precodegen"))
189 setHook("5.precodegen", PreCodeGenModuleHook);
190 if (SaveTempsArgs.contains(V: "combinedindex"))
191 CombinedIndexHook = SaveCombinedIndex;
192 }
193
194 return Error::success();
195}
196
197#define HANDLE_EXTENSION(Ext) \
198 llvm::PassPluginLibraryInfo get##Ext##PluginInfo();
199#include "llvm/Support/Extension.def"
200#undef HANDLE_EXTENSION
201
202static void RegisterPassPlugins(const Config &Conf, PassBuilder &PB) {
203#define HANDLE_EXTENSION(Ext) \
204 get##Ext##PluginInfo().RegisterPassBuilderCallbacks(PB);
205#include "llvm/Support/Extension.def"
206#undef HANDLE_EXTENSION
207
208 // Load requested pass plugins and let them register pass builder callbacks
209 for (auto &PluginFN : Conf.PassPluginFilenames) {
210 auto PassPlugin = PassPlugin::load(Filename: PluginFN);
211 if (!PassPlugin)
212 reportFatalUsageError(Err: PassPlugin.takeError());
213 PassPlugin->registerPassBuilderCallbacks(PB);
214 }
215
216 // Register already loaded plugins
217 for (auto *LoadedPlugin : Conf.LoadedPassPlugins)
218 LoadedPlugin->registerPassBuilderCallbacks(PB);
219}
220
221static std::unique_ptr<TargetMachine>
222createTargetMachine(const Config &Conf, const Target *TheTarget, Module &M) {
223 const Triple &TheTriple = M.getTargetTriple();
224 SubtargetFeatures Features;
225 Features.getDefaultSubtargetFeatures(Triple: TheTriple);
226 for (const std::string &A : Conf.MAttrs)
227 Features.AddFeature(String: A);
228
229 std::optional<Reloc::Model> RelocModel;
230 if (Conf.RelocModel)
231 RelocModel = *Conf.RelocModel;
232 else if (M.getModuleFlag(Key: "PIC Level"))
233 RelocModel =
234 M.getPICLevel() == PICLevel::NotPIC ? Reloc::Static : Reloc::PIC_;
235
236 std::optional<CodeModel::Model> CodeModel;
237 if (Conf.CodeModel)
238 CodeModel = *Conf.CodeModel;
239 else
240 CodeModel = M.getCodeModel();
241
242 TargetOptions TargetOpts = Conf.Options;
243 if (TargetOpts.MCOptions.ABIName.empty()) {
244 TargetOpts.MCOptions.ABIName = M.getTargetABIFromMD();
245 }
246
247 std::unique_ptr<TargetMachine> TM(TheTarget->createTargetMachine(
248 TT: TheTriple, CPU: Conf.CPU, Features: Features.getString(), Options: TargetOpts, RM: RelocModel,
249 CM: CodeModel, OL: Conf.CGOptLevel));
250
251 assert(TM && "Failed to create target machine");
252
253 if (std::optional<uint64_t> LargeDataThreshold = M.getLargeDataThreshold())
254 TM->setLargeDataThreshold(*LargeDataThreshold);
255
256 return TM;
257}
258
259static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
260 unsigned OptLevel, bool IsThinLTO,
261 ModuleSummaryIndex *ExportSummary,
262 const ModuleSummaryIndex *ImportSummary,
263 const DenseSet<StringRef> &BitcodeLibFuncs) {
264 std::optional<PGOOptions> PGOOpt;
265 if (!Conf.SampleProfile.empty())
266 PGOOpt = PGOOptions(Conf.SampleProfile, "", Conf.ProfileRemapping,
267 /*MemoryProfile=*/"", PGOOptions::SampleUse,
268 PGOOptions::NoCSAction,
269 PGOOptions::ColdFuncOpt::Default, true);
270 else if (Conf.RunCSIRInstr) {
271 PGOOpt = PGOOptions("", Conf.CSIRProfile, Conf.ProfileRemapping,
272 /*MemoryProfile=*/"", PGOOptions::IRUse,
273 PGOOptions::CSIRInstr, PGOOptions::ColdFuncOpt::Default,
274 Conf.AddFSDiscriminator);
275 } else if (!Conf.CSIRProfile.empty()) {
276 PGOOpt =
277 PGOOptions(Conf.CSIRProfile, "", Conf.ProfileRemapping,
278 /*MemoryProfile=*/"", PGOOptions::IRUse, PGOOptions::CSIRUse,
279 PGOOptions::ColdFuncOpt::Default, Conf.AddFSDiscriminator);
280 NoPGOWarnMismatch = !Conf.PGOWarnMismatch;
281 } else if (Conf.AddFSDiscriminator) {
282 PGOOpt = PGOOptions("", "", "", /*MemoryProfile=*/"", PGOOptions::NoAction,
283 PGOOptions::NoCSAction,
284 PGOOptions::ColdFuncOpt::Default, true);
285 }
286 TM->setPGOOption(PGOOpt);
287
288 LoopAnalysisManager LAM;
289 FunctionAnalysisManager FAM;
290 CGSCCAnalysisManager CGAM;
291 ModuleAnalysisManager MAM;
292
293 PassInstrumentationCallbacks PIC;
294 StandardInstrumentations SI(Mod.getContext(), Conf.DebugPassManager,
295 Conf.VerifyEach);
296 SI.registerCallbacks(PIC, MAM: &MAM);
297 PassBuilder PB(TM, Conf.PTO, PGOOpt, &PIC);
298
299 RegisterPassPlugins(Conf, PB);
300
301 std::unique_ptr<TargetLibraryInfoImpl> TLII(
302 new TargetLibraryInfoImpl(TM->getTargetTriple(), TM->Options.VecLib));
303 if (Conf.Freestanding)
304 TLII->disableAllFunctions();
305
306 // Determine whether or not its safe to emit calls to each libfunc. Libfuncs
307 // that might have been present in the current LTO unit, but are not, have
308 // lost their only opportunity to be defined, and calls must not be emitted to
309 // them.
310 // FIXME: BitcodeLibFuncs isn't yet set for distributed ThinLTO.
311 TargetLibraryInfo TLI(*TLII);
312 for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
313 ++I) {
314 LibFunc F = static_cast<LibFunc>(I);
315 if (BitcodeLibFuncs.contains(V: TLI.getName(F)))
316 TLII->setUnavailable(F);
317 }
318
319 FAM.registerPass(PassBuilder: [&] { return TargetLibraryAnalysis(*TLII); });
320
321 // Parse a custom AA pipeline if asked to.
322 if (!Conf.AAPipeline.empty()) {
323 AAManager AA;
324 if (auto Err = PB.parseAAPipeline(AA, PipelineText: Conf.AAPipeline)) {
325 report_fatal_error(reason: Twine("unable to parse AA pipeline description '") +
326 Conf.AAPipeline + "': " + toString(E: std::move(Err)));
327 }
328 // Register the AA manager first so that our version is the one used.
329 FAM.registerPass(PassBuilder: [&] { return std::move(AA); });
330 }
331
332 // Register all the basic analyses with the managers.
333 PB.registerModuleAnalyses(MAM);
334 PB.registerCGSCCAnalyses(CGAM);
335 PB.registerFunctionAnalyses(FAM);
336 PB.registerLoopAnalyses(LAM);
337 PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
338
339 if (Conf.PassBuilderCallback)
340 Conf.PassBuilderCallback(PB);
341
342 ModulePassManager MPM;
343
344 if (!Conf.DisableVerify)
345 MPM.addPass(Pass: VerifierPass());
346
347 OptimizationLevel OL;
348
349 switch (OptLevel) {
350 default:
351 llvm_unreachable("Invalid optimization level");
352 case 0:
353 OL = OptimizationLevel::O0;
354 break;
355 case 1:
356 OL = OptimizationLevel::O1;
357 break;
358 case 2:
359 OL = OptimizationLevel::O2;
360 break;
361 case 3:
362 OL = OptimizationLevel::O3;
363 break;
364 }
365
366 // Parse a custom pipeline if asked to.
367 if (!Conf.OptPipeline.empty()) {
368 if (auto Err = PB.parsePassPipeline(MPM, PipelineText: Conf.OptPipeline)) {
369 report_fatal_error(reason: Twine("unable to parse pass pipeline description '") +
370 Conf.OptPipeline + "': " + toString(E: std::move(Err)));
371 }
372 } else if (IsThinLTO) {
373 MPM.addPass(Pass: PB.buildThinLTODefaultPipeline(Level: OL, ImportSummary));
374 } else {
375 MPM.addPass(Pass: PB.buildLTODefaultPipeline(Level: OL, ExportSummary));
376 }
377
378 if (!Conf.DisableVerify)
379 MPM.addPass(Pass: VerifierPass());
380
381 if (PB.getPrintPipelinePasses()) {
382 std::string PipelineStr;
383 raw_string_ostream OS(PipelineStr);
384 MPM.printPipeline(OS, MapClassName2PassName: [&PIC](StringRef ClassName) {
385 auto PassName = PIC.getPassNameForClassName(ClassName);
386 return PassName.empty() ? ClassName : PassName;
387 });
388 outs() << "pipeline-passes: " << PipelineStr << '\n';
389 }
390
391 MPM.run(IR&: Mod, AM&: MAM);
392}
393
394static bool isEmptyModule(const Module &Mod) {
395 // Module is empty if it has no functions, no globals, no inline asm and no
396 // named metadata (aliases and ifuncs require functions or globals so we
397 // don't need to check those explicitly).
398 return Mod.empty() && Mod.global_empty() && Mod.named_metadata_empty() &&
399 Mod.getModuleInlineAsm().empty();
400}
401
402bool lto::opt(const Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
403 bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
404 const ModuleSummaryIndex *ImportSummary,
405 const std::vector<uint8_t> &CmdArgs,
406 ArrayRef<StringRef> BitcodeLibFuncs) {
407 llvm::TimeTraceScope timeScope("opt");
408 if (EmbedBitcode == LTOBitcodeEmbedding::EmbedPostMergePreOptimized) {
409 // FIXME: the motivation for capturing post-merge bitcode and command line
410 // is replicating the compilation environment from bitcode, without needing
411 // to understand the dependencies (the functions to be imported). This
412 // assumes a clang - based invocation, case in which we have the command
413 // line.
414 // It's not very clear how the above motivation would map in the
415 // linker-based case, so we currently don't plumb the command line args in
416 // that case.
417 if (CmdArgs.empty())
418 LLVM_DEBUG(
419 dbgs() << "Post-(Thin)LTO merge bitcode embedding was requested, but "
420 "command line arguments are not available");
421 llvm::embedBitcodeInModule(M&: Mod, Buf: llvm::MemoryBufferRef(),
422 /*EmbedBitcode*/ true, /*EmbedCmdline*/ true,
423 /*Cmdline*/ CmdArgs);
424 }
425 // No need to run any opt passes if the module is empty.
426 // In theory these passes should take almost no time for an empty
427 // module, however, this guards against doing any unnecessary summary-based
428 // analysis in the case of a ThinLTO build where this might be an empty
429 // regular LTO combined module, with a large combined index from ThinLTO.
430 if (!isEmptyModule(Mod)) {
431 DenseSet<StringRef> BitcodeLibFuncsSet(BitcodeLibFuncs.begin(),
432 BitcodeLibFuncs.end());
433 // FIXME: Plumb the combined index into the new pass manager.
434 runNewPMPasses(Conf, Mod, TM, OptLevel: Conf.OptLevel, IsThinLTO, ExportSummary,
435 ImportSummary, BitcodeLibFuncs: BitcodeLibFuncsSet);
436 }
437 return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
438}
439
440static void codegen(const Config &Conf, TargetMachine *TM,
441 AddStreamFn AddStream, unsigned Task, Module &Mod,
442 const ModuleSummaryIndex &CombinedIndex) {
443 llvm::TimeTraceScope timeScope("codegen");
444 if (Conf.PreCodeGenModuleHook && !Conf.PreCodeGenModuleHook(Task, Mod))
445 return;
446
447 if (EmbedBitcode == LTOBitcodeEmbedding::EmbedOptimized)
448 llvm::embedBitcodeInModule(M&: Mod, Buf: llvm::MemoryBufferRef(),
449 /*EmbedBitcode*/ true,
450 /*EmbedCmdline*/ false,
451 /*CmdArgs*/ std::vector<uint8_t>());
452
453 std::unique_ptr<ToolOutputFile> DwoOut;
454 SmallString<1024> DwoFile(Conf.SplitDwarfOutput);
455 if (!Conf.DwoDir.empty()) {
456 std::error_code EC;
457 if (auto EC = llvm::sys::fs::create_directories(path: Conf.DwoDir))
458 report_fatal_error(reason: Twine("Failed to create directory ") + Conf.DwoDir +
459 ": " + EC.message());
460
461 DwoFile = Conf.DwoDir;
462 sys::path::append(path&: DwoFile, a: std::to_string(val: Task) + ".dwo");
463 TM->Options.MCOptions.SplitDwarfFile = std::string(DwoFile);
464 } else
465 TM->Options.MCOptions.SplitDwarfFile = Conf.SplitDwarfFile;
466
467 if (!DwoFile.empty()) {
468 std::error_code EC;
469 DwoOut = std::make_unique<ToolOutputFile>(args&: DwoFile, args&: EC, args: sys::fs::OF_None);
470 if (EC)
471 report_fatal_error(reason: Twine("Failed to open ") + DwoFile + ": " +
472 EC.message());
473 }
474
475 Expected<std::unique_ptr<CachedFileStream>> StreamOrErr =
476 AddStream(Task, Mod.getModuleIdentifier());
477 if (Error Err = StreamOrErr.takeError())
478 report_fatal_error(Err: std::move(Err));
479 std::unique_ptr<CachedFileStream> &Stream = *StreamOrErr;
480 TM->Options.ObjectFilenameForDebug = Stream->ObjectPathName;
481
482 // Create the codegen pipeline in its own scope so it gets deleted before
483 // Stream->commit() is called. The commit function of CacheStream deletes
484 // the raw stream, which is too early as streamers (e.g. MCAsmStreamer)
485 // keep the pointer and may use it until their destruction. See #138194.
486 {
487 legacy::PassManager CodeGenPasses;
488 TargetLibraryInfoImpl TLII(Mod.getTargetTriple(), TM->Options.VecLib);
489 CodeGenPasses.add(P: new TargetLibraryInfoWrapperPass(TLII));
490 CodeGenPasses.add(P: new RuntimeLibraryInfoWrapper(
491 TM->Options.MCOptions.ABIName, TM->Options.VecLib));
492
493 // No need to make index available if the module is empty.
494 // In theory these passes should not use the index for an empty
495 // module, however, this guards against doing any unnecessary summary-based
496 // analysis in the case of a ThinLTO build where this might be an empty
497 // regular LTO combined module, with a large combined index from ThinLTO.
498 if (!isEmptyModule(Mod))
499 CodeGenPasses.add(
500 P: createImmutableModuleSummaryIndexWrapperPass(Index: &CombinedIndex));
501 if (Conf.PreCodeGenPassesHook)
502 Conf.PreCodeGenPassesHook(CodeGenPasses);
503 if (TM->addPassesToEmitFile(CodeGenPasses, *Stream->OS,
504 DwoOut ? &DwoOut->os() : nullptr,
505 Conf.CGFileType))
506 report_fatal_error(reason: "Failed to setup codegen");
507 CodeGenPasses.run(M&: Mod);
508
509 if (DwoOut)
510 DwoOut->keep();
511 }
512
513 if (Error Err = Stream->commit())
514 report_fatal_error(Err: std::move(Err));
515}
516
517static void splitCodeGen(const Config &C, TargetMachine *TM,
518 AddStreamFn AddStream,
519 unsigned ParallelCodeGenParallelismLevel, Module &Mod,
520 const ModuleSummaryIndex &CombinedIndex) {
521 DefaultThreadPool CodegenThreadPool(
522 heavyweight_hardware_concurrency(ThreadCount: ParallelCodeGenParallelismLevel));
523 unsigned ThreadCount = 0;
524 const Target *T = &TM->getTarget();
525
526 const auto HandleModulePartition =
527 [&](std::unique_ptr<Module> MPart) {
528 // We want to clone the module in a new context to multi-thread the
529 // codegen. We do it by serializing partition modules to bitcode
530 // (while still on the main thread, in order to avoid data races) and
531 // spinning up new threads which deserialize the partitions into
532 // separate contexts.
533 // FIXME: Provide a more direct way to do this in LLVM.
534 SmallString<0> BC;
535 raw_svector_ostream BCOS(BC);
536 WriteBitcodeToFile(M: *MPart, Out&: BCOS);
537
538 // Enqueue the task
539 CodegenThreadPool.async(
540 F: [&](const SmallString<0> &BC, unsigned ThreadId) {
541 LTOLLVMContext Ctx(C);
542 Expected<std::unique_ptr<Module>> MOrErr =
543 parseBitcodeFile(Buffer: MemoryBufferRef(BC.str(), "ld-temp.o"), Context&: Ctx);
544 if (!MOrErr)
545 report_fatal_error(reason: "Failed to read bitcode");
546 std::unique_ptr<Module> MPartInCtx = std::move(MOrErr.get());
547
548 std::unique_ptr<TargetMachine> TM =
549 createTargetMachine(Conf: C, TheTarget: T, M&: *MPartInCtx);
550
551 codegen(Conf: C, TM: TM.get(), AddStream, Task: ThreadId, Mod&: *MPartInCtx,
552 CombinedIndex);
553 },
554 // Pass BC using std::move to ensure that it get moved rather than
555 // copied into the thread's context.
556 ArgList: std::move(BC), ArgList: ThreadCount++);
557 };
558
559 // Try target-specific module splitting first, then fallback to the default.
560 if (!TM->splitModule(M&: Mod, NumParts: ParallelCodeGenParallelismLevel,
561 ModuleCallback: HandleModulePartition)) {
562 SplitModule(M&: Mod, N: ParallelCodeGenParallelismLevel, ModuleCallback: HandleModulePartition,
563 PreserveLocals: false);
564 }
565
566 // Because the inner lambda (which runs in a worker thread) captures our local
567 // variables, we need to wait for the worker threads to terminate before we
568 // can leave the function scope.
569 CodegenThreadPool.wait();
570}
571
572static Expected<const Target *> initAndLookupTarget(const Config &C,
573 Module &Mod) {
574 if (!C.OverrideTriple.empty())
575 Mod.setTargetTriple(Triple(C.OverrideTriple));
576 else if (Mod.getTargetTriple().empty())
577 Mod.setTargetTriple(Triple(C.DefaultTriple));
578
579 std::string Msg;
580 const Target *T = TargetRegistry::lookupTarget(TheTriple: Mod.getTargetTriple(), Error&: Msg);
581 if (!T)
582 return make_error<StringError>(Args&: Msg, Args: inconvertibleErrorCode());
583 return T;
584}
585
586Error lto::finalizeOptimizationRemarks(LLVMRemarkFileHandle DiagOutputFile) {
587 // Make sure we flush the diagnostic remarks file in case the linker doesn't
588 // call the global destructors before exiting.
589 if (!DiagOutputFile)
590 return Error::success();
591 DiagOutputFile.finalize();
592 DiagOutputFile->keep();
593 DiagOutputFile->os().flush();
594 return Error::success();
595}
596
597Error lto::backend(const Config &C, AddStreamFn AddStream,
598 unsigned ParallelCodeGenParallelismLevel, Module &Mod,
599 ModuleSummaryIndex &CombinedIndex,
600 ArrayRef<StringRef> BitcodeLibFuncs) {
601 llvm::TimeTraceScope timeScope("LTO backend");
602 Expected<const Target *> TOrErr = initAndLookupTarget(C, Mod);
603 if (!TOrErr)
604 return TOrErr.takeError();
605
606 std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf: C, TheTarget: *TOrErr, M&: Mod);
607
608 LLVM_DEBUG(dbgs() << "Running regular LTO\n");
609 if (!C.CodeGenOnly) {
610 if (!opt(Conf: C, TM: TM.get(), Task: 0, Mod, /*IsThinLTO=*/false,
611 /*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr,
612 /*CmdArgs*/ std::vector<uint8_t>(), BitcodeLibFuncs))
613 return Error::success();
614 }
615
616 if (ParallelCodeGenParallelismLevel == 1) {
617 codegen(Conf: C, TM: TM.get(), AddStream, Task: 0, Mod, CombinedIndex);
618 } else {
619 splitCodeGen(C, TM: TM.get(), AddStream, ParallelCodeGenParallelismLevel, Mod,
620 CombinedIndex);
621 }
622 return Error::success();
623}
624
625static void dropDeadSymbols(Module &Mod, const GVSummaryMapTy &DefinedGlobals,
626 const ModuleSummaryIndex &Index) {
627 llvm::TimeTraceScope timeScope("Drop dead symbols");
628 std::vector<GlobalValue*> DeadGVs;
629
630 for (auto &GV : Mod.global_values()) {
631 auto GUID = GV.getGUIDIfAssigned();
632 if (!GUID)
633 continue;
634
635 if (GlobalValueSummary *GVS = DefinedGlobals.lookup(Val: *GUID))
636 if (!Index.isGlobalValueLive(GVS)) {
637 DeadGVs.push_back(x: &GV);
638 convertToDeclaration(GV);
639 }
640 }
641
642 // Now that all dead bodies have been dropped, delete the actual objects
643 // themselves when possible.
644 for (GlobalValue *GV : DeadGVs) {
645 GV->removeDeadConstantUsers();
646 // Might reference something defined in native object (i.e. dropped a
647 // non-prevailing IR def, but we need to keep the declaration).
648 if (GV->use_empty())
649 GV->eraseFromParent();
650 }
651}
652
653Error lto::thinBackend(const Config &Conf, unsigned Task, AddStreamFn AddStream,
654 Module &Mod, const ModuleSummaryIndex &CombinedIndex,
655 const FunctionImporter::ImportMapTy &ImportList,
656 const GVSummaryMapTy &DefinedGlobals,
657 MapVector<StringRef, BitcodeModule> *ModuleMap,
658 bool CodeGenOnly, ArrayRef<StringRef> BitcodeLibFuncs,
659 AddStreamFn IRAddStream,
660 const std::vector<uint8_t> &CmdArgs) {
661 llvm::TimeTraceScope timeScope("Thin backend", Mod.getModuleIdentifier());
662 Expected<const Target *> TOrErr = initAndLookupTarget(C: Conf, Mod);
663 if (!TOrErr)
664 return TOrErr.takeError();
665
666 std::unique_ptr<TargetMachine> TM = createTargetMachine(Conf, TheTarget: *TOrErr, M&: Mod);
667
668 // Setup optimization remarks.
669 auto DiagFileOrErr = lto::setupLLVMOptimizationRemarks(
670 Context&: Mod.getContext(), RemarksFilename: Conf.RemarksFilename, RemarksPasses: Conf.RemarksPasses,
671 RemarksFormat: Conf.RemarksFormat, RemarksWithHotness: Conf.RemarksWithHotness, RemarksHotnessThreshold: Conf.RemarksHotnessThreshold,
672 Count: Task);
673 if (!DiagFileOrErr)
674 return DiagFileOrErr.takeError();
675 auto DiagnosticOutputFile = std::move(*DiagFileOrErr);
676
677 // Set the partial sample profile ratio in the profile summary module flag of
678 // the module, if applicable.
679 Mod.setPartialSampleProfileRatio(CombinedIndex);
680
681 LLVM_DEBUG(dbgs() << "Running ThinLTO\n");
682 if (CodeGenOnly) {
683 // If CodeGenOnly is set, we only perform code generation and skip
684 // optimization. This value may differ from Conf.CodeGenOnly.
685 codegen(Conf, TM: TM.get(), AddStream, Task, Mod, CombinedIndex);
686 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
687 }
688
689 if (Conf.PreOptModuleHook && !Conf.PreOptModuleHook(Task, Mod))
690 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
691
692 auto OptimizeAndCodegen =
693 [&](Module &Mod, TargetMachine *TM,
694 LLVMRemarkFileHandle DiagnosticOutputFile) {
695 // Perform optimization and code generation for ThinLTO.
696 if (!opt(Conf, TM, Task, Mod, /*IsThinLTO=*/true,
697 /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex,
698 CmdArgs, BitcodeLibFuncs))
699 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
700
701 // Save the current module before the first codegen round.
702 // Note that the second codegen round runs only `codegen()` without
703 // running `opt()`. We're not reaching here as it's bailed out earlier
704 // with `CodeGenOnly` which has been set in `SecondRoundThinBackend`.
705 if (IRAddStream)
706 cgdata::saveModuleForTwoRounds(TheModule: Mod, Task, AddStream: IRAddStream);
707
708 codegen(Conf, TM, AddStream, Task, Mod, CombinedIndex);
709 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
710 };
711
712 if (ThinLTOAssumeMerged)
713 return OptimizeAndCodegen(Mod, TM.get(), std::move(DiagnosticOutputFile));
714
715 // When linking an ELF shared object, dso_local should be dropped. We
716 // conservatively do this for -fpic.
717 bool ClearDSOLocalOnDeclarations =
718 Mod.getTargetTriple().isOSBinFormatELF() &&
719 TM->getRelocationModel() != Reloc::Static &&
720 Mod.getPIELevel() == PIELevel::Default;
721 renameModuleForThinLTO(M&: Mod, Index: CombinedIndex, ClearDSOLocalOnDeclarations);
722
723 dropDeadSymbols(Mod, DefinedGlobals, Index: CombinedIndex);
724
725 thinLTOFinalizeInModule(TheModule&: Mod, DefinedGlobals, /*PropagateAttrs=*/true);
726
727 if (Conf.PostPromoteModuleHook && !Conf.PostPromoteModuleHook(Task, Mod))
728 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
729
730 if (!DefinedGlobals.empty())
731 thinLTOInternalizeModule(TheModule&: Mod, DefinedGlobals);
732
733 if (Conf.PostInternalizeModuleHook &&
734 !Conf.PostInternalizeModuleHook(Task, Mod))
735 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
736
737 auto ModuleLoader = [&](StringRef Identifier) {
738 llvm::TimeTraceScope moduleLoaderScope("Module loader", Identifier);
739 assert(Mod.getContext().isODRUniquingDebugTypes() &&
740 "ODR Type uniquing should be enabled on the context");
741 if (ModuleMap) {
742 auto I = ModuleMap->find(Key: Identifier);
743 assert(I != ModuleMap->end());
744 return I->second.getLazyModule(Context&: Mod.getContext(),
745 /*ShouldLazyLoadMetadata=*/true,
746 /*IsImporting*/ true);
747 }
748
749 ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> MBOrErr =
750 llvm::MemoryBuffer::getFile(Filename: Identifier);
751 if (!MBOrErr)
752 return Expected<std::unique_ptr<llvm::Module>>(make_error<StringError>(
753 Args: Twine("Error loading imported file ") + Identifier + " : ",
754 Args: MBOrErr.getError()));
755
756 Expected<BitcodeModule> BMOrErr = findThinLTOModule(MBRef: **MBOrErr);
757 if (!BMOrErr)
758 return Expected<std::unique_ptr<llvm::Module>>(make_error<StringError>(
759 Args: Twine("Error loading imported file ") + Identifier + " : " +
760 toString(E: BMOrErr.takeError()),
761 Args: inconvertibleErrorCode()));
762
763 Expected<std::unique_ptr<Module>> MOrErr =
764 BMOrErr->getLazyModule(Context&: Mod.getContext(),
765 /*ShouldLazyLoadMetadata=*/true,
766 /*IsImporting*/ true);
767 if (MOrErr)
768 (*MOrErr)->setOwnedMemoryBuffer(std::move(*MBOrErr));
769 return MOrErr;
770 };
771
772 {
773 llvm::TimeTraceScope importScope("Import functions");
774 FunctionImporter Importer(CombinedIndex, ModuleLoader,
775 ClearDSOLocalOnDeclarations);
776 if (Error Err = Importer.importFunctions(M&: Mod, ImportList).takeError())
777 return Err;
778 }
779
780 // Do this after any importing so that imported code is updated.
781 updatePublicTypeTestCalls(M&: Mod, WholeProgramVisibilityEnabledInLTO: CombinedIndex.withWholeProgramVisibility());
782
783 if (Conf.PostImportModuleHook && !Conf.PostImportModuleHook(Task, Mod))
784 return finalizeOptimizationRemarks(DiagOutputFile: std::move(DiagnosticOutputFile));
785
786 return OptimizeAndCodegen(Mod, TM.get(), std::move(DiagnosticOutputFile));
787}
788
789BitcodeModule *lto::findThinLTOModule(MutableArrayRef<BitcodeModule> BMs) {
790 if (ThinLTOAssumeMerged && BMs.size() == 1)
791 return BMs.begin();
792
793 for (BitcodeModule &BM : BMs) {
794 Expected<BitcodeLTOInfo> LTOInfo = BM.getLTOInfo();
795 if (LTOInfo && LTOInfo->IsThinLTO)
796 return &BM;
797 }
798 return nullptr;
799}
800
801Expected<BitcodeModule> lto::findThinLTOModule(MemoryBufferRef MBRef) {
802 Expected<std::vector<BitcodeModule>> BMsOrErr = getBitcodeModuleList(Buffer: MBRef);
803 if (!BMsOrErr)
804 return BMsOrErr.takeError();
805
806 // The bitcode file may contain multiple modules, we want the one that is
807 // marked as being the ThinLTO module.
808 if (const BitcodeModule *Bm = lto::findThinLTOModule(BMs: *BMsOrErr))
809 return *Bm;
810
811 return make_error<StringError>(Args: "Could not find module summary",
812 Args: inconvertibleErrorCode());
813}
814
815bool lto::initImportList(const Module &M,
816 const ModuleSummaryIndex &CombinedIndex,
817 FunctionImporter::ImportMapTy &ImportList) {
818 if (ThinLTOAssumeMerged)
819 return true;
820 // We can simply import the values mentioned in the combined index, since
821 // we should only invoke this using the individual indexes written out
822 // via a WriteIndexesThinBackend.
823 for (const auto &GlobalList : CombinedIndex) {
824 // Ignore entries for undefined references.
825 if (GlobalList.second.getSummaryList().empty())
826 continue;
827
828 auto GUID = GlobalList.first;
829 for (const auto &Summary : GlobalList.second.getSummaryList()) {
830 // Skip the summaries for the importing module. These are included to
831 // e.g. record required linkage changes.
832 if (Summary->modulePath() == M.getModuleIdentifier())
833 continue;
834 // Add an entry to provoke importing by thinBackend.
835 ImportList.addGUID(FromModule: Summary->modulePath(), GUID, ImportKind: Summary->importType());
836 }
837 }
838 return true;
839}
840