1 //===-LTOBackend.cpp - LLVM Link Time Optimizer Backend -------------------===//
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
7 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/LTO/LTOBackend.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/CGSCCPassManager.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Bitcode/BitcodeReader.h"
22 #include "llvm/Bitcode/BitcodeWriter.h"
23 #include "llvm/IR/LegacyPassManager.h"
24 #include "llvm/IR/PassManager.h"
25 #include "llvm/IR/RemarkStreamer.h"
26 #include "llvm/IR/Verifier.h"
27 #include "llvm/LTO/LTO.h"
28 #include "llvm/MC/SubtargetFeature.h"
29 #include "llvm/Object/ModuleSymbolTable.h"
30 #include "llvm/Passes/PassBuilder.h"
31 #include "llvm/Support/Error.h"
32 #include "llvm/Support/FileSystem.h"
33 #include "llvm/Support/MemoryBuffer.h"
34 #include "llvm/Support/Path.h"
35 #include "llvm/Support/Program.h"
36 #include "llvm/Support/TargetRegistry.h"
37 #include "llvm/Support/ThreadPool.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetMachine.h"
40 #include "llvm/Transforms/IPO.h"
41 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
42 #include "llvm/Transforms/Scalar/LoopPassManager.h"
43 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
44 #include "llvm/Transforms/Utils/SplitModule.h"
49 LLVM_ATTRIBUTE_NORETURN
static void reportOpenError(StringRef Path
, Twine Msg
) {
50 errs() << "failed to open " << Path
<< ": " << Msg
<< '\n';
55 Error
Config::addSaveTemps(std::string OutputFileName
,
56 bool UseInputModulePath
) {
57 ShouldDiscardValueNames
= false;
60 ResolutionFile
= llvm::make_unique
<raw_fd_ostream
>(
61 OutputFileName
+ "resolution.txt", EC
, sys::fs::OpenFlags::OF_Text
);
63 return errorCodeToError(EC
);
65 auto setHook
= [&](std::string PathSuffix
, ModuleHookFn
&Hook
) {
66 // Keep track of the hook provided by the linker, which also needs to run.
67 ModuleHookFn LinkerHook
= Hook
;
68 Hook
= [=](unsigned Task
, const Module
&M
) {
69 // If the linker's hook returned false, we need to pass that result
71 if (LinkerHook
&& !LinkerHook(Task
, M
))
74 std::string PathPrefix
;
75 // If this is the combined module (not a ThinLTO backend compile) or the
76 // user hasn't requested using the input module's path, emit to a file
77 // named from the provided OutputFileName with the Task ID appended.
78 if (M
.getModuleIdentifier() == "ld-temp.o" || !UseInputModulePath
) {
79 PathPrefix
= OutputFileName
;
80 if (Task
!= (unsigned)-1)
81 PathPrefix
+= utostr(Task
) + ".";
83 PathPrefix
= M
.getModuleIdentifier() + ".";
84 std::string Path
= PathPrefix
+ PathSuffix
+ ".bc";
86 raw_fd_ostream
OS(Path
, EC
, sys::fs::OpenFlags::OF_None
);
87 // Because -save-temps is a debugging feature, we report the error
90 reportOpenError(Path
, EC
.message());
91 WriteBitcodeToFile(M
, OS
, /*ShouldPreserveUseListOrder=*/false);
96 setHook("0.preopt", PreOptModuleHook
);
97 setHook("1.promote", PostPromoteModuleHook
);
98 setHook("2.internalize", PostInternalizeModuleHook
);
99 setHook("3.import", PostImportModuleHook
);
100 setHook("4.opt", PostOptModuleHook
);
101 setHook("5.precodegen", PreCodeGenModuleHook
);
103 CombinedIndexHook
= [=](const ModuleSummaryIndex
&Index
) {
104 std::string Path
= OutputFileName
+ "index.bc";
106 raw_fd_ostream
OS(Path
, EC
, sys::fs::OpenFlags::OF_None
);
107 // Because -save-temps is a debugging feature, we report the error
108 // directly and exit.
110 reportOpenError(Path
, EC
.message());
111 WriteIndexToFile(Index
, OS
);
113 Path
= OutputFileName
+ "index.dot";
114 raw_fd_ostream
OSDot(Path
, EC
, sys::fs::OpenFlags::OF_None
);
116 reportOpenError(Path
, EC
.message());
117 Index
.exportToDot(OSDot
);
121 return Error::success();
126 std::unique_ptr
<TargetMachine
>
127 createTargetMachine(Config
&Conf
, const Target
*TheTarget
, Module
&M
) {
128 StringRef TheTriple
= M
.getTargetTriple();
129 SubtargetFeatures Features
;
130 Features
.getDefaultSubtargetFeatures(Triple(TheTriple
));
131 for (const std::string
&A
: Conf
.MAttrs
)
132 Features
.AddFeature(A
);
134 Reloc::Model RelocModel
;
136 RelocModel
= *Conf
.RelocModel
;
139 M
.getPICLevel() == PICLevel::NotPIC
? Reloc::Static
: Reloc::PIC_
;
141 Optional
<CodeModel::Model
> CodeModel
;
143 CodeModel
= *Conf
.CodeModel
;
145 CodeModel
= M
.getCodeModel();
147 return std::unique_ptr
<TargetMachine
>(TheTarget
->createTargetMachine(
148 TheTriple
, Conf
.CPU
, Features
.getString(), Conf
.Options
, RelocModel
,
149 CodeModel
, Conf
.CGOptLevel
));
152 static void runNewPMPasses(Config
&Conf
, Module
&Mod
, TargetMachine
*TM
,
153 unsigned OptLevel
, bool IsThinLTO
,
154 ModuleSummaryIndex
*ExportSummary
,
155 const ModuleSummaryIndex
*ImportSummary
) {
156 Optional
<PGOOptions
> PGOOpt
;
157 if (!Conf
.SampleProfile
.empty())
158 PGOOpt
= PGOOptions(Conf
.SampleProfile
, "", Conf
.ProfileRemapping
,
159 PGOOptions::SampleUse
, PGOOptions::NoCSAction
, true);
160 else if (Conf
.RunCSIRInstr
) {
161 PGOOpt
= PGOOptions("", Conf
.CSIRProfile
, Conf
.ProfileRemapping
,
162 PGOOptions::IRUse
, PGOOptions::CSIRInstr
);
163 } else if (!Conf
.CSIRProfile
.empty()) {
164 PGOOpt
= PGOOptions(Conf
.CSIRProfile
, "", Conf
.ProfileRemapping
,
165 PGOOptions::IRUse
, PGOOptions::CSIRUse
);
168 PassBuilder
PB(TM
, PipelineTuningOptions(), PGOOpt
);
171 // Parse a custom AA pipeline if asked to.
172 if (auto Err
= PB
.parseAAPipeline(AA
, "default"))
173 report_fatal_error("Error parsing default AA pipeline");
175 LoopAnalysisManager
LAM(Conf
.DebugPassManager
);
176 FunctionAnalysisManager
FAM(Conf
.DebugPassManager
);
177 CGSCCAnalysisManager
CGAM(Conf
.DebugPassManager
);
178 ModuleAnalysisManager
MAM(Conf
.DebugPassManager
);
180 // Register the AA manager first so that our version is the one used.
181 FAM
.registerPass([&] { return std::move(AA
); });
183 // Register all the basic analyses with the managers.
184 PB
.registerModuleAnalyses(MAM
);
185 PB
.registerCGSCCAnalyses(CGAM
);
186 PB
.registerFunctionAnalyses(FAM
);
187 PB
.registerLoopAnalyses(LAM
);
188 PB
.crossRegisterProxies(LAM
, FAM
, CGAM
, MAM
);
190 ModulePassManager
MPM(Conf
.DebugPassManager
);
191 // FIXME (davide): verify the input.
193 PassBuilder::OptimizationLevel OL
;
197 llvm_unreachable("Invalid optimization level");
199 OL
= PassBuilder::O0
;
202 OL
= PassBuilder::O1
;
205 OL
= PassBuilder::O2
;
208 OL
= PassBuilder::O3
;
213 MPM
= PB
.buildThinLTODefaultPipeline(OL
, Conf
.DebugPassManager
,
216 MPM
= PB
.buildLTODefaultPipeline(OL
, Conf
.DebugPassManager
, ExportSummary
);
219 // FIXME (davide): verify the output.
222 static void runNewPMCustomPasses(Module
&Mod
, TargetMachine
*TM
,
223 std::string PipelineDesc
,
224 std::string AAPipelineDesc
,
225 bool DisableVerify
) {
229 // Parse a custom AA pipeline if asked to.
230 if (!AAPipelineDesc
.empty())
231 if (auto Err
= PB
.parseAAPipeline(AA
, AAPipelineDesc
))
232 report_fatal_error("unable to parse AA pipeline description '" +
233 AAPipelineDesc
+ "': " + toString(std::move(Err
)));
235 LoopAnalysisManager LAM
;
236 FunctionAnalysisManager FAM
;
237 CGSCCAnalysisManager CGAM
;
238 ModuleAnalysisManager MAM
;
240 // Register the AA manager first so that our version is the one used.
241 FAM
.registerPass([&] { return std::move(AA
); });
243 // Register all the basic analyses with the managers.
244 PB
.registerModuleAnalyses(MAM
);
245 PB
.registerCGSCCAnalyses(CGAM
);
246 PB
.registerFunctionAnalyses(FAM
);
247 PB
.registerLoopAnalyses(LAM
);
248 PB
.crossRegisterProxies(LAM
, FAM
, CGAM
, MAM
);
250 ModulePassManager MPM
;
252 // Always verify the input.
253 MPM
.addPass(VerifierPass());
255 // Now, add all the passes we've been requested to.
256 if (auto Err
= PB
.parsePassPipeline(MPM
, PipelineDesc
))
257 report_fatal_error("unable to parse pass pipeline description '" +
258 PipelineDesc
+ "': " + toString(std::move(Err
)));
261 MPM
.addPass(VerifierPass());
265 static void runOldPMPasses(Config
&Conf
, Module
&Mod
, TargetMachine
*TM
,
266 bool IsThinLTO
, ModuleSummaryIndex
*ExportSummary
,
267 const ModuleSummaryIndex
*ImportSummary
) {
268 legacy::PassManager passes
;
269 passes
.add(createTargetTransformInfoWrapperPass(TM
->getTargetIRAnalysis()));
271 PassManagerBuilder PMB
;
272 PMB
.LibraryInfo
= new TargetLibraryInfoImpl(Triple(TM
->getTargetTriple()));
273 PMB
.Inliner
= createFunctionInliningPass();
274 PMB
.ExportSummary
= ExportSummary
;
275 PMB
.ImportSummary
= ImportSummary
;
276 // Unconditionally verify input since it is not verified before this
277 // point and has unknown origin.
278 PMB
.VerifyInput
= true;
279 PMB
.VerifyOutput
= !Conf
.DisableVerify
;
280 PMB
.LoopVectorize
= true;
281 PMB
.SLPVectorize
= true;
282 PMB
.OptLevel
= Conf
.OptLevel
;
283 PMB
.PGOSampleUse
= Conf
.SampleProfile
;
284 PMB
.EnablePGOCSInstrGen
= Conf
.RunCSIRInstr
;
285 if (!Conf
.RunCSIRInstr
&& !Conf
.CSIRProfile
.empty()) {
286 PMB
.EnablePGOCSInstrUse
= true;
287 PMB
.PGOInstrUse
= Conf
.CSIRProfile
;
290 PMB
.populateThinLTOPassManager(passes
);
292 PMB
.populateLTOPassManager(passes
);
296 bool opt(Config
&Conf
, TargetMachine
*TM
, unsigned Task
, Module
&Mod
,
297 bool IsThinLTO
, ModuleSummaryIndex
*ExportSummary
,
298 const ModuleSummaryIndex
*ImportSummary
) {
299 // FIXME: Plumb the combined index into the new pass manager.
300 if (!Conf
.OptPipeline
.empty())
301 runNewPMCustomPasses(Mod
, TM
, Conf
.OptPipeline
, Conf
.AAPipeline
,
303 else if (Conf
.UseNewPM
)
304 runNewPMPasses(Conf
, Mod
, TM
, Conf
.OptLevel
, IsThinLTO
, ExportSummary
,
307 runOldPMPasses(Conf
, Mod
, TM
, IsThinLTO
, ExportSummary
, ImportSummary
);
308 return !Conf
.PostOptModuleHook
|| Conf
.PostOptModuleHook(Task
, Mod
);
311 void codegen(Config
&Conf
, TargetMachine
*TM
, AddStreamFn AddStream
,
312 unsigned Task
, Module
&Mod
) {
313 if (Conf
.PreCodeGenModuleHook
&& !Conf
.PreCodeGenModuleHook(Task
, Mod
))
316 std::unique_ptr
<ToolOutputFile
> DwoOut
;
317 SmallString
<1024> DwoFile(Conf
.SplitDwarfOutput
);
318 if (!Conf
.DwoDir
.empty()) {
320 if (auto EC
= llvm::sys::fs::create_directories(Conf
.DwoDir
))
321 report_fatal_error("Failed to create directory " + Conf
.DwoDir
+ ": " +
324 DwoFile
= Conf
.DwoDir
;
325 sys::path::append(DwoFile
, std::to_string(Task
) + ".dwo");
326 TM
->Options
.MCOptions
.SplitDwarfFile
= DwoFile
.str().str();
328 TM
->Options
.MCOptions
.SplitDwarfFile
= Conf
.SplitDwarfFile
;
330 if (!DwoFile
.empty()) {
332 DwoOut
= llvm::make_unique
<ToolOutputFile
>(DwoFile
, EC
, sys::fs::OF_None
);
334 report_fatal_error("Failed to open " + DwoFile
+ ": " + EC
.message());
337 auto Stream
= AddStream(Task
);
338 legacy::PassManager CodeGenPasses
;
339 if (TM
->addPassesToEmitFile(CodeGenPasses
, *Stream
->OS
,
340 DwoOut
? &DwoOut
->os() : nullptr,
342 report_fatal_error("Failed to setup codegen");
343 CodeGenPasses
.run(Mod
);
349 void splitCodeGen(Config
&C
, TargetMachine
*TM
, AddStreamFn AddStream
,
350 unsigned ParallelCodeGenParallelismLevel
,
351 std::unique_ptr
<Module
> Mod
) {
352 ThreadPool
CodegenThreadPool(ParallelCodeGenParallelismLevel
);
353 unsigned ThreadCount
= 0;
354 const Target
*T
= &TM
->getTarget();
357 std::move(Mod
), ParallelCodeGenParallelismLevel
,
358 [&](std::unique_ptr
<Module
> MPart
) {
359 // We want to clone the module in a new context to multi-thread the
360 // codegen. We do it by serializing partition modules to bitcode
361 // (while still on the main thread, in order to avoid data races) and
362 // spinning up new threads which deserialize the partitions into
363 // separate contexts.
364 // FIXME: Provide a more direct way to do this in LLVM.
366 raw_svector_ostream
BCOS(BC
);
367 WriteBitcodeToFile(*MPart
, BCOS
);
370 CodegenThreadPool
.async(
371 [&](const SmallString
<0> &BC
, unsigned ThreadId
) {
372 LTOLLVMContext
Ctx(C
);
373 Expected
<std::unique_ptr
<Module
>> MOrErr
= parseBitcodeFile(
374 MemoryBufferRef(StringRef(BC
.data(), BC
.size()), "ld-temp.o"),
377 report_fatal_error("Failed to read bitcode");
378 std::unique_ptr
<Module
> MPartInCtx
= std::move(MOrErr
.get());
380 std::unique_ptr
<TargetMachine
> TM
=
381 createTargetMachine(C
, T
, *MPartInCtx
);
383 codegen(C
, TM
.get(), AddStream
, ThreadId
, *MPartInCtx
);
385 // Pass BC using std::move to ensure that it get moved rather than
386 // copied into the thread's context.
387 std::move(BC
), ThreadCount
++);
391 // Because the inner lambda (which runs in a worker thread) captures our local
392 // variables, we need to wait for the worker threads to terminate before we
393 // can leave the function scope.
394 CodegenThreadPool
.wait();
397 Expected
<const Target
*> initAndLookupTarget(Config
&C
, Module
&Mod
) {
398 if (!C
.OverrideTriple
.empty())
399 Mod
.setTargetTriple(C
.OverrideTriple
);
400 else if (Mod
.getTargetTriple().empty())
401 Mod
.setTargetTriple(C
.DefaultTriple
);
404 const Target
*T
= TargetRegistry::lookupTarget(Mod
.getTargetTriple(), Msg
);
406 return make_error
<StringError
>(Msg
, inconvertibleErrorCode());
413 finalizeOptimizationRemarks(std::unique_ptr
<ToolOutputFile
> DiagOutputFile
) {
414 // Make sure we flush the diagnostic remarks file in case the linker doesn't
415 // call the global destructors before exiting.
417 return Error::success();
418 DiagOutputFile
->keep();
419 DiagOutputFile
->os().flush();
420 return Error::success();
423 Error
lto::backend(Config
&C
, AddStreamFn AddStream
,
424 unsigned ParallelCodeGenParallelismLevel
,
425 std::unique_ptr
<Module
> Mod
,
426 ModuleSummaryIndex
&CombinedIndex
) {
427 Expected
<const Target
*> TOrErr
= initAndLookupTarget(C
, *Mod
);
429 return TOrErr
.takeError();
431 std::unique_ptr
<TargetMachine
> TM
= createTargetMachine(C
, *TOrErr
, *Mod
);
433 // Setup optimization remarks.
434 auto DiagFileOrErr
= lto::setupOptimizationRemarks(
435 Mod
->getContext(), C
.RemarksFilename
, C
.RemarksPasses
, C
.RemarksFormat
,
436 C
.RemarksWithHotness
);
438 return DiagFileOrErr
.takeError();
439 auto DiagnosticOutputFile
= std::move(*DiagFileOrErr
);
441 if (!C
.CodeGenOnly
) {
442 if (!opt(C
, TM
.get(), 0, *Mod
, /*IsThinLTO=*/false,
443 /*ExportSummary=*/&CombinedIndex
, /*ImportSummary=*/nullptr))
444 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
447 if (ParallelCodeGenParallelismLevel
== 1) {
448 codegen(C
, TM
.get(), AddStream
, 0, *Mod
);
450 splitCodeGen(C
, TM
.get(), AddStream
, ParallelCodeGenParallelismLevel
,
453 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
456 static void dropDeadSymbols(Module
&Mod
, const GVSummaryMapTy
&DefinedGlobals
,
457 const ModuleSummaryIndex
&Index
) {
458 std::vector
<GlobalValue
*> DeadGVs
;
459 for (auto &GV
: Mod
.global_values())
460 if (GlobalValueSummary
*GVS
= DefinedGlobals
.lookup(GV
.getGUID()))
461 if (!Index
.isGlobalValueLive(GVS
)) {
462 DeadGVs
.push_back(&GV
);
463 convertToDeclaration(GV
);
466 // Now that all dead bodies have been dropped, delete the actual objects
467 // themselves when possible.
468 for (GlobalValue
*GV
: DeadGVs
) {
469 GV
->removeDeadConstantUsers();
470 // Might reference something defined in native object (i.e. dropped a
471 // non-prevailing IR def, but we need to keep the declaration).
473 GV
->eraseFromParent();
477 Error
lto::thinBackend(Config
&Conf
, unsigned Task
, AddStreamFn AddStream
,
478 Module
&Mod
, const ModuleSummaryIndex
&CombinedIndex
,
479 const FunctionImporter::ImportMapTy
&ImportList
,
480 const GVSummaryMapTy
&DefinedGlobals
,
481 MapVector
<StringRef
, BitcodeModule
> &ModuleMap
) {
482 Expected
<const Target
*> TOrErr
= initAndLookupTarget(Conf
, Mod
);
484 return TOrErr
.takeError();
486 std::unique_ptr
<TargetMachine
> TM
= createTargetMachine(Conf
, *TOrErr
, Mod
);
488 // Setup optimization remarks.
489 auto DiagFileOrErr
= lto::setupOptimizationRemarks(
490 Mod
.getContext(), Conf
.RemarksFilename
, Conf
.RemarksPasses
,
491 Conf
.RemarksFormat
, Conf
.RemarksWithHotness
, Task
);
493 return DiagFileOrErr
.takeError();
494 auto DiagnosticOutputFile
= std::move(*DiagFileOrErr
);
496 if (Conf
.CodeGenOnly
) {
497 codegen(Conf
, TM
.get(), AddStream
, Task
, Mod
);
498 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
501 if (Conf
.PreOptModuleHook
&& !Conf
.PreOptModuleHook(Task
, Mod
))
502 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
504 renameModuleForThinLTO(Mod
, CombinedIndex
);
506 dropDeadSymbols(Mod
, DefinedGlobals
, CombinedIndex
);
508 thinLTOResolvePrevailingInModule(Mod
, DefinedGlobals
);
510 if (Conf
.PostPromoteModuleHook
&& !Conf
.PostPromoteModuleHook(Task
, Mod
))
511 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
513 if (!DefinedGlobals
.empty())
514 thinLTOInternalizeModule(Mod
, DefinedGlobals
);
516 if (Conf
.PostInternalizeModuleHook
&&
517 !Conf
.PostInternalizeModuleHook(Task
, Mod
))
518 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
520 auto ModuleLoader
= [&](StringRef Identifier
) {
521 assert(Mod
.getContext().isODRUniquingDebugTypes() &&
522 "ODR Type uniquing should be enabled on the context");
523 auto I
= ModuleMap
.find(Identifier
);
524 assert(I
!= ModuleMap
.end());
525 return I
->second
.getLazyModule(Mod
.getContext(),
526 /*ShouldLazyLoadMetadata=*/true,
527 /*IsImporting*/ true);
530 FunctionImporter
Importer(CombinedIndex
, ModuleLoader
);
531 if (Error Err
= Importer
.importFunctions(Mod
, ImportList
).takeError())
534 if (Conf
.PostImportModuleHook
&& !Conf
.PostImportModuleHook(Task
, Mod
))
535 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
537 if (!opt(Conf
, TM
.get(), Task
, Mod
, /*IsThinLTO=*/true,
538 /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex
))
539 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));
541 codegen(Conf
, TM
.get(), AddStream
, Task
, Mod
);
542 return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile
));