Make WriteIndexesThinBackend multi threaded (#109847)

We've noticed that for large builds executing thin-link can take on the
order of 10s of minutes. We are only using a single thread to write the
sharded indices and import files for each input bitcode file. While we
need to ensure the index file produced lists modules in a deterministic
order, that doesn't prevent us from executing the rest of the work in
parallel.

In this change we use a thread pool to execute as much of the backend's
work as possible in parallel. In local testing on a machine with 80
cores, this change makes a thin-link for ~100,000 input files run in ~2
minutes. Without this change it takes upwards of 10 minutes.

---------

Co-authored-by: Nuri Amari <[email protected]>

Author: NuriAmari

lld/COFF/LTO.cpp

@@ -118,6 +118,7 @@ BitcodeCompiler::BitcodeCompiler(COFFLinkerContext &c) : ctx(c) {
   if (ctx.config.thinLTOIndexOnly) {
     auto OnIndexWrite = [&](StringRef S) { thinIndices.erase(S); };
     backend = lto::createWriteIndexesThinBackend(
+        llvm::hardware_concurrency(ctx.config.thinLTOJobs),
         std::string(ctx.config.thinLTOPrefixReplaceOld),
         std::string(ctx.config.thinLTOPrefixReplaceNew),
         std::string(ctx.config.thinLTOPrefixReplaceNativeObject),

lld/ELF/LTO.cpp

@@ -179,6 +179,7 @@ BitcodeCompiler::BitcodeCompiler(Ctx &ctx) : ctx(ctx) {
   auto onIndexWrite = [&](StringRef s) { thinIndices.erase(s); };
   if (ctx.arg.thinLTOIndexOnly) {
     backend = lto::createWriteIndexesThinBackend(
+        llvm::hardware_concurrency(ctx.arg.thinLTOJobs),
         std::string(ctx.arg.thinLTOPrefixReplaceOld),
         std::string(ctx.arg.thinLTOPrefixReplaceNew),
         std::string(ctx.arg.thinLTOPrefixReplaceNativeObject),

lld/MachO/LTO.cpp

@@ -87,6 +87,7 @@ BitcodeCompiler::BitcodeCompiler() {
   auto onIndexWrite = [&](StringRef S) { thinIndices.erase(S); };
   if (config->thinLTOIndexOnly) {
     backend = lto::createWriteIndexesThinBackend(
+        llvm::hardware_concurrency(config->thinLTOJobs),
         std::string(config->thinLTOPrefixReplaceOld),
         std::string(config->thinLTOPrefixReplaceNew),
         std::string(config->thinLTOPrefixReplaceNativeObject),

lld/test/COFF/thinlto-emit-imports.ll

@@ -35,7 +35,7 @@
 ; RUN: not lld-link -entry:main -thinlto-index-only \
 ; RUN:     -thinlto-emit-imports-files %t1.obj %t2.obj %t3.obj \
 ; RUN:     -out:%t4.exe 2>&1 | FileCheck -DMSG=%errc_EACCES %s --check-prefix=ERR
-; ERR: cannot open {{.*}}3.obj.imports: [[MSG]]
+; ERR: 'cannot open {{.*}}3.obj.imports': [[MSG]]
 
 ; Ensure lld doesn't generate import files when thinlto-index-only is not enabled
 ; RUN: rm -f %t1.obj.imports

lld/test/ELF/lto/thinlto-cant-write-index.ll

@@ -10,7 +10,7 @@
 ; RUN: chmod u-w %t2.o.thinlto.bc
 ; RUN: not ld.lld --plugin-opt=thinlto-index-only -shared %t1.o %t2.o -o /dev/null 2>&1 | FileCheck -DMSG=%errc_EACCES %s
 ; RUN: chmod u+w %t2.o.thinlto.bc
-; CHECK: cannot open {{.*}}2.o.thinlto.bc: [[MSG]]
+; CHECK: 'cannot open {{.*}}2.o.thinlto.bc': [[MSG]]
 
 target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"

lld/test/ELF/lto/thinlto-emit-imports.ll

@@ -10,7 +10,7 @@
 ; RUN: touch %t3.o.imports
 ; RUN: chmod 400 %t3.o.imports
 ; RUN: not ld.lld --plugin-opt=thinlto-index-only --plugin-opt=thinlto-emit-imports-files -shared %t1.o %t2.o %t3.o -o /dev/null 2>&1 | FileCheck -DMSG=%errc_EACCES %s --check-prefix=ERR
-; ERR: cannot open {{.*}}3.o.imports: [[MSG]]
+; ERR: 'cannot open {{.*}}3.o.imports': [[MSG]]
 
 ; RUN: rm -f %t1.o.imports %t2.o.imports rm -f %t3.o.imports
 ; RUN: ld.lld --plugin-opt=thinlto-emit-imports-files -shared %t1.o %t2.o %t3.o -o %t4

lld/test/MachO/thinlto-emit-imports.ll

@@ -33,7 +33,7 @@
 ; RUN: chmod 400 %t3.o.imports
 ; RUN: not %lld --thinlto-index-only --thinlto-emit-imports-files -dylib %t1.o %t2.o %t3.o -o /dev/null 2>&1 \
 ; RUN:     | FileCheck -DMSG=%errc_EACCES %s --check-prefix=ERR
-; ERR: cannot open {{.*}}3.o.imports: [[MSG]]
+; ERR: 'cannot open {{.*}}3.o.imports': [[MSG]]
 
 ; Ensure lld doesn't generate import files when thinlto-index-only is not enabled
 ; RUN: rm -f %t1.o.imports

llvm/include/llvm/LTO/LTO.h

@@ -231,7 +231,8 @@ ThinBackend createInProcessThinBackend(ThreadPoolStrategy Parallelism,
 /// the objects with NativeObjectPrefix instead of NewPrefix. OnWrite is
 /// callback which receives module identifier and notifies LTO user that index
 /// file for the module (and optionally imports file) was created.
-ThinBackend createWriteIndexesThinBackend(std::string OldPrefix,
+ThinBackend createWriteIndexesThinBackend(ThreadPoolStrategy Parallelism,
+                                          std::string OldPrefix,
                                           std::string NewPrefix,
                                           std::string NativeObjectPrefix,
                                           bool ShouldEmitImportsFiles,

llvm/include/llvm/Support/Threading.h

@@ -188,6 +188,18 @@ constexpr bool llvm_is_multithreaded() { return LLVM_ENABLE_THREADS; }
     return S;
   }
 
+  /// Like hardware_concurrency() above, but builds a strategy
+  /// based on the rules described for get_threadpool_strategy().
+  /// If \p Num is invalid, returns a default strategy where one thread per
+  /// hardware core is used.
+  inline ThreadPoolStrategy hardware_concurrency(StringRef Num) {
+    std::optional<ThreadPoolStrategy> S =
+        get_threadpool_strategy(Num, hardware_concurrency());
+    if (S)
+      return *S;
+    return hardware_concurrency();
+  }
+
   /// Returns an optimal thread strategy to execute specified amount of tasks.
   /// This strategy should prevent us from creating too many threads if we
   /// occasionaly have an unexpectedly small amount of tasks.

llvm/include/llvm/Transforms/IPO/FunctionImport.h

@@ -417,9 +417,9 @@ void gatherImportedSummariesForModule(
     GVSummaryPtrSet &DecSummaries);
 
 /// Emit into \p OutputFilename the files module \p ModulePath will import from.
-std::error_code
-EmitImportsFiles(StringRef ModulePath, StringRef OutputFilename,
-                 const ModuleToSummariesForIndexTy &ModuleToSummariesForIndex);
+Error EmitImportsFiles(
+    StringRef ModulePath, StringRef OutputFilename,
+    const ModuleToSummariesForIndexTy &ModuleToSummariesForIndex);
 
 /// Based on the information recorded in the summaries during global
 /// summary-based analysis:

llvm/lib/LTO/LTO.cpp

@@ -1376,15 +1376,20 @@ class lto::ThinBackendProc {
   const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries;
   lto::IndexWriteCallback OnWrite;
   bool ShouldEmitImportsFiles;
+  DefaultThreadPool BackendThreadPool;
+  std::optional<Error> Err;
+  std::mutex ErrMu;
 
 public:
   ThinBackendProc(
       const Config &Conf, ModuleSummaryIndex &CombinedIndex,
       const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
-      lto::IndexWriteCallback OnWrite, bool ShouldEmitImportsFiles)
+      lto::IndexWriteCallback OnWrite, bool ShouldEmitImportsFiles,
+      ThreadPoolStrategy ThinLTOParallelism)
       : Conf(Conf), CombinedIndex(CombinedIndex),
         ModuleToDefinedGVSummaries(ModuleToDefinedGVSummaries),
-        OnWrite(OnWrite), ShouldEmitImportsFiles(ShouldEmitImportsFiles) {}
+        OnWrite(OnWrite), ShouldEmitImportsFiles(ShouldEmitImportsFiles),
+        BackendThreadPool(ThinLTOParallelism) {}
 
   virtual ~ThinBackendProc() = default;
   virtual Error start(
@@ -1393,13 +1398,19 @@ class lto::ThinBackendProc {
       const FunctionImporter::ExportSetTy &ExportList,
       const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
       MapVector<StringRef, BitcodeModule> &ModuleMap) = 0;
-  virtual Error wait() = 0;
-  virtual unsigned getThreadCount() = 0;
+  Error wait() {
+    BackendThreadPool.wait();
+    if (Err)
+      return std::move(*Err);
+    return Error::success();
+  }
+  unsigned getThreadCount() { return BackendThreadPool.getMaxConcurrency(); }
+  virtual bool isSensitiveToInputOrder() { return false; }
 
   // Write sharded indices and (optionally) imports to disk
   Error emitFiles(const FunctionImporter::ImportMapTy &ImportList,
                   llvm::StringRef ModulePath,
-                  const std::string &NewModulePath) {
+                  const std::string &NewModulePath) const {
     ModuleToSummariesForIndexTy ModuleToSummariesForIndex;
     GVSummaryPtrSet DeclarationSummaries;
 
@@ -1411,32 +1422,29 @@ class lto::ThinBackendProc {
     raw_fd_ostream OS(NewModulePath + ".thinlto.bc", EC,
                       sys::fs::OpenFlags::OF_None);
     if (EC)
-      return errorCodeToError(EC);
+      return createFileError("cannot open " + NewModulePath + ".thinlto.bc",
+                             EC);
 
     writeIndexToFile(CombinedIndex, OS, &ModuleToSummariesForIndex,
                      &DeclarationSummaries);
 
     if (ShouldEmitImportsFiles) {
-      EC = EmitImportsFiles(ModulePath, NewModulePath + ".imports",
-                            ModuleToSummariesForIndex);
-      if (EC)
-        return errorCodeToError(EC);
+      Error ImportFilesError = EmitImportsFiles(
+          ModulePath, NewModulePath + ".imports", ModuleToSummariesForIndex);
+      if (ImportFilesError)
+        return ImportFilesError;
     }
     return Error::success();
   }
 };
 
 namespace {
 class InProcessThinBackend : public ThinBackendProc {
-  DefaultThreadPool BackendThreadPool;
   AddStreamFn AddStream;
   FileCache Cache;
   DenseSet<GlobalValue::GUID> CfiFunctionDefs;
   DenseSet<GlobalValue::GUID> CfiFunctionDecls;
 
-  std::optional<Error> Err;
-  std::mutex ErrMu;
-
   bool ShouldEmitIndexFiles;
 
 public:
@@ -1447,9 +1455,9 @@ class InProcessThinBackend : public ThinBackendProc {
       AddStreamFn AddStream, FileCache Cache, lto::IndexWriteCallback OnWrite,
       bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles)
       : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
-                        OnWrite, ShouldEmitImportsFiles),
-        BackendThreadPool(ThinLTOParallelism), AddStream(std::move(AddStream)),
-        Cache(std::move(Cache)), ShouldEmitIndexFiles(ShouldEmitIndexFiles) {
+                        OnWrite, ShouldEmitImportsFiles, ThinLTOParallelism),
+        AddStream(std::move(AddStream)), Cache(std::move(Cache)),
+        ShouldEmitIndexFiles(ShouldEmitIndexFiles) {
     for (auto &Name : CombinedIndex.cfiFunctionDefs())
       CfiFunctionDefs.insert(
           GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
@@ -1546,18 +1554,6 @@ class InProcessThinBackend : public ThinBackendProc {
       OnWrite(std::string(ModulePath));
     return Error::success();
   }
-
-  Error wait() override {
-    BackendThreadPool.wait();
-    if (Err)
-      return std::move(*Err);
-    else
-      return Error::success();
-  }
-
-  unsigned getThreadCount() override {
-    return BackendThreadPool.getMaxConcurrency();
-  }
 };
 } // end anonymous namespace
 
@@ -1618,12 +1614,13 @@ class WriteIndexesThinBackend : public ThinBackendProc {
 public:
   WriteIndexesThinBackend(
       const Config &Conf, ModuleSummaryIndex &CombinedIndex,
+      ThreadPoolStrategy ThinLTOParallelism,
       const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
       std::string OldPrefix, std::string NewPrefix,
       std::string NativeObjectPrefix, bool ShouldEmitImportsFiles,
       raw_fd_ostream *LinkedObjectsFile, lto::IndexWriteCallback OnWrite)
       : ThinBackendProc(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
-                        OnWrite, ShouldEmitImportsFiles),
+                        OnWrite, ShouldEmitImportsFiles, ThinLTOParallelism),
         OldPrefix(OldPrefix), NewPrefix(NewPrefix),
         NativeObjectPrefix(NativeObjectPrefix),
         LinkedObjectsFile(LinkedObjectsFile) {}
@@ -1635,9 +1632,11 @@ class WriteIndexesThinBackend : public ThinBackendProc {
       const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
       MapVector<StringRef, BitcodeModule> &ModuleMap) override {
     StringRef ModulePath = BM.getModuleIdentifier();
-    std::string NewModulePath =
-        getThinLTOOutputFile(ModulePath, OldPrefix, NewPrefix);
 
+    // The contents of this file may be used as input to a native link, and must
+    // therefore contain the processed modules in a determinstic order that
+    // match the order they are provided on the command line. For that reason,
+    // we cannot include this in the asynchronously executed lambda below.
     if (LinkedObjectsFile) {
       std::string ObjectPrefix =
           NativeObjectPrefix.empty() ? NewPrefix : NativeObjectPrefix;
@@ -1646,33 +1645,49 @@ class WriteIndexesThinBackend : public ThinBackendProc {
       *LinkedObjectsFile << LinkedObjectsFilePath << '\n';
     }
 
-    if (auto E = emitFiles(ImportList, ModulePath, NewModulePath))
-      return E;
+    BackendThreadPool.async(
+        [this](const StringRef ModulePath,
+               const FunctionImporter::ImportMapTy &ImportList,
+               const std::string &OldPrefix, const std::string &NewPrefix) {
+          std::string NewModulePath =
+              getThinLTOOutputFile(ModulePath, OldPrefix, NewPrefix);
+          auto E = emitFiles(ImportList, ModulePath, NewModulePath);
+          if (E) {
+            std::unique_lock<std::mutex> L(ErrMu);
+            if (Err)
+              Err = joinErrors(std::move(*Err), std::move(E));
+            else
+              Err = std::move(E);
+            return;
+          }
+        },
+        ModulePath, ImportList, OldPrefix, NewPrefix);
 
     if (OnWrite)
       OnWrite(std::string(ModulePath));
     return Error::success();
   }
 
-  Error wait() override { return Error::success(); }
-
-  // WriteIndexesThinBackend should always return 1 to prevent module
-  // re-ordering and avoid non-determinism in the final link.
-  unsigned getThreadCount() override { return 1; }
+  bool isSensitiveToInputOrder() override {
+    // The order which modules are written to LinkedObjectsFile should be
+    // deterministic and match the order they are passed on the command line.
+    return true;
+  }
 };
 } // end anonymous namespace
 
 ThinBackend lto::createWriteIndexesThinBackend(
-    std::string OldPrefix, std::string NewPrefix,
-    std::string NativeObjectPrefix, bool ShouldEmitImportsFiles,
-    raw_fd_ostream *LinkedObjectsFile, IndexWriteCallback OnWrite) {
+    ThreadPoolStrategy Parallelism, std::string OldPrefix,
+    std::string NewPrefix, std::string NativeObjectPrefix,
+    bool ShouldEmitImportsFiles, raw_fd_ostream *LinkedObjectsFile,
+    IndexWriteCallback OnWrite) {
   return
       [=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
           const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
           AddStreamFn AddStream, FileCache Cache) {
         return std::make_unique<WriteIndexesThinBackend>(
-            Conf, CombinedIndex, ModuleToDefinedGVSummaries, OldPrefix,
-            NewPrefix, NativeObjectPrefix, ShouldEmitImportsFiles,
+            Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
+            OldPrefix, NewPrefix, NativeObjectPrefix, ShouldEmitImportsFiles,
             LinkedObjectsFile, OnWrite);
       };
 }
@@ -1854,7 +1869,8 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
           ResolvedODR[Mod.first], ThinLTO.ModuleMap);
     };
 
-    if (BackendProcess->getThreadCount() == 1) {
+    if (BackendProcess->getThreadCount() == 1 ||
+        BackendProcess->isSensitiveToInputOrder()) {
       // Process the modules in the order they were provided on the
       // command-line. It is important for this codepath to be used for
       // WriteIndexesThinBackend, to ensure the emitted LinkedObjectsFile lists

llvm/lib/LTO/ThinLTOCodeGenerator.cpp

@@ -837,9 +837,8 @@ void ThinLTOCodeGenerator::emitImports(Module &TheModule, StringRef OutputName,
       ModuleIdentifier, ModuleToDefinedGVSummaries,
       ImportLists[ModuleIdentifier], ModuleToSummariesForIndex, DecSummaries);
 
-  std::error_code EC;
-  if ((EC = EmitImportsFiles(ModuleIdentifier, OutputName,
-                             ModuleToSummariesForIndex)))
+  if (Error EC = EmitImportsFiles(ModuleIdentifier, OutputName,
+                                  ModuleToSummariesForIndex))
     report_fatal_error(Twine("Failed to open ") + OutputName +
                        " to save imports lists\n");
 }

llvm/lib/Transforms/IPO/FunctionImport.cpp

@@ -1553,20 +1553,21 @@ void llvm::gatherImportedSummariesForModule(
 }
 
 /// Emit the files \p ModulePath will import from into \p OutputFilename.
-std::error_code llvm::EmitImportsFiles(
+Error llvm::EmitImportsFiles(
     StringRef ModulePath, StringRef OutputFilename,
     const ModuleToSummariesForIndexTy &ModuleToSummariesForIndex) {
   std::error_code EC;
   raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::OF_Text);
   if (EC)
-    return EC;
+    return createFileError("cannot open " + OutputFilename,
+                           errorCodeToError(EC));
   for (const auto &ILI : ModuleToSummariesForIndex)
     // The ModuleToSummariesForIndex map includes an entry for the current
     // Module (needed for writing out the index files). We don't want to
     // include it in the imports file, however, so filter it out.
     if (ILI.first != ModulePath)
       ImportsOS << ILI.first << "\n";
-  return std::error_code();
+  return Error::success();
 }
 
 bool llvm::convertToDeclaration(GlobalValue &GV) {

llvm/tools/gold/gold-plugin.cpp

@@ -899,7 +899,7 @@ static std::unique_ptr<LTO> createLTO(IndexWriteCallback OnIndexWrite,
     std::string OldPrefix, NewPrefix;
     getThinLTOOldAndNewPrefix(OldPrefix, NewPrefix);
     Backend = createWriteIndexesThinBackend(
-        OldPrefix, NewPrefix,
+        llvm::hardware_concurrency(options::Parallelism) OldPrefix, NewPrefix,
         // TODO: Add support for optional native object path in
         // thinlto_prefix_replace option to match lld.
         /*NativeObjectPrefix=*/"", options::thinlto_emit_imports_files,

llvm/tools/llvm-lto2/llvm-lto2.cpp

@@ -346,7 +346,8 @@ static int run(int argc, char **argv) {
 
   ThinBackend Backend;
   if (ThinLTODistributedIndexes)
-    Backend = createWriteIndexesThinBackend(/*OldPrefix=*/"",
+    Backend = createWriteIndexesThinBackend(llvm::hardware_concurrency(Threads),
+                                            /*OldPrefix=*/"",
                                             /*NewPrefix=*/"",
                                             /*NativeObjectPrefix=*/"",
                                             ThinLTOEmitImports,