[clang-doc] Merge data into persistent memory (#190056)

We have a need for persistent memory for the final info. Since each group processes a single USR at a time, every USR is only ever processed by a single thread from the thread pool. This means that we can keep per thread persistent storage for all the info. There is significant duplicated data between all the serialized records, so we can just merge the final/unique items into the persistent arena, and clear out the scratch/transient arena as we process each record in the bitcode. The patch adds some APIs to help with managing the data, merging, and allocation of data in the correct arena. It also safely merges and deep copies data from the transient arenas into persistent storage that is never reset until the program completes. This patch reduces memory by another % over the previous patches, bringing the total savings over the baseline to 57%. Runtime performance and benchmarks stay mostly flat with modest improvements. | Metric | Baseline | Prev | This | Culm% | Seq% | | :--- | :--- | :--- | :--- | :--- | :--- | | Time | 920.5s | 991.5s | 987.2s | +7.2% | -0.4% | | Memory | 86.0G | 40.0G | 36.9G | -57.1% | -8.0% | | Benchmark | Baseline | Prev | This | Culm% | Seq% | | :--- | :--- | :--- | :--- | :--- | :--- | | BM_BitcodeReader_Scale/10 | 67.9us | 72.2us | 72.2us | +6.3% | -0.0% | | BM_BitcodeReader_Scale/10000 | 70.5ms | 22.5ms | 17.3ms | -75.5% |-23.2% | | BM_BitcodeReader_Scale/4096 | 23.2ms | 6.6ms | 7.1ms | -69.5% | +7.4%| | BM_BitcodeReader_Scale/512 | 509.4us | 898.7us | 550.5us | +8.1% |-38.7% | | BM_BitcodeReader_Scale/64 | 114.8us | 133.7us | 120.8us | +5.2% |-9.6% | | BM_EmitInfoFunction | 1.6us | 1.9us | 1.8us | +12.9% | -3.3% | | BM_Index_Insertion/10 | 2.3us | 4.1us | 3.5us | +52.4% | -14.7% | | BM_Index_Insertion/10000 | 3.1ms | 5.3ms | 4.8ms | +53.5% | -10.0% | | BM_Index_Insertion/4096 | 1.3ms | 2.1ms | 1.9ms | +50.8% | -9.2% | | BM_Index_Insertion/512 | 153.6us | 251.8us | 227.0us | +47.8% | -9.9%| | BM_Index_Insertion/64 | 18.1us | 30.2us | 26.7us | +47.9% | -11.7% | | BM_JSONGenerator_Scale/10000 | 89.6ms | 81.4ms | 83.4ms | -6.9% |+2.5% | | BM_JSONGenerator_Scale/4096 | 33.7ms | 31.0ms | 32.4ms | -3.9% | +4.5%| | BM_Mapper_Scale/10000 | 104.3ms | 112.3ms | 103.5ms | -0.8% | -7.9% | | BM_Mapper_Scale/4096 | 44.3ms | 45.0ms | 43.8ms | -1.2% | -2.5% | | BM_Mapper_Scale/512 | 7.6ms | 7.7ms | 7.5ms | -1.3% | -2.4% | | BM_Mapper_Scale/64 | 3.1ms | 3.0ms | 3.0ms | -1.9% | -0.3% | | BM_MergeInfos_Scale/10000 | 12.2ms | 575.6us | 500.1us | -95.9% |-13.1% | | BM_MergeInfos_Scale/2 | 1.9us | 1.8us | 1.8us | -4.4% | -1.7% | | BM_MergeInfos_Scale/4096 | 2.8ms | 205.3us | 200.4us | -92.8% | -2.4%| | BM_MergeInfos_Scale/512 | 68.9us | 20.5us | 19.5us | -71.7% | -5.1% | | BM_MergeInfos_Scale/64 | 10.3us | 3.8us | 4.0us | -60.9% | +4.8% | | BM_MergeInfos_Scale/8 | 2.8us | 1.9us | 1.9us | -31.7% | -1.8% | | BM_SerializeFunctionInfo | 25.5us | 25.8us | 26.1us | +2.2% | +1.3% |
2026-04-10 15:32:50 -07:00
parent 96b1ae7461
commit 21e0034c69
3 changed files with 166 additions and 35 deletions
--- a/clang-tools-extra/clang-doc/Representation.cpp
+++ b/clang-tools-extra/clang-doc/Representation.cpp
@@ -115,7 +115,31 @@ static void reduceChildren(llvm::simple_ilist<T> &Children,
    auto It = llvm::find_if(
        Children, [&](const T &C) { return C.USR == ChildToMerge->USR; });
    if (It == Children.end()) {
-      T *NewChild = allocatePtr<T>(PersistentArena, std::move(*ChildToMerge));
+      T *NewChild = allocatePtr<T>(PersistentArena, ChildToMerge->USR);
+      NewChild->merge(std::move(*ChildToMerge));
+      Children.push_back(*NewChild);
+    } else {
+      It->merge(std::move(*ChildToMerge));
+    }
+  }
+}
+
+template <>
+void reduceChildren<Reference>(
+    llvm::simple_ilist<Reference> &Children,
+    llvm::simple_ilist<Reference> &&ChildrenToMerge) {
+  while (!ChildrenToMerge.empty()) {
+    Reference *ChildToMerge = &ChildrenToMerge.front();
+    ChildrenToMerge.pop_front();
+
+    auto It = llvm::find_if(Children, [&](const Reference &C) {
+      return C.USR == ChildToMerge->USR;
+    });
+    if (It == Children.end()) {
+      Reference *NewChild = allocatePtr<Reference>(PersistentArena);
+      NewChild->USR = ChildToMerge->USR;
+      NewChild->RefType = ChildToMerge->RefType;
+      NewChild->merge(std::move(*ChildToMerge));
      Children.push_back(*NewChild);
    } else {
      It->merge(std::move(*ChildToMerge));
@@ -130,12 +154,106 @@ static void mergeUnkeyed(Container &Target, Container &&Source) {
    auto &Item = Source.front();
    Source.pop_front();
    if (llvm::none_of(Target, [&](const auto &E) { return E == Item; })) {
-      T *NewItem = allocatePtr<T>(PersistentArena, std::move(Item));
+      T *NewItem = allocatePtr<T>(PersistentArena, Item);
      Target.push_back(*NewItem);
    }
  }
 }

+template <>
+void mergeUnkeyed<OwningVec<CommentInfo>>(OwningVec<CommentInfo> &Target,
+                                          OwningVec<CommentInfo> &&Source) {
+  while (!Source.empty()) {
+    auto &Item = Source.front();
+    Source.pop_front();
+    if (llvm::none_of(Target, [&](const auto &E) { return E == Item; })) {
+      CommentInfo *NewItem =
+          allocatePtr<CommentInfo>(PersistentArena, Item, PersistentArena);
+      Target.push_back(*NewItem);
+    }
+  }
+}
+
+llvm::Error mergeSingleInfo(doc::OwnedPtr<doc::Info> &Reduced,
+                            doc::OwnedPtr<doc::Info> &&NewInfo,
+                            llvm::BumpPtrAllocator &Arena) {
+  if (!Reduced) {
+    switch (NewInfo->IT) {
+    case InfoType::IT_namespace:
+      Reduced = allocatePtr<NamespaceInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_record:
+      Reduced = allocatePtr<RecordInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_enum:
+      Reduced = allocatePtr<EnumInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_function:
+      Reduced = allocatePtr<FunctionInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_typedef:
+      Reduced = allocatePtr<TypedefInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_concept:
+      Reduced = allocatePtr<ConceptInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_variable:
+      Reduced = allocatePtr<VarInfo>(Arena, NewInfo->USR);
+      break;
+    case InfoType::IT_friend:
+      Reduced = allocatePtr<FriendInfo>(Arena, NewInfo->USR);
+      break;
+    default:
+      return llvm::createStringError(llvm::inconvertibleErrorCode(),
+                                     "unknown info type");
+    }
+  }
+
+  if (Reduced->IT != NewInfo->IT)
+    return llvm::createStringError(llvm::inconvertibleErrorCode(),
+                                   "info types mismatch");
+
+  switch (Reduced->IT) {
+  case InfoType::IT_namespace:
+    static_cast<NamespaceInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<NamespaceInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_record:
+    static_cast<RecordInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<RecordInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_enum:
+    static_cast<EnumInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<EnumInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_function:
+    static_cast<FunctionInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<FunctionInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_typedef:
+    static_cast<TypedefInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<TypedefInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_concept:
+    static_cast<ConceptInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<ConceptInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_variable:
+    static_cast<VarInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<VarInfo *>(getPtr(NewInfo))));
+    break;
+  case InfoType::IT_friend:
+    static_cast<FriendInfo *>(getPtr(Reduced))
+        ->merge(std::move(*static_cast<FriendInfo *>(getPtr(NewInfo))));
+    break;
+  default:
+    return llvm::createStringError(llvm::inconvertibleErrorCode(),
+                                   "unknown info type");
+  }
+
+  return llvm::Error::success();
+}
+
 // Dispatch function.
 llvm::Expected<OwnedPtr<Info>> mergeInfos(OwningPtrArray<Info> &Values) {
  if (Values.empty() || !Values[0])
@@ -293,6 +411,8 @@ void Reference::merge(Reference &&Other) {
    Name = Other.Name;
  if (Path.empty())
    Path = Other.Path;
+  if (QualName.empty())
+    QualName = Other.QualName;
  if (DocumentationFileName.empty())
    DocumentationFileName = Other.DocumentationFileName;
 }
@@ -340,8 +460,8 @@ void Info::mergeBase(Info &&Other) {
    Name = Other.Name;
  if (Path == "")
    Path = Other.Path;
-  if (Namespace.empty())
-    Namespace = std::move(Other.Namespace);
+  if (Namespace.empty() && !Other.Namespace.empty())
+    Namespace = allocateArray(Other.Namespace, PersistentArena);
  // Unconditionally extend the description, since each decl may have a comment.
  mergeUnkeyed(Description, std::move(Other.Description));
  if (ParentUSR == EmptySID)
@@ -374,6 +494,8 @@ void SymbolInfo::merge(SymbolInfo &&Other) {
  mergeBase(std::move(Other));
  if (MangledName.empty())
    MangledName = std::move(Other.MangledName);
+  if (!IsStatic)
+    IsStatic = Other.IsStatic;
 }

 NamespaceInfo::NamespaceInfo(SymbolID USR, StringRef Name, StringRef Path)
@@ -442,8 +564,8 @@ void RecordInfo::merge(RecordInfo &&Other) {
  reduceChildren(Children.Enums, std::move(Other.Children.Enums));
  reduceChildren(Children.Typedefs, std::move(Other.Children.Typedefs));
  SymbolInfo::merge(std::move(Other));
-  if (!Template)
-    Template = Other.Template;
+  if (!Template && Other.Template)
+    Template = TemplateInfo(*Other.Template, PersistentArena);
 }

 EnumValueInfo::EnumValueInfo(const EnumValueInfo &Other,
@@ -461,6 +583,8 @@ void EnumInfo::merge(EnumInfo &&Other) {
  assert(mergeable(Other));
  if (!Scoped)
    Scoped = Other.Scoped;
+  if (!BaseType && Other.BaseType)
+    BaseType = std::move(Other.BaseType);
  if (Members.empty() && !Other.Members.empty())
    Members = deepCopyArray(Other.Members, PersistentArena);
  SymbolInfo::merge(std::move(Other));
@@ -479,8 +603,8 @@ void FunctionInfo::merge(FunctionInfo &&Other) {
  if (Params.empty() && !Other.Params.empty())
    Params = allocateArray(Other.Params, PersistentArena);
  SymbolInfo::merge(std::move(Other));
-  if (!Template)
-    Template = Other.Template;
+  if (!Template && Other.Template)
+    Template = TemplateInfo(*Other.Template, PersistentArena);
 }

 void TypedefInfo::merge(TypedefInfo &&Other) {
@@ -489,8 +613,8 @@ void TypedefInfo::merge(TypedefInfo &&Other) {
    IsUsing = Other.IsUsing;
  if (Underlying.Type.Name == "")
    Underlying = Other.Underlying;
-  if (!Template)
-    Template = Other.Template;
+  if (!Template && Other.Template)
+    Template = TemplateInfo(*Other.Template, PersistentArena);
  SymbolInfo::merge(std::move(Other));
 }

--- a/clang-tools-extra/clang-doc/Representation.h
+++ b/clang-tools-extra/clang-doc/Representation.h
@@ -52,6 +52,7 @@ private:
 ConcurrentStringPool &getGlobalStringPool();

 extern thread_local llvm::BumpPtrAllocator TransientArena;
+extern thread_local llvm::BumpPtrAllocator PersistentArena;

 inline StringRef internString(const Twine &T) {
  if (T.isTriviallyEmpty())
@@ -776,6 +777,12 @@ struct Index : public Reference {
 // if they are different.
 llvm::Expected<OwnedPtr<Info>> mergeInfos(OwningPtrArray<Info> &Values);

+// Merges a single new Info into an existing Reduced Info (allocating it if
+// needed).
+llvm::Error mergeSingleInfo(doc::OwnedPtr<doc::Info> &Reduced,
+                            doc::OwnedPtr<doc::Info> &&NewInfo,
+                            llvm::BumpPtrAllocator &Arena);
+
 struct ClangDocContext {
  ClangDocContext(tooling::ExecutionContext *ECtx, StringRef ProjectName,
                  bool PublicOnly, StringRef OutDirectory, StringRef SourceRoot,
--- a/clang-tools-extra/clang-doc/tool/ClangDocMain.cpp
+++ b/clang-tools-extra/clang-doc/tool/ClangDocMain.cpp
@@ -29,6 +29,7 @@
 #include "clang/Tooling/CommonOptionsParser.h"
 #include "clang/Tooling/Execution.h"
 #include "llvm/ADT/APFloat.h"
+#include "llvm/ADT/ScopeExit.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Error.h"
 #include "llvm/Support/FileSystem.h"
@@ -358,15 +359,22 @@ Example usage for a project using a compile commands database:
        llvm::hardware_concurrency(ExecutorConcurrency));
    {
      llvm::TimeTraceScope TS("Reduce");
-      for (auto &Group : USRToBitcode) {
-        Pool.async([&, &Diags = Diags]() { // time trace decoding bitcode
-          if (FTimeTrace)
+      for (const auto &Group : USRToBitcode) {
+        StringRef Key = Group.getKey();
+        std::vector<StringRef> Bitcodes = Group.getValue();
+        Pool.async([Key, Bitcodes, &CDCtx, &Diags, &USRToInfo, &USRToInfoMutex,
+                    &IndexMutex, &DiagMutex, &Error, DiagIDBitcodeReading,
+                    DiagIDBitcodeMerging]() {
+          if (CDCtx.FTimeTrace)
            llvm::timeTraceProfilerInitialize(200, "clang-doc");

-          doc::OwningPtrVec<doc::Info> Infos;
+          doc::OwnedPtr<doc::Info> Reduced = nullptr;
          {
-            llvm::TimeTraceScope Red("decoding bitcode");
-            for (auto &Bitcode : Group.getValue()) {
+            llvm::TimeTraceScope Red("decoding and merging bitcode");
+            for (const auto &Bitcode : Bitcodes) {
+
+              llvm::scope_exit ArenaGuard(
+                  [] { clang::doc::TransientArena.Reset(); });
              llvm::BitstreamCursor Stream(Bitcode);
              doc::ClangDocBitcodeReader Reader(Stream, Diags);
              auto ReadInfos = Reader.readBitcode();
@@ -378,25 +386,17 @@ Example usage for a project using a compile commands database:
                Error = true;
                return;
              }
-              std::move(ReadInfos->begin(), ReadInfos->end(),
-                        std::back_inserter(Infos));
+              for (auto &I : *ReadInfos) {
+                if (auto Err = doc::mergeSingleInfo(
+                        Reduced, std::move(I), clang::doc::PersistentArena)) {
+                  std::lock_guard<llvm::sys::Mutex> Guard(DiagMutex);
+                  Diags.Report(DiagIDBitcodeMerging)
+                      << toString(std::move(Err));
+                  return;
+                }
+              }
            }
-          } // time trace decoding bitcode
-
-          doc::OwnedPtr<doc::Info> Reduced;
-
-          {
-            llvm::TimeTraceScope Merge("merging bitcode");
-            auto ExpReduced = doc::mergeInfos(Infos);
-
-            if (!ExpReduced) {
-              std::lock_guard<llvm::sys::Mutex> Guard(DiagMutex);
-              Diags.Report(DiagIDBitcodeMerging)
-                  << toString(ExpReduced.takeError());
-              return;
-            }
-            Reduced = std::move(*ExpReduced);
-          } // time trace merging bitcode
+          } // time trace decoding and merging bitcode

          // Add a reference to this Info in the Index
          {
@@ -408,7 +408,7 @@ Example usage for a project using a compile commands database:
          {
            llvm::TimeTraceScope Merge("USRToInfo");
            std::lock_guard<llvm::sys::Mutex> Guard(USRToInfoMutex);
-            USRToInfo[Group.getKey()] = std::move(Reduced);
+            USRToInfo[Key] = std::move(Reduced);
          }

          if (CDCtx.FTimeTrace)