[ThinLTO] Import composite types as declarations

Summary:
When reading the metadata bitcode, create a type declaration when
possible for composite types when we are importing. Doing this in
the bitcode reader saves memory. Also it works naturally in the case
when the type ODR map contains a definition for the same composite type
because it was used in the importing module (buildODRType will
automatically use the existing definition and not create a type
declaration).

For Chromium built with -g2, this reduces the aggregate size of the
generated native object files by 66% (from 31G to 10G). It reduced
the time through the ThinLTO link and backend phases by about 20% on
my machine.

Reviewers: mehdi_amini, dblaikie, aprantl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27775

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289993 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Bitcode/BitcodeReader.h

@@ -61,9 +61,10 @@ namespace llvm {
     friend Expected<std::vector<BitcodeModule>>
     getBitcodeModuleList(MemoryBufferRef Buffer);
 
-    Expected<std::unique_ptr<Module>>
-    getModuleImpl(LLVMContext &Context, bool MaterializeAll,
-                  bool ShouldLazyLoadMetadata);
+    Expected<std::unique_ptr<Module>> getModuleImpl(LLVMContext &Context,
+                                                    bool MaterializeAll,
+                                                    bool ShouldLazyLoadMetadata,
+                                                    bool IsImporting);
 
   public:
     StringRef getBuffer() const {
@@ -74,8 +75,11 @@ namespace llvm {
 
     /// Read the bitcode module and prepare for lazy deserialization of function
     /// bodies. If ShouldLazyLoadMetadata is true, lazily load metadata as well.
-    Expected<std::unique_ptr<Module>>
-    getLazyModule(LLVMContext &Context, bool ShouldLazyLoadMetadata);
+    /// If IsImporting is true, this module is being parsed for ThinLTO
+    /// importing into another module.
+    Expected<std::unique_ptr<Module>> getLazyModule(LLVMContext &Context,
+                                                    bool ShouldLazyLoadMetadata,
+                                                    bool IsImporting);
 
     /// Read the entire bitcode module and return it.
     Expected<std::unique_ptr<Module>> parseModule(LLVMContext &Context);
@@ -93,18 +97,20 @@ namespace llvm {
 
   /// Read the header of the specified bitcode buffer and prepare for lazy
   /// deserialization of function bodies. If ShouldLazyLoadMetadata is true,
-  /// lazily load metadata as well.
+  /// lazily load metadata as well. If IsImporting is true, this module is
+  /// being parsed for ThinLTO importing into another module.
   Expected<std::unique_ptr<Module>>
   getLazyBitcodeModule(MemoryBufferRef Buffer, LLVMContext &Context,
-                       bool ShouldLazyLoadMetadata = false);
+                       bool ShouldLazyLoadMetadata = false,
+                       bool IsImporting = false);
 
   /// Like getLazyBitcodeModule, except that the module takes ownership of
   /// the memory buffer if successful. If successful, this moves Buffer. On
-  /// error, this *does not* move Buffer.
-  Expected<std::unique_ptr<Module>>
-  getOwningLazyBitcodeModule(std::unique_ptr<MemoryBuffer> &&Buffer,
-                             LLVMContext &Context,
-                             bool ShouldLazyLoadMetadata = false);
+  /// error, this *does not* move Buffer. If IsImporting is true, this module is
+  /// being parsed for ThinLTO importing into another module.
+  Expected<std::unique_ptr<Module>> getOwningLazyBitcodeModule(
+      std::unique_ptr<MemoryBuffer> &&Buffer, LLVMContext &Context,
+      bool ShouldLazyLoadMetadata = false, bool IsImporting = false);
 
   /// Read the header of the specified bitcode buffer and extract just the
   /// triple information. If successful, this returns a string. On error, this

lib/Bitcode/Reader/BitcodeReader.cpp

@@ -488,7 +488,8 @@ class BitcodeReader : public BitcodeReaderBase, public GVMaterializer {
 
   /// \brief Main interface to parsing a bitcode buffer.
   /// \returns true if an error occurred.
-  Error parseBitcodeInto(Module *M, bool ShouldLazyLoadMetadata = false);
+  Error parseBitcodeInto(Module *M, bool ShouldLazyLoadMetadata = false,
+                         bool IsImporting = false);
 
   static uint64_t decodeSignRotatedValue(uint64_t V);
 
@@ -3084,9 +3085,10 @@ Error BitcodeReader::parseModule(uint64_t ResumeBit,
   }
 }
 
-Error BitcodeReader::parseBitcodeInto(Module *M, bool ShouldLazyLoadMetadata) {
+Error BitcodeReader::parseBitcodeInto(Module *M, bool ShouldLazyLoadMetadata,
+                                      bool IsImporting) {
   TheModule = M;
-  MDLoader = MetadataLoader(Stream, *M, ValueList,
+  MDLoader = MetadataLoader(Stream, *M, ValueList, IsImporting,
                             [&](unsigned ID) { return getTypeByID(ID); });
   return parseModule(0, ShouldLazyLoadMetadata);
 }
@@ -5220,7 +5222,7 @@ llvm::getBitcodeModuleList(MemoryBufferRef Buffer) {
 /// everything.
 Expected<std::unique_ptr<Module>>
 BitcodeModule::getModuleImpl(LLVMContext &Context, bool MaterializeAll,
-                             bool ShouldLazyLoadMetadata) {
+                             bool ShouldLazyLoadMetadata, bool IsImporting) {
   BitstreamCursor Stream(Buffer);
 
   std::string ProducerIdentification;
@@ -5243,7 +5245,8 @@ BitcodeModule::getModuleImpl(LLVMContext &Context, bool MaterializeAll,
   M->setMaterializer(R);
 
   // Delay parsing Metadata if ShouldLazyLoadMetadata is true.
-  if (Error Err = R->parseBitcodeInto(M.get(), ShouldLazyLoadMetadata))
+  if (Error Err =
+          R->parseBitcodeInto(M.get(), ShouldLazyLoadMetadata, IsImporting))
     return std::move(Err);
 
   if (MaterializeAll) {
@@ -5259,9 +5262,9 @@ BitcodeModule::getModuleImpl(LLVMContext &Context, bool MaterializeAll,
 }
 
 Expected<std::unique_ptr<Module>>
-BitcodeModule::getLazyModule(LLVMContext &Context,
-                             bool ShouldLazyLoadMetadata) {
-  return getModuleImpl(Context, false, ShouldLazyLoadMetadata);
+BitcodeModule::getLazyModule(LLVMContext &Context, bool ShouldLazyLoadMetadata,
+                             bool IsImporting) {
+  return getModuleImpl(Context, false, ShouldLazyLoadMetadata, IsImporting);
 }
 
 // Parse the specified bitcode buffer, returning the function info index.
@@ -5323,28 +5326,28 @@ static Expected<BitcodeModule> getSingleModule(MemoryBufferRef Buffer) {
 }
 
 Expected<std::unique_ptr<Module>>
-llvm::getLazyBitcodeModule(MemoryBufferRef Buffer,
-                           LLVMContext &Context, bool ShouldLazyLoadMetadata) {
+llvm::getLazyBitcodeModule(MemoryBufferRef Buffer, LLVMContext &Context,
+                           bool ShouldLazyLoadMetadata, bool IsImporting) {
   Expected<BitcodeModule> BM = getSingleModule(Buffer);
   if (!BM)
     return BM.takeError();
 
-  return BM->getLazyModule(Context, ShouldLazyLoadMetadata);
+  return BM->getLazyModule(Context, ShouldLazyLoadMetadata, IsImporting);
 }
 
-Expected<std::unique_ptr<Module>>
-llvm::getOwningLazyBitcodeModule(std::unique_ptr<MemoryBuffer> &&Buffer,
-                                 LLVMContext &Context,
-                                 bool ShouldLazyLoadMetadata) {
-  auto MOrErr = getLazyBitcodeModule(*Buffer, Context, ShouldLazyLoadMetadata);
+Expected<std::unique_ptr<Module>> llvm::getOwningLazyBitcodeModule(
+    std::unique_ptr<MemoryBuffer> &&Buffer, LLVMContext &Context,
+    bool ShouldLazyLoadMetadata, bool IsImporting) {
+  auto MOrErr = getLazyBitcodeModule(*Buffer, Context, ShouldLazyLoadMetadata,
+                                     IsImporting);
   if (MOrErr)
     (*MOrErr)->setOwnedMemoryBuffer(std::move(Buffer));
   return MOrErr;
 }
 
 Expected<std::unique_ptr<Module>>
 BitcodeModule::parseModule(LLVMContext &Context) {
-  return getModuleImpl(Context, true, false);
+  return getModuleImpl(Context, true, false, false);
   // TODO: Restore the use-lists to the in-memory state when the bitcode was
   // written.  We must defer until the Module has been fully materialized.
 }

lib/Bitcode/Reader/MetadataLoader.cpp

@@ -86,6 +86,12 @@
 
 using namespace llvm;
 
+/// Flag whether we need to import full type definitions for ThinLTO.
+/// Currently needed for Darwin and LLDB.
+static cl::opt<bool> ImportFullTypeDefinitions(
+    "import-full-type-definitions", cl::init(false), cl::Hidden,
+    cl::desc("Import full type definitions for ThinLTO."));
+
 namespace {
 
 static int64_t unrotateSign(uint64_t U) { return U & 1 ? ~(U >> 1) : U >> 1; }
@@ -399,7 +405,7 @@ class MetadataLoader::MetadataLoaderImpl {
         Stream(Stream), Context(TheModule.getContext()), TheModule(TheModule),
         getTypeByID(getTypeByID) {}
 
-  Error parseMetadata(bool ModuleLevel);
+  Error parseMetadata(bool ModuleLevel, bool IsImporting);
 
   bool hasFwdRefs() const { return MetadataList.hasFwdRefs(); }
   Metadata *getMetadataFwdRef(unsigned Idx) {
@@ -435,7 +441,8 @@ Error error(const Twine &Message) {
 
 /// Parse a METADATA_BLOCK. If ModuleLevel is true then we are parsing
 /// module level metadata.
-Error MetadataLoader::MetadataLoaderImpl::parseMetadata(bool ModuleLevel) {
+Error MetadataLoader::MetadataLoaderImpl::parseMetadata(bool ModuleLevel,
+                                                        bool IsImporting) {
   if (!ModuleLevel && MetadataList.hasFwdRefs())
     return error("Invalid metadata: fwd refs into function blocks");
 
@@ -709,18 +716,38 @@ Error MetadataLoader::MetadataLoaderImpl::parseMetadata(bool ModuleLevel) {
       Metadata *File = getMDOrNull(Record[3]);
       unsigned Line = Record[4];
       Metadata *Scope = getDITypeRefOrNull(Record[5]);
-      Metadata *BaseType = getDITypeRefOrNull(Record[6]);
+      Metadata *BaseType = nullptr;
       uint64_t SizeInBits = Record[7];
       if (Record[8] > (uint64_t)std::numeric_limits<uint32_t>::max())
         return error("Alignment value is too large");
       uint32_t AlignInBits = Record[8];
-      uint64_t OffsetInBits = Record[9];
+      uint64_t OffsetInBits = 0;
       DINode::DIFlags Flags = static_cast<DINode::DIFlags>(Record[10]);
-      Metadata *Elements = getMDOrNull(Record[11]);
+      Metadata *Elements = nullptr;
       unsigned RuntimeLang = Record[12];
-      Metadata *VTableHolder = getDITypeRefOrNull(Record[13]);
-      Metadata *TemplateParams = getMDOrNull(Record[14]);
+      Metadata *VTableHolder = nullptr;
+      Metadata *TemplateParams = nullptr;
       auto *Identifier = getMDString(Record[15]);
+      // If this module is being parsed so that it can be ThinLTO imported
+      // into another module, composite types only need to be imported
+      // as type declarations (unless full type definitions requested).
+      // Create type declarations up front to save memory. Also, buildODRType
+      // handles the case where this is type ODRed with a definition needed
+      // by the importing module, in which case the existing definition is
+      // used.
+      if (IsImporting && !ImportFullTypeDefinitions &&
+          (Tag == dwarf::DW_TAG_enumeration_type ||
+           Tag == dwarf::DW_TAG_class_type ||
+           Tag == dwarf::DW_TAG_structure_type ||
+           Tag == dwarf::DW_TAG_union_type)) {
+        Flags = Flags | DINode::FlagFwdDecl;
+      } else {
+        BaseType = getDITypeRefOrNull(Record[6]);
+        OffsetInBits = Record[9];
+        Elements = getMDOrNull(Record[11]);
+        VTableHolder = getDITypeRefOrNull(Record[13]);
+        TemplateParams = getMDOrNull(Record[14]);
+      }
       DICompositeType *CT = nullptr;
       if (Identifier)
         CT = DICompositeType::buildODRType(
@@ -1281,17 +1308,19 @@ MetadataLoader &MetadataLoader::operator=(MetadataLoader &&RHS) {
   return *this;
 }
 MetadataLoader::MetadataLoader(MetadataLoader &&RHS)
-    : Pimpl(std::move(RHS.Pimpl)) {}
+    : Pimpl(std::move(RHS.Pimpl)), IsImporting(RHS.IsImporting) {}
 
 MetadataLoader::~MetadataLoader() = default;
 MetadataLoader::MetadataLoader(BitstreamCursor &Stream, Module &TheModule,
                                BitcodeReaderValueList &ValueList,
+                               bool IsImporting,
                                std::function<Type *(unsigned)> getTypeByID)
     : Pimpl(llvm::make_unique<MetadataLoaderImpl>(Stream, TheModule, ValueList,
-                                                  getTypeByID)) {}
+                                                  getTypeByID)),
+      IsImporting(IsImporting) {}
 
 Error MetadataLoader::parseMetadata(bool ModuleLevel) {
-  return Pimpl->parseMetadata(ModuleLevel);
+  return Pimpl->parseMetadata(ModuleLevel, IsImporting);
 }
 
 bool MetadataLoader::hasFwdRefs() const { return Pimpl->hasFwdRefs(); }

lib/Bitcode/Reader/MetadataLoader.h

@@ -36,12 +36,14 @@ class Type;
 class MetadataLoader {
   class MetadataLoaderImpl;
   std::unique_ptr<MetadataLoaderImpl> Pimpl;
+  /// True if metadata is being parsed for a module being ThinLTO imported.
+  bool IsImporting = false;
   Error parseMetadata(bool ModuleLevel);
 
 public:
   ~MetadataLoader();
   MetadataLoader(BitstreamCursor &Stream, Module &TheModule,
-                 BitcodeReaderValueList &ValueList,
+                 BitcodeReaderValueList &ValueList, bool IsImporting,
                  std::function<Type *(unsigned)> getTypeByID);
   MetadataLoader &operator=(MetadataLoader &&);
   MetadataLoader(MetadataLoader &&);

lib/LTO/LTO.cpp

@@ -252,7 +252,8 @@ Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) {
   // ModuleSymbolTable.
   for (auto BM : *BMsOrErr) {
     Expected<std::unique_ptr<Module>> MOrErr =
-        BM.getLazyModule(File->Ctx, /*ShouldLazyLoadMetadata*/ true);
+        BM.getLazyModule(File->Ctx, /*ShouldLazyLoadMetadata*/ true,
+                         /*IsImporting*/ false);
     if (!MOrErr)
       return MOrErr.takeError();
 
@@ -415,7 +416,8 @@ Error LTO::addRegularLTO(BitcodeModule BM, const SymbolResolution *&ResI,
     RegularLTO.Mover = llvm::make_unique<IRMover>(*RegularLTO.CombinedModule);
   }
   Expected<std::unique_ptr<Module>> MOrErr =
-      BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true);
+      BM.getLazyModule(RegularLTO.Ctx, /*ShouldLazyLoadMetadata*/ true,
+                       /*IsImporting*/ false);
   if (!MOrErr)
     return MOrErr.takeError();
 

lib/LTO/LTOBackend.cpp

@@ -356,7 +356,8 @@ Error lto::thinBackend(Config &Conf, unsigned Task, AddStreamFn AddStream,
     auto I = ModuleMap.find(Identifier);
     assert(I != ModuleMap.end());
     return I->second.getLazyModule(Mod.getContext(),
-                                   /*ShouldLazyLoadMetadata=*/true);
+                                   /*ShouldLazyLoadMetadata=*/true,
+                                   /*IsImporting*/ true);
   };
 
   FunctionImporter Importer(CombinedIndex, ModuleLoader);

lib/LTO/ThinLTOCodeGenerator.cpp

@@ -168,12 +168,13 @@ static void promoteModule(Module &TheModule, const ModuleSummaryIndex &Index) {
 
 static std::unique_ptr<Module>
 loadModuleFromBuffer(const MemoryBufferRef &Buffer, LLVMContext &Context,
-                     bool Lazy) {
+                     bool Lazy, bool IsImporting) {
   SMDiagnostic Err;
   Expected<std::unique_ptr<Module>> ModuleOrErr =
-      Lazy ? getLazyBitcodeModule(Buffer, Context,
-                                  /* ShouldLazyLoadMetadata */ true)
-           : parseBitcodeFile(Buffer, Context);
+      Lazy
+          ? getLazyBitcodeModule(Buffer, Context,
+                                 /* ShouldLazyLoadMetadata */ true, IsImporting)
+          : parseBitcodeFile(Buffer, Context);
   if (!ModuleOrErr) {
     handleAllErrors(ModuleOrErr.takeError(), [&](ErrorInfoBase &EIB) {
       SMDiagnostic Err = SMDiagnostic(Buffer.getBufferIdentifier(),
@@ -191,7 +192,7 @@ crossImportIntoModule(Module &TheModule, const ModuleSummaryIndex &Index,
                       const FunctionImporter::ImportMapTy &ImportList) {
   auto Loader = [&](StringRef Identifier) {
     return loadModuleFromBuffer(ModuleMap[Identifier], TheModule.getContext(),
-                                /*Lazy=*/true);
+                                /*Lazy=*/true, /*IsImporting*/ true);
   };
 
   FunctionImporter Importer(Index, Loader);
@@ -787,7 +788,8 @@ void ThinLTOCodeGenerator::run() {
         Context.setDiscardValueNames(LTODiscardValueNames);
 
         // Parse module now
-        auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false);
+        auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false,
+                                              /*IsImporting*/ false);
 
         // CodeGen
         ProducedBinaries[count] = codegen(*TheModule);
@@ -933,7 +935,8 @@ void ThinLTOCodeGenerator::run() {
         }
 
         // Parse module now
-        auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false);
+        auto TheModule = loadModuleFromBuffer(ModuleBuffer, Context, false,
+                                              /*IsImporting*/ false);
 
         // Save temps: original file.
         saveTempBitcode(*TheModule, SaveTempsDir, count, ".0.original.bc");

lib/Object/IRObjectFile.cpp

@@ -127,7 +127,8 @@ IRObjectFile::create(MemoryBufferRef Object, LLVMContext &Context) {
   std::vector<std::unique_ptr<Module>> Mods;
   for (auto BM : *BMsOrErr) {
     Expected<std::unique_ptr<Module>> MOrErr =
-        BM.getLazyModule(Context, /*ShouldLazyLoadMetadata*/ true);
+        BM.getLazyModule(Context, /*ShouldLazyLoadMetadata*/ true,
+                         /*IsImporting*/ false);
     if (!MOrErr)
       return MOrErr.takeError();
 

test/ThinLTO/X86/Inputs/debuginfo-compositetype-import.ll

@@ -0,0 +1,13 @@
+; ModuleID = 'debuginfo-compositetype-import2.c'
+source_filename = "debuginfo-compositetype-import2.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind uwtable
+define i32 @main() {
+entry:
+  call void (...) @foo()
+  ret i32 0
+}
+
+declare void @foo(...) #1