Run the resolver in parallel with the reader.

[ruiu]

Summary

This patch makes File::parse() multi-thread safe. If one thread is running
File::parse(), other threads will block if they try to call the same method.
File::parse() is idempotent, so you can safely call multiple times.

With this change, we don't have to wait for all worker threads to finish
in Driver::link(). Previously, Driver::link() calls TaskGroup::sync() to
wait for all threads running File::parse(). This was not ideal because
we couldn't start the resolver until we parse all files.

This patch increase parallelism by making Driver::link() to not wait for
worker threads. The resolver calls parse() to make sure that the file
being read has been parsed, and then uses the file. In this approach,
the resolver can run with the parser threads in parallel.

(This patch also opens up a possibility to read files from archive files
speculatively. Currently, reading member files from an archive file is
pretty slow because it's serialized, since the resolver is a single-
threaded routine. We could speed this thing up by instantiating member
files that are likely to be needed. As long as a file instantiation and
its parse() method are side-effect free, that operation does not affect
correctness. I don't do that in this patch, but will try in future.)

[atanasyan]

LGTM

[Eugene.Zelenko]

Committed in r226281.

include/lld/Core/File.h

 #include "llvm/Support/ErrorHandling.h"
 #include <functional>
 #include <memory>
+#include <mutex>
 #include <vector>
 
 namespace lld {

   /// (because YAML reader does not read blobs but structured data).
   void setLastError(std::error_code err) { _lastError = err; }
 
-  std::error_code parse() {
-    if (!_lastError.hasValue())
-      _lastError = doParse();
-    return _lastError.getValue();
-  }
+  std::error_code parse();
 
   // Usually each file owns a std::unique_ptr<MemoryBuffer>.
   // However, there's one special case. If a file is an archive file,

   static atom_collection_empty<UndefinedAtom>     _noUndefinedAtoms;
   static atom_collection_empty<SharedLibraryAtom> _noSharedLibraryAtoms;
   static atom_collection_empty<AbsoluteAtom>      _noAbsoluteAtoms;
-  llvm::Optional<std::error_code>                 _lastError;
   mutable llvm::BumpPtrAllocator                  _allocator;
 
 private:

   Kind              _kind;
   mutable uint64_t  _ordinal;
   std::shared_ptr<MemoryBuffer> _sharedMemoryBuffer;
+  llvm::Optional<std::error_code> _lastError;
+  std::mutex _parseMutex;
 };
 
 /// \brief A mutable File.

include/lld/Core/LinkingContext.h

 #include "lld/Core/Error.h"
 #include "lld/Core/LLVM.h"
 #include "lld/Core/Node.h"
+#include "lld/Core/Parallel.h"
 #include "lld/Core/Reference.h"
 #include "lld/Core/range.h"
 #include "lld/ReaderWriter/Reader.h"

   // Derived classes may use that chance to rearrange the input files.
   virtual void maybeSortInputFiles() {}
 
+  TaskGroup &getTaskGroup() { return _taskGroup; }
+
   /// @}
 protected:
   LinkingContext(); // Must be subclassed

 private:
   /// Validate the subclass bits. Only called by validate.
   virtual bool validateImpl(raw_ostream &diagnostics) = 0;
+  TaskGroup _taskGroup;
 };
 
 } // end namespace lld

include/lld/Core/Resolver.h

   void maybeAddSectionGroupOrGnuLinkOnce(const DefinedAtom &atom);
 
   /// \brief The main function that iterates over the files to resolve
-  void resolveUndefines();
+  bool resolveUndefines();
   void updateReferences();
   void deadStripOptimize();
   bool checkUndefines();

lib/Core/File.cpp

 
 #include "lld/Core/File.h"
 #include "lld/Core/LLVM.h"
+#include <mutex>
 
 namespace lld {
 

 File::atom_collection_empty<SharedLibraryAtom> File::_noSharedLibraryAtoms;
 File::atom_collection_empty<AbsoluteAtom>      File::_noAbsoluteAtoms;
 
+std::error_code File::parse() {
+  std::lock_guard<std::mutex> lock(_parseMutex);
+  if (!_lastError.hasValue())
+    _lastError = doParse();
+  return _lastError.getValue();
+}
+
 } // namespace lld

lib/Core/Resolver.cpp

   for (int i = begin; i < end; ++i)
     if (FileNode *node = dyn_cast<FileNode>(inputs[i].get()))
       if (_newUndefinesAdded[node->getFile()])
         return true;
   return false;
 }
 

 
 // Keep adding atoms until _context.getNextFile() returns an error. This
 // function is where undefined atoms are resolved.
-void Resolver::resolveUndefines() {
+bool Resolver::resolveUndefines() {
   ScopedTask task(getDefaultDomain(), "resolveUndefines");
   int index = 0;
   int groupLevel = 0;

     bool undefAdded = false;
     File *file = getFile(index, groupLevel);
     if (!file)
-      return;
+      return true;
+    if (std::error_code ec = file->parse()) {
+      llvm::errs() << "Cannot open " + file->path()
+                   << ": " << ec.message() << "\n";
+      return false;
+    }
     switch (file->kind()) {
     case File::kindObject:
       if (groupLevel > 0)

 }
 
 bool Resolver::resolve() {
-  resolveUndefines();
+  if (!resolveUndefines())
+    return false;
   updateReferences();
   deadStripOptimize();
   if (checkUndefines())

lib/Driver/Driver.cpp

   if (context.getNodes().empty())
     return false;
 
-  bool fail = false;
-
-  // Read inputs
-  ScopedTask readTask(getDefaultDomain(), "Read Args");
-  TaskGroup tg;
-  std::mutex diagnosticsMutex;
-  for (std::unique_ptr<Node> &ie : context.getNodes()) {
-    tg.spawn([&] {
-      // Writes to the same output stream is not guaranteed to be thread-safe.
-      // We buffer the diagnostics output to a separate string-backed output
-      // stream, acquire the lock, and then print it out.
-      std::string buf;
-      llvm::raw_string_ostream stream(buf);
-
-      if (FileNode *node = dyn_cast<FileNode>(ie.get())) {
-       if (File *file = node->getFile()) {
-         if (std::error_code ec = file->parse()) {
-           stream << "Cannot open " + file->path()
-                  << ": " << ec.message() << "\n";
-           fail = true;
-         }
-       }
-      }
-
-      stream.flush();
-      if (!buf.empty()) {
-        std::lock_guard<std::mutex> lock(diagnosticsMutex);
-        diagnostics << buf;
-      }
-    });
-  }
-  tg.sync();
-  readTask.end();
-
-  if (fail)
-    return false;
+  for (std::unique_ptr<Node> &ie : context.getNodes())
+    if (FileNode *node = dyn_cast<FileNode>(ie.get()))
+      context.getTaskGroup().spawn([node] { node->getFile()->parse(); });
 
   std::vector<std::unique_ptr<File>> internalFiles;
   context.createInternalFiles(internalFiles);

lib/ReaderWriter/ELF/ELFFile.h

 
 public:
   ELFFile(StringRef name)
-      : File(name, kindObject), _ordinal(0), _doStringsMerge(false) {}
+      : File(name, kindObject), _ordinal(0), _doStringsMerge(false) {
+    setLastError(std::error_code());
+  }
 
   ELFFile(std::unique_ptr<MemoryBuffer> mb, bool atomizeStrings = false)
       : File(mb->getBufferIdentifier(), kindObject), _mb(std::move(mb)),

lib/ReaderWriter/PECOFF/PECOFFLinkingContext.cpp

       "<command line option /include>");
 }
 
+static int getGroupStartPos(std::vector<std::unique_ptr<Node>> &nodes) {
+  for (int i = 0, e = nodes.size(); i < e; ++i)
+    if (GroupEnd *group = dyn_cast<GroupEnd>(nodes[i].get()))
+      return i - group->getSize();
+  llvm::report_fatal_error("internal error");
+}
+
 void PECOFFLinkingContext::addLibraryFile(std::unique_ptr<FileNode> file) {
   GroupEnd *currentGroupEnd;
   int pos = -1;

   // Create a file for the entry point function.
   std::unique_ptr<FileNode> entry(new FileNode(
       llvm::make_unique<pecoff::EntryPointFile>(*this, syms)));
-  members.insert(members.begin(), std::move(entry));
+  members.insert(members.begin() + getGroupStartPos(members), std::move(entry));
 
   // Create a file for __ImageBase.
   std::unique_ptr<FileNode> fileNode(new FileNode(

 void pecoff::ResolvableSymbols::readAllSymbols() {
   std::lock_guard<std::mutex> lock(_mutex);
   for (File *file : _queue) {
+    if (file->parse())
+      return;
     if (auto *archive = dyn_cast<ArchiveLibraryFile>(file)) {
       for (const std::string &sym : archive->getDefinedSymbols())
        _defined.insert(sym);