[Lldb-commits] [PATCH] D122975: parallelize module loading in DynamicLoaderPOSIXDYLD()

[Luboš Luňák via Phabricator via lldb-commits]

llunak created this revision.
llunak added reviewers: clayborg, labath.
llunak added a project: LLDB.
Herald added subscribers: usaxena95, JDevlieghere, kadircet.
Herald added a project: All.
llunak requested review of this revision.
Herald added subscribers: lldb-commits, ilya-biryukov.

If LLDB index cache is enabled and everything is cached, then loading of debug info is essentially single-threaded. This patch parallelizes module loading in DynamicLoaderPOSIXDYLD::LoadAllCurrentModules(), which may greatly reduce the load time if the debugged program uses a large number of binaries (as opposed to monolithic programs where this presumably doesn't make a difference). In my specific case of LibreOffice Calc this reduces startup time from 6s to 2s.

One problem with this change is that it creates a threadpool-within-threadpool situation in ManualDWARFIndex::Index(), and so the number of threads running at the same time may not be properly limited. I think the threadpool in ManualDWARFIndex::Index() is still useful even with this change because of the case of a monolithic program that's not cached.

If not limiting thread count properly is considered a problem, I think a solution to that is using a global semaphore to limit tasks from both places. There's a simple Semaphore class in clangd (with clangd-specific trace code, so I guess it'd need to be copy&pasted into lldb) and the semaphore object would need to be stored somewhere (and I don't know where).

I have not checked if all the relevant code is thread-safe, but since all of it is already running in another thread I assume that to be the case.

Presumably a similar change could be done for other platforms (I have no plans to do so).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D122975

Files:
  lldb/source/Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.cpp


Index: lldb/source/Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.cpp
===================================================================
--- lldb/source/Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.cpp
+++ lldb/source/Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.cpp
@@ -24,6 +24,7 @@
 #include "lldb/Utility/LLDBLog.h"
 #include "lldb/Utility/Log.h"
 #include "lldb/Utility/ProcessInfo.h"
+#include "llvm/Support/ThreadPool.h"
 
 #include <memory>
 
@@ -602,9 +603,24 @@
   m_process->PrefetchModuleSpecs(
       module_names, m_process->GetTarget().GetArchitecture().GetTriple());
 
-  for (I = m_rendezvous.begin(), E = m_rendezvous.end(); I != E; ++I) {
-    ModuleSP module_sp =
+  std::vector<DYLDRendezvous::iterator> infos;
+  for (I = m_rendezvous.begin(), E = m_rendezvous.end(); I != E; ++I)
+    infos.push_back(I);
+  std::vector<ModuleSP> loaded_modules;
+  loaded_modules.resize(infos.size());
+  llvm::ThreadPool pool(llvm::optimal_concurrency(infos.size()));
+  auto load_module_fn = [&](size_t idx) {
+    DYLDRendezvous::iterator I = infos[idx];
+    loaded_modules[idx] =
         LoadModuleAtAddress(I->file_spec, I->link_addr, I->base_addr, true);
+  };
+  for (size_t i = 0; i < infos.size(); ++i)
+    pool.async(load_module_fn, i);
+  pool.wait();
+
+  for (size_t i = 0; i < infos.size(); ++i) {
+    DYLDRendezvous::iterator I = infos[i];
+    ModuleSP module_sp = loaded_modules[i];
     if (module_sp.get()) {
       LLDB_LOG(log, "LoadAllCurrentModules loading module: {0}",
                I->file_spec.GetFilename());


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D122975.419973.patch
Type: text/x-patch
Size: 1582 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/lldb-commits/attachments/20220402/bdffce99/attachment-0001.bin>