A demo of ClangBuildAnalyzer as a follow up to Ep 338 (#39)

[BenjaminNavarro]

Channel

C++Weekly

Topics

-ftime-trace was covered in the last episode (338, #39) and how to analyze the results with Chrome tracing capabilities.

But there is a tool, called ClangBuildAnalyzer, that gather the generated data to give a good, and IMO helpful, overview of the most expensive things to build. Taking from their readme:

Analyzing build trace from 'artifacts/FullCapture.bin'...
**** Time summary:
Compilation (1761 times):
  Parsing (frontend):         5167.4 s
  Codegen & opts (backend):   7576.5 s

**** Files that took longest to parse (compiler frontend):
 19524 ms: artifacts/Modules_TLS_0.o
 18046 ms: artifacts/Editor_Src_4.o
 17026 ms: artifacts/Modules_Audio_Public_1.o
 16581 ms: artifacts/Runtime_Camera_4.o
 
**** Files that took longest to codegen (compiler backend):
145761 ms: artifacts/Modules_TLS_0.o
123048 ms: artifacts/Runtime_Core_Containers_1.o
 56975 ms: artifacts/Runtime_Testing_3.o
 52031 ms: artifacts/Tools_ShaderCompiler_1.o

**** Templates that took longest to instantiate:
 19006 ms: std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::... (2665 times, avg 7 ms)
 12821 ms: std::__1::map<core::basic_string<char, core::StringStorageDefault<ch... (250 times, avg 51 ms)
  9142 ms: std::__1::map<core::basic_string<char, core::StringStorageDefault<ch... (432 times, avg 21 ms)
  8515 ms: std::__1::map<int, std::__1::pair<List<ListNode<Behaviour> > *, List... (392 times, avg 21 ms) 

**** Functions that took longest to compile:
  8710 ms: yyparse(glslang::TParseContext*) (External/ShaderCompilers/glslang/glslang/MachineIndependent/glslang_tab.cpp)
  4580 ms: LZ4HC_compress_generic_dictCtx (External/Compression/lz4/lz4hc_quarantined.c)
  4011 ms: sqlite3VdbeExec (External/sqlite/sqlite3.c)
  2737 ms: ProgressiveRuntimeManager::Update() (artifacts/Editor_Src_GI_Progressive_0.cpp)

*** Expensive headers:
136567 ms: /BuildEnvironment/MacOSX10.14.sdk/System/Library/Frameworks/Foundation.framework/Headers/Foundation.h (included 92 times, avg 1484 ms), included via:
  CocoaObjectImages.o AppKit.h  (2033 ms)
  OSXNativeWebViewWindowHelper.o OSXNativeWebViewWindowHelper.h AppKit.h  (2007 ms)
  RenderSurfaceMetal.o RenderSurfaceMetal.h MetalSupport.h Metal.h MTLTypes.h  (2003 ms)
  OSXWebViewWindowPrivate.o AppKit.h  (1959 ms)
  ...

112344 ms: Runtime/BaseClasses/BaseObject.h (included 729 times, avg 154 ms), included via:
  PairTests.cpp TestFixtures.h  (337 ms)
  Stacktrace.cpp MonoManager.h GameManager.h EditorExtension.h  (312 ms)
  PlayerPrefs.o PlayerSettings.h GameManager.h EditorExtension.h  (301 ms)
  Animation.cpp MaterialDescription.h  (299 ms)
  ...

103856 ms: Runtime/Threads/ReadWriteLock.h (included 478 times, avg 217 ms), included via:
  DownloadHandlerAssetBundle.cpp AssetBundleManager.h  (486 ms)
  LocalizationDatabase.cpp LocalizationDatabase.h LocalizationAsset.h StringTable.h  (439 ms)
  Runtime_BaseClasses_1.o MonoUtility.h ScriptingProfiler.h  (418 ms)
  ...

IIRC you can also get the templates that are the most expensive to compile, not just a particular instantiation.

Length

I guess it can be quite short since you already covered the -ftime-trace functionality in a previous episode and so maybe you could just show how to set up the tool and use it to find the same compilation hot spot in chaiscript

EDIT: missing link

[lefticus]

Good idea, I had forgotten about this tool after I got comfortable with just browsing the flame graphs