Skip to content


Repository files navigation

Clang Build Analyzer Build Status

Clang C/C++ build analysis tool when using Clang 9+ -ftime-trace. The -ftime-trace compiler flag (see blog post or Clang 9 release notes) can be useful to figure out what takes time during compilation of one source file. This tool helps to aggregate time trace reports from multiple compilations, and output "what took the most time" summary:

  • Which files are slowest to parse? i.e. spend time in compiler lexer/parser front-end
  • Which C++ templates took the most time to instantiate?
  • Which files are slowest to generate code for? i.e. spend time in compiler backend doing codegen and optimizations
  • Which functions are slowest to generate code for?
  • Which header files are included the most in the whole build, how much time is spent parsing them, and what are the include chains of them?


  1. Start the build capture: ClangBuildAnalyzer --start <artifacts_folder>
    This will write current timestamp in a ClangBuildAnalyzerSession.txt file under the given artifacts_folder. The artifacts folder is where the compiled object files (and time trace report files) are expected to be produced by your build.
  2. Do your build. Does not matter how; an IDE build, a makefile, a shell script, whatever. As long as it invokes Clang and passes -ftime-trace flag to the compiler (Clang 9.0 or later is required for this).
  3. Stop the build capture: ClangBuildAnalyzer --stop <artifacts_folder> <capture_file>
    This will load all Clang time trace compatible *.json files under the given artifacts_folder that were modified after --start step was done (Clang -ftime-trace produces one JSON file next to each object file), process them and store data file into a binary capture_file.
  4. Run the build analysis: ClangBuildAnalyzer --analyze <capture_file>
    This will read the capture_file produced by --stop step, calculate the slowest things and print them. If a ClangBuildAnalyzer.ini file exists in the current folder, it will be read to control how many of various things to print.

Aternatively, instead of doing --start and --stop steps, you can do ClangBuildAnalyzer --all <artifacts_folder> <capture_file> after your build; that will include all the compatible *.json files for analysis, no matter when they were produced.

Analysis Output

The analysis output will look something like this:

Analyzing build trace from 'artifacts/FullCapture.bin'...
**** Time summary:
Compilation (7664 times):
  Parsing (frontend):         2118.9 s
  Codegen & opts (backend):   1204.1 s

**** Files that took longest to parse (compiler frontend):
  5084 ms:
  4471 ms:
  4225 ms:
  4121 ms:
**** Files that took longest to codegen (compiler backend):
 47123 ms:
 39617 ms:
 37488 ms:
 30676 ms:

**** Templates that took longest to instantiate:
 11172 ms: fmt::detail::vformat_to<char> (142 times, avg 78 ms)
  6662 ms: std::__scalar_hash<std::_PairT, 2>::operator() (3549 times, avg 1 ms)
  6281 ms: std::__murmur2_or_cityhash<unsigned long, 64>::operator() (3549 times, avg 1 ms)
  5757 ms: std::basic_string<char>::basic_string (3597 times, avg 1 ms)
  5541 ms: blender::CPPType::to_static_type_tag<float, blender::VecBase<float, ... (70 times, avg 79 ms)

**** Template sets that took longest to instantiate:
 32421 ms: std::unique_ptr<$> (30461 times, avg 1 ms)
 30098 ms: Eigen::MatrixBase<$> (8639 times, avg 3 ms)
 27524 ms: Eigen::internal::call_assignment_no_alias<$> (2397 times, avg 11 ms)

**** Functions that took longest to compile:
 28359 ms: gpu_shader_create_info_init (source/blender/gpu/intern/
  4090 ms: ccl::GetConstantValues(ccl::KernelData const*) (intern/cycles/device/metal/
  3996 ms: gpu_shader_dependency_init (source/blender/gpu/intern/

**** Function sets that took longest to compile / optimize:
 10606 ms: bool openvdb::v10_0::tree::NodeList<$>::initNodeChildren<$>(openvdb:... (470 times, avg 22 ms)
  9640 ms: void tbb::interface9::internal::dynamic_grainsize_mode<$>::work_bala... (919 times, avg 10 ms)
  9459 ms: void tbb::interface9::internal::dynamic_grainsize_mode<$>::work_bala... (715 times, avg 13 ms)
  7279 ms: blender::Vector<$>::realloc_to_at_least(long long) (1840 times, avg 3 ms)
**** Expensive headers:
261580 ms: /Developer/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/algorithm (included 3389 times, avg 77 ms), included via:
  341x: BKE_context.h BLI_string_ref.hh string 
  180x: DNA_mesh_types.h BLI_math_vector_types.hh array 
  125x: DNA_space_types.h DNA_node_types.h DNA_node_tree_interface_types.h BLI_function_ref.hh BLI_memory_utils.hh 

188777 ms: /Developer/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/string (included 3447 times, avg 54 ms), included via:
  353x: BKE_context.h BLI_string_ref.hh 
  184x: DNA_mesh_types.h BLI_offset_indices.hh BLI_index_mask.hh BLI_linear_allocator.hh BLI_string_ref.hh 
  131x: DNA_node_types.h DNA_node_tree_interface_types.h BLI_span.hh 

174792 ms: source/blender/makesdna/DNA_node_types.h (included 1653 times, avg 105 ms), included via:
  316x: ED_screen.hh DNA_space_types.h 
  181x: DNA_space_types.h 
  173x: <direct include>

Granularity and amount of most expensive things (files, functions, templates, includes) that are reported can be controlled by having an ClangBuildAnalyzer.ini file in the working directory. Take a look at ClangBuildAnalyzer.ini for an example.

Building it

  • Windows: Visual Studio 2019 solution at projects/vs2019/ClangBuildAnalyzer.sln.
  • Mac: Xcode 10.x project at projects/xcode/ClangBuildAnalyzer.xcodeproj.
  • Linux: Makefile for gcc (tested with 7.4), build with make -f projects/make/Makefile.
  • You can also use provided CMakeLists.txt, if you want to build using CMake.


  • Does not capture anything related to linking or LTO right now.
  • May or may not scale to huge builds (I haven't tried on something ginormous like a Chrome build). However I have tried it on Unity editor and Blender builds and it worked fine.


License for the Clang Build Analyzer itself is Unlicense, i.e. public domain. However, the source code includes several external library source files (all under src/external), each with their own license: