123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163 |
- =============================================================
- How To Build Clang and LLVM with Profile-Guided Optimizations
- =============================================================
- Introduction
- ============
- PGO (Profile-Guided Optimization) allows your compiler to better optimize code
- for how it actually runs. Users report that applying this to Clang and LLVM can
- decrease overall compile time by 20%.
- This guide walks you through how to build Clang with PGO, though it also applies
- to other subprojects, such as LLD.
- Using the script
- ================
- We have a script at ``utils/collect_and_build_with_pgo.py``. This script is
- tested on a few Linux flavors, and requires a checkout of LLVM, Clang, and
- compiler-rt. Despite the the name, it performs four clean builds of Clang, so it
- can take a while to run to completion. Please see the script's ``--help`` for
- more information on how to run it, and the different options available to you.
- If you want to get the most out of PGO for a particular use-case (e.g. compiling
- a specific large piece of software), please do read the section below on
- 'benchmark' selection.
- Please note that this script is only tested on a few Linux distros. Patches to
- add support for other platforms, as always, are highly appreciated. :)
- This script also supports a ``--dry-run`` option, which causes it to print
- important commands instead of running them.
- Selecting 'benchmarks'
- ======================
- PGO does best when the profiles gathered represent how the user plans to use the
- compiler. Notably, highly accurate profiles of llc building x86_64 code aren't
- incredibly helpful if you're going to be targeting ARM.
- By default, the script above does two things to get solid coverage. It:
- - runs all of Clang and LLVM's lit tests, and
- - uses the instrumented Clang to build Clang, LLVM, and all of the other
- LLVM subprojects available to it.
- Together, these should give you:
- - solid coverage of building C++,
- - good coverage of building C,
- - great coverage of running optimizations,
- - great coverage of the backend for your host's architecture, and
- - some coverage of other architectures (if other arches are supported backends).
- Altogether, this should cover a diverse set of uses for Clang and LLVM. If you
- have very specific needs (e.g. your compiler is meant to compile a large browser
- for four different platforms, or similar), you may want to do something else.
- This is configurable in the script itself.
- Building Clang with PGO
- =======================
- If you prefer to not use the script, this briefly goes over how to build
- Clang/LLVM with PGO.
- First, you should have at least LLVM, Clang, and compiler-rt checked out
- locally.
- Next, at a high level, you're going to need to do the following:
- 1. Build a standard Release Clang and the relevant libclang_rt.profile library
- 2. Build Clang using the Clang you built above, but with instrumentation
- 3. Use the instrumented Clang to generate profiles, which consists of two steps:
- - Running the instrumented Clang/LLVM/lld/etc. on tasks that represent how
- users will use said tools.
- - Using a tool to convert the "raw" profiles generated above into a single,
- final PGO profile.
- 4. Build a final release Clang (along with whatever other binaries you need)
- using the profile collected from your benchmark
- In more detailed steps:
- 1. Configure a Clang build as you normally would. It's highly recommended that
- you use the Release configuration for this, since it will be used to build
- another Clang. Because you need Clang and supporting libraries, you'll want
- to build the ``all`` target (e.g. ``ninja all`` or ``make -j4 all``).
- 2. Configure a Clang build as above, but add the following CMake args:
- - ``-DLLVM_BUILD_INSTRUMENTED=IR`` -- This causes us to build everything
- with instrumentation.
- - ``-DLLVM_BUILD_RUNTIME=No`` -- A few projects have bad interactions when
- built with profiling, and aren't necessary to build. This flag turns them
- off.
- - ``-DCMAKE_C_COMPILER=/path/to/stage1/clang`` - Use the Clang we built in
- step 1.
- - ``-DCMAKE_CXX_COMPILER=/path/to/stage1/clang++`` - Same as above.
- In this build directory, you simply need to build the ``clang`` target (and
- whatever supporting tooling your benchmark requires).
- 3. As mentioned above, this has two steps: gathering profile data, and then
- massaging it into a useful form:
- a. Build your benchmark using the Clang generated in step 2. The 'standard'
- benchmark recommended is to run ``check-clang`` and ``check-llvm`` in your
- instrumented Clang's build directory, and to do a full build of Clang/LLVM
- using your instrumented Clang. So, create yet another build directory,
- with the following CMake arguments:
- - ``-DCMAKE_C_COMPILER=/path/to/stage2/clang`` - Use the Clang we built in
- step 2.
- - ``-DCMAKE_CXX_COMPILER=/path/to/stage2/clang++`` - Same as above.
- If your users are fans of debug info, you may want to consider using
- ``-DCMAKE_BUILD_TYPE=RelWithDebInfo`` instead of
- ``-DCMAKE_BUILD_TYPE=Release``. This will grant better coverage of
- debug info pieces of clang, but will take longer to complete and will
- result in a much larger build directory.
- It's recommended to build the ``all`` target with your instrumented Clang,
- since more coverage is often better.
- b. You should now have a few ``*.profraw`` files in
- ``path/to/stage2/profiles/``. You need to merge these using
- ``llvm-profdata`` (even if you only have one! The profile merge transforms
- profraw into actual profile data, as well). This can be done with
- ``/path/to/stage1/llvm-profdata merge
- -output=/path/to/output/profdata.prof path/to/stage2/profiles/*.profraw``.
- 4. Now, build your final, PGO-optimized Clang. To do this, you'll want to pass
- the following additional arguments to CMake.
- - ``-DLLVM_PROFDATA_FILE=/path/to/output/profdata.prof`` - Use the PGO
- profile from the previous step.
- - ``-DCMAKE_C_COMPILER=/path/to/stage1/clang`` - Use the Clang we built in
- step 1.
- - ``-DCMAKE_CXX_COMPILER=/path/to/stage1/clang++`` - Same as above.
- From here, you can build whatever targets you need.
- .. note::
- You may see warnings about a mismatched profile in the build output. These
- are generally harmless. To silence them, you can add
- ``-DCMAKE_C_FLAGS='-Wno-backend-plugin'
- -DCMAKE_CXX_FLAGS='-Wno-backend-plugin'`` to your CMake invocation.
- Congrats! You now have a Clang built with profile-guided optimizations, and you
- can delete all but the final build directory if you'd like.
- If this worked well for you and you plan on doing it often, there's a slight
- optimization that can be made: LLVM and Clang have a tool called tblgen that's
- built and run during the build process. While it's potentially nice to build
- this for coverage as part of step 3, none of your other builds should benefit
- from building it. You can pass the CMake options
- ``-DCLANG_TABLEGEN=/path/to/stage1/bin/clang-tblgen
- -DLLVM_TABLEGEN=/path/to/stage1/bin/llvm-tblgen`` to steps 2 and onward to avoid
- these useless rebuilds.
|