Docker: Add a clang install step to dockerfile #10
Conversation
Dockerfile
Outdated
| @@ -25,6 +25,9 @@ RUN sh /home/modules/install_cmake.sh | |||
| COPY install_ninja.sh /home/modules/install_ninja.sh | |||
| RUN sh /home/modules/install_ninja.sh | |||
|
|
|||
| COPY install_clang.sh /home/modules/install_clang.sh | |||
There was a problem hiding this comment.
As mentioned in #3, I think this should be before cmake and ninja so that bumping the "small" projects doesn't require rebuilding clang. We can install a system-wide cmake and ninja-build for that. We should probably also just modify PATH using ENV commands rather than a wrapper script so that it's always available.
| cmake -G Ninja \ | ||
| -DCMAKE_BUILD_TYPE=Release \ | ||
| -DCMAKE_INSTALL_PREFIX="$HOME/misc/root/clang" \ | ||
| -DLLVM_ENABLE_PROJECTS='clang' \ |
There was a problem hiding this comment.
I think we also want a shared build. That should certainly reduce the installation tree size. I've found that -DLLVM_PARALLEL_LINK_JOBS:STRING=1 was required to not blow memory limits when linking. Adding -DLLVM_INCLUDE_TESTS:BOOL=OFF would also likely reduce build times.
There was a problem hiding this comment.
What's the option to enable shared build only? I've heard something about libllvm.so.
There may be other options that may reduce build time and image size, like selecting only the host machine's architecture (I think by default it's building a bunch of cross-compiling support as I can see it building stuff related to many archs in the log). Other options I don't know may limit which languages are supported.
There was a problem hiding this comment.
-DBUILD_SHARED_LIBS:BOOL=ON
I think removing cross compilation makes sense. Removing TableGen compilation rules would certainly help compilation time. Likely link time/memory too.
There was a problem hiding this comment.
OK. I got the install size of clang reduced by 10 times! (from 1.5GB down to 155MB). I played with the LVM_INCLUDE_TESTS and LVM_BUILD_TESTS options but it triggers a bug where a link commands wants the gtest lib, and it's not there. Anyway, the doc says LVM_BUILD_TESTS is OFF by default (the LVM_INCLUDE_TESTS seems to only control whether or not the generated ninja file includes rules for the test target or not).
There was a problem hiding this comment.
Sweet! I'm looking at reducing the GCC install size now. Probably don't need to compile non-C and C++ compilers…
…image. Also, automatically determine an appropriate number of jobs to avoid overwhelming the small CI machines.
johan-boule
force-pushed
the
docker
branch
7 times, most recently
from
ddea032
to
25ac3b2
Compare
|
I'm going to cherry-pick bits down into the branch I'm working on right now to respin. I have GCC down to < 250 MB :) . |
|
Merged as part of #13. Thanks! |
These commits add a clang install step to dockerfile.
The script downloads a specific commit hash of the llvm-project git repository (clang being part of it).
That commit was taken from the tip of the master branch as it was yesterday, so it's a pre-release of version 9.
I also changed other build scripts so they determine an appropriate number for job parallelism.