-
Notifications
You must be signed in to change notification settings - Fork 12.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Strip unexpected debuginfo from libLLVM.so
and librustc_driver.so
when not requesting any debuginfo
#114305
Conversation
r? @ozkanonur (rustbot has picked a reviewer for you, use r? to override) |
This PR changes how LLVM is built. Consider updating src/bootstrap/download-ci-llvm-stamp. |
@bors r+ rollup |
Strip unexpected debuginfo from `libLLVM.so` and `librustc_driver.so` when not requesting any debuginfo As seen in rust-lang#114175 and in [this zulip discussion](https://rust-lang.zulipchat.com/#narrow/stream/247081-t-compiler.2Fperformance/topic/Artifact.20sizes/near/379302655), there's still some small amount of debuginfo in LLVM's shared library on linux, even when not requesting it (nightly CI), coming from `libstdc++`. ``` $ readelf --debug-dump=info ~/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/libLLVM-16-rust-1.73.0-nightly.so | grep DW_TAG_compile_unit -A5 | grep DW_AT_comp_dir | cut -d ":" -f 2- | counts 101 counts ( 1) 39 (38.6%, 38.6%): (indirect string, offset: 0x7): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/libsupc++ ( 2) 38 (37.6%, 76.2%): (indirect string, offset: 0x43fb2): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src/c++11 ( 3) 23 (22.8%, 99.0%): (indirect string, offset: 0x18ed8): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src/c++98 ( 4) 1 ( 1.0%,100.0%): (indirect string, offset: 0x53f04): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src ``` Similarly, here's `librustc_driver.so` when not requesting debuginfo from either rustc or the tools (nightly CI), coming e.g. from our LLVM wrapper: ``` $ readelf --debug-dump=info ~/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/librustc_driver-e534b3a316089f5f.so | grep DW_TAG_compile_unit -A5 | grep DW_AT_comp_dir | cut -d ":" -f 2- | counts 116 counts ( 1) 34 (29.3%, 29.3%): (indirect string, offset: 0x3c11): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/libsupc++ ( 2) 32 (27.6%, 56.9%): (indirect string, offset: 0x9753c): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src/c++11 ( 3) 25 (21.6%, 78.4%): (indirect string, offset: 0x393bd): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src/c++98 ( 4) 23 (19.8%, 98.3%): (indirect string, offset: 0x33ed3): /cargo/registry/src/index.crates.io-6f17d22bba15001f/compiler_builtins-0.1.98 ( 5) 1 ( 0.9%, 99.1%): (indirect string, offset: 0xaffff): /rustc/0d95f9132909ae7c5f2456748d0ffd1c3ba4a8e8 ( 6) 1 ( 0.9%,100.0%): (indirect string, offset: 0xb604a): /tmp/gcc-build/x86_64-pc-linux-gnu/libstdc++-v3/src ``` To reduce the size of distributed artifacts, this PR strips debuginfo from the LLVM and `rustc_driver` shared libraries, when: - no debuginfo is requested when building LLVM: `link-shared` is true, `optimize` is true and `release-debuginfo` is false - no debuginfo is requested when building the rustc driver: - `debuginfo-level-rustc` and `debuginfo-level-tools` are off. - when building with a stage != 0 compiler: since this is about the distributed artifacts, there's no need to do this at other stages. - for both: on a x64 linux host and target where `strip -g` is available and fixes the issue (I don't know how to strip debuginfo from a `.dylib` on mac). The LLVM BOLTed .so, and `librustc_driver.so` are big there, and this will help a little. Other targets/hosts can be added in the future if we want to. rust-lang#114175 did the same thing unconditionally in `opt-dist`, prior to BOLTing LLVM. But this should only be used in conjunction with the other config options mentioned above, and which `opt-dist` doesn't know about. Therefore, it makes more sense as in bootstrap when building LLVM and rustc when applicable and no debuginfo is requested. This shouldn't interact badly with CI caching builds and artifacts, right? --- From the other PR, `libLLVM-16-rust-1.73.0-nightly.so` prior to rust-lang#114141: - master: 173.13 MiB - stripped debuginfo: 165.12 MiB (-8 MiB, -4.6%) `libLLVM-16-rust-1.73.0-nightly.so` after rust-lang#114141: - master: 121.13 MiB - stripped debuginfo: 113.12 MiB (still -8 MiB, -6.6%) `librustc_driver.so`: - master: 118.58 MiB - stripped debuginfo: 106.46 MiB (-12 MiB, -10.2%) (Results are also available in this most recent [perf run's artifact sizes](https://perf.rust-lang.org/compare.html?start=b321edd1b2d4bd00c7b4611e8f20a03ee7b77023&end=810ab570d5d27facb91806e5d9847815d9dac22a&stat=instructions%3Au&tab=artifact-size))
@bors rollup=never |
☀️ Test successful - checks-actions |
Finished benchmarking commit (a7caaae): comparison URL. Overall result: ❌✅ regressions and improvements - no action needed@rustbot label: -perf-regression Instruction countThis is a highly reliable metric that was used to determine the overall result at the top of this comment.
Max RSS (memory usage)ResultsThis is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.
CyclesResultsThis is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.
Binary sizeThis benchmark run did not return any relevant results for this metric. Bootstrap: 649.049s -> 647.584s (-0.23%) |
Prevent constant rebuilds of `rustc-main` (and thus everything else) PR rust-lang#114305 changed bootstrap to run `strip -g` on `librustc_driver.so` and `libllvm.so` on Linux when no debuginfo was requested. Unfortunately, that PR resulted in bootstrap always rebuilding everything starting from stage 1 `rustc-main` (including stage 1 libraries and tests) when invoking bootstrap multiple times. We noticed this because Ferrocene's CI times increased to between 2x and 3x total execution time, but the regression can also be reproduced locally by running `./x build library/sysroot --stage 1` twice. The explanation of the problem is in the code comments. r? `@lqd` cc `@ozkanonur`
Prevent constant rebuilds of `rustc-main` (and thus everything else) PR rust-lang#114305 changed bootstrap to run `strip -g` on `librustc_driver.so` and `libllvm.so` on Linux when no debuginfo was requested. Unfortunately, that PR resulted in bootstrap always rebuilding everything starting from stage 1 `rustc-main` (including stage 1 libraries and tests) when invoking bootstrap multiple times. We noticed this because Ferrocene's CI times increased to between 2x and 3x total execution time, but the regression can also be reproduced locally by running `./x build library/sysroot --stage 1` twice. The explanation of the problem is in the code comments. r? ``@lqd`` cc ``@ozkanonur``
Prevent constant rebuilds of `rustc-main` (and thus everything else) PR rust-lang#114305 changed bootstrap to run `strip -g` on `librustc_driver.so` and `libllvm.so` on Linux when no debuginfo was requested. Unfortunately, that PR resulted in bootstrap always rebuilding everything starting from stage 1 `rustc-main` (including stage 1 libraries and tests) when invoking bootstrap multiple times. We noticed this because Ferrocene's CI times increased to between 2x and 3x total execution time, but the regression can also be reproduced locally by running `./x build library/sysroot --stage 1` twice. The explanation of the problem is in the code comments. r? ```@lqd``` cc ```@ozkanonur```
As seen in #114175 and in this zulip discussion, there's still some small amount of debuginfo in LLVM's shared library on linux, even when not requesting it (nightly CI), coming from
libstdc++
.Similarly, here's
librustc_driver.so
when not requesting debuginfo from either rustc or the tools (nightly CI), coming e.g. from our LLVM wrapper:To reduce the size of distributed artifacts, this PR strips debuginfo from the LLVM and
rustc_driver
shared libraries, when:link-shared
is true,optimize
is true andrelease-debuginfo
is falsedebuginfo-level-rustc
anddebuginfo-level-tools
are off.strip -g
is available and fixes the issue (I don't know how to strip debuginfo from a.dylib
on mac). The LLVM BOLTed .so, andlibrustc_driver.so
are big there, and this will help a little. Other targets/hosts can be added in the future if we want to.#114175 did the same thing unconditionally in
opt-dist
, prior to BOLTing LLVM. But this should only be used in conjunction with the other config options mentioned above, and whichopt-dist
doesn't know about. Therefore, it makes more sense as in bootstrap when building LLVM and rustc when applicable and no debuginfo is requested.This shouldn't interact badly with CI caching builds and artifacts, right?
From the other PR,
libLLVM-16-rust-1.73.0-nightly.so
prior to #114141:libLLVM-16-rust-1.73.0-nightly.so
after #114141:librustc_driver.so
:(Results are also available in this most recent perf run's artifact sizes)