Enable AutoFDO. #87918

mikebenfield · 2021-08-10T16:35:01Z

This largely involves implementing the options debug-info-for-profiling
and profile-sample-use and forwarding them on to LLVM.

AutoFDO can be used on x86-64 Linux like this:
rustc -O -Clink-arg='Wl,--no-rosegment' -Cdebug-info-for-profiling main.rs -o main
perf record -b ./main
create_llvm_prof --binary=main --out=code.prof
rustc -O -Cprofile-sample-use=code.prof main.rs -o main2

Now main2 will have feedback directed optimization applied to it.

The create_llvm_prof tool can be obtained from this github repository:
https://github.com/google/autofdo

The option -Clink-arg='Wl,--no-rosegment' is necessary to avoid lld
putting an extra RO segment before the executable code, which would make
the binary silently incompatible with create_llvm_prof.

rust-highfive · 2021-08-10T16:41:04Z

Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @cjgillot (or someone else) soon.

Please see the contribution instructions for more information.

mikebenfield · 2021-08-10T17:32:48Z

I ran clang-format on the diff in PassWrapper.cpp, which reordered the includes. I can revert that part if desired.

compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

bjorn3 · 2021-08-11T12:58:22Z

The option -Clink-arg='Wl,--no-rosegment' is necessary to avoid lld putting an extra RO segment before the executable code, which would make the binary silently incompatible with create_llvm_prof.

This will make read-only data executable, which depending on what program you have may or may not be a bad idea.

mikebenfield · 2021-08-11T15:58:56Z

This will make read-only data executable, which depending on what program you have may or may not be a bad idea.

Indeed. Unfortunately it's the only way to deal with a current limitation of the create_llvm_prof tool, which can't deal with that RO segment. I've been told they plan to remove that limitation eventually, but that it won't happen in the immediate future.

mikebenfield · 2021-08-11T16:31:44Z

Rebased, undid formatting in PassWrapper.cpp, fixed formatting in config.rs.

mikebenfield · 2021-08-11T16:33:23Z

And fixed the commit message.

cjgillot · 2021-08-14T12:51:23Z

I don't know anything about LLVM.
r? @nikic

bors · 2021-08-26T05:27:08Z

☔ The latest upstream changes (presumably #88069) made this pull request unmergeable. Please resolve the merge conflicts.

mikebenfield · 2021-08-30T20:56:12Z

A fix is in the works for the create_llvm_prof tool so it will no longer require --no-rosegment, so I removed that from the commit message.

mikebenfield · 2021-09-07T20:01:52Z

@nikic Would you mind taking a look at this?

nikic

This looks pretty straightforward and seems to mirror what clang does. Some high-level notes:

New options are always introduced as unstable -Z options first, and then stabilized as -C options. You should switch the new options to use -Z and open a tracking issue for this feature.
Not strictly required, but adding some basic docs on how to use this to src/doc/unstable-book/src/compiler-flags wouldn't hurt. You'll find profile.md and instrument_coverage.md there for related functionality. This could basically just be what you have in your PR description, maybe with an introduction on what this is actually about (it's easy to get lost in all these profiling, coverage and instrumentation options...)
It would be great to have some kind of testing for this. I see that the dependence on create_llvm_prof would make an end-to-end test inconvenient. If nothing else, we can test that -Cdebuginfo-for-profiling adds the function attribute using a test in src/test/codegen.

compiler/rustc_session/src/options.rs

src/bootstrap/compile.rs

compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

compiler/rustc_codegen_llvm/src/callee.rs

compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

mikebenfield · 2021-09-12T23:54:30Z

I made these into -Z options and added some documentation.

Testing seems more difficult. debug-info-for-profiling actually doesn't add a function attribute. profile-sample-use does, but it requires a profile, and I'm not sure whether it's wise to include a profile. Please let me know what you think.

mikebenfield · 2021-09-23T22:15:15Z

@nikic Mind taking another look?

compiler/rustc_codegen_ssa/src/back/write.rs

src/bootstrap/config.rs

mikebenfield · 2021-10-04T17:28:26Z

I addressed both recent comments.

nikic

Looks good, can you please squash the commits?

mikebenfield · 2021-10-06T19:35:59Z

Squashed.

This largely involves implementing the options debug-info-for-profiling and profile-sample-use and forwarding them on to LLVM. AutoFDO can be used on x86-64 Linux like this: rustc -O -Cdebug-info-for-profiling main.rs -o main perf record -b ./main create_llvm_prof --binary=main --out=code.prof rustc -O -Cprofile-sample-use=code.prof main.rs -o main2 Now `main2` will have feedback directed optimization applied to it. The create_llvm_prof tool can be obtained from this github repository: https://github.com/google/autofdo Fixes rust-lang#64892.

mikebenfield · 2021-10-06T19:37:40Z

And fixed the commit message.

nikic · 2021-10-06T19:38:25Z

@bors r+

bors · 2021-10-06T19:38:26Z

📌 Commit a17193d has been approved by nikic

bors · 2021-10-06T19:38:27Z

🌲 The tree is currently closed for pull requests below priority 100. This pull request will be tested once the tree is reopened.

Enable AutoFDO. This largely involves implementing the options debug-info-for-profiling and profile-sample-use and forwarding them on to LLVM. AutoFDO can be used on x86-64 Linux like this: rustc -O -Clink-arg='Wl,--no-rosegment' -Cdebug-info-for-profiling main.rs -o main perf record -b ./main create_llvm_prof --binary=main --out=code.prof rustc -O -Cprofile-sample-use=code.prof main.rs -o main2 Now `main2` will have feedback directed optimization applied to it. The create_llvm_prof tool can be obtained from this github repository: https://github.com/google/autofdo The option -Clink-arg='Wl,--no-rosegment' is necessary to avoid lld putting an extra RO segment before the executable code, which would make the binary silently incompatible with create_llvm_prof.

…ingjubilee Rollup of 8 pull requests Successful merges: - rust-lang#87918 (Enable AutoFDO.) - rust-lang#88137 (On macOS, make strip="symbols" not pass any options to strip) - rust-lang#88772 (Fixed confusing wording on Result::map_or_else.) - rust-lang#89025 (Implement `#[link_ordinal(n)]`) - rust-lang#89082 (Implement rust-lang#85440 (Random test ordering)) - rust-lang#89288 (Wrapper for `-Z gcc-ld=lld` to invoke rust-lld with the correct flavor) - rust-lang#89476 (Correct decoding of foreign expansions during incr. comp.) - rust-lang#89622 (Use correct edition for panic in [debug_]assert!().) Failed merges: r? `@ghost` `@rustbot` modify labels: rollup

mikebenfield force-pushed the pr-afdo branch from e19506e to f9dd5c4 Compare August 10, 2021 16:40

rust-highfive assigned cjgillot Aug 10, 2021

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Aug 10, 2021