Assorted bootstrap LLVM refactors (part 4/N) - #161290
Conversation
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
e33910a to
c2469f0
Compare
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
c2469f0 to
8a5b940
Compare
|
This PR modifies If appropriate, please update This PR changes how LLVM is built. Consider updating src/bootstrap/download-ci-llvm-stamp. |
|
Ok, should be ready for a review now. |
There was a problem hiding this comment.
There is one annoyance related to that, and that is that
download-ci-llvmnow applies to all targets for which you try to build LLVM (d'uh), but that also means that if (for whatever reason) LLVM fails to be downloaded from CI, the build will fail. So if you build for target T2 from target T1:
- If you want to download T1, but build T2, that's not possible to express.
- If T2 fails to be downloaded, the build fails, even if it could be built locally.
I think that we mostly have four options how to deal with this:
- Just ignore it and wait to see if someone complains.
- Revert the change and always download only for the host target. Worked so far. However, downloading LLVM for non-host targets would be quite useful for further bootstrap improvements and refactorings, because the current logic around sysroots and libdirs is.. convoluted, to say the last, and making cross-compilation easier would help with that a lot.
- Allow specifying
download-ci-llvmper target in the target config section. So that you can say that you want to download for T1, but build for T2.- Make download failures non-fatal, and cause them to trigger a local build. This would also help with removing the hacky
is_ci_llvm_available_for_targetlogic, which hard-codes a bunch of targets to "know" which ones offer LLVM and which don't. We could just try to download, and if the result is 404, then we print a warning and continue with building (but this is slightly orthogonal, we can do this even if we don't make LLVM build failures non-fatal).I think that 3 or 4 would be the best solution, perhaps slightly opting for 4. If we get a 404, there's no way we can download, so we build instead. If we get a different error, we still make the failed download fail the build. And only if someone has a use-case for 3, we'd add the new config.
I'm on board with option (4) but with a caveat:
- If user did not specify their preference (so getting a cascading default to
download-ci-llvm = "if-unchanged") or if they explicitly specifieddownload-ci-llvm = "if-unchanged", then "try to download, build if missing" fallback feels reasonable. - If user specified
download-ci-llvm = true, we should fail the build if no CI LLVM is available.
I think this matches what currently download-ci-llvm = "if-unchanged" for host targets do?
To clarify, I think what you mean here is for download-ci-llvm = "if-unchanged"?
There was a problem hiding this comment.
I think this matches what currently download-ci-llvm = "if-unchanged" for host targets do?
Actualy, it doesn't :) true and if-unchanged are not different in this behavior currently. If you use either of them, and the given target is unavailable on CI, bootstrap will just build LLVM locally. It would seem surprising to me if those two modes had a different behavior in this specific aspect.
- if-unchanged:
rust/src/bootstrap/src/core/config/config.rs
Line 2489 in 287084b
- true:
rust/src/bootstrap/src/core/config/config.rs
Line 2513 in 287084b
(By the way, I'm considering merging those two modes together, same as we do for gcc. So that you only say true, and if there are local changes, you would build. But if the LLVM submodule is checked out, you wouldn't check git changes, ofc.)
There was a problem hiding this comment.
Hm... I guess that sometimes I do want to download-ci-llvm to fail the build if CI LLVM cannot be used for whatever reason and never checkout the llvm submodule. (In that, for specific cases like testing in-tree tools I prefer to fail the build rather than trying to recover.)
But yeah if this matches existing behavior then we can consider revisiting this behavior later.
| // CI-rustc can't be used without CI-LLVM. If LLVM Ci is requested, but the | ||
| // LLVM submodule has changes, it is an error. |
| "setting dictionary value" | ||
| ); | ||
| assert!(!config.llvm_ci_mode.download_from_ci()); | ||
| assert!(matches!(config.llvm_ci_mode, LlvmCiMode::BuildLocally)); |
There was a problem hiding this comment.
Oh, I always forget it's already stable :)
| // Sanity check | ||
| check_llvm_version(builder, llvm_ci.output.llvm_config()); | ||
| if builder.host_target == self.target { | ||
| check_llvm_version(builder, llvm_ci.output.llvm_config()); | ||
| } |
There was a problem hiding this comment.
Do we only check host LLVM version via host llvm-config here? Since target LLVM's llvm-config on cross-compile isn't guaranteed to be runnable on host? Maybe worth a comment.
There was a problem hiding this comment.
Yeah, we can't execute the binary if it's not built for the proper target. I'll add a comment.
| // Set a flag for `check`/`clippy`/`fix`, so that certain build | ||
| // scripts can do less work (i.e. not building/requiring LLVM). | ||
| if matches!(cmd_kind, Kind::Check | Kind::Clippy | Kind::Fix) { | ||
| // Set a flag for `check`/`clippy`/`fix`, so that te rustc_llvm build |
This comment has been minimized.
This comment has been minimized.
8a5b940 to
9cc2384
Compare
|
This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed. Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers. |
|
Rebased. @rustbot ready |
| assert!(self.is_running_on_ci()); | ||
| panic!("ERROR: LLVM submodule has changes, `download-rustc` can't be used."); |
There was a problem hiding this comment.
If we are local, won't the assertion swallow the actual useful panic message here?
| // CI-rustc can't be used without CI-LLVM. If LLVM Ci is requested, but the | ||
| // LLVM submodule has changes, it is an error. |
| "setting dictionary value" | ||
| ); | ||
| assert!(!config.llvm_ci_mode.download_from_ci()); | ||
| assert!(matches!(config.llvm_ci_mode, LlvmCiMode::BuildLocally)); |
| if builder.config.llvm_ci_mode.requests_download_from_ci() | ||
| && let Some(config) = builder.config.target_config.get(&target) | ||
| { | ||
| if config.llvm_config.is_some() { | ||
| panic!( | ||
| "Cannot configure `llvm-config` for {target} when using `llvm.download-ci-llvm`", | ||
| ); | ||
| } | ||
| if config.llvm_filecheck.is_some() { | ||
| panic!( | ||
| "Cannot configure `llvm-filecheck` for {target} when using `llvm.download-ci-llvm`" | ||
| ); | ||
| } | ||
| } |
There was a problem hiding this comment.
Without per-target download-ci-llvm, with this logic here, we can't mix:
- I want to download CI LLVM for host, where it is available
- I want to provide an external LLVM for the target, where CI LLVM is unavailable
Previously, did we just use external LLVM for the target when host != target, when CI LLVM was available for host and requested? Or did this just never work?
This logic seems a bit fishy, won't in-tree default profiles (compiler/tools/libs) also hit this with if-changed if they try to cross-compile?
View all comments
Continuation on #161247.
This PR completely removes handling of git changes or LLVM downloads from config parsing, and moves it into the
LlvmFromCistep. Thanks to that, we now also allow downloading LLVM for non-host targets.There is one annoyance related to that, and that is that
download-ci-llvmnow applies to all targets for which you try to build LLVM (d'uh), but that also means that if (for whatever reason) LLVM fails to be downloaded from CI, the build will fail. So if you build for target T2 from target T1:I think that we mostly have four options how to deal with this:
download-ci-llvmper target in the target config section. So that you can say that you want to download for T1, but build for T2.is_ci_llvm_available_for_targetlogic, which hard-codes a bunch of targets to "know" which ones offer LLVM and which don't. We could just try to download, and if the result is 404, then we print a warning and continue with building (but this is slightly orthogonal, we can do this even if we don't make LLVM build failures non-fatal).I think that 3 or 4 would be the best solution, perhaps slightly opting for 4. If we get a 404, there's no way we can download, so we build instead. If we get a different error, we still make the failed download fail the build. And only if someone has a use-case for 3, we'd add the new config.
Already before this PR, we did this:
so if LLVM wasn't available, we just silently reverted from
download-ci-llvm=truetodownload-ci-llvm=false. The 4. proposal would just generalize that, to actually check whether the LLVM files are present on the CDN or not.The problem with 4. is that you can't really set any custom build options for LLVM though, because if you also enable
download-ci-llvm, the config sanity check will tell you to GTFO :( So we would probably need to make some changes there.r? jieyouxu