Description
I tried this code:
./configure --enable-llvm-link-shared --enable-llvm-plugins --enable-llvm-enzyme --release-channel=nightly --enable-llvm-assertions --enable-clang --enable-lld --enable-option-checking --enable-ninja --disable-docs --enable-debug
./x.py build --stage 1 library
I expected to see this happen: builds rustc
Instead, this happened:
thread 'rustc' panicked at compiler/rustc_codegen_ssa/src/codegen_attrs.rs:57:9:
unexpected `def_kind` in `codegen_fn_attrs`: GlobalAsm
[...]
query stack during panic:
#0 [codegen_fn_attrs] computing codegen attributes of `probestack::{global_asm#0}`
#1 [collect_and_partition_mono_items] collect_and_partition_mono_items
... and 1 other queries... use `env RUST_BACKTRACE=1` to see the full query stack
error: could not compile `compiler_builtins` (lib)
Meta
rustc --version --verbose
:
build from source
I know that debug builds used to work a few months ago.
So one could try to debug this by looking at the code. A much simpler solution probably is just
to do some binary search by building some older PR from the tracking issue with debug=true until finding one that builds. From there it should be easy to figure out which PR introduced the bug. Once that's clear it is hopefully easier to understand which change introduced the bug, and either directly fix it, or ask for help here or in the zulip channel. I'm marking it as easy, since it takes a bit of time to test multiple PRs but it isn't overly complex, and we had support for it and didn't introduce overly complex code changes in the last months. So the fix is likely also comparably simple.
This bug prevents us from enabling autodiff in CI and thus nightly, so any help would be appreciated!
Backtrace
<backtrace>
Activity
Shourya742 commentedon Apr 12, 2025
@rustbot claim
ZuseZ4 commentedon Apr 19, 2025
@Shourya742 Since you still have the other open issue and I had some time today, I looked into it.
I could reproduce the bug starting with #133429, together with the error message it was easy from there to find the one-line fix. We want to prevent source functions to be inlined before autodiff, however we were incorrectly also checking for the other two MonoItem types, Static and GlobalAsm. Marking them as no-inline wouldn't be sensible.
Shourya742 commentedon Apr 20, 2025
That’s cool—I can help test and review it then. Naive me was going through all the commits that changed that file one by one, and each build took around 20 minutes to complete. Is there any way we can reduce the build time?
ZuseZ4 commentedon Apr 20, 2025
Oh it's just a one line fix, so no need to review or test, oli should approve it soon. I am looking into some ways to improve bootstrap time and another person also wanted to check something related, I think it's better if not too many people work on it to avoid duplications.
If you look for other issues, it looks like no one is currently working on this one: EnzymeAD#193 (support for
&dyn type
) It's definitely a bit harder, but by now there should also be a lot of documentation. You'd need to work with opt and llvm-ir, if you want to give it a try. Once you figured out how to support it by hand, you can document it and we can add Rust wrapers which do it automatically. But I can also look for a pure Rust issue if you prefer?Shourya742 commentedon Apr 20, 2025
I can take this up. While more challenging issues might take some time, they will provide a great learning opportunity, and I'm ready for them. Pure Rust tasks can also be assigned — I have some time next week
Rollup merge of rust-lang#140030 - EnzymeAD:autodiff-debug, r=jieyouxu
Unrolled build for rust-lang#140030