regression: misaligned pointer dereference: address must be a multiple ... #111487
Comments
Mark-Simulacrum
added
the
regression-from-stable-to-beta
rustbot
added
the
I-prioritize
Mark-Simulacrum
added
T-compiler
|
#98112 doesn't seem to have an FCP for doing this, which makes some sense (it's just "exploiting" UB). @saethlin did we file PRs against the relevant crates which break the most transitive tests at least? e.g., rand_core 0.2.x, 0.3.x, 0.4.x seem to be a good fraction of this list. There's a few more crates that seem to stand out too (e.g., plotters-bitmap). |
|
We did not make any effort to file fixes for the relevant crates. The The |
|
The only things I think are significant are: An issue in an old version (almost 4 years old) of Multiple crates hitting an issue in I don't think these are worth doing anything with. All we could do here is ask maintainers to backport fixes. |
|
Sounds like that's right. The main thing we usually try to ensure is that there is some upgrade path, especially for common code, and I think this covers most of that. Plus it technically won't affect non-debug builds, so users have that escape hatch too. I'm fine leaving this as-is. |
Mark-Simulacrum
removed
the
I-prioritize
Don't check for misaligned raw pointer derefs inside Rvalue::AddressOf From rust-lang#112026 (comment): rustc 1.70 (stable next week) added a Mir pass to add pointer alignment checks in debug mode. Adding these checks caused some crates to break, but that was expected, since they contain broken code (rust-lang#111487) for tracking that. However, the checks added are slightly more aggressive than they should have been. Specifically, they also check the place in an `addr_of!` expression. Whether lack of alignment there is or isn't UB is unclear. This PR modifies the pass to not affect those cases. I spot checked the crater regressions and the ones I saw were not the case that this PR is modifying. It still seems good to not land anything overaggressive though
Don't check for misaligned raw pointer derefs inside Rvalue::AddressOf From rust-lang/rust#112026 (comment): rustc 1.70 (stable next week) added a Mir pass to add pointer alignment checks in debug mode. Adding these checks caused some crates to break, but that was expected, since they contain broken code (rust-lang/rust#111487) for tracking that. However, the checks added are slightly more aggressive than they should have been. Specifically, they also check the place in an `addr_of!` expression. Whether lack of alignment there is or isn't UB is unclear. This PR modifies the pass to not affect those cases. I spot checked the crater regressions and the ones I saw were not the case that this PR is modifying. It still seems good to not land anything overaggressive though
|
I think |
|
@Mark-Simulacrum Should this issue stay open? I don't think we have an actual regression here. |
|
Yeah, I think we can close this. |
A bunch of these seem to come from rand_core 0.2.1, 0.3.0, and 0.4.0 code (not sure if we can fix it there?)
Not rand_core:
cc #98112
The text was updated successfully, but these errors were encountered: