Rustc Trait System Refactor Initiative Update: Stabilizing `-Znext-solver=coherence`

Dec. 4, 2024 · lcnr on behalf of The Rustc Trait System Refactor Initiative

It's been half a year since we last summarized our progress in the Types Team update blog post. With the next-generation trait solver now getting used by default in coherence checking on beta1, it's time for another update. The next-generation trait solver is intended to fully replace the existing type system components responsible for proving trait bounds, normalizing associated types, and much more. This should fix many long-standing (soundness) bugs, enable future type system improvements, and improve compile-times. See this previous blog post for more details.

The FCP to stabilize the use of the new solver in coherence already passed in April. Checking auto-trait bounds of deeply recursive types unfortunately caused exponential blowup with the new implementation, resulting in a hang caught by CI. Fixing this hang required us to add a complex caching optimization to efficiently handle cycles while still being correct2. We then merged the stabilization PR only to revert it 5 days later as it resulted in a separate hang in nalgebra. Luckily the fix this time was a lot simpler, which allowed us to stabilize -Znext-solver=coherence again soon after.

While the new solver will only be used in coherence checking for now, this already removes some inconsistencies and theoretical soundness issues, see the stabilization report for a complete summary. In case you encounter any weird hangs, unexpected overlap errors, or unhelpful diagnostics, please open a GitHub issue.

We then moved our focus to bootstrap using exclusively the new solver again. We've made significant progress here and are able to fully compile rustc and cargo, allowing try-builds to succeed. We are currently looking into the remaining compile-time benchmark failures. We intend to start to optimize the new implementation and to test more and more crates until a full crater run will be useful.

Looking ahead

To unblock the stabilization of the next-generation trait solver we're required to fix multiple issues which also affect the current implementation. This is due to subtle differences between the implementations causing the new solver to trigger bugs which are avoided by the old solver. The cases where the new implementation works correctly while the old one does not are unfortunately far less relevant as nobody has written code which relies on the behavior of the next-generation trait solver yet.

We will also have to spend a significant effort towards improving error messages when using the new trait solver. There are still many cases where they are significantly worse than the status quo.

We are still working towards our first crater run and are aware of multiple significant issues. We expect to resolve most of them during the next few months. So while we're not quite at the point where bug reports are useful, we intend to explicitly ask for testing of -Znext-solver=globally soon. To follow our progress, check out the -Znext-solver tracking issue or join us on zulip.

  1. Which will get stabilized as version 1.84. ↩

  2. The old implementation avoids these hangs by incorrectly reusing cache-entries even if they should not be applicable. Due to subtle reasons™ we believe this to not be an exploitable soundness issue. ↩