Next-generation trait solver

[nikomatsakis]

Metadata
Point of contact	@lcnr
Team(s)	types
Goal document	2025h1/next-solver

Summary

Continue work towards the stabilization of -Znext-solver=globally, collecting and resolving remaining blockers. Extend its use in lints and rustdoc.

Tasks and status

• Discussion and moral support (types )
• Implementation (@lcnr, @compiler-errors)
• Standard reviews (types )
• FCP decision(s) (types )

[nikomatsakis]

This issue is intended for status updates only.

For general questions or comments, please contact the owner(s) directly.

[lcnr]

I am currently working on caching in the trait solver. The current implementation results in hangs when encountering complex auto-trait cycles. See https://rust-lang.zulipchat.com/#narrow/stream/364551-t-types.2Ftrait-system-refactor/topic/Fuchsia.20hanging.20on.20.60-Znext-solver.3Dcoherence.60 for the progress here. We now have an approach which fixes the performance regression. It'll still take at least multiple weeks until it's ready for merge however.

The librarification of the trait solver is in a very good state thanks to a lot of work by @compiler-errors. We're currently discussing its use in r-a on zulip: https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/New.20Trait.20Solver. While there are still components we intend to move from rustc into the shared library to enable their reuse, all major ones have already been moved by now.

[lcnr]

Since the last update, we've landed the new caching optimization in rust-lang/rust#128828 and hope to land the -Znext-solver=coherence stabilization PR this week.

There has been one other bug fix in the last month, see https://github.com/rust-lang/rust/pulls?q=is%3Apr+is%3Amerged+label%3AWG-trait-system-refactor+-label%3Arollup+closed%3A%3C2024-09-02+-closed%3A%3C2024-08-01 for the full list of PRs.

[lcnr]

We've tried to stabilize -Znext-solver=coherence in rust-lang/rust#121848 which we had to revert later (rust-lang/rust#130249) as it caused a hang in nalgebra.

We mostly fixed this hang with rust-lang/rust#130617 and rust-lang/rust#130821. As nalgebra still ends up hitting the recursion limit with a quite wide search graph, the performance impact of enabling -Znext-solver=coherence is quite bad rust-lang/rust#130654 (comment).

There haven't been any other significant changes to the new solver during the last month.

[lcnr]

We stabilized -Znext-solver=coherence again in rust-lang/rust#130654. It's looking like the stabilization will actually go through this time.

We're currently refactoring the way the current "typing mode" is tracked, cc rust-lang/types-team#128. This is necessary to fix rust-lang/trait-system-refactor-initiative#106. I also started an FCP to clean up the way we merge candidates when proving trait goals: rust-lang/rust#132325. I intend to pretty much exactly match this behavior in the new solver as well, so I expect all of our issues there to be resolved.

[lcnr]

The -Znext-solver=coherence stabilization didn't encounter any issues and is now on beta. Published a new update blogpost. We've made significant progress on bootstrap with -Znext-solver=globally this last month. I finished the core of the "typing mode" refactor with rust-lang/rust#133242 which then allowed me to unblock bootstrap via rust-lang/rust#133501. There has been a lot of work by @compiler-errors and myself with significant help from @lqd since then.

We're now able to bootstrap using exclusively the new solver. This allowed us to finish the first @bors try-run and run the benchmark suite for the first time. There are 3 remaining benchmarks which fail to compile and quite a few with very significant performance regressions. I am now working towards proper support for defining uses of opaque types inside of closures to fix wg-grammar.

[nikomatsakis]

This is a continuing project goal, and the updates below this comment will be for the new period 2025h1

[lcnr]

We've stabilized -Znext-solver=coherence in version 1.84 and started to track the remaining issues in a project board.

Fixing the opaque types issue breaking wg-grammar is difficult and requires a more in-depth change for which there is now an accepted Types MCP. This likely also unblocks the TAIT stabilization using the old solver.

While waiting on the MCP I've started to look into the errors when compiling nalgebra. @lqd minimized the failures. They have been caused by insufficiencies in our cycle handling. With rust-lang/rust#136824 and rust-lang/rust#137314 cycles should now be fully supported.

We also fully triaged all UI tests with the new solver enabled with @compiler-errors, @BoxyUwU, and myself fixing multiple less involved issues.

[lcnr]

The two cycle handling PRs mentioned in the previous update have been merged, allowing nalgebra to compile with the new solver enabled. I have now started to work on opaque types in borrowck again. This is a quite involved issue and will likely take a few more weeks until it's fully implemented.

[lcnr]

We've made a lot of progress over the last 1.5 months. My change to opaque types in borrowck is pretty much done now: rust-lang/rust#139587. It still needs some cleanup and an FCP to actually merge. We've already merged multiple cleanups on the way here.

We then started to test crater with the -Znext-solver=globally. @compiler-errors and me encountered and merged the fixes for 13 issues since then: rust-lang/rust#139791 rust-lang/rust#139798 rust-lang/rust#140236 rust-lang/rust#139900 rust-lang/rust#139828 rust-lang/rust#139774 rust-lang/rust#139762 rust-lang/rust#139789 rust-lang/rust#138845 rust-lang/rust#140306 rust-lang/rust#140305 rust-lang/rust#140276 rust-lang/rust#140302. @Nadrieril was also helpful by minimizing an encountered issue.

With these improvements and multiple in-flight changes we're now at significantly less than 100 remaining regressions in the top 10000 crates and have started the first complete crater run today. We are using a single PR for all crater runs. Check out rust-lang/rust#133502 for the current status and the stack of in-flight changes.