Stabilize the next-generation trait solver

Summary

Stabilize -Znext-solver=globally, replacing the existing trait solver implementation entirely.

Motivation

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.

The status quo

There are multiple type system unsoundnesses blocked on the next-generation trait solver: [project board][unsoundnesses]. Desirable features such as coinductive trait semantics and perfect derive, where-bounds on binders, and better handling of higher-ranked bounds and types are also stalled due to shortcomings of the existing implementation.

Since starting to work on this at the EOY 2022, we’ve:

• stabilized its use in coherence checking in Rust 1.84
• replaced the use of chalk in rust-analyzer
• done full crater runs and all large regressions are either fixed or intended
• improved performance and diagnostics, even if they is still more work to do

What we propose to do about it

Stabilize -Znext-solver=globally and rip out the old implementation

• finish triaging the crater and and resolve remaining issues tracked in the project board
• achieve performance parity with the old solver
• write an RFC for the cycle semantics used by the new trait solver
• document the behavior of the new trait solver and write stabilization reports

Work items over the next year

Team asks

Frequently asked questions

What do I do with this space?