Trait solver re-derives `Send` proofs per impl when an outlives where-bound is in scope

[ashi009]

One trait, M impls, each wrapping the same Shared struct (150 fields of Arc<Mutex<Vec<…>>> nested 10 deep); every method's future borrows &self. The two files differ only in how the method's lifetimes are spelled:

// outlives_*.rs — the shape #[async_trait] emits for every method
fn check<'life0, 'at>(&'life0 self, x: &'at [u8])
    -> Pin<Box<dyn Future<Output = u64> + Send + 'at>>
where Self: 'at, 'life0: 'at;

// unified_*.rs — same borrows under one lifetime, no outlives clause
fn check<'a>(&'a self, x: &'a [u8]) -> Pin<Box<dyn Future<Output = u64> + Send + 'a>>;

$ git clone https://github.com/ashi009/region-cache-repro && cd region-cache-repro
$ rustc -V
rustc 1.96.0 (ac68faa20 2026-05-25)
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta outlives_150.rs
real	0m3.326s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta outlives_300.rs
real	0m6.627s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta unified_150.rs
real	0m0.134s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta unified_300.rs
real	0m0.180s

$ rustc -V
rustc 1.98.0-nightly (8954863c8 2026-06-05)
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta outlives_300.rs
real	0m6.612s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -o /tmp/r.rmeta unified_300.rs
real	0m0.188s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -Znext-solver=globally -o /tmp/r.rmeta outlives_300.rs
real	0m14.539s
$ time rustc --edition 2021 --crate-type=lib --emit=metadata -Znext-solver=globally -o /tmp/r.rmeta unified_300.rs
real	0m0.273s

97% of the outlives time is evaluate_obligation, re-deriving the same Shared: Send proof once per impl. The outlives bound puts a region in caller_bounds. The query canonicalizes the ParamEnv, so that region comes back as an infer var. can_use_global_caches then bails on param_env.has_infer(), and the proof never reaches tcx.evaluation_cache. #[async_trait] emits that bound on every method, so large async codebases pay this per method × impl.

Same diagnosis and fix as #92044 (validated on #87012), closed unmerged over selection soundness (impl<T: 'static> bounds participate in selection); it reproduces unchanged today.

(-Znext-solver=globally not removing the gap is unexpected; cause not investigated.)

[oli-obk]

#157463 (comment)

[ashi009]

@oli-obk the issue is rewritten by hand. PTAL and reopen.

[Dandandan]

I found three small changes / issues in the current code that fix the issue / remove the evaluate_obligation explosion issue, perhaps it can help someone to complete this (I used Claude to help me identify these issues as I'm no active Rust contributor).

• Perf: Only scan the affected suffix of the provisional cache #161130 As explained in the PR description, instead of fully walking all (growing list of) items each we can only walk the final entries, removing the O(n²) behaviour
• Perf: Fix the check in the provisional cache's on_failure #161133 A typo, !eval.from_dfn >= dfn doesn't fire, dropping the entries in the cache doesn't have an effect on the async_trait usage
• Perf: Clear the param env for more global trait goals #161126 Widening the conditions for the global cache which @ashi009 already identified as one of the issues

The changes together makes it linear (avoiding the blowup).

Hopefully this helps someone with active contribution history

[Dandandan]

When I have some time I'll create some PRs with the fixes.