performance regression with -Znext-solver=globally

The new solver struggles more than the old one with (somewhat pathologically) large bodies, which happen in generated code quite a bit, I imagine. This is readily reproduced with:

// lib.rs — one function, N repetitions of the push line
pub fn big() {
    let mut v = Vec::new();
    v.push(Default::default());
    // ... repeat N times ...
    v.push(0u8);
}

To create the test cases:

python3 -c "
n = 4000
print('pub fn big() {'); print('    let mut v = Vec::new();')
print('    v.push(Default::default());\n' * n, end='')
print('    v.push(0u8);'); print('}')" > lib.rs

Run (typeck only, no codegen):

rustc +nightly --crate-type=lib --edition=2024 --emit=metadata --out-dir /tmp/out lib.rs
rustc +nightly --crate-type=lib --edition=2024 --emit=metadata --out-dir /tmp/out -Znext-solver=globally lib.rs

Measurements (wall clock)

Both are clearly non-linear in behavior, due to the O(n) passes x O(n) pending obligations, but the new solver has a much higher constant factor, it seems. The work done in #158249 has made each handling of a stalled obligation cheaper, but did not remove the scan. My friendly neighborhood AI points to FulfillmentCtxt::try_evaluate_obligations gathering work with drain_pending(|_, _| true) in compiler/rustc_trait_selection/src/solve/fulfill.rs.

The following might well be related:

• rust-lang/rust#131411 (cranelift-codegen, ISLE-generated functions, 10×, open and undiagnosed)
• rust-lang/trait-system-refactor-initiative#228 (wg-grammar, closed as "acceptable" after constant-factor fast-path work)
• rust-lang/trait-system-refactor-initiative#254 (crater perf triage)

I've run into this in rust-analyzer, which shares this code, its inference makes more fulfillment passes per body.

Meta

rustc 1.99.0-nightly (89c61a754 2026-07-23)
Disclosure: the investigation and benchmarking was supported by AI

@rustbot label +I-compiletime +WG-trait-system-refactor

Start with the supplied generated lib.rs reproducer and the two rustc commands, then inspect compiler/rustc_trait_selection/src/solve/fulfill.rs, especially try_evaluate_obligations and drain_pending. Compare timings across the listed push counts and trace the pending-obligation work; done means identifying and addressing the next-solver regression without changing the reproducer's expected behavior.