stupid fix for coercion hack perf regression

[lcnr]

cc #136127 #138542

r? @ghost

[lcnr]

@bors try @rust-timer queue

[bors]

⌛ Trying commit e57808d with merge b6303fd...

[bors]

☀️ Try build successful - checks-actions
Build commit: b6303fd (b6303fdafa669e76fca4cabfa942c07295d3f457)

[lcnr]

@rust-timer build b6303fd

[compiler-errors]

@rust-timer build b6303fd

[compiler-errors]

Oh, this is the perf results: https://perf.rust-lang.org/compare.html?start=10a76d634781180b4f5402c519f0c237d3be6ee6&end=b6303fdafa669e76fca4cabfa942c07295d3f457

That's why rust-timer isn't requeueing the job, b/c it's already done

[lcnr]

it seems like the perf results are pretty much equivalent between this and #138542. Would prefer going ahead with this PR @compiler-errors

it has a better vibe for me. I think the big part is avoiding to use may_coerce in the happy path and keeping casts as

• try coerce
• if that fails, cast

[compiler-errors]

I'm still mostly in favor of the approach to relocate a (solely) cast-related hack to casting code, which is the primary motivation of #138542.

I know that you seem to have a negative opinion of the added complexity of the casting code in that PR, but IMO that added complexity (which is now bubbled up into check_cast) reflects the complexity of the change being introduced in #138542 and its interaction with the strangely (possibly load bearing) false-positive-biased behavior of coerce_unsized.

There's quite a few usages of coercion around HIR typeck, and none of them care about *const dyn Tr -> *const W<dyn Tr>-like coercions other than as casts, so I don't see why we wouldn't want to localize this behavior as closely to the code that concerns it to avoid its behavioral side-effects even needing to be considered for other types of coercion.

Making coerce_unsized less complicated is definitely a benefit that this PR doesn't really have, even if it recovers most of the perf loss in the original PR.