perf slowdown from #70831 (remove a stack frame from .await calls)

[nikomatsakis]

Post-merge perf results are 100% restricted to await-call-tree, and it's more clearly >1% now, post-rust-lang/rustc-perf#645 (cc @Mark-Simulacrum).

Presumably we consider that test a stress test and we're fine with the 1-5% slowdown?
cc @rust-lang/wg-async-foundations

Originally posted by @eddyb in #70831 (comment)

[nikomatsakis]

cc @sfackler, author of #70831

[sfackler]

It does seem a bit strange to me that there's a slowdown here - the complexity of the overall code post expansion seems like it should be just about the same as before, right?

[eddyb]

IIRC, a poll_with_context call was replaced with two calls, to Future::poll and get_context.

That would presumably double type-checking time in an .await-heavy function.

I'm actually a bit surprised all of the unsafety (i.e. get_context) isn't contained entirely in the HIR lowering, but I suppose there's several casts involved.

[tmandry]

Context for wg-prioritization: here's the benchmark. It's sort of a stress test, but not an unrealistic one in that it only goes 7 levels deep (I'm sure we have some call trees that deep, and maybe with a 3x factor, in Fuchsia, but only a few).

[spastorino]

Assigning P-medium as discussed as part of the Prioritization Working Group process and removing I-prioritize.