[HACK] metadata/rmeta: skip sorting impls by DefPathHashes.

[eddyb]

See #69060 (comment) for some background.

I'm mostly opening this to do some experimentation (which will require multiple perf runs using the same base, so I'll probably start try builds in parallel using a second PR).

I'd prefer the solution described in #69060 (comment), as it should result in better accuracy as well, but that's more work than this quick hack, at least for testing purposes.

cc @nikomatsakis @michaelwoerister @Mark-Simulacrum

[rust-highfive]

r? @petrochenkov

(rust_highfive has picked a reviewer for you, use r? to override)

[eddyb]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[eddyb]

@bors try (oops, started the other build too soon)

[bors]

⌛ Trying commit d6b49f9 with merge 335898da9076089c34ec0596e12cc7de600ff03b...

[eddyb]

(try builds both seem to be running, common merge parent 697d6b3)

[bors]

☀️ Try build successful - checks-azure
Build commit: 335898da9076089c34ec0596e12cc7de600ff03b (335898da9076089c34ec0596e12cc7de600ff03b)

[rust-timer]

Queued 335898da9076089c34ec0596e12cc7de600ff03b with parent 697d6b3, future comparison URL.

[rust-timer]

Finished benchmarking try commit 335898da9076089c34ec0596e12cc7de600ff03b, comparison URL.

[eddyb]

Comparison between "this PR" and "this PR + empty commit".

Nothing I've looked at has any query count differences, which is nice, but there's still a lot of noise from codegen/LLVM, so that needs to be looked into separately. I suppose I was wrong to assume that the query counts and noisy codegen/LLVM were related.

[petrochenkov]

r? @ghost