Dogfood new_uninit and maybe_uninit_slice in rustc_arena

[est31]

Dogfoods a few cool MaybeUninit related features in the compiler's rustc_arena crate.

Split off from #76821

r? @oli-obk

[oli-obk]

@bors r+

[bors]

📌 Commit 7b995232b6a6a02bbd4680e8b3c102623eed966c has been approved by oli-obk

[est31]

Does this need a perf run? I see some regressions in the microbenchmarks, but being microbenchmarks they might lie. Did ./x.py bench --stage 0 compiler/rustc_arena.

Output with base commit from bors:

test tests::bench_copy              ... bench:           5 ns/iter (+/- 0)
test tests::bench_copy_nonarena     ... bench:           0 ns/iter (+/- 0)
test tests::bench_noncopy           ... bench:          64 ns/iter (+/- 1)
test tests::bench_noncopy_nonarena  ... bench:          34 ns/iter (+/- 1)
test tests::bench_typed_arena_clear ... bench:           5 ns/iter (+/- 0)

Output with only the first commit:

test tests::bench_copy              ... bench:           5 ns/iter (+/- 0)
test tests::bench_copy_nonarena     ... bench:           0 ns/iter (+/- 0)
test tests::bench_noncopy           ... bench:          65 ns/iter (+/- 1)
test tests::bench_noncopy_nonarena  ... bench:          36 ns/iter (+/- 1)
test tests::bench_typed_arena_clear ... bench:           7 ns/iter (+/- 0)

Output with both commits:

test tests::bench_copy              ... bench:           5 ns/iter (+/- 0)
test tests::bench_copy_nonarena     ... bench:           0 ns/iter (+/- 0)
test tests::bench_noncopy           ... bench:          63 ns/iter (+/- 1)
test tests::bench_noncopy_nonarena  ... bench:          33 ns/iter (+/- 1)
test tests::bench_typed_arena_clear ... bench:           6 ns/iter (+/- 0

There is tons of noise in the benchmarks but I made multiple runs each time and there is a consistent difference. There might be a regression in bench_typed_arena_clear. Again, it's a microbenchmark and the usage pattern of the arena in the compiler might not reflect it at all.

If I allocate 100 items and do the same clearance I can't find a regression, in fact it's a bit faster now. Benchmark:

#[bench]
pub fn bench_typed_arena_clear_100(b: &mut Bencher) {
    let mut arena = TypedArena::default();
    b.iter(|| {
        for _ in 0..100 {
            arena.alloc(Point { x: 1, y: 2, z: 3 });
        }
        arena.clear();
    })
}

Before this PR:

test tests::bench_typed_arena_clear_100 ... bench:         177 ns/iter (+/- 5)

After this PR:

test tests::bench_typed_arena_clear_100 ... bench:         173 ns/iter (+/- 12)

Again, lot's of noise.

I've added the benchmark to this PR. I've also amended the first commit to do a RawVec → Box change in one of the comments. re-r? @oli-obk

[oli-obk]

@bors r+ rollup=never

I don't think this will have an effect on perfbot, but let's not roll this up to make sure

[bors]

📌 Commit 2805a05 has been approved by oli-obk

[bors]

⌛ Testing commit 2805a05 with merge a2c82df...

[bors]

☀️ Test successful - checks-actions, checks-azure
Approved by: oli-obk
Pushing a2c82df to master...