Good news, this worked!

Longer: I followed @matklad's steps to reproduce this issue:

$ cargo install cargo-llvm-lines
$ git clone rust-analyzer/rust-analyzer
$ cd rust-analyzer
[patching the tracing and tracing-core dependencies to this branch]
$ cd crates/ra_hir_ty
$ cargo llvm-lines

tracing_core::dispatcher::get_default is now significantly less hot, according to cargo llvm-lines:

  Lines           Copies        Function name
  -----           ------        -------------
  1202887 (100%)  46615 (100%)  (TOTAL)
    41106 (3.4%)    102 (0.2%)  alloc::raw_vec::RawVec<T,A>::grow
    31151 (2.6%)   3544 (7.6%)  core::ptr::drop_in_place
    31076 (2.6%)    539 (1.2%)  core::option::Option<T>::map
    30687 (2.6%)    317 (0.7%)  core::iter::traits::iterator::Iterator::try_fold
    27867 (2.3%)    429 (0.9%)  core::result::Result<T,E>::map_err
    25488 (2.1%)     22 (0.0%)  salsa::derived::slot::Slot<Q,MP>::read_upgrade
    24456 (2.0%)     44 (0.1%)  salsa::derived::slot::Slot<Q,MP>::probe
    17258 (1.4%)    409 (0.9%)  core::ptr::swap_nonoverlapping_one
    16742 (1.4%)     22 (0.0%)  salsa::derived::slot::Slot<Q,MP>::maybe_changed_since
    15634 (1.3%)    202 (0.4%)  alloc::vec::Vec<T>::extend_desugared

I've removed the unneeded dispatcher imports.

I observe meaningful change in cargo-llvm lines:

1687120 (100%) 52706 (100%) (TOTAL)
>>>
1587714 (100%) 51297 (100%) (TOTAL)

wall-clock time compilation impact is harder to estimate -- variance within single setup is stupidly high, and each run takes about a minute, so using proper statistic will require too much time. So, "it didn't make compilation 50% faster or slower" is all I can say.

Another thing I've noticed is that there's a bunch of calls to 14632 (0.9%) 236 (0.5%) std::sync::once::Once::call_once. (presumably from here). This shouldn't be too bad (as Once dynamizes the cold path anyway), but still I am wondering if Once is an overkill. If I am reading the code correctly, there's a mutex inside anyway, so we incur double synchronisation. It seems that Once is replaceable with Acquire/Release AtomicBool, at the cost of calling the same register more than once, which should be OK, as we lock a mutex there anyway.

@matklad I'm glad you're able to observe a meaningful change in cargo-llvm lines! No worries about the compile times on your end. I've personally observed a (relatively) consistent 1-2 second change on my beefy MacBook Pro when compiling ra_hir_ty.

Another thing I've noticed is that there's a bunch of calls to 14632 (0.9%) 236 (0.5%) std::sync::once::Once::call_once. (presumably from here). This shouldn't be too bad (as Once dynamizes the cold path anyway), but still I am wondering if Once is an overkill. If I am reading the code correctly, there's a mutex inside anyway, so we incur double synchronisation. It seems that Once is replaceable with Acquire/Release AtomicBool, at the cost of calling the same register more than once, which should be OK, as we lock a mutex there anyway.

That's interesting! Since Eliza wrote that code, I'll let her comment on it, but if you're correct, I'll open another issue + PR to fix that.

Yeah, the Once can probably be removed with a second atomic. That path hasn't been a major focus for optimization, since it only happens once per callsite, so there's definitely room for improvement. I think this may also let us remove the vendored Once dependency for no-std targets, which would probably be the biggest benefit in my book.