Use mimalloc in rustc

[iSwapna]

No description provided.

[rustbot]

r? @pnkfelix

rustbot has assigned @pnkfelix.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[fee1-dead]

We've had several experiments in #92249 and #92317. Any reasons for how the performance of mimalloc could have changed to justify replacing jemalloc with it?

[iSwapna]

We've had several experiments in #92249 and #92317. Any reasons for how the performance of mimalloc could have changed to justify replacing jemalloc with it?

My goal is to measure the performance and see if it's worth to do anything further. That's why the PR is a draft. I needed the CI to run. Let me know if that is something I can do.

[fee1-dead]

There are some version updates, and it doesn't hurt to try. But please follow some additional things I have done in #92249 too, namely replacing it for rustdoc too. Because we also benchmark rustdoc in the perf suite it is helpful to see the impact there also.

afterwards I think it is okay to request a perf run.

[iSwapna]

There are some version updates, and it doesn't hurt to try. But please follow some additional things I have done in #92249 too, namely replacing it for rustdoc too. Because we also benchmark rustdoc in the perf suite it is helpful to see the impact there also.

afterwards I think it is okay to request a perf run.

I will do that! Thank you for the pointers.

[klensy]

There was more recent #103944 and #122369

[iSwapna]

There was more recent #103944 and #122369

Will run the perf and then compare with last run

[iSwapna]

@fee1-dead I will get back to this after exams (late Dec).

[bors]

☔ The latest upstream changes (presumably #133770) made this pull request unmergeable. Please resolve the merge conflicts.

[kamulos]

there now exists a v3.0.1-alpha which might be worth a test

[lqd]

We've done recent measurements of mimalloc, and things haven't changed all that much, as far as I can tell.

Here are the results with the latest libmimalloc-sys, like in this PR (mimalloc 2.1.7 from last May):

• 2% mean improvement on icounts. That used to be bigger in the past but jemalloc has also improved in the mean time.
• cycles look very meh here
• max-rss regressions on realistic crates like cargo, ripgrep, etc of over 20%. This maximum has improved over time: it used to be a +33% regression on regex and is now at +26%.
• wall-times are mixed (and there's some noise and not enough runs), but most primary benchmarks seem to show regressions on full runs, but some improvements on incremental runs.

The tradeoffs have stayed relatively similar over the years (with a negative trend). The max-rss regressions look too big to be worth the change, and the runtimes are not even that clear of a win in the first place anymore.

Note on the v3.0.1 revision mentioned above: I did try to run it through our perf infra and it segfaulted during PGO. While we don't have numbers yet, there's probably no need to rush testing these alphas though, it seems unlikely to yield vastly different results once again.

This PR can probably be closed, as we've done the requested measurements already. Unless something about the max-rss significantly changes in mimalloc, it doesn't seem like a better fit than jemalloc for the workloads rustc throws at it.