1328954 - stylo: Investigate any performance win from using opt-level 3 on Rust code

Status: RESOLVED WONTFIX

(Core :: CSS Parsing and Computation, defect, P3)

Description

[Bobby Holley (:bholley)]

Rust code currently builds with opt-level 2, since according to Nathan opt-level 3 bloats code with minimal performance impact.

According to bug 1328497 comment 15, switching from 2 to 3 bloats the rust code by about 230k. This is certainly 230k we'd rather not take, but not off the table if it turns out to significantly increase style system performance.

We should measure when we're getting close to shipping and evaluate the tradeoff.

Comment 1

[J. Ryan Stinnett [:jryans] (Use needinfo, replies may be slow)]

Here's a Perfherder comparison of opt-level 2 (the current setting) vs. opt-level 3:

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=8fd0766283fcd335d79741d6a97fe8fe651b04e0&newProject=try&newRevision=29d9d38997cc247977649ffe26641ef765e79806&framework=1&showOnlyImportant=0&showOnlyConfident=1

The results are fairly noisy, but there appear to be no strong gains for Stylo enabled results with opt-level 3.  Curiously, there are a few medium confidence regressions and improvements for Stylo _disabled_ runs, which is a bit odd, but perhaps they are affected by the larger binary size in some way.  The main_startup_fileio test shows that opt-level 3 does indeed add ~204k of file IO (undoubtedly from the larger binary).

So, like we suspected in comment 0, opt-level 3 doesn't appear to be worth it when looking at our current performance metrics.