1728265 - 8.89 - 7.78% perf_reftest_singletons bidi-resolution-1.html / perf_reftest_singletons bidi-resolution-1.html (Linux) regression on Fri August 27 2021

Status: RESOLVED WONTFIX

(Core :: Layout: Text and Fonts, defect)

Description

[Andra Esanu (needinfo me)]

Perfherder has detected a talos performance regression from push c36347bc9a8337114d762d0cf6ec8918dc9a24bd. As author of one of the patches included in that push, we need your help to address this regression.

Regressions:

Ratio	Test	Platform	Options	Absolute values (old vs new)
9%	perf_reftest_singletons bidi-resolution-1.html	linux1804-64-shippable-qr	e10s stylo webrender	335.22 -> 365.02
8%	perf_reftest_singletons bidi-resolution-1.html	linux1804-64-shippable-qr	e10s stylo webrender	337.83 -> 364.12

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
58%	perf_reftest bidi-resolution-1.html	linux1804-64-shippable-qr	e10s stylo webrender	20.53 -> 8.62

Details of the alert can be found in the alert summary, including links to graphs and comparisons for each of the affected tests. Please follow our guide to handling regression bugs and let us know your plans within 3 business days, or the offending patch(es) will be backed out in accordance with our regression policy.

For more information on performance sheriffing please see our FAQ.

Comment 1

[Jonathan Kew [:jfkthame]]

Looking at bidi-resolution-1.html, it makes sense we could see a small regression there: the test creates a <div> containing 24000 copies of a mixed Latin- and Arabic-script string, and then reflows it, which will mean creating around 48000 continuation frames because of the bidi direction boundaries. We're now doing a little extra work to set the back-pointer to the primary frame for each nsContinuingTextFrame we create.

Although this shows up on the micro-test here, I believe the extra work is essentially insignificant in the overall context of reflow for more typical real-world content, where such a constant flip-flopping of text directions would be highly unlikely. Given that the regression we see here in perf_reftest_singletons is apparently offset by an improvement in perf_reftest, and more importantly given the dramatic benefit bug 1725555 offers in real-world user experience (around a 90-95% speedup for the example profiled there), I think we should accept this.

So my proposal is to close this as WONTFIX. ni? to Emilio for any other thoughts.

Comment 2

[Emilio Cobos Álvarez (:emilio)]

Yeah, that seems reasonable.