1917031 - Reconsider ProbablyShortLivingWrapper annotation

Status: NEW

(Core :: DOM: Bindings (WebIDL), task)

Description

[Matthew Gaudet (he/him) [:mgaudet]]

Over here I experimented with making all wrappers "ProbablyShortLivingWrapper" on top of the DOM Allocation Site patch.

This showed some nice improvements (up to 10% on the ReactSVG subtest). Which makes me think maybe once we have the DOM allocationSites we can consider removing this annotation and marking everything as nursery allocatable, relying on the allocSite to pre-tenure where possible.

Now, there are some regressions too that probably need to be considered more deeply. It may be that we actually need to flip the sense of the annotation from "ProbablyShortLivingWrapper" to "ProbablyLongLivingWrapper"

Comment 1

[Olli Pettay [:smaug][[e-mail]]]

This could be very effective in sp3, I think, iff we can store the information about allocation sites across page loads. We usually manage to avoid minorGCs during subtest runs, so without knowledge about previous runs this might not help too much - or this would mean basically just allocating always from nursery.

Comment 2

[Matthew Gaudet (he/him) [:mgaudet]]

Opened Bug 1917338 about saving AllocSite data

Comment 3

[Matthew Gaudet (he/him) [:mgaudet]]

Eyeballing this code, it seems as if all binding generated finalizers are generated here. Some notes:

• I would be unsurprised if there were challenges with observable arrays (used by adopted style sheets.
• Same about globals

Comment 4

[Matthew Gaudet (he/him) [:mgaudet]]

A more recent try push here

• Wins on TodoMVC and ChartObservable, but it's a regression overall.

I'll put this on pause for now pending Bug 1917338 though I think there is a space here for using ProbablyShortLivingWrapper more.