1125101 - Incrementalise compacting GC by zones

Status: RESOLVED FIXED

(Core :: JavaScript: GC, defect)

Description

[Jon Coppeard (:jonco)]

We should be able to split up compacting GC by running it a zone at a time until we run out of slice budget.  The is still going to be fairly coarse-grained but better than what we have now.

Comment 1

[Jon Coppeard (:jonco)]

Attachment: incremental-compacting-1 WIP

Work in progress.  Reorganise compacting to work a zone at a time.

Comment 2

[Jon Coppeard (:jonco)]

Attachment: incremental-compacting-2-budget

Work in progress.  Budget compacting and yield appropriately.

Comment 3

[Jon Coppeard (:jonco)]

Attachment: incremental-compacting-1

Comment 4

[Jon Coppeard (:jonco)]

Attachment: incremental-compacting-2-budget

Comment 5

[Terrence Cole [:terrence]]

Comment on attachment 8567913 [details] [diff] [review]
incremental-compacting-1

If you export your patches with a .patch suffix, bugzilla will recognize them as patches and set the "patch" flag automatically.

Comment 6

[Terrence Cole [:terrence]]

(In reply to Terrence Cole [:terrence] from comment #5)
> Comment on attachment 8567913 [details] [diff] [review]
> incremental-compacting-1
> 
> If you export your patches with a .patch suffix, bugzilla will recognize
> them as patches and set the "patch" flag automatically.

Derp, I said patch too many times in that sentence: it should be a .diff suffix.

Comment 7

[Terrence Cole [:terrence]]

Comment on attachment 8567913 [details] [diff] [review]
incremental-compacting-1

Review of attachment 8567913 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jsgc.cpp
@@ +2468,5 @@
> +                 arena;
> +                 arena = arena->next)
> +            {
> +                return arena;
> +                resumePoint:;

This is a label so it should stay unindented by 2 spaces?

Comment 8

[Jon Coppeard (:jonco)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/f7104b435b47

Comment 9

[Jon Coppeard (:jonco)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/4eab437d2b51

Comment 10

[Ryan VanderMeulen [:RyanVM]]

Backed out for being the likely culprit behind a big spike in Linux64 debug mochitest-bc timeouts/test slowdowns. Retriggers pinned it down to this push, and bug 1130439 seemed unlikely since it landed before and didn't cause any issues like these from my recollection.
https://hg.mozilla.org/integration/mozilla-inbound/rev/dcb0abbfa423

https://treeherder.mozilla.org/#/jobs?repo=mozilla-inbound&revision=4eab437d2b51

Comment 11

[Ryan VanderMeulen [:RyanVM]]

The backout appears to have reduced the failure rate from nearly 100% back to it's previously-crummy state, so I guess that means victory here.

Comment 12

[Jon Coppeard (:jonco)]

Try runs are looking green again so attempting to re-land:

https://hg.mozilla.org/integration/mozilla-inbound/rev/4d0cce795ce0
https://hg.mozilla.org/integration/mozilla-inbound/rev/77671f29d256

Comment 13

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/4d0cce795ce0
https://hg.mozilla.org/mozilla-central/rev/77671f29d256