1833114 - Use separate mark stacks for black and gray marking

Status: RESOLVED FIXED

(Core :: JavaScript: GC, task, P3)

Description

[Jon Coppeard (:jonco) (PTO until 14th September)]

I know we've flip-flopped on this a few times before but using separate stacks simplifies the code and should make some parallel marking optimisaitons easier. However we need to do this without regressing memory usage.

Comment 1

[Jon Coppeard (:jonco) (PTO until 14th September)]

Attachment: Bug 1833114 - Use separate mark stacks for black and gray marking r?sfink

This adds a separate auxililary mark stack and switches to it when necessary.
Switching is done by std::swap so that |stack| is always the stack for the
current mark color. We avoid swapping if both stacks are empty so that we don't
always use the auxiliary stack.

This removes a bunch of state from MarkStack and assertions around ensuring
that state is valid.

Comment 2

[Jon Coppeard (:jonco) (PTO until 14th September)]

Attachment: Bug 1833114 - Simplify marking code now |stack| represents the mark stack for the current color r?sfink

We can now make a few simplifications now |stack| is always the current stack.

Depends on D178040

Comment 3

[Jon Coppeard (:jonco) (PTO until 14th September)]

I'm never sure how to interpret these results, but all the Windows ones are same or negative on size, so I don't think this will cause a regression:

https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=3c889d30efb3772c900861f2a782f359e5186e64&newProject=try&newRevision=e762711f87a7616ec2e8cda7d3aab3ae87ca2c53&framework=4&page=1

Comment 4

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/8b4f4da78faa
Use separate mark stacks for black and gray marking r=sfink
https://hg.mozilla.org/integration/autoland/rev/0d1f0b668a79
Simplify marking code now |stack| represents the mark stack for the current color r=sfink

Comment 5

[Norisz Fay [:noriszfay]]

https://hg.mozilla.org/mozilla-central/rev/8b4f4da78faa
https://hg.mozilla.org/mozilla-central/rev/0d1f0b668a79

Comment 6

[Steve Fink [:sfink] [:s:]]

(In reply to Jon Coppeard (:jonco) from comment #3)

I'm never sure how to interpret these results, but all the Windows ones are same or negative on size, so I don't think this will cause a regression:

Ok, I'm not sure why my version was so problematic, but I definitely prefer the simplicity of separate stacks and so I'm happy to rely on these results.