Bug 129741 – JSCell::m_gcData should encode its information differently

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Mark Hahnenberg]

We want to keep track of three states for an object:

(1) Not marked (which implies not in the remembered set)
(2) Marked but not in the remembered set
(3) Marked and in the remembered set

Currently we only indicate marked vs. not marked in JSCell::m_gcData. During a write barrier, we only want to take the slow path if the object being stored to is in state (2). We'd like to make the test for state (2) as fast as possible, which probably means making it a compare against 0.

[Mark Hahnenberg]

Created attachment 225902 [details]
Patch

[Mark Hahnenberg]

Created attachment 225903 [details]
Patch

[Mark Hahnenberg]

Created attachment 225905 [details]
Patch

[Geoffrey Garen]

Comment on attachment 225905 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=225905&action=review

r=me

Is speedup?

> Source/JavaScriptCore/runtime/JSCell.h:150
> +    void mark() { m_gcData = Marked; }

I would call this "setMarked".

> Source/JavaScriptCore/runtime/JSCell.h:151
> +    void remember() { ASSERT(m_gcData == Marked); m_gcData = MarkedAndRemembered; }

"setRemembered(true)"

> Source/JavaScriptCore/runtime/JSCell.h:152
> +    void forget() { ASSERT(m_gcData == MarkedAndRemembered); m_gcData = Marked; }

"setRemembered(false)"

[Mark Hahnenberg]

(In reply to comment #4)
> (From update of attachment 225905 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=225905&action=review
> 
> r=me
> 
> Is speedup?
No visible effect on benchmarks.

[Mark Hahnenberg]

Committed r165135: <http://trac.webkit.org/changeset/165135>