413345 - Replacing ParsedObjectBox with chunked lists

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, enhancement)

Description

[Igor Bukanov]

[This is a spin-off from an initial prototype of patch for bug 406356. Since this is useful on its own, I prefer to have a separated bug for it.]

Currently to root an object created during parsing the code allocates the special JSParsedObjectBox struct. The struct holds a pointer to the object plus extra 2 fields. The first field is for linking all the structs so the GC can trace the objects. The second field is for the emitter.

But this mixing of the emitter and GC-tracing fields complicates the code and requires an extra level of indirection to access the object. So as an alternative I suggest to use separated lists for the emitter and GC rooting. This should simplify the code and, in fact, save memory as such lists can be implemented as a lists of chunks with each chunk holding few pointers.

For example, currently the parser allocates 3*N words to put N objects on the sinle JSparseContext.rootList list. With 2 separated chunked lsists where each chunk holds 4 object pointers and one linking field, the space requrement would be ceil(N/4) * (4 + 1) * 2 or 2.5*N words. In addition the space allocated for the emitter list would be reused after compiling each top-level function in a script.

Comment 1

[Igor Bukanov]

Attachment: backup of a working prototype

Comment 2

[Ryan VanderMeulen [:RyanVM]]

Is this still wanted?

Comment 3

[Igor Bukanov]

(In reply to Ryan VanderMeulen from comment #2)
> Is this still wanted?

Probably - AFAIK the parser still uses a few list structures that could benefit from the optimization.

Comment 4

[Mayank Bansal]

Been 13 years - Engine has changed a lot since then.
Closing this, but please reopen if still needed.

Comment 5

[Iain Ireland [:iain]]

I confirm that this is WONTFIX. The parser no longer touches GC things.