1246186 - Figure out the right semantics for Allocate<JSObject, NoGC>

Status: RESOLVED INCOMPLETE

(Core :: JavaScript: GC, defect)

Description

[Terrence Cole [:terrence]]

After bug 1245965 lands, CreateMatchResultFallbackFunc will be the only caller of Allocate<JSObject, NoGC>.

Because of the NewObjectCache (the prior only user), the current semantics are that Allocate<JSObject, NoGC> will just fail outright if the Nursery is full, rather than falling back to Tenured and filling up memory. This behavior is quite weird and frequently not what is intended. Is this actually the behavior you wanted? Now that CreateMatchResultFallbackFunc is the only caller, we could shift the semantics however we want.

Comment 1

[Tooru Fujisawa [:arai]]

wow, thank you for pointing that out.
I checked the return value of Allocate<JSObject, NoGC> and it's 0 in most of time (it returns non-null only 2% while running octane/regexp)

commenting out |if (!allowGC) return nullptr;| doesn't improve the octane score, but regress.

octane/regexp score:

  current m-c:                                      3559
  current m-c + comment out !allowGC branch:        3003
  current m-c + remove fallback for createGCObject: 3539

  bug 887016 patched:                               3212
  bug 887016 + comment out !allowGC branch:         2991
  bug 887016 + remove fallback for createGCObject:  3215

so, at least fallback to Tenured won't be what we want there.
anyway, maybe, we just don't have to fallback to Allocate when createGCObject fails?
(looks like fallback for newGCString/newGCFatInlineString was sufficient...)