932800 - IonMonkey always bails for string[string] GETELEM

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Jan de Mooij [:jandem]]

The jQuery event benchmark in bug 928318 has a "trigger" function that's bailing out from Ion to Baseline every time it's called:

event = event[ jQuery.expando ] ?
	event :
	new jQuery.Event( type, typeof event === "object" && event );

The problem is that |event| is a string, and we try to convert jQuery.expando (also a string) to int32, but this will always bail because string to int32 is complicated. Instead, we should only try to optimize string[x] if x is int32.

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Patch

This also fixes another issue that Nicolas fixed in bug 799818 but apparently got lost: don't optimize string[index] if the index has been out-of-bounds before to avoid frequent bailouts.

Comment 2

[Hannes Verschore [:h4writer]]

Was actually just looking into PDF.js and saw the same issue. This improves pdf.js with 4% (didn't test elaborate, so wait for awfy result.)

Comment 3

[Hannes Verschore [:h4writer]]

Comment on attachment 824675 [details] [diff] [review]
Patch

Review of attachment 824675 [details] [diff] [review]:
-----------------------------------------------------------------

Looks fine. The test for Undefined was unexpected, but does the job. Isn't Out-of-Bound access something that the baseline keeps track of? I think we only do this for writes now. But why not for reads?

Comment 4

[Jan de Mooij [:jandem]]

(In reply to Hannes Verschore [:h4writer] from comment #3)
> Looks fine. The test for Undefined was unexpected, but does the job.

See also jsop_getelem_typed for instance.

> Isn't
> Out-of-Bound access something that the baseline keeps track of? I think we
> only do this for writes now. But why not for reads?

An out-of-bounds read will produce |undefined| (ignoring indexed properties on the prototype) so there's no need to track this separately.

Comment 5

[Jan de Mooij [:jandem]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/29b8e502a296

(Try: https://tbpl.mozilla.org/?tree=Try&rev=f43d7e939533)

Comment 6

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/29b8e502a296