850534 - Yarr performance improvements

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Hannes Verschore [:h4writer]]

This is mainly by looking at the diff between our yarr folder and webkit's. It consist of reserving the memory of vector before doing append. A second trick is using "swap" to get the contents of the vector, instead of copying it. (This is possible as the original vector isn't used anymore. We already did a similar trick for one of the two vectors, but instead used copy and clear).

This is good for 6% on v8-regexp.

Comment 1

[Hannes Verschore [:h4writer]]

Attachment: Patch

Annotate pointed into your direction, Dave. Feel free to nominate somebody else.

Comment 2

[Hannes Verschore [:h4writer]]

Thanks,

https://hg.mozilla.org/integration/mozilla-inbound/rev/9f39547cfab3

Comment 3

[Hannes Verschore [:h4writer]]

I'll investigate tomorrow, but awfy somehow doesn't show this 6% improvement ?!

Comment 4

[Hannes Verschore [:h4writer]]

When measuring on my computer I took an average. And looks like before this commit v8-regexp was really jumping up and down here. With this patch it is dead stable. Somehow awfy didn't had this jumping and therefore also not the performance increase, booh, but it will definitely decrease the memory usage, yeay.

Comment 5

[Hannes Verschore [:h4writer]]

Looks like octane-regexp did get this improvement, a 4% one :D

Comment 6

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/9f39547cfab3

Comment 7

[Hannes Verschore [:h4writer]]

I think this caused the crashes mentioned in bug 825243. Backout just to be sure.

https://hg.mozilla.org/integration/mozilla-inbound/rev/3e1241c6d2ce

Comment 8

[Ryan VanderMeulen [:RyanVM]]

Backout:
https://hg.mozilla.org/mozilla-central/rev/3e1241c6d2ce

However, we still hit the same crash after backing out, so I think you're OK to re-land this.

Comment 9

[Hannes Verschore [:h4writer]]

Attachment: Use MOZ_CRASH when failing to allocate

Waldo suggested to use this for when we are failing allocating in Vector...

Comment 10

[Hannes Verschore [:h4writer]]

Because this wasn't the reason for the crashes, I pushed again.

https://hg.mozilla.org/integration/mozilla-inbound/rev/a7ea5daa2dc9

Comment 11

[David Anderson [:dvander] - inactive, e-mail if emergency]

Comment on attachment 725359 [details] [diff] [review]
Use MOZ_CRASH when failing to allocate

Review of attachment 725359 [details] [diff] [review]:
-----------------------------------------------------------------

I guess this will be better than what's there, but I can guarantee we'll see OOM crash reports. Is there any way we can propagate OOM?

Comment 12

[Hannes Verschore [:h4writer]]

Looks like we indeed see OOM crashes there. Original bug to propagate OOM is bug 574459 open since FF4.0 :(

https://hg.mozilla.org/integration/mozilla-inbound/rev/2c41bf87b4e5

Comment 13

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/a7ea5daa2dc9
https://hg.mozilla.org/mozilla-central/rev/2c41bf87b4e5