Bug 125944

Summary:

REGRESSION(r156214): 11-13% regressions on Dromaeo/jslib-traverse-jquery

Product:

WebKit

Reporter:

Ryosuke Niwa <rniwa>

Component:

JavaScriptCore

Assignee:

Nobody <webkit-unassigned>

Status:

NEW ---

Severity:

Normal

CC:

barraclough, ggaren, mhahnenberg, oliver, webkit-bug-importer

Priority:

Keywords:

InRadar, Regression

Version:

528+ (Nightly build)

Hardware:

Unspecified

OS:

Unspecified

Bug Depends on:

121717

Bug Blocks:

Attachments:

Description	Flags
A/B test result	none

Description Ryosuke Niwa 2013-12-18 13:47:36 PST

See https://perf.webkit.org/#mode=charts&chartList=%5B%5B%22mac-mountainlion%22%2C%22Dromaeo%2Fjslib-traverse-jquery%3ARuns%22%5D%2C%5B%22mac-mountainlion%22%2C%22Dromaeo%2Fjslib-traverse-prototype%3ARuns%22%5D%2C%5B%22mac-lion%22%2C%22Dromaeo%2Fjslib-traverse-jquery%3ARuns%22%5D%5D&days=218&zoom=%5B1379161460728.1538%2C1380252801171.3845%5D

http://trac.webkit.org/log/?rev=156215&stop_rev=156206&verbose=on

Comment 1 Radar WebKit Bug Importer 2013-12-20 11:57:46 PST

<rdar://problem/15710364>

Comment 2 Ryosuke Niwa 2013-12-20 17:33:04 PST

I've bisected locally.  This regression is caused by http://trac.webkit.org/changeset/156214.

Comment 3 Ryosuke Niwa 2013-12-20 17:38:06 PST

Created attachment 219817 [details]
A/B test result

Comment 4 Ryosuke Niwa 2013-12-20 17:40:39 PST

*** Bug 125953 has been marked as a duplicate of this bug. ***

Comment 5 Mark Hahnenberg 2013-12-20 17:55:47 PST

The original bug was that we would be in the middle of copying, see and hole and we realize we should have done something else, and then we start copying all over again.

The fix was a simple one: first check if there's a hole, then do the copy. So you're always doing two passes over the array.

If we want to fix this regression then we need a way to start copying and then gracefully fail when we see a hole without forgetting that we had already started copying.

Comment 6 Mark Hahnenberg 2013-12-20 18:01:07 PST

(In reply to comment #5)
> The original bug was that we would be in the middle of copying, see and hole and we realize we should have done something else, and then we start copying all over again.
> 
..see a hole, realize we should...

Jeez.

Comment 7 Ryosuke Niwa 2013-12-20 18:37:23 PST

I think we need to come up with some strategy to improve the performance of jslib-traverse-jquery because we've regressed this particular test thrice:

https://perf.webkit.org/#mode=charts&chartList=%5B%5B%22mac-mountainlion%22%2C%22Dromaeo%2Fjslib-traverse-jquery%3ARuns%22%5D%5D&days=200

We used to get 22 runs/s but now we're getting 17 runs/s, a hopping 30% regression in the total :(

Perhaps we can improve elsewhere to compensate?