Bug 120963

Summary:	REGRESSION (r155069?!): Lots of reftests fail with single line or single pixel differences
Product:	WebKit	Reporter:	Alexey Proskuryakov <ap>
Component:	Layout and Rendering	Assignee:	Nobody <webkit-unassigned>
Status:	RESOLVED DUPLICATE
Severity:	Normal	CC:	benjamin, rniwa, simon.fraser, thorton
Priority:	P2
Version:	528+ (Nightly build)
Hardware:	Unspecified
OS:	Unspecified

Description Alexey Proskuryakov 2013-09-07 13:04:18 PDT

See e.g. <http://build.webkit.org/results/Apple%20MountainLion%20Release%20WK1%20(Tests)/r155249%20(12626)/results.html> - there are lots of tests that are marked flaky because they failed on first run and passed on second run. It's most noticeable on fast/regions and multicolumn tests, but it's not limited to these.

I've been looking into this for a while, and I couldn't find any way to reproduce locally short of running the whole test suite, and even that is not 100% reliable.

So far, I could reproduce locally with r155076, and couldn't with r155065, r155064 or r155056. But there seems to be absolutely nothing suspicious between r155065 and r155076.

Comment 1 Alexey Proskuryakov 2013-09-07 22:33:34 PDT

Surprising as it is, bisection says that this regressed in <http://trac.webkit.org/changeset/155069> (PDFImageDocument changes).

Tim, any thoughts as to how to confirm or deny these results? Should we temporarily roll out the change to see if that helps?

Comment 2 Tim Horton 2013-09-07 22:46:16 PDT

(In reply to comment #1)
> Surprising as it is, bisection says that this regressed in <http://trac.webkit.org/changeset/155069> (PDFImageDocument changes).
> 
> Tim, any thoughts as to how to confirm or deny these results? Should we temporarily roll out the change to see if that helps?

Whaaaat. That change was extraordinarily self contained; that would be extremely surprising. Regions and multicol have nothing at all to do with PDFDocumentImage...

Comment 3 Tim Horton 2013-09-07 22:46:44 PDT

(In reply to comment #1)
> Surprising as it is, bisection says that this regressed in <http://trac.webkit.org/changeset/155069> (PDFImageDocument changes).
> 
> Tim, any thoughts as to how to confirm or deny these results? Should we temporarily roll out the change to see if that helps?

You can do that, but if it succeeds I give up on computers :(

Comment 4 Alexey Proskuryakov 2013-09-08 11:12:26 PDT

Rolled out in <http://trac.webkit.org/changeset/155306>, we'll know soon if that helped.

Comment 5 Alexey Proskuryakov 2013-09-08 12:20:50 PDT

Yes, the rollout fixed the tests.

*** This bug has been marked as a duplicate of bug 121008 ***

Comment 6 Simon Fraser (smfr) 2013-09-09 11:29:56 PDT

It looks to me like PDFImageDocument is changing CG state somehow.