470868 - Occasional few-pixel reftest failure in reftests/svg/smil/container/ (when moving elements between SVG document fragments)

Status: NEW

(Core :: SVG, defect, P5)

Description

[Daniel Holbert [:dholbert]]

When running the reftests added in bug 216462's latest patch, I ran into an occasional few-pixel reftest failure, which doesn't seem to be SMIL-specific. (I managed to reproduce it even after taking out all <animate> elements and all SMIL API calls)

It seems to be a random failure with low probability, though I often can reproduce it multiple times in a row and then can never reproduce it at all.  It seems to depend on environmental conditions.

I'll attach some reftest-snapshots of the failure, to demonstrate the problem.

I'm not sure how to resolve or debug this issue, since it's so hard to reproduce -- I'm mostly filing this bug to have a place to discuss & track it.

Comment 1

[Daniel Holbert [:dholbert]]

Attachment: snapshot 1a: enveloped-tree-1 testcase

Comment 2

[Daniel Holbert [:dholbert]]

Attachment: snapshot 1b: enveloped-tree-1 reference

Comment 3

[Daniel Holbert [:dholbert]]

Attachment: snapshot 1c: enveloped-tree-1 DIFFS

Comment 4

[Daniel Holbert [:dholbert]]

Attachment: snapshot 2a: moved-tree-1 testcase

Comment 5

[Daniel Holbert [:dholbert]]

Attachment: snapshot 2b: moved-tree-1 reference

Comment 6

[Daniel Holbert [:dholbert]]

Attachment: snapshot 2c: moved-tree-1 DIFFS

Comment 7

[Daniel Holbert [:dholbert]]

Attachment: snapshot 3a: enveloped-tree no-animation testcase

Comment 8

[Daniel Holbert [:dholbert]]

Attachment: snapshot 3b: enveloped-tree no-animation reference

Comment 9

[Daniel Holbert [:dholbert]]

Attachment: snapshot 3c: enveloped-tree no-animation DIFFS

Comment 10

[Daniel Holbert [:dholbert]]

Attachment: testcase 1 (enveloped-tree no-animation)

Here's the reftest testcase behind "snapshot 3a: enveloped-tree no-animation testcase" (attachment 354257 [details]).  It's just a modified version of "enveloped-tree-1", with one circle removed for simplicity and with all animations & SMIL API calls removed.

Comment 11

[Daniel Holbert [:dholbert]]

Attachment: reference 1 (enveloped-tree no-animation)

Here's the reftest reference case behind "snapshot 3b: enveloped-tree no-animation
reference" (attachment 354258 [details]).

Comment 12

[Daniel Holbert [:dholbert]]

Attachment: failure log for "moved-tree-1.xhtml" on mac

I've hit this bug with "moved-tree-1.xhtml" on the mac TryServer tinderboxen twice over the last week. e.g.
http://tinderbox.mozilla.org/showlog.cgi?log=MozillaTry/1254686026.1254695109.22469.gz

I'm attaching the failure log -- this can be pasted / loaded into Mozilla's "reftest-analyzer.xhtml" tool to see the few-pixel differences (available at http://hg.mozilla.org/mozilla-central/raw-file/tip/layout/tools/reftest/reftest-analyzer.xhtml )

Note that the differences in this failure are a few pixels at the left edge of the snapshot, whereas the differences shown in attachment "snapshot 2c" (using the same reftest) are at the right edge.  That may be a platform difference -- this new failure is from a mac tinderbox, whereas the old one was from my own Linux box.

Comment 13

[Daniel Holbert [:dholbert]]

I've marked "moved-tree-1.xhtml" as being currently random, due to this bug:
http://hg.mozilla.org/mozilla-central/rev/f511a5abb19d

Comment 15

[(not currently active) Ted Mielczarek]

deferred-anim-1.xhtml (from the duped bug 524098) was failing pretty frequently on mac.

Comment 16

[Daniel Holbert [:dholbert]]

Yeah, that's not out of the ordinary for this bug.  (per comment 0, sometimes this can happen multiple times in a row / frequently, whereas other times I can't get it to happen at all)

Comment 17

[Daniel Holbert [:dholbert]]

Attachment: reftst output from Cedar OpenGL failures (deferred-tree-1.xhtml, deferred-tree-2a.xhtml, deferred-tree-2b.xhtml)

The Birch project branch (which is currently tracking some OpenGL-backend work) is hitting this failure (12-16px difference) frequently on a few more tests:
deferred-tree-1.xhtml, deferred-tree-2a.xhtml, deferred-tree-2b.xhtml
http://tinderbox.mozilla.org/showlog.cgi?log=Birch/1284407086.1284408730.26415.gz

reftest output from that log is attached.

It may turn out that the OpenGL backend makes this bug easier to reproduce -- if so, that'll greatly assist in debugging it! :) In the meantime, though, we might as well mark those three tests as random (or random-on-mac), as we have with other tests here.

Comment 18

[Joe Drew (not getting mail)]

I marked some of these reftests as random on OpenGL in http://hg.mozilla.org/mozilla-central/rev/eaba42694849

Comment 19

[Robert Longson [:longsonr]]

Attachment: WIP

Comment 20

[Robert Longson [:longsonr]]

Daniel the WIP patch fixes everything except moved-tree. It looks like something has completely broken this since it was written as with or without this patch the test and the reference look completely different. Any ideas?

Comment 21

[Robert Longson [:longsonr]]

Attachment: patch

Comment 22

[Daniel Holbert [:dholbert]]

Comment on attachment 535951 [details] [diff] [review]
patch

>-random == moved-tree-1.xhtml moved-tree-1-ref.xhtml  # bug 470868
[...]
>+== moved-tree-1.xhtml moved-tree-1-ref.xhtml

At least on my system, moved-tree still fails completely after this patch, as you mentioned in comment 20 here.  (The circle is at opposite ends of the green bar in testcase vs. reference.)

We probably need to replace "random" with "fails" for that one, rather than just removing 'random', right?  Could you do that & file a followup on that regression, to figure out what's going on there?

r=dholbert with that.  Thanks for fixing these!

Comment 23

[Daniel Holbert [:dholbert]]

(and note the followup bug number in reftest.list, so that the the 'fails' annotation will point to something)

Comment 24

[Cameron McCormack (:heycam)]

Robert, is this patch still good for fixing up (per comment 22) and landing?

Comment 25

[Robert Longson [:longsonr]]

As far as I know it is. I can't remember why I abandoned work on it now.