503283 - "unbounded" globalCompositeOperations are terribly broken

Status: RESOLVED FIXED

(Core :: Graphics: Canvas2D, defect)

Description

[Masatoshi Kimura [:emk]]

Build ID:
Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv:1.9.2a1pre) Gecko/20090708 Minefield/3.6a1pre
Steps to reproduce:
See the above URL.
Actual result:
source-in, source-out, destination-in, and destination-atop are not rendered at all.
Expected result:
Some sort of composition result should be rendered.

Fx 3.5 (Mozilla/5.0 (Windows; U; Windows NT 5.1; ja; rv:1.9.1) Gecko/20090624 Firefox/3.5) also fails. Fx 3.0.11 (Mozilla/5.0 (Windows; U; Windows NT 5.1; ja; rv:1.9.0.11) Gecko/2009060215 Firefox/3.0.11) seems to work.

Comment 1

[Masatoshi Kimura [:emk]]

Another failing example:
http://uupaa-js-spinoff.googlecode.com/svn/trunk/uuCanvas.js/demo1/5_1_canvas_savestate+shadow.html

Comment 2

[Masatoshi Kimura [:emk]]

http://hg.mozilla.org/mozilla-central/rev/dcec193ba5d7 (20081105): works
http://hg.mozilla.org/mozilla-central/rev/f3993df71c29 (20081106): broken

Comment 3

[Masatoshi Kimura [:emk]]

pushlog in regression range:
http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=dcec193ba5d7&tochange=f3993df71c29
regression from bug 462938?

Comment 4

[Takeshi Kurosawa :takenspc]

Attachment: testcase using drawImage() and shadow (broken)

drawImage() or shadows cause this problem.
# fill() or stroke() look OK.

I attached modified version (using fill()) of "URL" testcase.
That doesn't cause the problem.

Comment 5

[Takeshi Kurosawa :takenspc]

Attachment: testcase using fill() (not broken)

Oops, I attached shadow testcase!
This one uses fill().

Comment 6

[Takeshi Kurosawa :takenspc]

Comment on attachment 387637 [details]
testcase using drawImage() and shadow (broken)

This uses shadow not fill()

Comment 7

[Masatoshi Kimura [:emk]]

Comment #1 seems to be a different problem.
2007072505 not rendered at all (me.mozDrawText is not a function)
2007072517 cvs trunk: works
2008022002 mozilla-central: not rendered at all (ctx.fillText is not a function)
2008022005 cvs trunk: works
2008052006 cvs trunk: works
2008052103 mozilla-central: not rendered at all
2008061012 mozilla-central: not rendered at all
2008061107 mozilla-central: broken
I was unable to spot a regression range due to mercurial transition...

Comment 8

[Takeshi Kurosawa :takenspc]

Sorry, I confused a bit...

drawImage() causes the problem.
shadows does not cause the problem.

(In reply to comment #1)
> Another failing example:
> http://uupaa-js-spinoff.googlecode.com/svn/trunk/uuCanvas.js/demo1/5_1_canvas_savestate+shadow.html

This is not failed.

http://www.whatwg.org/specs/web-apps/current-work/multipage/the-canvas-element.html#dom-context-2d-globalcompositeoperation
> source-out
>     A out B. Display the source image wherever the source image is opaque and the destination image is transparent. Display transparency elsewhere.

Both "green shadow" and "red shadow" are rendered twice, but first one of them are cleared when rendering "blue shadow".

Comment 9

[Masatoshi Kimura [:emk]]

(In reply to comment #8)
> Both "green shadow" and "red shadow" are rendered twice, but first one of them
> are cleared when rendering "blue shadow".
All other browsers (Opera 9, Safari 4, Chrome 2, and even Fx3) doesn't clear the first "green shadow" and "red shadow". Is it an incorrect behaviour per spec?

Comment 10

[Takeshi Kurosawa :takenspc]

(In reply to comment #9)
> (In reply to comment #8)
> > Both "green shadow" and "red shadow" are rendered twice, but first one of them
> > are cleared when rendering "blue shadow".
> All other browsers (Opera 9, Safari 4, Chrome 2, and even Fx3) doesn't clear
> the first "green shadow" and "red shadow". Is it an incorrect behaviour per
> spec?

In these browsers source-in, source-out,... operators are "bounded".
# ... Display transparency elsewhere covered by the source image.
But after bug 462938 (cairo 1.8.x?) these are "unbounded".
# ... Display transparency elsewhere.

http://cairographics.org/operators/

Comment 11

[Masatoshi Kimura [:emk]]

I found SVG Compositing calls them as clip-to-self property.
http://www.w3.org/TR/SVGCompositing/#clip-to-self
However, even if the operators are "unbounded", the "URL" testcase doesn't look correct to me.
If the operaters are "unbounded", It should get the similar result as Opera.
http://d.hatena.ne.jp/chaperatta/20080614/1213454680

Comment 12

[Takeshi Kurosawa :takenspc]

(In reply to comment #11)
> However, even if the operators are "unbounded", the "URL" testcase doesn't look
> correct to me.
> If the operaters are "unbounded", It should get the similar result as Opera.
> http://d.hatena.ne.jp/chaperatta/20080614/1213454680

Yes. The "URL" testcase doesn't work well for Firefox 3.5 and later.
It uses drawImage() (source is an image) and drawImage() seems cause this problem.
# 
# The testcase I attached uses fill() (source is a path) and its result seems OK.
# -> It seems like the cause of this bug is the handling of images.

Comment 13

[Masatoshi Kimura [:emk]]

nsCanvasRenderingContext2D::DrawImage() calls UpdateSurfaceClip()
http://hg.mozilla.org/releases/mozilla-1.9.1/annotate/f223409207c0/content/canvas/src/nsCanvasRenderingContext2D.cpp#l2983
which trys to fill the empty rect.
http://hg.mozilla.org/releases/mozilla-1.9.1/annotate/f223409207c0/gfx/thebes/src/gfxContext.cpp#l627
Then the entire canvas will be cleared out because the operator is "unbounded".

Comment 14

[Masatoshi Kimura [:emk]]

Attachment: remove old workarounds

-    // XXX cairo bug workaround; force a clip update on mThebes.
This workaround is now useless or even harmful. Bug 346421 and bug 347379 didn't regress.

-        // XXX certain operators cause 2d.composite.uncovered.* tests to fail
I don't know the exact reason why OperatorAffectsUncoveredAreas was temporarily disabled, but 2d.composite.uncovered.* tests failed without enabling it (as the comment indicates).

Comment 15

[Masatoshi Kimura [:emk]]

Fixed by bug 599698 and bug 600120.

Comment 16

[Masatoshi Kimura [:emk]]

Comment on attachment 388920 [details] [diff] [review]
remove old workarounds

Canceling review request of an obsolete patch.