cc: Outset image rects by 1 for the purposes of the rtree.

cc: Outset image rects by 1 for the purposes of the rtree.

This patch ensures that we outset the visual rects used by images by 1
since at raster time, we may query/try to use images that are touching
the raster bounds, but not intersecting. This is (at least in part)
due to Skia returning device bounds that are outset by 1.

This ensures that we create proper dependency graphs for the image
tasks.

R=enne
CQ_INCLUDE_TRYBOTS=master.tryserver.blink:linux_trusty_blink_rel

Review-Url: https://codereview.chromium.org/2743203006
Cr-Commit-Position: refs/heads/master@{#457235}
Committed: https://chromium.googlesource.com/chromium/src/+/3530c014a885f3195f6cfad735a297dabe99ebc3

[vmpstr, 2017-03-14 20:58:34]

Description was changed from

==========
cc: Outset image rects by 1 for the purposes of the rtree.

This patch ensures that we outset the visual rects used by images by 1
since at raster time, we may query/try to use images that are touching
the raster bounds, but not intersecting. This is (at least in part)
due to Skia returning device bounds that are outset by 1.

This ensures that we create proper dependency graphs for the image
tasks.

R=enne
==========

to

==========
cc: Outset image rects by 1 for the purposes of the rtree.

This patch ensures that we outset the visual rects used by images by 1
since at raster time, we may query/try to use images that are touching
the raster bounds, but not intersecting. This is (at least in part)
due to Skia returning device bounds that are outset by 1.

This ensures that we create proper dependency graphs for the image
tasks.

R=enne
CQ_INCLUDE_TRYBOTS=master.tryserver.blink:linux_trusty_blink_rel
==========

[vmpstr, 2017-03-14 21:04:03]

Please take a look.

[enne (OOO), 2017-03-14 21:12:35]

(1) Can we just clip with the right rect? If so, does that fix the issue?

(2) Is there any way to DCHECK that we don't do on demand decode? <_<

[vmpstr, 2017-03-14 22:38:50]

On 2017/03/14 21:12:35, enne wrote:
> (1) Can we just clip with the right rect? If so, does that fix the issue?
> 

Not in all cases. I kind of tried to explain that case in the comment. We could
have a situation where the display item's visual rect is padded for whatever
reason (pdr@/wkorman@ mentioned that sometimes we have images + other stuff in
one skp), so even if we use the correct clip, it might still "select" that item
at raster and pass the whole shebang to skia, which will use its internal clips
to quickreject or in this case quicknotreject it).

So to answer your question, in this specific instance, yes it would fix it, but
I think my proposed solution covers more cases. 

> (2) Is there any way to DCHECK that we don't do on demand decode? <_<

Unfortunately, we have an optimization in place that doesn't create an image
hijack canvas if there are no (perceived) images in the tile, so this dcheck
would have to be some sort of a flag on an SkCanvas like
SkDcheckOnAnyImageCallsCanvas, which is kind of equivalent of
SkImageHijackCanvas except every drawImage call is a dcheck, but that adds the
overhead of the sknwaycanvas as a result, which I'm trying to avoid. 

In the medium term, CDL potentially addresses this, depending on how we process
everything. That is, if the decoded state is stored on the cdl or something like
that then we can dcheck that we either (1) know it's at raster decode or (2)
it's decoded. All other cases would dcheck.

[enne (OOO), 2017-03-15 20:54:45]

lgtm

[vmpstr, 2017-03-15 20:59:51]

The CQ bit was checked by vmpstr@chromium.org

[commit-bot: I haz the power, 2017-03-15 21:00:35]

CQ is trying da patch. Follow status at
 
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2017-03-15 22:25:25]

CQ is committing da patch.

Bot data: {"patchset_id": 1, "attempt_start_ts": 1489611591924590, "parent_rev":
"9fc66c2cd53582af4f0ad3f4c6ebf169a16678bf", "commit_rev":
"3530c014a885f3195f6cfad735a297dabe99ebc3"}

[commit-bot: I haz the power, 2017-03-15 22:26:27]

Description was changed from

==========
cc: Outset image rects by 1 for the purposes of the rtree.

This patch ensures that we outset the visual rects used by images by 1
since at raster time, we may query/try to use images that are touching
the raster bounds, but not intersecting. This is (at least in part)
due to Skia returning device bounds that are outset by 1.

This ensures that we create proper dependency graphs for the image
tasks.

R=enne
CQ_INCLUDE_TRYBOTS=master.tryserver.blink:linux_trusty_blink_rel
==========

to

==========
cc: Outset image rects by 1 for the purposes of the rtree.

This patch ensures that we outset the visual rects used by images by 1
since at raster time, we may query/try to use images that are touching
the raster bounds, but not intersecting. This is (at least in part)
due to Skia returning device bounds that are outset by 1.

This ensures that we create proper dependency graphs for the image
tasks.

R=enne
CQ_INCLUDE_TRYBOTS=master.tryserver.blink:linux_trusty_blink_rel

Review-Url: https://codereview.chromium.org/2743203006
Cr-Commit-Position: refs/heads/master@{#457235}
Committed:
https://chromium.googlesource.com/chromium/src/+/3530c014a885f3195f6cfad735a2...
==========

[commit-bot: I haz the power, 2017-03-15 22:26:35]

Committed patchset #1 (id:1) as
https://chromium.googlesource.com/chromium/src/+/3530c014a885f3195f6cfad735a2...