Hello Jim,
I have completely changed the fix while trying to answer your questions :)
I agree that the check for composite type is incorrect: in fact we
have to check
whether the gdi surface data can be used as a destination for
rendering primitives
which the 'loops' filed refers to.
However, the main question is why the 'loops' is non-null after the
d3d->gdi fallback.
It happens because d3d surface leaves d3d-specific loops and pipes
registered in
the graphics instance in the case of XOR, and just disables the
acceleration for given
peer. In particular, it leaves rendering loops as is, and it leads to
fall into generic (any-to-any)
loop in the case of gdi surface. To avoid this, we have to invalidate
the graphics
at some point.
This invalidation probably can be done in 'revalidateAll' method: if
original surface data,
and it's replacement have different types, we probably have to
invalidate the graphics object
in order to avoid the situation described above.
Please take a look to updated fix:
http://cr.openjdk.java.net/~bae/8036022/9/webrev.01/
Thanks,
Andrew
On 3/14/2014 4:36 AM, Jim Graham wrote:
Hi Andrew,
It looks like you are covering the case of the existing loops being
Solid and we switch to XOR so you get new loops. What about the case
where we used to be XOR and then we switch to Solid and the loops
might be stale, but in the other way (i.e. XOR when we want Solid
rather than Solid when we want XOR)? Also, if we were XOR and we
remain XOR, does this force us to fetch the loops on every validate?
I forget the strategy for the loops variable, was it supposed to be
set to null to force a refetch somewhere, but some invalidation case
missed setting the loops=null for XOR?
...jim
On 3/13/14 4:54 AM, Andrew Brygin wrote:
Hello,
could you please review a fix for CR 8036022?
Bug: http://bugs.sun.com/view_bug.do?bug_id=8036022
Webrev: http://cr.openjdk.java.net/~bae/8036022/9/webrev.00/
In case of XOR composite, rendering pipeline falls back from
d3d to gdi. However, gdi surface data does not re-set rendering
loops during validation, that leads to usage of the software
loops (due to dst type mismatch), and results in observed internal
error.
Suggested fix forces the re-set of the rendering loops,
at least for the case of XOR composite.
This change does not trigger any existing regression test.
Supplied regression test demonstrates the problem.
Please take a look.
Thanks,
Andrew.
