[OpenJDK 2D-Dev] [9] request for review: 8036022: D3D: rendering with XOR composite causes InternalError.

[Andrew Brygin]

Hi Jim,

   we probably may want to avoid the unconditional invalidation here
   due to performance reasons. Potentially it saves some work if the
   replacement type is exactly same, and the current set of pipes can
   be re-used.

   However, as soon as there seems to be no reliable way to check
   whether current set of pipes corresponds to the replacement surface,
   it makes sense to invalidate it unconditionally.

   Please take a loot to updated webrev:
   http://cr.openjdk.java.net/~bae/8036022/9/webrev.02/

Thanks,
Andrew

On 3/19/2014 4:39 AM, Jim Graham wrote:
> Interesting. I don't know how this impacts the original issue, but I 
> think we do need some form of invalidation there.  Is there any reason 
> not to make it unconditional?  I think that code was logically 
> intending to start over from scratch so an unconditional reset of all 
> of the pipes makes sense.
>
> The following comment is probably obsolete given what I just said in 
> the last paragraph, but I wanted to also point out that equals() on a 
> SurfaceType may not be the right thing to use since we may have 
> multiple ST's with the same "type", but they are distinct objects 
> because they inherit from different types at a higher level.  In 
> essense the ST is a chain of related types, not just a single 
> description of one way of looking at the surface (which is all an 
> individual object can do is describe it one way and link to another 
> description).  In the end, I probably should have distinguished 
> between ST and "collection of STs to describe a surface", but got 
> clever with chaining instead.  As a result, I think that == is the 
> more proper way to compare STs?  (Or outlaw comparing them?)  But, 
> given that I think it makes sense to make the invalidate 
> unconditional, I don't think this matters...
>
>             ...jim
>
> On 3/17/14 2:22 AM, Andrew Brygin wrote:
>> 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.
>>