[PULL xserver 1.10] GLX backport

[Adam Jackson]

On 4/5/11 4:17 PM, Ian Romanick wrote:
> On 04/05/2011 11:15 AM, Jeremy Huddleston wrote:
>> There is still a ton of fat left in there... but most of the fat is
>> rather benign (whitespace, deleting dead code, etc), so I don't have
>> a problem with it if it's part of the whole package.

I tried a first cut series that narrowed down to just the really key 
bits (redoing _glapi_add_dispatch and the context tag rewrite), but 
there was enough change in context from other patches in the series that 
this was both mechanically easier and more obviously corresponds to master.

>> What I do have a problem with is the fragility of GLX.  This has
>> traditionally been one of the more sensitive areas of the server, and
>> IIRC both 1.7 and 1.8 saw regressions introduced because of changes
>> in GLX.  I'd prefer to merge something like this in before an RC1 to
>> get the most exposure possible.
>
> Eric and I have been talking about this a lot lately.  The big problem
> is that we have basically no regression tests for GLX.  Until recently,
> when failures have been found we haven't made tests that reproduce them.

You'll note I've been writing piglit glx tests recently.  The first few 
are admittedly incredibly basic, but that's because they found some 
incredibly basic bugs in the GLX code.  I'm only too happy to write 
more.  If there's specific cases I should be covering, please, let me know.

> Do we have a list of the failures from 1.7 and 1.8?  It's never too late
> to make regression tests for old failures. :)

I'll compare the piglit and clutter runs against some older server 
builds, and against nvidia's driver (since, in a practical sense, 
they're the reference implementation for GLX whenever the spec is silent 
or ambiguous).

>> How would you feel about sitting on this until after 1.10.1 is out?
>> If there is a pressing reason why this should be part of 1.10.1 (such
>> as fixing clutter failures that were not in 1.9.5), then I'll
>> certainly consider merging it for RC2 and pushing out 1.10.1 by a
>> allow more time for testing feedback.

I can wait until 1.10.2.  There's at least one more GLX fix that 
logically belongs with this set, which I haven't sent a pull for master 
for yet so I didn't feel right including it in the stable branch.

I mean, I think the changes are all perfect, but I wrote them, so that's 
sort of the definition of observer bias.

- ajax