[PATCH] d3d11/tests: Trace the name of the tests in single threaded mode.

[Francois Gouget]

On Tue, 11 May 2021, Henri Verbeet wrote:

> On Mon, 10 May 2021 at 16:14, Francois Gouget <fgouget at codeweavers.com> wrote:
> > On Mon, 10 May 2021, Henri Verbeet wrote:
> > > On Fri, 7 May 2021 at 19:39, Francois Gouget <fgouget at codeweavers.com> wrote:
> > > > This allows figuring out which function the crash happens in when there
> > > > is no backtrace. Just set WINETEST_NO_MT_D3D=1 WINETEST_DEBUG=2.
> > > >
> > > I'm not necessarily opposed to this (you do seem to have misplaced a
> > > number of *'s, as well as some other minor style issues though),
> >
> > I can resubmit with the *s moved over but I don't know what the other
> > issues are.
> >
> By convention:
>     _queue_test_fl() -> queue_test_fl_()
>     _queue_test() -> queue_test_()
>     _queue_for_each_feature_level_in_range() ->
> queue_for_each_feature_level_in_range_()
>     _queue_for_each_feature_level() -> queue_for_each_feature_level_()
> as well as preferring %u over %d in most cases.

I will resubmit with these changes (though technically, being an enum, I 
doubt D3D_FEATURE_LEVEL is unsigned).


[...]
> > So it's not a problem with a specific test, nor with a specific Windows
> > version. If someone knows what's going on...
> >
> Without having debugged it, my first guess would be that perhaps the
> tls_data structure is getting corrupted somehow. Does that happen
> consistently for that particular crash?

Yes.
More specifically:
* Only cw-gtx560 crashes almost systematically in the d3d11:d3d11 test.
* When running Windows 10 cw-gtx560 never gets the 'last trace' message.
* When running Windows 8.1 cw-gtx560 sometimes gets the 'last trace' 
  message but that looks like a different crash: the last trace line 
  number does not match.
* Machines with rx460 and rx550 GPUs also crash but a bit less often 
  and they never get the 'last trace' message either.


> Also, does that still happen with WINETEST_NO_MT_D3D=1?

Yes.


> Still, even if this particular functionality happens to be broken in 
> this case, for debug purposes WINETEST_REPORT_SUCCESS=1 could serve 
> much of the same purpose.

True.


-- 
Francois Gouget <fgouget at codeweavers.com>