[Mesa-dev] X test suite regression due to gallivm change

[Jose Fonseca]

On 23/02/17 21:54, Adam Jackson wrote:
> On Thu, 2017-02-23 at 21:59 +0100, Roland Scheidegger wrote:
>
>> So, what does the failing test do?
>
> Not much, curiously. There are five that fail and they're all fairly
> trivial, although the xts harness makes that hard to see.
>
> XClearArea/6 and XClearWindow/4 set the window background pixmap to
> None and calls XClearArea. This should not change any pixels;
> apparently it does.
>
> XClearArea/7 and XClearWindow/5 create a parent window, then a child,
> sets the parent bg to None, the child to ParentRelative, and clears the
> child. Again, this should not change anything, but does.
>
> XCopyArea/1 creates two windows, tiles the background of the first one
> (equivalent to ClearWindow), copies from one window to the other, and
> verifies that the second one has the right content.
>
> The first four there are a little strange because, in principle, the
> thing they're testing is that no rendering happens. Presumably what's
> being measured is that some _other_ rendering prior to the action under
> test is being done incorrectly. The CopyArea test is probably closest
> to the root cause. I'll drill down a bit on exactly what that turns
> into in terms of GL draws.
>
> - ajax
>

I'm not sure if apitrace w/ glamor is practical / feasible, but being 
able to repro the issue outside X would be great.


Another tack on this issue would be to revert parts of 
320d1191c61a0a82444605c12e5c4b2ee0b241eb separately and see which 
trigger/fix the problem.

In particular:
- src/gallium/drivers/llvmpipe/lp_bld_interp.c for interpolation
- src/gallium/auxiliary/draw/draw_llvm.c for viewport / clipping
- src/gallium/auxiliary/gallivm/lp_bld_tgsi_action.c for TGSI translation

I tried to use fmad everywhere I could, but it's possible a couple or so 
is simply a bad idea.


Jose