[Mesa-dev] [PATCH] st/mesa: remove copy-propagation pass

[Ilia Mirkin]

On Wed, Mar 25, 2015 at 4:44 PM, Tom Stellard <tom at stellard.net> wrote:
> On Wed, Mar 25, 2015 at 04:35:03PM -0400, Ilia Mirkin wrote:
>> On Wed, Mar 25, 2015 at 4:27 PM, Dave Airlie <airlied at gmail.com> wrote:
>> > On 26 March 2015 at 06:07, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> >> So.... what do you do when someone goes to shadertoy.com which on
>> >> average uses 1000 temps?
>> >
>> > Fall over in a heap, like every other driver on shadertoy,
>>
>> nouveau (nv50/nvc0) tends to do pretty well. There's one particular
>> shader that ends up being shifted over for reasons unknown, but no
>> compilation failures. i965 tends to do OK too, although I've seen
>> weird rendering artifacts, although those could well be due to the
>> shaders relying on unspecified behaviour.
>>
>> >
>> > This would be raelly bad for r600g in its present state, since it
>> > doesn't go TGSI->SB yet, and thus has a TEMP limit that this would
>> > only make worse.
>>
>> D'oh, right, r600 relies on there being no more than 124 or 128
>> registers. Forgot about that.
>>
>> >
>> > so Nak from me.
>> >
>> > Dave.
>>
>> So I suppose you want to fix the copy_propagation issue? :)
>>
>
> Can you add a flag for drivers to disable it?

Well, the present reality is that the pass is broken. I was hoping
that simply disabling it would be an acceptable fix for the issue, but
it sounds like there are still a few use-cases that want it around. I
think that reducing the amount of different TGSI output that st/mesa
can produce would be best, so I'd rather explore other options before
making it an optional pass. Even if I make it optional for nouveau,
it'll still break code when feeding to r600/etc...

  -ilia