[Mesa-dev] [PATCH] Add an ir_triop_mid3 expression and lower it to mins and maxs.

[Petri Latvala]

On 04/29/2014 09:57 PM, Matt Turner wrote:
> On Tue, Apr 29, 2014 at 6:01 AM, Petri Latvala <petri.latvala at intel.com> wrote:
>> For the record, tested this with the following shader:
>>
>>
> Cool. Please submit this as a piglit test.

Sent to piglit mailing list, with accompanying tests for min3 and max3.

> Wouldn't it be simpler to detect constant arguments in opt_algebraic 
> and do the optimization there, and just perform the standard lowering 
> here? It seems cleaner to me. I don't think we generate different code 
> based on the arguments in any other lowering pass. 

I was thinking about that, but ended up optimizing on lowering. My 
reasoning was that if mid3(x, 1, 3) was transformed to min(max(x, 1), 3) 
in opt_algebraic, backends with theoretical native support for mid3 
would then need to recognize that and transform it back to a mid3.

(Are there even any GPUs that can do mid3 directly? AMD?)

Of course, it can be said that it's not a regression as mid3 doesn't get 
to backends as mid3 before this patch either.

I'll change the patch to do this optimization in opt_algebraic.


-- 
Petri Latvala