[Mesa-dev] [PATCH 2/2] nir/lower_vec_to_movs: Don't emit unneeded movs

[Eduardo Lima Mitev]

On 09/23/2015 07:18 AM, Jason Ekstrand wrote:
> 
> On Sep 22, 2015 10:01 PM, "Jason Ekstrand" <jason at jlekstrand.net
> <mailto:jason at jlekstrand.net>> wrote:
>>
>> It's possible that, if a vecN operation is involved in a phi node, that we
>> could end up moving from a register to itself.  If swizzling is involved,
>> we need to emit the move but.  However, if there is no swizzling, then the
>> mov is a no-op and we might as well not bother emitting it.
>>
>> Shader-db results on Haswell:
>>
>>    total instructions in shared programs: 6262536 -> 6259558 (-0.05%)
>>    instructions in affected programs:     184780 -> 181802 (-1.61%)
>>    helped:                                838
>>    HURT:                                  0
> 
> By the way, I have absolutely no idea why this helps more than Matt's
> patch to delete these in register coalesce.
> 

Also in HSW, and with current master, I get this:

total instructions in shared programs: 6260467 -> 6259804 (-0.01%)
instructions in affected programs:     84738 -> 84075 (-0.78%)
helped:                                375
HURT:                                  0

Something must have overlapped improvements in the mean, likely
Alejandro's register coalescing modifiers patch.

The two patches are:

Reviewed-by: Eduardo Lima Mitev <elima at igalia.com>

>> ---
>>  src/glsl/nir/nir_lower_vec_to_movs.c | 19 ++++++++++++++++++-
>>  1 file changed, 18 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/glsl/nir/nir_lower_vec_to_movs.c
> b/src/glsl/nir/nir_lower_vec_to_movs.c
>> index 287f2bf..2039891 100644
>> --- a/src/glsl/nir/nir_lower_vec_to_movs.c
>> +++ b/src/glsl/nir/nir_lower_vec_to_movs.c
>> @@ -83,7 +83,24 @@ insert_mov(nir_alu_instr *vec, unsigned start_idx,
> nir_shader *shader)
>>        }
>>     }
>>
>> -   nir_instr_insert_before(&vec->instr, &mov->instr);
>> +   /* In some situations (if the vecN is involved in a phi-web), we
> can end
>> +    * up with a mov from a register to itself.  Some of those
> channels may end
>> +    * up doing nothing and there's no reason to have them as part of
> the mov.
>> +    */
>> +   if (src_matches_dest_reg(&mov->dest.dest, &mov->src[0].src) &&
>> +       !mov->src[0].abs && !mov->src[0].negate) {
>> +      for (unsigned i = 0; i < 4; i++) {
>> +         if (mov->src[0].swizzle[i] == i)
>> +            mov->dest.write_mask &= ~(1 << i);
>> +      }
>> +   }
>> +
>> +   /* Only emit the instruction if it actually does something */
>> +   if (mov->dest.write_mask) {
>> +      nir_instr_insert_before(&vec->instr, &mov->instr);
>> +   } else {
>> +      ralloc_free(mov);
>> +   }
>>
>>     return mov->dest.write_mask;
>>  }
>> --
>> 2.5.0.400.gff86faf
>>
> 
> 
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>