[Mesa-dev] [PATCH] i965/fs_nir: Put the immediate in src1 for commutative ALU ops

[Matt Turner]

On Fri, May 8, 2015 at 10:04 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> Shader-db results for fragment shaders on Broadwell:
>
>    total instructions in shared programs: 4310987 -> 4310663 (-0.01%)
>    instructions in affected programs:     43242 -> 42918 (-0.75%)
>    helped:                                142
>    HURT:                                  0
>
> Shader-db results for vertex shaders on Broadwell:
>
>    total instructions in shared programs: 2889581 -> 2844309 (-1.57%)
>    instructions in affected programs:     1418720 -> 1373448 (-3.19%)
>    helped:                                6139
>    HURT:                                  0
> ---
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 12 ++++++++++++
>  1 file changed, 12 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index 555987d..161a262 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -21,6 +21,8 @@
>   * IN THE SOFTWARE.
>   */
>
> +#include <algorithm>
> +
>  #include "glsl/ir.h"
>  #include "glsl/ir_optimization.h"
>  #include "glsl/nir/glsl_to_nir.h"
> @@ -662,6 +664,16 @@ fs_visitor::nir_emit_alu(nir_alu_instr *instr)
>        op[i] = offset(op[i], instr->src[i].swizzle[channel]);
>     }
>
> +   /* Immediates can only be used as the second source of two-source
> +    * instructions.  We have code in opt_algebraic to flip them as needed
> +    * for most instructions.  However, it doesn't hurt anything to just do
> +    * the right thing if we can detect it at the NIR level.
> +    */
> +   if ((nir_op_infos[instr->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) &&
> +       nir_src_as_const_value(instr->src[0].src)) {
> +      std::swap(op[0], op[1]);

We don't use STL.