[Mesa-dev] [PATCH 2/2] i965/fs: Try to avoid generating extra MOVs to do saturates.

[Kenneth Graunke]

On 03/13/2012 02:37 PM, Eric Anholt wrote:
> shader-db results:
> Total instructions: 212648 ->  206044
> 614/1246 programs affected (49.3%)
> 178350 ->  171746 instructions in affected programs (3.7% reduction)
> ---
>   src/mesa/drivers/dri/i965/brw_fs_visitor.cpp |   24 +++++++++++++++++++++---
>   1 files changed, 21 insertions(+), 3 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> index eb129ce..e7af75b 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> @@ -172,12 +172,30 @@ fs_visitor::try_emit_saturate(ir_expression *ir)
>      if (!sat_val)
>         return false;
>
> +   fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
> +
>      sat_val->accept(this);
>      fs_reg src = this->result;
>
> -   this->result = fs_reg(this, ir->type);
> -   fs_inst *inst = emit(BRW_OPCODE_MOV, this->result, src);
> -   inst->saturate = true;
> +   fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
> +
> +   /* If the last instruction from our accept() didn't generate our
> +    * src, generate a saturated MOV
> +    */
> +   if (last_inst == pre_inst ||
> +       last_inst->predicated ||
> +       last_inst->force_uncompressed ||
> +       last_inst->force_sechalf ||
> +       !src.equals(&last_inst->dst) ||
> +       last_inst->regs_written() != 1) {

This duplicates a bunch of conditions from try_rewrite_rhs_to_dest.  It 
also seems like something we're going to want to do more of.  Could we 
capture some of this in a predicate like "last_inst->writes(src)"?

> +      this->result = fs_reg(this, ir->type);
> +      fs_inst *inst = emit(BRW_OPCODE_MOV, this->result, src);
> +      inst->saturate = true;
> +   } else {
> +      last_inst->saturate = true;
> +      this->result = src;
> +   }
> +
>
>      return true;
>   }