[Mesa-dev] [PATCH v2 3/6] i965/fs: copy propagate 'NOT' instruction when used with logical operation

[Matt Turner]

On Thu, Jun 5, 2014 at 11:05 AM, Abdiel Janulgue
<abdiel.janulgue at linux.intel.com> wrote:
> On Broadwell, this reduces the instruction to a single operation when NOT is used with
> a logical instruction.
>
> Signed-off-by: Abdiel Janulgue <abdiel.janulgue at linux.intel.com>
> ---
>  src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp | 17 +++++++++++++----
>  1 file changed, 13 insertions(+), 4 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp b/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> index aa506f5..54d2cb4 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> @@ -341,7 +341,11 @@ fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
>        return false;
>
>     if (brw->gen >= 8) {
> -      if (entry->src.negate) {
> +      if (entry->opcode == BRW_OPCODE_NOT) {
> +         if (!is_logic_op(inst->opcode)) {
> +            return false;
> +         }
> +      } else if (entry->src.negate) {
>           if (is_logic_op(inst->opcode)) {
>              return false;
>           }
> @@ -359,6 +363,10 @@ fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
>        inst->src[arg].negate ^= entry->src.negate;
>     }
>
> +   if (brw->gen >=8 && entry->opcode == BRW_OPCODE_NOT) {

Space between >= and 8.

Also, there's a bit of code immediately before this:

   if (!inst->src[arg].abs) {
      inst->src[arg].abs = entry->src.abs;
      inst->src[arg].negate ^= entry->src.negate;
   }

I wonder how this might interfere? The Bspec explicitly describes how
abs and negate work before Broadwell, and how negate works after
Broadwell, but no mention of abs on Broadwell. I guess we'll need to
check.

> +      inst->src[arg].negate ^= !entry->src.negate;
> +   }
> +
>     return true;
>  }
>
> @@ -498,9 +506,10 @@ fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
>  }
>
>  static bool
> -can_propagate_from(fs_inst *inst)
> +can_propagate_from(struct brw_context *brw, fs_inst *inst)
>  {
> -   return (inst->opcode == BRW_OPCODE_MOV &&
> +   return ((inst->opcode == BRW_OPCODE_MOV ||
> +            (inst->opcode == BRW_OPCODE_NOT && brw->gen >=8)) &&

Space between >= and 8.

>             inst->dst.file == GRF &&
>             ((inst->src[0].file == GRF &&
>               (inst->src[0].reg != inst->dst.reg ||
> @@ -566,7 +575,7 @@ fs_visitor::opt_copy_propagate_local(void *copy_prop_ctx, bblock_t *block,
>        /* If this instruction's source could potentially be folded into the
>         * operand of another instruction, add it to the ACP.
>         */
> -      if (can_propagate_from(inst)) {
> +      if (can_propagate_from(brw, inst)) {
>          acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
>          entry->dst = inst->dst;
>          entry->src = inst->src[0];
> --
> 1.9.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev