[Mesa-dev] [PATCH 3/4] nir: support to flatten_all in peephole-select

[Matt Turner]

On Tue, Mar 31, 2015 at 3:57 PM, Rob Clark <robdclark at gmail.com> wrote:
> From: Rob Clark <robclark at freedesktop.org>
>
> Freedreno and vc4 want this behavior for the time being (until we have
> real flow control).  Even after that, we probably want to turn this into
> some sort of driver tunable threshold, since for at least some hardware,
> reasonably large if/else is best flattend rather than having divergent
> flow control.
>
> NOTE: wasn't sure about some of the additional restrictions in
> block_check_for_allowed_instrs()..  are there some other cases where
> I might need to fix things up in order to be guaranteed to be able to
> flatten?
>
> NOTE: drop vc4 hunk if this is merged first, ofc
>
> Signed-off-by: Rob Clark <robclark at freedesktop.org>
> ---
>  src/gallium/drivers/vc4/vc4_program.c    |  2 +-
>  src/glsl/nir/nir.h                       |  2 +-
>  src/glsl/nir/nir_opt_peephole_select.c   | 21 +++++++++++++++------
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp |  2 +-
>  4 files changed, 18 insertions(+), 9 deletions(-)
>
> diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
> index db51665..09896ce 100644
> --- a/src/gallium/drivers/vc4/vc4_program.c
> +++ b/src/gallium/drivers/vc4/vc4_program.c
> @@ -1673,7 +1673,7 @@ vc4_optimize_nir(struct nir_shader *s)
>                  progress = nir_copy_prop(s) || progress;
>                  progress = nir_opt_dce(s) || progress;
>                  progress = nir_opt_cse(s) || progress;
> -                progress = nir_opt_peephole_select(s) || progress;
> +                progress = nir_opt_peephole_select(s, true) || progress;

I really hate boolean arguments. Could we make this parameter a
"flags" enum and pass a readable name?