[Mesa-dev] [PATCH 2/2] nir: allow propagation of if evaluation to all alu instructions
Timothy Arceri
tarceri at itsqueeze.com
Wed Oct 31 22:21:00 UTC 2018
On 1/11/18 1:28 am, Jason Ekstrand wrote:
> On Tue, Oct 30, 2018 at 9:17 PM Timothy Arceri <tarceri at itsqueeze.com
> <mailto:tarceri at itsqueeze.com>> wrote:
>
> With the simplifications to this pass in a3b4cb34589e2f1a68 we
> can allow any alu instruction to be processed. For one this can
> potentially help with bcsels.
>
>
> Do we want to? I believe that this patch is correct and I agree that we
> now can but then we get into some heuristics as to when it is or isn't
> useful. With iand, ior, and inot it's obviously useful because, for
> boolean values, those instructions can be sort of half-folded, i.e. we
> know that opt_algebraic will look at them and replace them with either
> the other source or a constant. With bcsel that's true if the condition
> is used in src[0] but not for src[1] or src[2]. For other instructions,
> it's even less obvious that it helps.
>
> I'm not saying I'm opposed. Just thinking out loud.
I'm not sure either. Ian (Ccing) was asking about support for bcsel when
I first wrote the series, I held off sending this patch in the past (due
to no shader-db change) but since I was looking at the code again I
thought I might as well send it out and get comments.
I'm happy to hold it back for now but I think we still have many
opportunities for improving bcsels so maybe it will be useful in future.
>
> --Jason
>
> No shader-db change.
> ---
> src/compiler/nir/nir_opt_if.c | 20 ++++----------------
> 1 file changed, 4 insertions(+), 16 deletions(-)
>
> diff --git a/src/compiler/nir/nir_opt_if.c
> b/src/compiler/nir/nir_opt_if.c
> index 1fe95e53766..38489676e6b 100644
> --- a/src/compiler/nir/nir_opt_if.c
> +++ b/src/compiler/nir/nir_opt_if.c
> @@ -448,7 +448,7 @@ propagate_condition_eval(nir_builder *b, nir_if
> *nif, nir_src *use_src,
> if (!evaluate_if_condition(nif, b->cursor, &bool_value))
> return false;
>
> - nir_ssa_def *def[2] = {0};
> + nir_ssa_def *def[4] = {0};
> for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
> if (alu->src[i].src.ssa == use_src->ssa) {
> def[i] = nir_imm_bool(b, bool_value);
> @@ -456,7 +456,7 @@ propagate_condition_eval(nir_builder *b, nir_if
> *nif, nir_src *use_src,
> def[i] = alu->src[i].src.ssa;
> }
> }
> - nir_ssa_def *nalu = nir_build_alu(b, alu->op, def[0], def[1],
> NULL, NULL);
> + nir_ssa_def *nalu = nir_build_alu(b, alu->op, def[0], def[1],
> def[2], def[3]);
>
> /* Rewrite use to use new alu instruction */
> nir_src new_src = nir_src_for_ssa(nalu);
> @@ -469,19 +469,6 @@ propagate_condition_eval(nir_builder *b, nir_if
> *nif, nir_src *use_src,
> return true;
> }
>
> -static bool
> -can_propagate_through_alu(nir_src *src)
> -{
> - if (src->parent_instr->type == nir_instr_type_alu &&
> - (nir_instr_as_alu(src->parent_instr)->op == nir_op_ior ||
> - nir_instr_as_alu(src->parent_instr)->op == nir_op_iand ||
> - nir_instr_as_alu(src->parent_instr)->op == nir_op_inot ||
> - nir_instr_as_alu(src->parent_instr)->op == nir_op_b2i))
> - return true;
> -
> - return false;
> -}
> -
> static bool
> evaluate_condition_use(nir_builder *b, nir_if *nif, nir_src *use_src,
> bool is_if_condition)
> @@ -502,7 +489,8 @@ evaluate_condition_use(nir_builder *b, nir_if
> *nif, nir_src *use_src,
> progress = true;
> }
>
> - if (!is_if_condition && can_propagate_through_alu(use_src)) {
> + if (!is_if_condition &&
> + use_src->parent_instr->type == nir_instr_type_alu) {
> nir_alu_instr *alu = nir_instr_as_alu(use_src->parent_instr);
>
> nir_foreach_use_safe(alu_use, &alu->dest.dest.ssa) {
> --
> 2.17.2
>
More information about the mesa-dev
mailing list