[Mesa-dev] [PATCH 10/13] nir: verify destination bit size when checking algebraic optimizations

[Jason Ekstrand]

On Apr 12, 2016 1:06 AM, "Samuel Iglesias Gonsálvez" <siglesias at igalia.com>
wrote:
>
> Some instructions (like flrp in i965) cannot be lowered depending on the
> bit size because it doesn't support all bit sizes.
>
> If the bit size field is defined in nir_opt_algebraic.py, take it into
> account.

I'm sorry I keep punting on this patch.  I've never really like tacking yet
another argument on to the end of the expression but for a long time i
couldn't cone up with anything better.  How about this:  Add a bit_size
field to nir_search_value and allow you to write 'flrp at 64' to indicate a
64-bit destination on an ALU op, 'x at 64' to indicate a 64-bit "variable",
and '1.0 at 64' to indicate a 64-bit constant. Then it could be used both to
check for a particular bit-size in the search expression and to allow you
to specify a bit-size when constructing a value in case it's ambiguous.

If you're amenable to the idea, I'll hack on it a bit on Monday.  I've been
meaning to add a bit-size validator to nir_algebraic.py for some time now
anyway.

--Jason

> Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
> ---
>  src/compiler/nir/nir_algebraic.py | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
>
> diff --git a/src/compiler/nir/nir_algebraic.py
b/src/compiler/nir/nir_algebraic.py
> index 53a7907..f3630e0 100644
> --- a/src/compiler/nir/nir_algebraic.py
> +++ b/src/compiler/nir/nir_algebraic.py
> @@ -180,6 +180,11 @@ class SearchAndReplace(object):
>        else:
>           self.condition = 'true'
>
> +      if len(transform) > 3:
> +         self.bit_size = transform[3]
> +      else:
> +         self.bit_size = 0
> +
>        if self.condition not in condition_list:
>           condition_list.append(self.condition)
>        self.condition_index = condition_list.index(self.condition)
> @@ -208,6 +213,7 @@ struct transform {
>     const nir_search_expression *search;
>     const nir_search_value *replace;
>     unsigned condition_offset;
> +   unsigned bit_size;
>  };
>
>  struct opt_state {
> @@ -226,7 +232,7 @@ struct opt_state {
>
>  static const struct transform ${pass_name}_${opcode}_xforms[] = {
>  % for xform in xform_list:
> -   { &${xform.search.name}, ${xform.replace.c_ptr},
${xform.condition_index} },
> +   { &${xform.search.name}, ${xform.replace.c_ptr},
${xform.condition_index}, ${xform.bit_size}},
>  % endfor
>  };
>  % endfor
> @@ -249,6 +255,9 @@ ${pass_name}_block(nir_block *block, void *void_state)
>        case nir_op_${opcode}:
>           for (unsigned i = 0; i <
ARRAY_SIZE(${pass_name}_${opcode}_xforms); i++) {
>              const struct transform *xform =
&${pass_name}_${opcode}_xforms[i];
> +            if (xform->bit_size != 0 &&
> +                alu->dest.dest.ssa.bit_size != xform->bit_size)
> +               continue;
>              if (state->condition_flags[xform->condition_offset] &&
>                  nir_replace_instr(alu, xform->search, xform->replace,
>                                    state->mem_ctx)) {
> --
> 2.5.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160423/4d8dc5cd/attachment.html>