[Mesa-dev] [PATCH 29/59] intel/compiler: don't propagate HF immediates to 3-src instructions

[Jason Ekstrand]

I don't care too much when in the series it happens.  Disallowing it
temporarily is probably fine.

On Mon, Dec 10, 2018 at 2:44 AM Iago Toral <itoral at igalia.com> wrote:

> Yes, they are at the bottom of the series since I only worked on that more
> recently. I kept this here because at this point in the series we need
> this, but I should be able to esily pull those patches here instead and
> remove this one if that is preferred. I don't think they depend on anything
> else.
>
> On Fri, 2018-12-07 at 12:22 -0600, Jason Ekstrand wrote:
>
> Seems reasonable though I thought you had patches to the constant
> combining to fix this.  Maybe they'll be ready in time that we won't need
> this?
>
> On Tue, Dec 4, 2018 at 1:18 AM Iago Toral Quiroga <itoral at igalia.com>
> wrote:
>
> 3-src instructions don't support immediates, but since 36bc5f06dd22,
> we allow them on MAD and LRP relying on the combine constants pass to
> fix it up later. However, that pass is specialized for 32-bit float
> immediates and can't handle HF constants at present, so this patch
> ensures that copy-propagation only does this for 32-bit constants.
> ---
>  src/intel/compiler/brw_fs_copy_propagation.cpp | 12 ++++++++++--
>  1 file changed, 10 insertions(+), 2 deletions(-)
>
> diff --git a/src/intel/compiler/brw_fs_copy_propagation.cpp
> b/src/intel/compiler/brw_fs_copy_propagation.cpp
> index ab34b63748e..58d5080b4e9 100644
> --- a/src/intel/compiler/brw_fs_copy_propagation.cpp
> +++ b/src/intel/compiler/brw_fs_copy_propagation.cpp
> @@ -741,8 +741,16 @@ fs_visitor::try_constant_propagate(fs_inst *inst,
> acp_entry *entry)
>
>        case BRW_OPCODE_MAD:
>        case BRW_OPCODE_LRP:
> -         inst->src[i] = val;
> -         progress = true;
> +         /* 3-src instructions can't take IMM registers, however, for
> 32-bit
> +          * floating instructions we rely on the combine constants pass
> to fix
> +          * it up. For anything else, we shouldn't be promoting immediates
> +          * until we can make the pass capable of combining constants of
> +          * different sizes.
> +          */
> +         if (val.type == BRW_REGISTER_TYPE_F) {
> +            inst->src[i] = val;
> +            progress = true;
> +         }
>           break;
>
>        default:
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20181210/d57184ea/attachment.html>