[Mesa-dev] [PATCH] nir/lower_vec_to_movs: Don't coalesce the mmove when src and dest are the same reg

[Jason Ekstrand]

After giving it some thought, I don't think this patch is quite strong
enough to fix the real bug.  The problem isn't that we're reswizzling a
register.  The problem is that we're trying to coalesce something like

ssa_1 = fadd r1, r2
/* Some stuff */
r3 = vec4(ssa_1, ssa_1.y, ...)

coalescing this together moves the write to r3 up to the fadd even though
there may have been other writes to r3 in between.

None the less, very good job tracking this down! :-)

--Jason

On Thu, Mar 22, 2018 at 4:30 AM, vadym.shovkoplias <
vadim.shovkoplias at gmail.com> wrote:

> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105440
> Fixes: 2458ea95c56 "nir/lower_vec_to_movs: Coalesce movs on-the-fly when
> possible"
> Signed-off-by: Andriy Khulap <andriy.khulap at globallogic.com>
> Signed-off-by: Vadym Shovkoplias <vadym.shovkoplias at globallogic.com>
> ---
>  src/compiler/nir/nir_lower_vec_to_movs.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/src/compiler/nir/nir_lower_vec_to_movs.c
> b/src/compiler/nir/nir_lower_vec_to_movs.c
> index 711ddd3..4758f7d 100644
> --- a/src/compiler/nir/nir_lower_vec_to_movs.c
> +++ b/src/compiler/nir/nir_lower_vec_to_movs.c
> @@ -166,9 +166,14 @@ try_coalesce(nir_alu_instr *vec, unsigned start_idx)
>        /* If we are going to reswizzle the instruction, we can't have any
>         * non-per-component sources either.
>         */
> -      for (unsigned j = 0; j < nir_op_infos[src_alu->op].num_inputs; j++)
> +      for (unsigned j = 0; j < nir_op_infos[src_alu->op].num_inputs;
> j++) {
>           if (nir_op_infos[src_alu->op].input_sizes[j] != 0)
>              return 0;
> +
> +         /* Don't coalesce the mmove when src and dest are the same reg */
> +         if (src_matches_dest_reg(&vec->dest.dest, &src_alu->src[j].src))
> +            return 0;
> +      }
>     }
>
>     /* Stash off all of the ALU instruction's swizzles. */
> --
> 2.7.4
>
> _______________________________________________
> 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/20180323/093a8223/attachment.html>