broadcom/compiler: move stores to the end of shader

It is possible that shader comes with output stores executed before
loading inputs. As the memory to read the inputs and store the outputs
is the same, this mean it could be overwriting the inputs before reading
them.

This move avoids this situation.

This partially improves
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33053.

Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33310>

src/broadcom/ci/traces-broadcom.yml

@@ -198,9 +198,9 @@ traces:
 
   neverball/neverball-v2.trace:
     broadcom-rpi4:
-      checksum: 4f4b4b6f37c124fdda6a9efcad577257
+      checksum: c8e8ee352bdb303e4ed144b69272575e
     broadcom-rpi5:
-      checksum: 174394638c6f774948e7aac91c12f84d
+      checksum: 56a0adb0efdf799f269da2d734a6817c
 
   nheko/nheko-colors.trace:
     broadcom-rpi4:

src/broadcom/compiler/vir.c

@@ -1051,6 +1051,14 @@ v3d_nir_lower_vs_early(struct v3d_compile *c)
         NIR_PASS(_, c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
                  type_size_vec4,
                  (nir_lower_io_options)0);
+
+        /* For geometry stages using the same segment for inputs and outputs
+         * we need to read all inputs before writing any output. If we switch
+         * to separate segments in the future this may not longer be strictly
+         * required.
+         */
+        NIR_PASS(_, c->s, nir_move_output_stores_to_end);
+
         /* clean up nir_lower_io's deref_var remains and do a constant folding pass
          * on the code it generated.
          */