Mesa (master): radeonsi: Only call nir_lower_var_copies at the end of the opt loop

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 472a20c5fc0feda0f074b4ff95fd7c7a6305c8cd
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=472a20c5fc0feda0f074b4ff95fd7c7a6305c8cd

Author: Jason Ekstrand <jason at jlekstrand.net>
Date:   Wed Sep 23 13:14:26 2020 -0500

radeonsi: Only call nir_lower_var_copies at the end of the opt loop

In 283ad85944b5d, radeonsi started using nir_find_var_copies.  However,
it was also calling nir_lower_var_copies in the optimization loop and
the two can end up fighting.  The simple solution is to wait to lower
copies until the end of the optimization loop.

Fixes: 283ad85944b5d
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3550

Reviewed-by: Marek Olšák <marek.olsak at amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6841>

---

 src/gallium/drivers/radeonsi/si_shader_nir.c | 18 ++++--------------
 1 file changed, 4 insertions(+), 14 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index 70b2d6991e5..99be45ce5b0 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -496,23 +496,11 @@ static void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool
       bool lower_phis_to_scalar = false;
 
       if (first) {
-         bool opt_find_array_copies = false;
-
          NIR_PASS(progress, nir, nir_split_array_vars, nir_var_function_temp);
          NIR_PASS(lower_alu_to_scalar, nir, nir_shrink_vec_array_vars, nir_var_function_temp);
-         NIR_PASS(opt_find_array_copies, nir, nir_opt_find_array_copies);
-         NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
-
-         /* Call nir_lower_var_copies() to remove any copies introduced
-          * by nir_opt_find_array_copies().
-          */
-         if (opt_find_array_copies)
-            NIR_PASS(progress, nir, nir_lower_var_copies);
-         progress |= opt_find_array_copies;
-      } else {
-         NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
+         NIR_PASS(progress, nir, nir_opt_find_array_copies);
       }
-
+      NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
       NIR_PASS(progress, nir, nir_opt_dead_write_vars);
 
       NIR_PASS(lower_alu_to_scalar, nir, nir_opt_trivial_continues);
@@ -564,6 +552,8 @@ static void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool
       if (sscreen->info.has_packed_math_16bit)
          NIR_PASS(progress, nir, nir_opt_vectorize, NULL, NULL);
    } while (progress);
+
+   NIR_PASS_V(nir, nir_lower_var_copies);
 }
 
 static int type_size_vec4(const struct glsl_type *type, bool bindless)