[Bug tree-optimization/101925] [10/11/12 Regression] reversed storage order when compiling with -O3 only since r10-4742-g9b75f56d4b7951c6

rguenth at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Mon Aug 16 10:45:14 GMT 2021


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101925

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
So surprisingly the following is not enough to fix it.  We still get some
vectorizations but those are obviously not with reverse storage accesses.

Forcing be_net_addr to be -O0 works around the issue so it's vectorizing
of this function.  Likewise not inlining be_ip6_addr fixes it as well.

Disabling SRA fixes it also, and I think that SRA drops the rev storage order
access attribute.  Oddly enough for be_ip6_addr I see the rc.u.addr8[]
accesses do _not_ result in reverse_storage_order_for_component_p being true.
Why's that so?  How should I detect this is subject to re-ordering?

diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
index d594c0a1b1e..9591d8b9d1a 100644
--- a/gcc/tree-vect-data-refs.c
+++ b/gcc/tree-vect-data-refs.c
@@ -4291,6 +4291,27 @@ vect_analyze_data_refs (vec_info *vinfo, poly_uint64
*min_vf, bool *fatal)
       stmt_info->dr_aux.dr = dr;
       stmt_info->dr_aux.stmt = stmt_info;

+      /* We cannot preserve reverse storage order accesses when
+         vectorizing.  */
+      if (reverse_storage_order_for_component_p (DR_REF (dr)))
+       {
+         if (dump_enabled_p ())
+           dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                            "not vectorized: ref with reverse storage "
+                            "order in %G", stmt_info->stmt);
+         if (is_a <bb_vec_info> (vinfo))
+           {
+             /* In BB vectorization the ref can still participate
+                in dependence analysis, we just can't vectorize it.  */
+             STMT_VINFO_VECTORIZABLE (stmt_info) = false;
+             continue;
+           }
+         return opt_result::failure_at (stmt_info->stmt,
+                                        "not vectorized:"
+                                        " ref with reverse storage order: %G",
+                                        stmt_info->stmt);
+       }
+
       /* Check that analysis of the data-ref succeeded.  */
       if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT 


I've used for debuggging

diff --git a/gcc/gimple-pretty-print.c b/gcc/gimple-pretty-print.c
index d6e63d6e57f..ac5d760efeb 100644
--- a/gcc/gimple-pretty-print.c
+++ b/gcc/gimple-pretty-print.c
@@ -665,6 +665,9 @@ dump_gimple_assign (pretty_printer *buffer, const gassign
*gs, int spc,
        {
          dump_generic_node (buffer, gimple_assign_lhs (gs), spc, flags,
false);
          pp_space (buffer);
+
+         if (reverse_storage_order_for_component_p (gimple_assign_lhs (gs)))
+           pp_string (buffer, "{rev}");
          pp_equal (buffer);

          if (gimple_assign_nontemporal_move_p (gs))
@@ -673,6 +676,8 @@ dump_gimple_assign (pretty_printer *buffer, const gassign
*gs, int spc,
          if (gimple_has_volatile_ops (gs))
            pp_string (buffer, "{v}");

+         if (reverse_storage_order_for_component_p (gimple_assign_rhs1 (gs)))
+           pp_string (buffer, "{rev}");
          pp_space (buffer);
        }


More information about the Gcc-bugs mailing list