[Bug target/54781] [4.8 Regression] ICE in refs_may_alias_p_1, at tree-ssa-alias.c:1124

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54781

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-11-20
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |ramana at gcc dot gnu.org
          Component|middle-end                  |target
   Target Milestone|---                         |4.8.0
     Ever Confirmed|0                           |1

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-11-20 13:50:08 UTC ---
Looks like a target bug to me, neon_dereference_pointer creates an invalid
MEM_REF (the first operand on it doesn't satisfy the required
is_gimple_mem_ref_addr predicate) and this invalid MEM_REF is then added by
expand_normal into the MEM_EXPR of the MEM, and the aliasing code is then upset
about it.

In particular, the MEM_REF operand is &z[x_5(D)].  The reason why &z[x_5(D)] is
in the CALL_EXPR_ARG of the builtin is expand_call_stmt optimization:
      /* TER addresses into arguments of builtin functions so we have a
         chance to infer more correct alignment information.  See PR39954.  */
Best would be if in this case the backend could find out the original argument,
which was a SSA_NAME, and use that as the MEM_REF operand instead, but I'm
afraid we don't have such a way right now.  So perhaps the backend could clear
MEM_EXPR
if it is invalid:
    case NEON_ARG_MEMORY:
...
      if (MEM_EXPR (op[argc]) && TREE_CODE (MEM_EXPR (op[argc])) == MEM_REF
          && !is_gimple_mem_ref_addr (TREE_OPERAND (MEM_EXPR (op[argc]), 0)))
        /* Make sure MEM_EXPR created by neon_dereference_pointer isn't
invalid.  */
        set_mem_expr (op[argc], NULL_RTX);