[Bug c++/95768] [11 Regression] -march=sandybridge -O2 -Wall crashes as 'during GIMPLE pass: uninit ... Segmentation fault'

[msebor at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95768

Martin Sebor <msebor at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |msebor at gcc dot gnu.org
          Component|tree-optimization           |c++

--- Comment #3 from Martin Sebor <msebor at gcc dot gnu.org> ---
This is the same case of the pretty printer not handling an expression as
pr95580, except this one is in the C++ front end (the code is duplicated almost
verbatim between the two).  tree-ssa-uninit.c calls:

      if (warning_at (location, OPT_Wuninitialized,
                      "%G%qE is used uninitialized", stmt, rhs))

with this STMT and its RHS:

  b$4$g_7 = MEM <int> [(const struct m &)_6 + 4];

expr_to_string() in cp/error.c then does this with the RHS:

    case MEM_REF:
      if (TREE_CODE (TREE_OPERAND (t, 0)) == ADDR_EXPR
          && integer_zerop (TREE_OPERAND (t, 1)))
        dump_expr (pp, TREE_OPERAND (TREE_OPERAND (t, 0), 0), flags);
      else
        {
          pp_cxx_star (pp);
          if (!integer_zerop (TREE_OPERAND (t, 1)))
            {
              pp_cxx_left_paren (pp);
              if (!integer_onep (TYPE_SIZE_UNIT
                                 (TREE_TYPE (TREE_TYPE (TREE_OPERAND (t,
0))))))

but because the MEM_REF first operand is not a ADDR_EXPR the else branch trips
the ICE in integer_zerop() because because the RHS type is void* whose
TYPE_SIZE_UNIT is null that integer_zerop() is not prepared to handle.