[Bug middle-end/71002] [6/7 Regression] -fstrict-aliasing breaks Boost's short string optimization implementation

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71002

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
So optimize_bit_field_compare does make a difference implementation-wise
(but not really fixes the undefinedness in the boost code wrt the access
of the non-active union-member when reading from is_short).

optimize_bit_field_compare turns

  ((const struct string *) this)->m_repr.s.h.is_short != 0

into

  (BIT_FIELD_REF <*(const struct string *) this, 8, 0> & 1) != 0

where the former memory access uses alias-set zero because of
c-common.c:c_common_get_alias_set handling of union-accesses:

  /* Permit type-punning when accessing a union, provided the access
     is directly through the union.  For example, this code does not
     permit taking the address of a union member and then storing
     through it.  Even the type-punning allowed here is a GCC
     extension, albeit a common and useful one; the C standard says
     that such accesses have implementation-defined behavior.  */
  for (u = t;
       TREE_CODE (u) == COMPONENT_REF || TREE_CODE (u) == ARRAY_REF;
       u = TREE_OPERAND (u, 0))
    if (TREE_CODE (u) == COMPONENT_REF
        && TREE_CODE (TREE_TYPE (TREE_OPERAND (u, 0))) == UNION_TYPE)
      return 0;

(which is something I am trying to get rid of for quite some time ...)

And the latter gets alias-set (that of type 'string') (which I think
is the better behavior, even if "miscompiling" this case).

Not applying optimize_bit_field_compare also gets us better optimization,
removing dead code already at the early GIMPLE level.