[Bug tree-optimization/68835] [5/6 Regression] ICE in set_value_range, at tree-vrp.c:387, with __int128 bit field

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68835

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Testcase that is miscompiled starting with r210113 at -O2:
__attribute__((noinline, noclone)) unsigned __int128
foo (void)
{
  unsigned __int128 x = (unsigned __int128) 0xffffffffffffffffULL;
  struct { unsigned __int128 a:65; } w;
  w.a = x;
  w.a += x;
  return w.a;
}

int
main ()
{
  unsigned __int128 x = foo ();
  if ((unsigned long long) x != 0xfffffffffffffffeULL
      || (unsigned long long) (x >> 64) != 1)
    __builtin_abort ();
  return 0;
}

The function foo returns 0xfffffffffffffffffffffffffffffffe instead of
0x1fffffffffffffffe.  So, how do we want to represent such large unsigned
values?  Force in that case get_len () > 1 representation, or tweak all
consumers including tree-pretty-print.c and other spots that if the
representation is negative number and the type is unsigned non-power of two
precision > HOST_BITS_PER_WIDE_INT, treat negative values differently?