[Bug tree-optimization/87415] [9 Regression] wrong code at -O1 and above on x86_64-linux-gnu

[aldyh at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87415

--- Comment #3 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
Created attachment 44753
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=44753&action=edit
untested patch

1-bit signed fields are weird in that 0 - (-MIN) is still -MIN.  In any other
world, it is an impossible result.  For instance, in an 8-bit signed world, 0 -
(-128) is invalid.

One option would be to treat signed bit fields as TYPE_OVERFLOW wraps, since
they seem to have wrapping properties, but I'm afraid such a heavy handed
approach would yield latent bugs across the compiler.  What VRP seems to
currently be doing in set_and_canonicalize_value_range(), is just special
casing the wrapping of 1-bit fields:

  /* Wrong order for min and max, to swap them and the VR type we need
     to adjust them.  */
  if (tree_int_cst_lt (max, min))
    {
      tree one, tmp;

      /* For one bit precision if max < min, then the swapped
         range covers all values, so for VR_RANGE it is varying and
         for VR_ANTI_RANGE empty range, so drop to varying as well.  */
      if (TYPE_PRECISION (TREE_TYPE (min)) == 1)
        {
          set_value_range_to_varying (vr);
          return;
        }
...
    }

I've done the same in set_value_range_with_overflow().  The attached untested
patch fixes the problem on my end.