[Bug rtl-optimization/104675] [9/10/11/12 Regression] ICE: in expand_expr_real_2, at expr.cc:9773 at -O with __real__ + __imag__ extraction

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104675

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|10.3.1, 11.2.1, 12.0,       |
                   |6.5.0, 7.5.0, 8.5.0, 9.4.1  |

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
It isn't just that match.pd change though, consider:
_Complex int foo (_Complex int a) { return (-1 + -1i) - a; }
_Complex int bar (_Complex int a) { return - a - (1 + 1i); }
_Complex int baz (_Complex int a) { _Complex int b = -1 + -1i; return b - a; }
_Complex int qux (_Complex int a) { _Complex int b = 1 + 1i; return - a - b; }

I think both:
  /* -A - 1 -> ~A */
  (simplify
   (minus (convert? (negate @0)) integer_each_onep)
   (if (!TYPE_OVERFLOW_TRAPS (type)
        && tree_nop_conversion_p (type, TREE_TYPE (@0)))
    (bit_not (convert @0))))

  /* -1 - A -> ~A */
  (simplify
   (minus integer_all_onesp @0)
   (bit_not @0))
needs to punt for COMPLEX_TYPE.