The attached patch treats the intrinsic SIGN in the same way as MOD and 
DIM as it has the same arguments.

Tested using make -j 8 check-fortran on x86_64

Conditional compilation using #ifdef __GFC_REAL_16__ has been employed 
where appropriate in the test cases so should be OK on platforms without 
REAL(16).

Change logs:

gcc/fortran

     Mark Eggleston  <mark.eggleston@codethink.com>

     PR fortran/89286
         * check.c (gfc_check_sign): Deleted.
     * intrinsic.c (add_functions): Call add_sym_2 with gfc_check_a_p
     instead of gfc_check_sign for "sign".
     * intrinsic.h: Remove declaration of gfc_check_sign.
     * iresolve.c (gfc_resolve_sign): Check for largest kind of the actual
     arguments and convert the smaller. Set return kind to be the largest.
     * simplify.c (gfc_simplify_sign): Use the largest kind of the actual
     arguments for return
     * intrinsic.texi: Add GNU extension notes for return value to SIGN.

gcc/testsuite

     Mark Eggleston <mark.eggleston@codethink.com>

     PR fortran/89240
     * gfortran.dg/sign_gnu_extension_1.f90: New test.
     * gfortran.dg/sign_gnu_extension_2.f90: New test.
     * gfortran.dg/pr78619.f90: Check for "must have" instead of "must be".

If OK please can someone commit as I do not have the privileges.

-- 
https://www.codethink.co.uk/privacy.html