With the trunk, there are three issues:
(a) With bind(C), the callee side handles deallocation with intent(out).
This should produce the code:
if (cfi.0 != 0B)
{
__builtin_free (cfi.0);
cfi.0 = 0B;
}
This fails as cfi.0 (of type 'void*') is dereferenced and
*cfi.0 = 0B' (i.e. assignment of type 'void') causes the ICE.
(b) With that fixed, one gets:
sub (cfi.4);
_gfortran_cfi_desc_to_gfc_desc (&a, &cfi.4);
if (cfi.4 != 0B)
__builtin_free (cfi.4);
... code using "a" ...
That also won't shine as 'a.data' == 'cfi.4'; hence, one
accesses already freed memory.
I don't see whether freeing the cfi memory makes sense at all;
as I didn't come up with a reason, I removed it for good.
Those issues, I have solved. The third issue is now PR fortran/92189:
(c) When allocating memory in a Fortran-written Bind(C) function, the
shape/bounds changes are not propagated back to Fortran.
Namely, "sub" lacks some _gfortran_gfc_desc_to_cfi_desc call at the end!
The issue pops up, if you change 'dg-do compile' into 'dg-do run'. For
using a C-written function, that's a non-issue. Hence, it makes sense
to fix (a)+(b) of the bug separately.
OK for the trunk and GCC 9? (At least the ICE is a regression.)
Tobias