I'm decently sure you have to compile separately for each architecture, and if you elect to compile for multiple architectures up front, you'll have excruciating compile times. You'd think that would be annoying, but it ends up not really mattering since AMD completely switches out the toolchain about every graphics generation anyway. That's not a good reason to not have forwards compatibility, but it is a reason.

The reason I'm not completely sure is because I'm just doing this as a hobby, and I only have a single card, and that single card has never seen a revision. I think that's generally the best way to be happy with ROCM. Accept that it's at the abstraction level of embedded programming, any change in the hardware will have to result in a change in the software.