🚀 go-pugleaf

>However, leading-edge hardware seems hard-pressed to maintain binary
>compatibility very far backwards.

Examples, please?  Yes, MIPS, SPARC, and HP-PA have all introduced
additional user-mode-visible features to their architectures, but they
haven't, as far as I know, invalidated any *old* stuff (other than
blowing off "extended precision" in favor of "quad" floating point in
SPARC, or something such as that, but I don't know that anybody was
using the "extended precision" stuff, and don't know that anybody'd
implemented it in hardware).

Or are you thinking of changes that, while they don't require programs
to be recompiled in order to run at all, require them to be recompiled
in order to get the full performance boost from a new implementation
(e.g., different instruction scheduling)?