I wrote:
The "MI" layer, which is the lowest
supported "machine code" level of
the AS/400 (sort of like P-code or Java bytecodes) is very CISCy, but
that's implemented in software on both CISC and RISC AS/400s.
I should probably further add that the MI layer is translated into
native code, so there generally isn't a performance penalty for
interpretation. The point is, however, that IBM does not support users
writing actual machine code on the AS/400 boxes. Also, the MI code is
portable across all AS/400 models, while the internal code is not.