The MIT Lisp Machine microcode feels like writing in a modern 3-address RISC instruction set to me. I see a lot of reluctance to modifying the source in old mailing lists, maybe everyone was told that it was too hard so didn't try.
EDIT: An example: The CADR has a nice API from Lisp to the CHAOSNET hardware, the microcode wakes up a stack group (thread) and passes it a buffer to process. Later machines had Ethernet but there isn't any microcode support for the hardware, Lisp code just polls the status of the ethernet controller and copies packets around a byte at a time. The microcode buffer handling routines for CHAOSNET could have been reused for Ethernet.
Issue with the CADR (and Lambda ..) microcode isn't that it is hard to modify, it is that there is a very deep snake pit, and lots of complex interaction between the microcode and the Lisp Machine.
The CADR already had support for (pre-)Ethernet via microcode very early (~1979) and did it more or less the same way as for Chaosnet. The Lambda I think modified this quite heavily though to something else ...
EDIT: An example: The CADR has a nice API from Lisp to the CHAOSNET hardware, the microcode wakes up a stack group (thread) and passes it a buffer to process. Later machines had Ethernet but there isn't any microcode support for the hardware, Lisp code just polls the status of the ethernet controller and copies packets around a byte at a time. The microcode buffer handling routines for CHAOSNET could have been reused for Ethernet.