[lispm]

Pharo runs on top of something, running on top of something else, ...

Genera on the Lisp Machine runs on the metal. It is the process scheduler, it does handle the bus interrupts, it receives the network packets, it writes the bytes to the disk controller, it sets the bits in the graphics card, it writes the sound bytes to audio interface, the network packets are Lisp arrays, ...

[cpr]

Yes, but don't forget the microcode level that runs the Lisp code. Some very small bits of the OS were written in microcode (as few as possible), IIRC. (It's been a long time since I was hanging around the AI Lab. ;-)

[lispm]

The processor instruction set was implemented in Microcode. It was optimized to run compiled Lisp code. There was also a Lisp to Microcode compiler, though I haven't used it.

[rjsw]

The microarchitecture of the MIT/LMI/TI machines wasn't particularly complicated, it had pipelined 3-address load-store instructions just like the first RISC CPUs. Apart from the functions to handle each Lisp instruction the microcode was basically a simple RTOS. In my view, the clever part of the design was in picking the minimal features needed from the microcode to allow everything else to be writtin in Lisp.

There was a proposal from Sun for a Java OS that seemed to me to copy the same split.

It is a pity that all the Ivory papers that were in journals still seem to be behind paywalls.

[rbanffy]

I remember a guy once implemented a Smalltalk machine in an FPGA

http://siliconsqueak.org/ seems to be related.

[rjsw]

There is also a microcoded Java CPU [1] that runs in an FPGA.

[1] http://www.jopdesign.com/