Bookmarked "toread" -- it looks to be a really insightful look at byte-code that compares and analyzes a bunch of language implementations.

What an interesting guy. He runs a blog masquerading as a mailing list (that's what this link is from: "Kragen-tol" is a mailing list for "Kragen thinking out loud") which looks to be full of lots of interesting and incredibly detailed essays across a wide range of subjects.

His homepage (http://www.canonical.org/~kragen/) doesn't mention the mailing list at all (?), but it does have a bunch of other intriguing and random things, like a list of all the people he knows, for the purpose of telling search engines about it: http://www.canonical.org/~kragen/pik/

I am not sure about that insightful. He complains that literals on the F21 take two instructions. The Forth way is to define constants you need often like this:

    : 1 1 ; 

(that denies a function called 1 that 'returns' the integer 1)

With that function in place (3 instructions, I guess), you can replace n usages of the literal 1 (2n instructions) by n calls of that function (n instructions), for a total of n+3 instructions. So, this pays of if n>=3.

Oh, and subroutine calls are cheap on Forth hardware (one cycle for a call and zero cycles for a return is easily achievable)

I think requiring the interpretation of fib to be recursive is a little biased toward stack-based language VMs.

Non-recursive ARM (thumb) solution, 14 bytes:

   0:	2100      	movs	r1, #0
   2:	3801      	subs	r0, #1
   4:	dd01      	ble.n	0xa
   6:	4401      	add	r1, r0
   8:	e7fb      	b.n	0x2
   a:	4608      	mov	r0, r1
   c:	4770      	bx	lr

Self-reply: I'm an idiot, this just computes the sum of 1 to n. Doing fib requires a few more instructions.

I didn't see any mention of the Infocom z-machine bytecode (http://www.gnelson.demon.co.uk/zspec/); it would be interesting to see how it compares.

Dubé and Feeley went back to rework BIT after PICBIT: http://www.scribd.com/doc/49352760/BIT-A-very-compact-Scheme...

With the R4RS library it only needs 13KiB ROM and they claim you can run programs in less than 5KiB RAM. And it comes with a real-time GC. Not hand-coded assembler small, but the kind of development, testing and prototyping environment you get with Scheme is awesome.

The MuP21 sounds really impressive. 6000 transistors including video and DRAM is hard to believe.

Hurray for the mention of P-Code.

Is it blasphemy to want Go to have a horse in this race?

An equivalence beyond eval anyway... http://golang.org/pkg/exp/eval/

The Parallax BASIC Stamp used NRZ encoded bit tokens. The ROM for the interpreter was 1K words (14 bit words) as I recall. It is really efficient in terms of space.

Build a bit aligned FPGA CPU would be pretty cool, and using a serial EEPROM quite efficient in terms of I/O busses but like the PDP-8/S probably very very slow.