2 ms·
I was very familiar with references #1 and #2, those are good papers. Three minor issues: 1) Both this article and reference #3 make claims about not using or
by VLM 13y ago
I was very familiar with references #1 and #2, those are good papers. Three minor issues:
1) Both this article and reference #3 make claims about not using or needing adders to avoid self modifying code, although both implement indexed addressing where you take a base address and a offset and er... um... argh... ADD them together. So you're using an adder but playing word games. Only reference #2 actually "does it" without an adder. If you're doing it for "fun" then self modifying code is plenty of "fun".
2) The claim the compiler is going to be the icky part is not accurate. To a crude approximation a compiler would do something about as hard as a FPGA synth of a microcode decoder for a soft CPU. Not as simple as implementing a HTML BLINK tag but its just no big deal breaker. The icky part is going to be performance. Just adding a little cache isn't going to help if your code stalls it inherently and you're going to be pretty well memory starved. Good luck doing a fast floating point multiply-add for a DSP with this architecture.
3) A minor issue but it should be possible to get instantaneous power very low with this design. No one writes about this interesting issue (unless its in reference 3 which I don't have access to) Of course "power per FLOP" type of metrics are (Probably? Inevitably?) going to suck, unless I'm wrong.
I've done a lot of daydreaming about actually implementing a system outta reference #1 for fun. Its about the simplest hardware "processor" you can possibly build. After all, you're going to need (memory mapped) I/O anyway and who cares if one of the memory mapped I/O is a 32 bit wide full adder, or whatever other CPU-like functional module you'd like to entertain. I have two amusing daydreams, one is a 256 bit wide move machine (why 256 bits wide? Why the heck not? Difficult, but barely inside the limit of buildability) the other strange daydream is vaguely like an old IBM1620 but why not try floating point instead of BCD? Yes I'm no noob I know all about floating point rounding issues and I consider it a huge part of the "fun". Also I'm well aware a binary adder would be a wee bit quicker for the program counter than a floating point adder... that's the whole point I think a pure float machine would be hilarious fun. I'm not talking about merely sticking a floating point copro on a chip, but literally using IEEE 754 on the address bus (obviously with a bit of rounding) Quad precision floating point should be enough digits for an address bus to have a heck of a lot of fun with some slack for rounding issues.
- mietek 13y ago> 1) Both this article and reference #3 make claims about not using or needing adders to avoid self modifying code, although both implement indexed addressing where you take a base address and a offset and er... um... argh... ADD them together. So you're using an adder but playing word games. Your concern is addressed in this paper: "It would appear that we are cheating slightly by allowing arithmetic in addresses. However, our "arithmetic" is of a very restricted form: the indexed instructions may only be used when R_{src} (or R_{dest} for stores) is an even-numbered memory address. Since offsets are always 0 or 1, then our "arithmetic" can be implemented as bitwise OR, or even bitstring concatenation."
- adsche 13y agoHmm, I still don't see it. Can bitwise OR be implemented with just MOV?