Why didn't anyone notice that earlier? I've been playing around a lot recently with a Z80 to 6502 project that is along a similar vein, but I'm not ready to share yet, and when and if I do, it'll be in a fresh thread. Does a tiny token-threaded 4th-cousin of Forth belong in the Forth sub-forum, or the generic programming sub-forum?
I know what you mean, and I'm finding myself struggling to stay focused on the intent of the Z80 code I'm trying to grok. I really want to make my 6502 version look tight and efficient, not just a lazy translation. I'll know that I'm done when I no longer have any ZP variables called BC, DE, IX, IY and/or HL. I wouldn't even attempt a project this large with x86 source ... my brain cells would mutiny in a matter of a few hours.

_________________
Got a kilobyte lying fallow in your 65xx's memory map? Sprinkle some VTL02C on it and see how it grows on you!

Mike B. (about me) (learning how to github)

_________________
http://WilsonMinesCo.com/ lots of 6502 resources
The "second front page" is http://wilsonminesco.com/links.html .
What's an additional VIA among friends, anyhow?

Token threaded, direct threaded, indirect threaded, subroutine threaded, bit threaded, bit-token threaded ... I don't see that that makes a big difference, but if it is a fourth-cousin of Forth, that would be more of an edge case than SectorForth, which is a "minimal Forth" in which one can define a full fledged Forth.

Of course, as noted further up in this thread, like all minimal Forths, it will sacrifice operating speed compared to a full fledged implementation, and since it is substantially more minimal than eForth, it is likely to be substantially slower than eForth as well.

Since I am not pursuing the "smallest possible" prize (and I believe those in the chase are doing it more for the honor of the prize than for the honorarium of $0.00), I don't really mind if I end up with something that takes a whole 2KB to get to the point of being able to interpret file scripts and to save the result in an executable image.

Edit: Even putting a headless DOLIT in the loadtime section of my source, adding the STATE smart handling of literals pushed the runtime back over 2 binary pages. However, the whole assembled binary remains under five binary pages.

What I aim to do with this is to debug it to see if I can make it work ... which won't be happening until after Thanksgiving week, so this brief flurry of activity will pause ... and if I can, to convert it from the SectorForth to something more like a subset of Camel Forth, with RP@ RP! SP@ and SP! so that I can dump the clunky "stacks as indexes and also as pointer" approach that SectorForth requires.

Then I can convert to having a return stack in the ZP and a data stack with a TOS in the zero page and rest of the data stack on the hardware stack. That works for the DOLIST that I already have here, which relies on JSR DOLIST to get the list address onto the hardware stack and avoids juggling if it just pushes the IP into the return stack and then pops the return vector, incrementing it along the way to convert it to a jump vector. I can also add a more reasonable wordset for a more efficient minimal Forth ... including */ as a base for defining multiply and divide words when needed.

I've picked this up again. I didn't get it up beyond showing the prompt, and wasn't interested enough in actually having a pure SectorForth model forth to dig into it.

So what I'm doing is abandoning the "pure" SectorForth, using rp@ rp! sp@ sp! and aim to climb up the compiled Sector Forth model to define as primitives the things that required the return stack pointer and stack pointer to be an actual integer pointer.

I am using a "JMP (addr,X)" direct threaded model where the operand of the JMP is the zero page Instruction "base" address. I walk the X forward until it passes $7F, then update the IP base address if it gets there ... but since the IP base address gets updated with every call to high level compiled code and every DOLIT operation, and well written Forth words are rarely over 127 bytes long anyway, it should only very rarely require the "half-page fix" part of the routine.

DONEXT can be reduced to 15 clock cycles, including the three clock jmp overhead, if the entire DONEXT routine is placed in the zero page.

An alternative model is the split pointer model, where the low byte of IP is $00 and X holds the low byte address, but the trade-off is that you need to check for page overflow with each INX, so you simplify and speed up EXIT and DOLIST but add two cycles to the DONEXT routine.



I think the split-pointer version goes:

I've had a closer look at keeping it "more or less minimal" (as opposed to the bleeding edge minimalist of the SectorForth / MilliForth implementations for PC-DOS), but more closely aligned with the actual implementation model, and what I'm coming around to for the stack operations is to look at the R stack, S stack and TOS register as three distinct sources or destinations, and focus the primitives on the operations on one, or between two of them.

So, for example, >R is not a primitive, because with this stack implementation, you decrement the RNDX, copy from TOS to R, and then pull from S into TOS ... so all three are involved in the operation.

The primitive is "DUP>R", where you decrement RNDX, and copy from TOS to R, and that's that. Pulling from S to TOS is already provided by "DROP".



Then I'm thinking that you have something like:

: SWAP ( a b -- b a ) 2ND>R DUP DROP-R@ RDROP ;
: R@ ( R: a -- a R: a ) DUP DROP-R@ ;
: R> ( R: a -- a R: ) DUP DROP-R@ RDROP ;
: >R ( a -- R: a ) DUP>R DROP ;
: ROT ( a b c -- b c a ) DUP>R 2ND>R DROP R@>2ND RDROP R@>2ND RDROP ;
: 2DROP ( a b -- ) NIP DROP ;
; OVER ( a b -- a b a ) 2ND>R DUP DROP-R@ RDROP ;
; 2DUP ( a b -- a b a b ) 2nD>R R@>2ND DUP R@>2ND RDROP ;

... however, part of the point of the SectorForth approach is that if a compiled word is not needed in a specific application, it can simply be omitted, to reduce the total footprint.

Now, for logic and math:

Especially since I am going to have a "preliminary" hex literal evaluation if the token fails to be found as a word, I am fine with just "+" as the underlying primitive for 1+, 2+, -, 1- etc.

However, after working out / looking at the implementation of NOT, AND OR and XOR using NAND alone, I decided that I am OK with:

: NOT ( x1 -- x2 ) DUP NAND ;
: AND ( x1 x2 -- x3 ) NAND DUP NAND ;

... but the definition of OR and XOR with NAND alone was just too much for me, so I've added "OR" as a primitive, which then also allows a simpler XOR since:



So rather than the theoretical minimum NAND (or alternatively NOR) from which to build all logic, I got with the more practical set of NAND and OR as primitives.

Another big change I've made is going from embedded dictionary to a block dictionary. Length+Flags Byte up front, then as many characters as the length byte says, then the code field address, then the preceding entry in the dictionary. LATEST points to the bottom of the dictionary block, and adding a new entry is done by subtracting two from LASTEST, writing the address in HERE there, then subtracting the length +1 from LATEST, writing the counted string there, and finally setting any flags if need be. On loading, the dictionary block would be written to the top of RAM, so LATEST grows down from the top of RAM while HERE grows up toward the top of RAM. That copy is simplified because the dictionary contain any addresses located within the dictionary itself -- it has no link fields, so there is no issue in whether the link field is absolute or a slower relative offset.

I'm allowing up to 31 byte names, I've got an immediate flag, I reserving a flag for compile-only, so I can use the third flag set with a 0 length to mark the end of the dictionary block. My notion is if I ever need to handing situations where a simple block dictionary isn't enough, a non-zero length field with the third flag set can be used to indicate those special cases, but I'm not going to spend any time elaborating that at the moment ... this implementation will simply BRK on third flag set and a non-zero length field.

EDIT: Scratch that, it turns out that with this dictionary, I need the third flag as a smudge bit. However, it also turns out that a simple NUL works just fine as an end of dictionary marker. Further, it turns out that an immediate flag with a nul length works just fine for a "bridge" over arbitrary data embedded in the dictionary, with the following integer being the address of the next dictionary entry, so I'm going with that approach.

SO I've got to rewrite the dictionary search to shift it from an embedded linked list to a vocabulary built on a dictionary block.

The milliforth-6502 IS done.

Both Direct Thread Code (640 bytes)
and Minimal Thread Code (623 bytes)
models working with my_hello_work.FORTH.

Milliforth is a reduced sectorforth, both for x86, now ported for 6502.

Using ca65 compiler and run6502 emulator.

All tips are welcome.

https://github.com/agsb/milliforth-6502

The overhead of MTC over DTC IS about 0.23% for instructions and 1.59% for cycles, in absolute counters, compiling the file my_hello_world.FORTH.