Actually no, the '816 is more efficient than that, because in Forth, you leave the accumulator in 16-bit mode and the index registers in 8-bit mode almost full-time. There are very few REPs and SEPs in my '816 Forth kernel. I know people are afraid of those REPs and SEPs, but they're a rare occurrence in this application.

You can go further than that and:

• push a literal on the hardware stack in a single instruction with PEA which is "push 16-bit immediate data."
• If you want to fetch the 16-bit number contained at that address and push it on the hardware stack, use PEI.
• PER pushes onto the stack a 16-bit address relative to where the PC is at the moment, with the offset being specified in the operand. This is helpful in writing self-relocatable code.

None of these affect the flags or the accumulator, and they are always 16-bit regardless of the M or X flags, so REP and SEP are never necessary.

So how I would do the above would be:

NEXT left Y at 0 anyway, so there's no need to re-zero it, and we almost always deal in 16-bit cells, so I just leave the M and X flags alone.

The above is only if you want to leave that address on the return stack though. Otherwise we're back to just LDA addr, a single instruction for the whole operation.


Having the data stack (not the return stack) in ZP, "subroutines" (which could be anything in Forth) never step on the variables of other pending routines (that's the nature of a stack), so I never had any reason to change the DP register from zero in my Forth kernel. The best use of changing the DP register appears to be true multitasking, so every independent task can have its own "zero page." I have managed to do all the multitasking I need with interrupts, including servicing interrupts in high-level Forth with no overhead as my article describes, so I have not gotten into true multitasking. I call it "pseudomultitasking" for lack of a better term. The interrupt performance of the 65xx processors allows this to be practical.

Yes, 12 actual clocks (assuming zero wait-state memory). It doesn't save all registers, though. Only a handful of registers are saved, which allows you to use an interrupt routine written in C (or another high level language). If you need more registers, the function has to save them itself, like any other function.

The idea is that you can write a standard C function, and plug the address of the function into the vector table, and it will called when that interrupt occurs, using a standard calling context. All the normal 'scratch' registers that a function can use are saved by the interrupt hardware. I think there are 5 of those. The interrupt handler also allows 32 programmable priorities and automatic nesting of interrupts. On a typical controller, you'll have a few dozen different interrupt sources.

Also, the Cortex has an optimization when two interrupts overlap. As soon as the first one is handled, it jumps to the second one, skipping the part where it restores and saves the registers.

Here are some ideas:

lots of registers - avoids spilling to/from zero page a lot
memory bus width - 4x difference
cache - allows for harvard-like architecture internally, overlap of data and instruction activity
predication - avoids a branch penalty in many cases
more powerful instructions and addressing modes - get more work done for each instruction

------
I had a look to the Samuel tables. It is strange to see (first table) that z80 performs as the 6502 does (in letterature 6502 is claimed to be more efficient). Anyway, let's consider 0.02DMIPS/Mhz.

Giving that I tried to figure-out a different parameter considering also the gate count. In fact, gate count is fundamental for fpga user. I defined the DMIPS/Kates i.s.o. DMIPS/Mhz. From internet I found: arm7tdmi = 48kgates; 0.74DMIPS/Mhz in thumb mode (I'm interested in code density). For the 6502 I toke 3.5kgates (still from internet) and 0.02DMIPS/Mhz.

With this number we get: 0.015DMIPS/Kgates for the arm7; 0.006DMIPS/kgates for the 6502. I think that this figures represent better the architecture efficiency when the gate count is important. Giving these numbers the ratio between 6502 and arm7 is about 1:35 considering the DMIPS/Mhz and it decreases to 1:3 consideirng the DMIPS/kgates. No bad considering that arm7 can manipulate data in 32bit mode!

So, I thinks this a more fair way to compare the cores in order to find out a compromise between efficiency and gate count. May be having a 32 bit 6502 this gap could be even reduced.
I hope it make sense.

I'm still looking for code density statistics for 6502 but I didn't find nothing on internet at the moment.

Cheers

This topic is now the top hit for 6502 code density, so: search for a pdf titled "Code Density Concerns for New Architectures" and you should find a diagram like this:

Attachment:

(There are more breakdowns for different elements of the benchmark - all the sources are here and there are comments on the architectures here)

(Fair use exception to copyright, of course. Credit to the authors: Vincent M. Weaver and Sally A. McKee)

Cheers
Ed

Edit: replaced lost image with new screenshot

To answer some more points:

For this, and for pointers to some other variations on the 6502 theme, and for André's own 65k project, have a look at his pages here. His project allows for 32bit, and more besides.

On the question of 6502 gate count, there's a thread here with a summary of 10 implementations of 6502 for FPGA.

On this point:
I feel I should clarify: the 9-page topic was the latest of several brainstorms which covered all sorts of enhancement ideas for 6502. It resulted in a couple of specific ideas which seemed simple and attractive: the 16 and 32 bit versions which Garth mentions.

I coded up the 16-bit version, using Arlet's 6502 core as a base. I've implemented that core in an FPGA module, and TeamTempest and BitWise have both ported their 6502 assemblers to target the core. Electric_Eye has taken a fork of the core and started work on a spartan6 development board. I don't believe it would be much work to create the 32-bit variant (which would actually be simpler.)

No, I don't think 65816 has been tackled: there's a fair chance it's protected.

Cheers
Ed

PEA pushes its operand to the stack, which is fine if the address (which it might not be—the operand could be anything, including a simple 16 bit constant) is static and known at assembly time, unless using self-modifying code. PEA isn't of any particular value if double-indirection is involved.

Consider the case where a subroutine is to print a text string selected by an index value in one of the registers. The first step (after validating the index value) would be to get the appropriate string pointers from a look-up table. With the 65(c)02, you'd store the pointers into a ZP location. With the '816, you could push them to the stack and access the string with LDA (1,S),Y. PEA could push the starting address of the look-up table to the stack, but you'd still be faced with loading the actual pointers into a register in order to prime the stack with the starting address of the string. Hence my code example.

_________________
x86?  We ain't got no x86.  We don't NEED no stinking x86!

Not sure this is pertinent to this thread and I've never tryed to work
out details to know if it really makes sense.

Keep two copies of memory and make writes go to both but allow
reads in parallel. (so you get something like Harvard and possibly
get some of the advantage of registers for zp)

Dedicate a couple of bits in the instruction to jumps/subroutines
so that every instruction is (potentially) a jump or call to a subroutine
or a return. (so you can do something like microcoding without so
much subroutine time penalty or macro space penalty)
(would't necessarily have to be every instruction)

Yes, I could see that working - like a pair of mirrored disks. Thing is, everything is a tradeoff of cost/benefit, and this has cost a second memory interface and a second memory. Probably the pin count would be the thing I'd worry about most. An on-FPGA cache would have a similar effect, I think? And one could dedicate some part of that to zero page or stack, if it turned out worthwhile.

Not sure what this idea is: if you dedicate one or two bits to flag a jump or jsr, where do you get the destination from? Another operand I suppose? So this is an idea for packing a pair of operations into a single instruction fetch and decode... a bit like predication, but that case theres no need for an extra operand. The thing about this idea is that jmp/jsr may not be so common as to make this much of a saving. Common actions like increment/decrement, so you get compound actions like post or pre increment/decrement, are perhaps more likely to earn their keep?

Cheers
Ed

In my case, I would like to build a platform into an FPGA. The platform should be generic enough to develop firmware for simple system to perform industrial control (including control of LCD) but also simple DSP algorithm, so a low demanding platform.
I would prefer investing on 6502 (that is in fact what I would like to do since a love this processor for its simplicity) but I don't want to discover after few months that a need more DMIPS to do something that today I don't foresee. Having an ARM7 clone I could even reuse AHB peripherals that I can found on internet speeding-up the assembly and the verification of the FPGA.

BTW, I think to understand your advice: 6502 is better for simple applications when hand made code is required. 65816 would be bettter but no clones are available today in RTL format.

So, the major doubt remain around the DMIPS that I will require according to the specific application but probably I could mitigate this potential problem assembling two 6502 cores into the FPGA. Its gate count is so small that I could fit at least a couple of them.

Thank you!

Really interesting project, congratulation to Andre'! Anyway I think it needs tens and tens of man years to finalize this roadmap .... we shall to wait.




Thank you Ed, really useful. According to your experience are all these cores mature to be used or is there a version that you would suggest to be more stable respect to the other?




Congratulation for this 16-bit working derivative....!

Cheers

I don't know if they are all thoroughly excellent - one might have to do some additional research. The T65 and Syntiac, and retromaster cores have been used in games which I think makes them pretty thoroughly exercised. I have a lot of faith in Arlet's and I gather he's used it in a real-world project. (I also have a strong preference for verilog.)

(The ones I don't mention in the above paragraph might also be splendid - I don't mean anything by not naming them.)

There's a test suite by Wolfgang Lorenz - it's a good sign to pass that - but correctly responding to RDY and interrupts in all cycles of all instructions is another level of detail. Having precise timing and the right bus activity in each cycle is necessary for games. I collected some test suite info in a wiki page.


Thanks! (Shoulders of giants...) (Edit: both for the core and the emulator, I only changed a couple hundred lines of code in pre-existing 6502 projects, mostly trivial edits to allow for 16-bit bytes.)

Cheers
Ed

I just posted elsewhere a link to an article with a few well written paragraphs on ARM's feature set, performance and code density.

Just to add my 2 cents :-)

The 65k way with its two new registers E and B could be something like:

Another alternative would be something like:

... which made me think that actually an opcode "LDB", load base register is missing.

André

I notice an indication that this CPU might in fact end up open-sourced, which would be great!



Cheers
Ed