HN Theater @HNTheaterMonth

I'd recommend watching this CppCon talk [1] about compiler optioning. It isn't just that recompilation is hard, or that it is difficult for humans to do. At a fundamental level, the information is not there at all, because the compiler can make very impressive optimizations. Machine learning can deal with weak signals, but it can't deal with no signal.

[1] https://www.youtube.com/watch?v=nLv_INgaLq8

Your second link reminded me of https://www.youtube.com/watch?v=nLv_INgaLq8 (which definitely builds on some of Matt's work).

https://www.youtube.com/watch?v=zBkNBP00wJE - C++17 for the Commodore 64.

Probable language warnings for my other suggestions:

* Rescuing Prince of Persia from the Sands of Time https://www.youtube.com/watch?v=FnEWBtCnFs8 (was this talk ever given elsewhere?)

* And You Shall Know Me By My Trail of Documentation - https://www.youtube.com/watch?v=fEgHdHdUDaA

* The History and Evolution of Computer Viruses - https://www.youtube.com/watch?v=s2g9lgYrYJM

⬐ richard_todd

My eyes couldn't look away from the extra "= 0x01" at the top; it was driving me nuts. Glad he noticed it before the end of the video.

The middle part got me thinking: how helpful is it really, to write 50+ lines of high-level code, just to get 8 assembly instructions? I get that it's more readable and whatever, but seriously, I might rather write the 8 assembly lines and be done with it. Maybe I've spent too much time on comp.lang.forth lately :)

⬐ unsignedqword

Well, you said it yourself: it's more readable; not to mention more easily maintainable and extensible. If we all sought to create the most terse code possible with no regard to any of the other important factors in creating 'good code', then we all might as well be writing everything in Pyth or APL.

EDIT: I wouldn't say that all assembly is unreadable or anything like that (for instance, I find inline assembly easier and more readable than using compiler intrinsics for stuff like AMD64 SIMD instructions). I'm just saying that having more code around isn't exactly always a bad thing.

⬐ TillE

Really watchable video. Nothing fancy, just pleasantly concise.

The transparent optimization of modern (and not-so-modern) C++ features is not surprising, but it's cool to see in action.

⬐ egypturnash

Why is he doing this with an x86 compiler. The c64 used a 6510. None of this code would ever work.

...Oh, if you jump to about halfway through he starts using an x86->65xx translator. Okay. Nice.

⬐ Hydraulix989

There already seems to be a 6502 backend for LLVM specifically geared for c64:

https://github.com/c64scene-ar/llvm-6502

Which means you can just use the clang frontend and even write C++17 code which compiles down to 6502 ML.

⬐ cmrdporcupine

6502 is notoriously difficult to target a C compiler to. Limited stack depth, only 3 registers, and I believe some other limitations to the ISA that make hand it hard to target C to it.

cc65 seems to be the only viable one around last I looked. And it doesn't support the 65816, only the 6502. WDC offers their own compiler, but it's not fully open/free.

⬐ lefticus

I was unable to get any of the existing 3 llvm 6502 backends to produce reasonable code. Which is why I went this route. I meant to mention that in the video...

⬐ smcl

By reasonable code do you mean performant, or just plain working?

⬐ Hydraulix989

If I had to guess, now knowing that there's a crushing stack size limitation for 6502 CPU, probably "just plain working."

⬐ hermanhermitage

There's also https://github.com/puppeh/gcc-6502 and https://github.com/puppeh/gcc-6502-bits to get 6502 directly out of gcc.

⬐ lefticus

I don't believe I found those, thank you.

⬐ vitd

Oh man, I remember trying to understand AppleDOS and 6502 Assembly on the Apple II+ when I was a wee lad. It was so confusing what all these hex memory locations were. Being able to give them actual function names (or even just named constants!) would have made things so much easier!

⬐ jandrese

Nothing gives the warm fuzzies like seeing the translator step couldn't parse anything except the code he specifically wrote for the demo...

⬐ lefticus

Functional calls are not handled, which is part of the problem. There are several working examples in the x86-to-6502 repository. Please add more examples and help with additional translation.

⬐ ktRolster

The spec for C++14 wouldn't fit on a Commodore 64. At 1300 pages, it wouldn't even fit on external storage (170k floppies).

⬐ plexchat

Yes because the code emitted by a compiler is directly proportional to the size of the specification? I fail to see the point.

⬐ ktRolster

  >I fail to see the point.

That's ok, you don't need to see everything.

⬐ banachtarski

Humor me then. What's the point?

⬐ Feneric

Really cool, and this same approach looks like it'd work out-of-the-box with the C128 as a target, too.

⬐ arunc

Pleasant and interesting. Has anyone tried similar stuff with D?

⬐ lefticus

This was another reason I went the x86->6502 route instead of just a 6502 backend for llvm - it could plausibly work with any code. Let me know what you find.

⬐ white-flame

Static optimization to inlined code that ONLY handles immediate numbers and addresses is fine, and interesting to see the full pipeline working. But that possibility is there really for any language and any asm platform. I remember seeing Forth compile down to 6502 or other asm code, where it compiled down to the asm instructions a human would write for the same task, with basically zero language overhead.

HOWEVER, that was because the code written only did things similar to assembly instructions, similar to this example.

The biggest problem with compiling C family languages to 6502 is the stack. They generally assume something like a local stack frame for variables and parameters, and there's no way to directly express that style of programming at the 6502 instruction level. Note that there were no local variables in these examples, much less any actual function calls. Certainly the ABI used by C would not be machine-code-translatable from x86 to 6502; there would be an ABI written for 6502 and higher level targeting could work.

Instead of generating & trying to translate x86 assembly code, compiling C++14 down to plain C and using a 6502 compiler like cc65 could work significantly better for more idiomatic C++.

While it's an interesting novelty, why write 50+ lines of C++ code when writing 7 lines of commented, well-understood assembly instructions suffice? Add some variables and macros to take care of naming the various bits & locations, or target some codegen directly to 6502 if we're talking about programming "in the large" relatively speaking. Sure, you don't want to touch x64 code for human writing in the general case, but 6502 was designed with hand coding in mind.

⬐ to3m

Yes. This doesn't strike me as an especially productive line of attack, but maybe I'm missing something (hopefully not obvious), maybe I'm just not imaginative enough, and/or maybe I'm underestimating how much effort people might be willing to put into such an endeavour.

This did get me thinking about the problem a bit this evening though. The 6502 is an awful target for many higher-level languages, but I think you could compile something like C, if you treat zero page as the stack for locals, and index it by X. The programmer would need to be aware of the target's rather tiresome limitations, but you'd get somewhat passable code at least.

(Unlike absolute addressing, you do pay a 1-cycle penalty for indexed zero page accesses, but what can you do? You need to store things in zero page to handle pointers.)

Suppose you had a function with a couple of locals, like this:

    void strip(char *p,char c) {
        unsigned char *dest=p;
        do {
            if(*p!=c) *dest++=*p;
        } while(*p++!=0);
    }

So you'd store c (1 byte), then p (2 bytes), and then once the space for locals is reserved, zero page starting from the address in X would look like this:

    +0, +1 = dest
    +2, +3 = p
    +4     = c

So the ideal generated code would go along these lines:

    DEX:DEX                         ; reserve space for locals
    LDA 2,X:STA 0,X:LDA 3,X:STA 1,X ; dest=p
    .L0
    LDA (2,X)                       ; *p
    CMP 4,X
    BEQ L1                          ; taken when *p==c
    STA (0,X)                       ; *dest=*p
    INC 0,X:BNE L2:INC 1,X:.L2      ; ++dest
    .L1
    LDA (2,X)                       ; *p
    INC 2,X:BNE L3:INC 3,X:.L3      ; ++p
    CMP #0
    BNE L0                          ; taken when *p!=0
    INX:INX
    RTS

I don't think many people would write code like that by hand, but it's not completely insane given the requirements.

This has actually now piqued my interest enough to make me think about actually coding something, so I guess that answers the question about how much effort somebody might be willing to put into such an endeavour.

Other thoughts:

- since ZP,X costs like ABS,Y, maybe have a second stack in main memory for non-pointers and index it with the Y register (the RMW instructions don't have an appropriate addressing mode though)

- you could statically allocate zero page for many programs, which would make it easier to use the (ZP),Y addressing mode

- turning canned snippets (like the INC/BNE/INC) into routines would probably be rarely beneficial, since getting the address into X is rather involved

- I've never used a C compiler for the 6502 so maybe they all do this anyway ;)

⬐ ktRolster

I don't think many people would write code like that by hand, but it's not completely insane given the requirements

Typically if you were programming for the 6502 back in the day, you would build up a rather large macro library to make things simpler.

⬐ white-flame

cc65 (http://cc65.github.io/cc65/) is probably the most popular C compiler for the 6502, and its has support for many of the old home computers. I only use its assembler, since it's one of the most powerful, but I'm somewhat familiar with its C side of things.

I believe it uses a manual 16-bit stack pointer for frames, and zeropage for globals and statics. Pointers have to be copied into zeropage to be dereferenced, and I think it might have a numeric stack somewhere for intermediate results when evaluating numeric expressions. It's not super fast by any measure (simply due to impedance mismatch between C and 6502), but it does work and is quite mature.

As a side note, most 6502 Forths do use zeropage for the operand stack, with heavy use of zp,X addressing in their operators.

⬐ tcdent

Personally, I'm interested in programming for the sound chip; maintaining multiple harmonies with varying ADSR and noise to try to get as much depth as possible.

I'm not a low-level developer, so my progress thus far trying to fit an experiment-able design into 6502 machine code has been pretty slow going. Any chance to get closer to C syntax, real variables names and nested structure is a welcome relief to me. Thanks for mentioning cc65, I had not discovered it yet.

⬐ white-flame

cc65 will certainly scratch the itch you've got. Since control of the sound chip doesn't require super fast code, the significant overhead of running C on a 6502 won't be an issue, and you'll be in a familiar environment.

However, there are also lots of music editors for the C64, both native and running on PCs, that can give you very fine grained control of the SID without programming. I believe GoatTracker is highly regarded, but there are lots.