vbcc optimizing C compiler now with Oric Atmos support

[vbc]

Hello
I've tried this setup and tap file generated isn't properly defined as a basic file (with 123 call #50e) but as a binary file
Is it possible to change it. Not sure but i think that startupr bend: variable should be modified

The tape header is generated by the linker which currently always sets it as machine code. It did seem to work, however. Does it matter if the header says machine-code? In this case we could add an option for the linker to set the byte accordingly.

[coco.oric]

The tape header is generated by the linker which currently always sets it as machine code. It did seem to work, however. Does it matter if the header says machine-code? In this case we could add an option for the linker to set the byte accordingly.
Top

Yes, it's working i agree ; but as it's a basic file with a code attached, it should be more accurate to put the correct flag for a basic file and, i don't know prevent any discrepancy by have different "basic" files chained. As there's only one atmos configuration using basic&code, it's not an option. All files using this atmosr configuration should have this flag (but i've not yet read all your pdf files with the possibilities of your asm & c code process).

[iss]

Does it matter if the header says machine-code?

Yes, definitely when basic startup code is used the header byte must be set to $00. The ROM loading routine checks it and uses different handling of the loaded data dependent on this byte.
If it's machine code then Oric simply jumps in the void:

Code: Select all

JMP   ($02A9)

hoping ...

... to work, however.

And it really seams to work because:
1. the TAP isn't autostart and you have to type RUN which calls properly the compiled code at $50E;
2. we are lucky enough If the header was set to be autostart then 6502 lands at $501:

Screenshot_20230716_140250.jpg

then it eats some invalid but harmless OPcodes, a couple of BRK's, and continues happily at $50E .

In this case we could add an option for the linker to set the byte accordingly.

So, in any case if the compiled binary will be preceded with basic caller then the header byte must be set to $00, else the byte must be set to $80. IMHO adding another option to linker for this is pointless.

[iss]

...

Please, read the 4 previous posts!

From formal point of view the byte should be $00 to indicate BASIC code (full stop) .
But let's push things further - what if we left it "as is" i.e. $80 and make the TAP autostart ...

meme-contest.jpg

I am announcing a prize-less just for fun BASIC/assembler contest:
Current startup BASIC code is:

Code: Select all

123 CALL #50E

which loaded in ORIC memory looks so:

Code: Select all

$501: 0b 05          - next BASIC line address
$503: 7b 00          - line number i.e. 123
$505: bf             - 'CALL' token
$506: 23 35 30 45 00 - ascii of '#50E', zero terminated
$50B: 00 00          - end of BASIC program

THE GOAL: Replace above bytes ($501..$50C), so they are BOTH VALID:

• BASIC program in memory which can be RUNned and LISTed;

• 6502 OP codes, at-least first few to jump to $50E, so the $80 in header is OK!

No other limitations, the start address of the binary code can be any other (i.e. $50E is not a requirement).
Shorter solutions are better!

Don't be lazy and challenge yourself. Good luck.

PS: The meme picture contains a clue which can be used: register A is loaded from $2B1 (which contains normally $00) just before the jump to $501.

[waskol]

Hello Dr vbc,
IF you need, you can get inspired by my little utility "taptap" (wich is written in object pascal : Delphi and Lazarus), and the pdf doc wich synthetizes the Oric Tape format.
You'll find everything on my github :
https://github.com/DJChloe/taptap/tree/master

[jsk]

I've been trying to use variious compilers, but only really been able to get cc65 and vbcc to work on Android/Termux, others are quite difficult to compile for a new platform.

What strikes me is that cc65 is quite capable and is very reliable. But the code generated is maybe slower than others. vbcc, in my tests inline everything so a simple recursive fib() becomes 130+ bytes whereas cc65 gave 56 bytes.

I think, and here I'm guessing, that a cycles/codesize could be an interesting measure. All the data is there, buti it's difficult to compare them and get a bigger picture. 20% faster by more than doubling the codesize isn't always a good idea.

Of course we want both speed and smaller code, if one could annotate per function, if code is time critical that would be beneficial.

For my own fun I maintain a simple benchmark suite to compare some 6502 C compilers.
Results and samples sources can be found HERE.

Is there some documentation how you build and run those benchmarks?

IMHO, it would be very useful for Oric fans to be able somehow to use XA assembler sources with vbcc...
My solution so far looks like this :
[...]
... it works fine but it's grausam.

I am not sure I understand what that script is doing exactly, but from looking at the XA man-page, most of its directives should be handled by vasm using the options -dotdir -sect.

vlink supports the o65 object format, so it might also be possible to link object code assembled by XA together with code created by vbcc.

[vbc]

What strikes me is that cc65 is quite capable and is very reliable. But the code generated is maybe slower than others. vbcc, in my tests inline everything so a simple recursive fib() becomes 130+ bytes whereas cc65 gave 56 bytes.

It is true that vbcc will usually generate direct code instead of calling lots of library functions. There surely are some opportunities to improve the size of the vbcc generated code in this fashion. However, vbcc usually generates more optimized code which makes using fixed library functions more difficult. Also, in such a small example, you have to take into account the size of the library functions as well.

If the size difference is as large as in your example, the code is probably not very friendly to the 6502. In such situations, the best option for code size probably would be to use a virtual machine with a compact instruction set.

I think, and here I'm guessing, that a cycles/codesize could be an interesting measure. All the data is there, buti it's difficult to compare them and get a bigger picture. 20% faster by more than doubling the codesize isn't always a good idea.

This is not representative of the results I usually see or get reported by users. More typical is a code size difference of about +/-20% and a speed improvement of 0-200%.

[jsk]

I applaud any attempts to write compilers to the 6502 that is a notarious difficult target platform.

I've personally looked at the code cc65 generates and it feels more like it generates a "byte code" and then generates JSR calls "mostly" to save space. So your comment about using byte code interpreter hits home!

My test case is a recursive function, which, as I understand it, isn't typical or the best target for the given platform as the stack allocation cannot be removed. (My compiler source is lisp-code, thus it's congruent with that choice and target, to look at this example.)

unsigned int ultfib(unsigned int n) {
if (n<2) return n;
return ultfib(n-1) + ultfib(n-2);
}

I have trouble reading the IC1, IC2 output as it's quite verbose, I've not been able to run the code (android/termux) But this is relatively simple code, and the parameter can be passed in register, and pushed if need be saved for recursion.

But, it feels that 2 lines of code (ok 3 generates 56 bytes of code on cc65 using mostly JSR calls to library functions, but vbcc generates 130+ bytes. It seems a lot of double loading and movement between registers is taking place. Maybe because of the "unfortunate recursive" function.

Looking at the generated code, my own code-generator for a "lisp" can get it down to 39 bytes, using JSR calls. Mostly just avoiding redundant calls to load n when it's already in AX!

Now, cc65 is good at using "char" (at times) to optimized index access to arrays < 256 bytes, however, in parameter passing, not really.

Inlining all the functions in cc65, by hand, I get 112 bytes...

So, I'd *really* like to go for a hybrid, where functions called once are bytecompiledJSR compiled and others which are timecritical are inlined for maximum performance.

I'm sad that (standard) C and no platform provides a way to say INLINE(funccall(a,b,c)) at the *callsite*, all codes aren't created equal, and much of it is only run once or quite seldomly.

What strikes me is that cc65 is quite capable and is very reliable. But the code generated is maybe slower than others. vbcc, in my tests inline everything so a simple recursive fib() becomes 130+ bytes whereas cc65 gave 56 bytes.

It is true that vbcc will usually generate direct code instead of calling lots of library functions. There surely are some opportunities to improve the size of the vbcc generated code in this fashion. However, vbcc usually generates more optimized code which makes using fixed library functions more difficult. Also, in such a small example, you have to take into account the size of the library functions as well.

If the size difference is as large as in your example, the code is probably not very friendly to the 6502. In such situations, the best option for code size probably would be to use a virtual machine with a compact instruction set.

I think, and here I'm guessing, that a cycles/codesize could be an interesting measure. All the data is there, buti it's difficult to compare them and get a bigger picture. 20% faster by more than doubling the codesize isn't always a good idea.

This is not representative of the results I usually see or get reported by users. More typical is a code size difference of about +/-20% and a speed improvement of 0-200%.

[Dbug]

I remember that some compilers (like Visual Studio) had options like "optimize for speed" and "optimize for space", maybe it would be useful to have this kind of options in vbcc?

In small projects having the code as fast as it can be is nice, but for large projects keeping the size under control is quite important.

[vbc]

But, it feels that 2 lines of code (ok 3 generates 56 bytes of code on cc65 using mostly JSR calls to library functions, but vbcc generates 130+ bytes. It seems a lot of double loading and movement between registers is taking place. Maybe because of the "unfortunate recursive" function.

Not sure what options you were using, but I get 103 bytes for that function:

Code: Select all

$ vc +atmos -O -c r.c
$ vobjdump r.o | head

------------------------------------------------------------------------------
VOBJ 6502 (little endian), 8 bits per byte, 2 bytes per word.
42 symbols.
1 section.

------------------------------------------------------------------------------
0000022b: SECTION "text" (attributes="acrx")
Flags: 1         Alignment: 1      Total size: 103       File size: 103      
33 Relocations present.

The code uses 16bit ints throughout which, when using direct code, just needs double the instructions. Changing int to char gets the code down to 55 bytes.

So, I'd *really* like to go for a hybrid, where functions called once are bytecompiledJSR compiled and others which are timecritical are inlined for maximum performance.

I understand that, but providing both options with good quality means pretty mouch double the effort in the compiler for a pretty niche situation. After all, if your space-critical code reaches a certain size, you are probably better of using a different instruction set and interpreting that.

E.g. for your code vbcc generates 46 bytes for 65816 or 68k and 36 bytes for 6809.

I'm sad that (standard) C and no platform provides a way to say INLINE(funccall(a,b,c)) at the *callsite*, all codes aren't created equal, and much of it is only run once or quite seldomly.

I am not sure I understand what you mean with "at the callsite". If I have one callsite where I need a version of a function compiled for speed, it does not seem to make sense to have another version anywhere else.

[vbc]

I have uploaded version 4 of the vbcc distribution for 6502 targets at http://www.compilers.de/vbcc.html

Major changes since last update:

- several bug fixes and code improvements
- small library improvements
- updated compiler core
- updated vasm
- updated vlink
- new target Apple I
- new target Apple II (consider vbcc65816 for IIgs)
- new target SNES (better use vbcc65816 version)
- automated banking for Commander X16
- file I/O for Commander X16
- clock() support for PET

[iss]

I have uploaded version 4 of the vbcc distribution for 6502 targets at http://www.compilers.de/vbcc.html

Thanks for update, @vbc!
I saw some nice improvements and everything seams to work fine for me!

Just one thing:

Code: Select all

#include <stdio.h>
int main(void)
{
  putchar(12);
  snprintf((char*)0xbb80+2,40,"Hello %s %c","world",'!');   // Works OK !
  sprintf((char*)0xbb80+40+2,40,"Hello %s %c","world",'!'); // Doesn't work ?
  return 0;
}

'sprintf' doesn't put any chars on screen. Is this a bug or I'm missing something?
Attached is above code compiled with +atmosr and added -autox.

[Dbug]

sprintf((char*)0xbb80+40+2,40,"Hello %s %c","world",'!'); // Doesn't work ?

Should that not be

sprintf((char*)0xbb80+40+2,"Hello %s %c","world",'!');

Are you sure the first 40 is legal, it's only for the sn version, no?

[iss]

Ups, sorry my bad, this was a typo while posting the message.

But still, the described 'sprintf' issue is real!
The code "as is":

Code: Select all

#include <stdio.h>
#include <string.h>

#ifdef __VBCC__
static const char* toolchain = "VBCC";
#endif

#ifdef __CC65__
static const char* toolchain = "CC65";
#endif

int main(void)
{
  memset((char*)0xbb80,0x20,40*28);

  snprintf((char*)0xbb80,   40,"Hello %s %d%c (%s) [%s]","world",0,'!',"snprintf", toolchain);
  sprintf ((char*)0xbb80+40,   "Hello %s %d%c (%s) [%s]","world",1,'!',"sprintf",  toolchain);

  while(1);

  return 0;
}

Attached are source, 2 TAPs: main-vbcc.tap and main-cc65.tap.

20250303_101922.jpg

20250303_101835.jpg

[vbc]

Thanks for update, @vbc!
I saw some nice improvements and everything seams to work fine for me!

Nice to hear!

'sprintf' doesn't put any chars on screen. Is this a bug or I'm missing something?

It is a bug (and an embarrassing one as it had been reported on a previous release for another 6502 target).

I have uploaded patch 1 to my site. Can you please check if that fixes the issue?

Thanks for the report.