GCC 6.2 toolchain for AmigaOS 3

[Samurai_Crow]

I think you're supposed to build with custom startup code.

Try typing "make simple.library_r" and look at the -resident command line option.

[Hewitson]

Originally Posted by bebbo

For those interested in my recent versions, I am providing now

• a setup for windows (64 bit)
• an archive for linux(64 bit) built on centos
• no osx - sorry - well, not really^^

Would you be willing to host an OSX macOS version if I built one? I am happy to maintain it.

[alkis]

Originally Posted by pipper

Does anyone know how to build a .library file with the gcc toolchain? I tried to follow the example at https://github.com/bebbo/amigaos-cro...mple-library.c but had no success. It built _something_ but I don’t think it’s a valid library file

Not pure gcc, but this might be of help
https://github.com/alexalkis/library

[pipper]

Thanks, alkis! Yeah, this is not really what I’m looking for as you’re using asm to setup the library structure. If everything fails, I can use this approach, though. I was looking for a C only solution if possible. I thought that -ramiga-lib and the macros in <stabs.h> would do the magic.

[pipper]

I think I got it working

Two issues
1) I was a moron, producing essentially .o file (I had -c left over in the gcc commandline)
2) libnix seems to have a 14 year(!) old bug in libinit.c#L183 :

in LibInit() :
if (!__UserLibInit(&lib->LibNode,lib->DataSeg)) {
    FreeMem((UBYTE *)lib-lib->LibNode.lib_NegSize,lib->LibNode.lib_NegSize+lib->LibNode.lib_PosSize);
    lib = NULL;
  }

All sources and examples on the internet claim that __UserLibInit() needs to return 0 to indicate success. But libnix interprets it as failure. Returning !=0 from my __UserLibInit eventually got me past the initial hurdle (OpenLibrary claimed it couldn't open my library).

[asymetrix]

Originally Posted by bebbo

For those interested in my recent versions, I am providing now

• a setup for windows (64 bit)
• an archive for linux(64 bit) built on centos
• no osx - sorry - well, not really^^

Installed on Windows 10, 64bit.

I used ./bin to get direct access to gcc.
Im getting a command line : error .

<C:\amiga-gcc>.\bin\m68k-amigaos-gcc.exe hello.c hello -Os -noixemul
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0x68): multiple definition of `_exit'; C:/amiga-gcc/m68k-amigaos/libnix/lib/ncrt0.o:C:/amiga-gcc/m68k-amigaos/libnix/lib/ncrt0.o:(.text+0x68): first defined here
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0x68): multiple definition of `exit'
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0xe8): multiple definition of `main'; /Users/angel/AppData/Local/Temp/ccYVojmD.o:(.text+0x1c): first defined here
collect2: error: ld returned 1 exit status

C:\amiga-gcc>

[bebbo]

Originally Posted by asymetrix

Installed on Windows 10, 64bit.

I used ./bin to get direct access to gcc.
Im getting a command line : error .

<C:\amiga-gcc>.\bin\m68k-amigaos-gcc.exe hello.c hello -Os -noixemul
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0x68): multiple definition of `_exit'; C:/amiga-gcc/m68k-amigaos/libnix/lib/ncrt0.o:C:/amiga-gcc/m68k-amigaos/libnix/lib/ncrt0.o:(.text+0x68): first defined here
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0x68): multiple definition of `exit'
/amiga-gcc/lib/gcc/m68k-amigaos/6.4.1b/../../../../m68k-amigaos/bin/ld: hello:(.text+0xe8): multiple definition of `main'; /Users/angel/AppData/Local/Temp/ccYVojmD.o:(.text+0x1c): first defined here
collect2: error: ld returned 1 exit status

C:\amiga-gcc>

Your problem is a missing '-o' since that commandline tries to create a file named 'a.out' using 'hello.c' plus 'hello' as input files.

(And you should add that bin folder to the path or use the provided cmdline.bat which does this for you.)

[ReadOnlyCat]

Originally Posted by bebbo

For those interested in my recent versions, I am providing now

• no osx - sorry - well, not really^^

I second Hewitson's initiative to provide a macOS port and have you host it, this would be really nice.

Originally Posted by bebbo

I also wrote a script to easily benchmark different compilers. Here is the chart for
[...]
You may provide more benchmarks and compilers (prebuild for linux) and I'll update the chart and maybe also the compiler explorer

Am I correct when interpreting that this means that this produces code more than twice faster than the previous best C compilers?
If so, that is mightily impressive.

I do not recall the exact version of GCC Samuel Devulder and I used when we were fiddling with the Quake sources around 1996-97 (Samuel published them on Aminet at the time) but this makes me very eager to recompile them and measure the difference.

[alkis]

Originally Posted by ReadOnlyCat

I do not recall the exact version of GCC Samuel Devulder and I used when we were fiddling with the Quake sources around 1996-97 (Samuel published them on Aminet at the time) but this makes me very eager to recompile them and measure the difference.

My guess would be, little or no difference, since I'd assume the graphics bandwidth would be the bottleneck.

[asymetrix]

Originally Posted by bebbo

Your problem is a missing '-o' since that commandline tries to create a file named 'a.out' using 'hello.c' plus 'hello' as input files.

(And you should add that bin folder to the path or use the provided cmdline.bat which does this for you.)

Works perfect, thanks bebbo !

[Hewitson]

Originally Posted by asymetrix

<C:\amiga-gcc>.\bin\m68k-amigaos-gcc.exe

That's a strange way of doing things. "bin\m68k-amigaos-gcc" is sufficient.

[ReadOnlyCat]

Originally Posted by alkis

My guess would be, little or no difference, since I'd assume the graphics bandwidth would be the bottleneck.

You are assuming the compiled code was attaining the maximum bandwidth of AGA Chip RAM.

I doubt that was the case since the profiler runs I collected at the time revealed that most of the time was spent in

fmul

functions (*) followed only a bit further down the line by the span drawing functions.
Moreover, even then, there is no guarantee that these drawing functions were also reaching maximum bandwidth.
For both types of tasks, a better optimizing compiler would probably help maximize the throughput and end up having a positive influence on the frame rate, even if probably will be only by a small fraction.

Of course, switching from floating point maths to fixed point in the absence of an FPU would probably the first step to take before optimizing anything else.

(*) my 1200 was equipped with a 68030 but I do not recall if it was also fitted with a 68881/2 FPU at the time, I suspect it was not since the

fmul

functions would have been replaced by the corresponding assembly instructions then.

[pipper]

@bebbo: do you know if (and how good) the optimizer knows how to optimize for 060? I.e. shift instructions around to make use of dual-issue?
Also do -m68020-60 and -m68060 avoid instructions that were removed from the 060?
Thanks!

[bebbo]

Originally Posted by pipper

@bebbo: do you know if (and how good) the optimizer knows how to optimize for 060? I.e. shift instructions around to make use of dual-issue?
Also do -m68020-60 and -m68060 avoid instructions that were removed from the 060?
Thanks!

There are insns which are not used if code is generated for 68060. Some mulu/muls and some fpu insns - no idea if it's sufficient^^

And there is no insn reordering for m68k* in gcc6.

[bebbo]

Originally Posted by ReadOnlyCat

I second Hewitson's initiative to provide a macOS port and have you host it, this would be really nice.

Linux is ok, Windows needs a crutch and macOS is a PITA.

• Windows https://franke.ms/download/setup-amiga-gcc.exe
• Linux https://franke.ms/download/amiga-gcc.tgz
• MacOS https://franke.ms/download/amiga-gcc-macos.tgz

There you go...

Originally Posted by ReadOnlyCat

Am I correct when interpreting that this means that this produces code more than twice faster than the previous best C compilers?
If so, that is mightily impressive.

That's the result for this benchmark, if you benchmark a hello-world the program with the smallest startup overhead will win.

So propose a different benchmark and we'll see.

[ReadOnlyCat]

Originally Posted by bebbo

Linux is ok, Windows needs a crutch and macOS is a PITA.

• Windows https://franke.ms/download/setup-amiga-gcc.exe
• Linux https://franke.ms/download/amiga-gcc.tgz
• MacOS https://franke.ms/download/amiga-gcc-macos.tgz

There you go...

Thanks!

Originally Posted by bebbo

That's the result for this benchmark, if you benchmark a hello-world the program with the smallest startup overhead will win.

So propose a different benchmark and we'll see.

I was thinking of using Wolfenstein 3D as a benchmark since the code runs a relatively varied array of routines and should test memory bandwidth (for texture mapping) as well as geometric fixed points computations and some general purpose code (gameplay loop).
The issue being that as such, the Wolfenstein 3D code cannot be directly ran as a benchmark and would require modification to run tests straight from the command line.
I would love to do that but unfortunately I am and will be time starved for at least one full week.

[alkis]

get https://www.lua.org/ftp/lua-5.3.5.tar.gz
tar xvf the thing.
cd lua-5.3.5/src
make generic CC=m68k-amigaos-gcc CFLAGS="-O3 -fomit-frame-pointer" AR="m68k-amigaos-ar r" RANLIB=m68k-amigaos-ranlib LDFLAGS="-noixemul -s"

BOOM! Lua from upstream. No source changes. No makefile changes.
Thanks bebbo.

[Leffmann]

But does it pass the test suite? Lua is highly dependent on the standard C library to function correctly.

[alkis]

Frankly I haven't tried to do any tests (besides simple little lua scripts).

I took a quick look at https://www.lua.org/tests/ and there is no 5.3.5. I could try with 5.3.4 when I get some time

[Hedeon]

Hello,

I made a library with this great tool so there are no startup files used. However (/because of that), my code crashes on float conversion stuff like floatunsidf. Looking at the code I see jumps to a library offset (-36 etc) from an uninitialized library pointer. I guess I need to manually open some libraries... But which?

Edit: nvm. It was late. Looked at nm and I see I am missing MathIeeeDoubBasBase, MathIeeeDoubTransBase and MathIeeeSingBasBase