GCC 6.2 toolchain for AmigaOS 3

[bebbo]

I just added the (requested) feature to optimize the shift instructions:

Superfluous signed/unsigned extensions are suppressed and the smallest possible shift is used.

Have a look here: https://franke.ms/cex/z/3q19bh.

[alkis]

Something is not pushed to repository?
with
-O2 -fomit-frame-pointer -mregparm -m68020
I get

_lsl_ubb:
        and.l #255,d0
        lsl.l d1,d0
        rts
_lsl_sbb:
        lsl.b d1,d0
        rts
_lsr_ubb:
        and.l #255,d0
        asr.l d1,d0
        rts
_lsr_sbb:
        asr.b d1,d0
        rts
_lsl_usb:
        and.l #65535,d0
        lsl.l d1,d0
        rts
_lsl_ssb:
        lsl.w d1,d0
        rts
_lsr_usb:
        and.l #65535,d0
        asr.l d1,d0
        rts
_lsr_ssb:
        asr.w d1,d0
        rts

for 68000, 68010,68030,68040 I get the "correct" shorter assembly.
If I switch to -O3 I get the correct code for all cpus.
So only for 68020 with -O2 there seems to be a glitch.

(I've cleaned and rm'd the target directory. All compiled from scratch on pi400)

[bebbo]

Originally Posted by alkis

Something is not pushed to repository?
with
-O2 -fomit-frame-pointer -mregparm -m68020
I get

_lsl_ubb:
        and.l #255,d0
        lsl.l d1,d0
         rts
...

for 68000, 68010,68030,68040 I get the "correct" shorter assembly.
If I switch to -O3 I get the correct code for all cpus.
So only for 68020 with -O2 there seems to be a glitch.

(I've cleaned and rm'd the target directory. All compiled from scratch on pi400)

did the same on an odroid and it's ok.
gcc version 6.5.0b 230914152536


the result on the pi400 is still better than before :-)

[jotd]

Bebbo something should be done for code short branches to the next instruction

   bne.b 0f
0:

this generates incorrect assembly and crashes because short branch cannot be of offset 0. jne probably fixes that, but a compile-time error would be nice.

[bebbo]

Originally Posted by jotd

Bebbo something should be done for code short branches to the next instruction

   bne.b 0f
0:

this generates incorrect assembly and crashes because short branch cannot be of offset 0. jne probably fixes that, but a compile-time error would be nice.

example code please

Disassembly of section .text:

00000000  .text:
   0:   6600 0000       bne.w 0x2

[NorthWay]

I just stumbled into this shift discussion here, and possibly dancing into a minefield (especially at this hour), wouldn't you rather do

moveq #$3f,d1
and.b 11(sp),d1

Or did I possibly miss out on something bigtime?

[bebbo]

Originally Posted by NorthWay

I just stumbled into this shift discussion here, and possibly dancing into a minefield (especially at this hour), wouldn't you rather do

moveq #$3f,d1
and.b 11(sp),d1

Or did I possibly miss out on something bigtime?

That AND is implicitly executed by the CPU, no need to add it into the asm code.

2. Register—The shift count is the value in the data register specified in the instruction modulo 64.

Also the extension to long can be omitted for the shift count. Thus the original code

_lsr_sbb:
        move.b 7(sp),d1
        ext.w d1
        ext.l d1
        move.b 11(sp),d0
        ext.w d0
        ext.l d0
        asr.l d0,d1
        move.l d1,d0
        rts

is shorter now:

_lsr_sbb:
        move.b (7,sp),d0
        move.b (11,sp),d1
        asr.b d1,d0
        rts

[jotd]

main:
	bra.b	0f
0:
	rts

yields

	SECTION S_0,CODE

__stext:
	DC.W	$6000			;0
__etext:
	RTS				;2: 4e75
	END

K:\jff\AmigaHD\PROJETS\arcade_remakes\asm_test>c:\amiga-gcc\bin\m68k-amigaos-as --version
GNU assembler (GNU Binutils) 2.39.0.230129-131808
Copyright (C) 2022 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or later.
This program has absolutely no warranty.
This assembler was configured for a target of `m68k-amigaos'.

I also had a problem with add.b #1,a0 generating an illegal instruction without errors but now I can't reproduce it.

not using any optimization, and using -m68000 flag.

[bebbo]

Originally Posted by jotd

[CODE]
...
K:\jff\AmigaHD\PROJETS\arcade_remakes\asm_test>c:\amiga-gcc\bin\m68k-amigaos-as --version
GNU assembler (GNU Binutils) 2.39.0.230129-131808
Copyright (C) 2022 Free Software Foundation, Inc.

wow, msys2...


... fix is on the way - no idea when msys2 builds successfully again...

[jotd]

Don't look at me, I positively hate MSYS2 and its fake linux compatibility

I always use python to control my toolchains, no damn fake shells that can call .bat files (called by .bat files).

Take your time. I'm aware of this issue and I'm using jcc a lot anyway.

[jotd]

Oh I found the other bug

main:
	subq.b	#1,a1 
	rts

assembles without errors, when sub to Ax register should be w or l not b

disassembly: junk

        SECTION S_0,CODE

__stext:
        DC.W    $5309                   ;0
__etext:
        RTS                             ;2: 4e75
        END

[bebbo]

Originally Posted by jotd

Oh I found the other bug

main:
    subq.b    #1,a1 
    rts

assembles without errors, when sub to Ax register should be w or l not b

that's a binutils issue... anyway, it's fixed now. And I managed to create a new MSYS2 version.

[jotd]

you rock!

where can it be downloaded?

[bebbo]

Originally Posted by jotd

you rock!
where can it be downloaded?

there you go: https://franke.ms/download/setup-amiga-gcc.exe


... keep a backup^^

[jotd]

going to burn this on a CD

[jotd]

bingo

bagman.68k:6737: Error: operands mismatch -- statement `subq.b #1,a2' ignored

[girv]

I'm trying to build master on macOS Monterey. I'd managed to do this before without issue.

isl@0.18 is disabled in homebrew now, but I managed to get it installed by changing the system date temporarily.

Using the script posted in this thread previously: https://gist.github.com/rrs42/5e5b7b...b89d0f4dd3a2a8
with CC=gcc-12 CXX=g++-12 gmake all SHELL=$(brew --prefix)/bin/bash

It fails on building gcc with:

...
make use of stdint.h in include/isl/stdint.h (assuming C99 compatible system)
checking which gmp to use... system
checking gmp.h usability... no
checking gmp.h presence... no
checking for gmp.h... no
configure: error: gmp.h header not found
gmake[1]: *** [Makefile:4921: configure-isl] Error 1
make gcc...failed

bison and texinfo paths are set.
isl@0.18 and gmp are linked in brew.
gmp.h exists at /usr/local/Cellar/gmp/6.3.0/include/gmp.h

I'm guessing theres some issue with isl@0.18 and the installed version of gmp? Any ideas?

Also - is the gcc 13 branch gone now?

[Samurai_Crow]

@girv
What version of gmake are you using? Gmake 4.4.1 is known to work. 3.x is known not to work.

[girv]

$ gmake -v
GNU Make 4.4.1
Built for x86_64-apple-darwin21.6.0
...

[Samurai_Crow]

Ok. What about Bash?