LLVM / Clang Progress

[wawa]

looks like someone is working on clang/llvm for aros, there seems to be some preparational commits, but i havent seen it commited itself. propbably would need further patches though to produce amiga-m68k code.

[Samurai_Crow]

By patches I assume you mean an entire backend. A backend in LLVM is a third of the compiler.

[wawa]

im completely unfamiliar with the subject matter. i simply noticed someone made some preparatory commits to the aros repo (mschulz?) which might indicate that he may experiment with llvm locally.

i understand that you have been working on a 68k backend but it has never become actually usable, or simply none used it. given that aros would receive llvm backend (for contemporary platforms) perhaps some adoptation for amiga-m68k backend might be made, based on your experience. even if it seems unlikely. if it would practically mean to write a backend from the scratch, then its plain impossible of course.

[Samurai_Crow]

It's not impossible but the backend in my repo is incomplete. The work on AROS is for Mac hosted and uses an x86.

[wawa]

Originally Posted by Samurai_Crow

It's not impossible but the backend in my repo is incomplete. The work on AROS is for Mac hosted and uses an x86.

it looks like schulz uses a mac as well, he commited some patch(es) for osx hosted, but it doesnt make so much difference to linux hosted i guess, the host is anyway x86 no matter what i guess. so maybe the effort could be bundled if you dag up your sources and teamed up with him.

just kidding

Originally Posted by Samurai_Crow

Backend code for LLVM requires knowledge of how to implement the IR in native code. From there, you would have to implement the peephole optimization yourself.

The reason that the backend in my GitHub repo has languished is that it is based on a very old version of LLVM and merging it with current code would be an ordeal in itself.

Do you remember what clang version it was for? BTW, is that hard to maintain backends across LLVM versions? I thought the LLVM design was done with ease of maintenance in mind...

Also, do you have any list of what things were done and what things were pending?

[Samurai_Crow]

Originally Posted by ardi

Do you remember what clang version it was for? BTW, is that hard to maintain backends across LLVM versions? I thought the LLVM design was done with ease of maintenance in mind...

Also, do you have any list of what things were done and what things were pending?

It's version 3.2. GetElementPtr still not implemented and C runtime library needs to borrow from GCC.

GetElementPtr is the equivalent of all the non-register addressing modes but can probably be implemented as LEA during the earlier stages and just merged later when the needed addressing mode is available and the pointer is used once.

Originally Posted by Samurai_Crow

Backend code for LLVM requires knowledge of how to implement the IR in native code. From there, you would have to implement the peephole optimization yourself.

The reason that the backend in my GitHub repo has languished is that it is based on a very old version of LLVM and merging it with current code would be an ordeal in itself.

I just saw there's another fork, with one commit from last September. The commit adds new code, but I'm not sure if it can be considered more complete than yours, or the opposite: https://github.com/mntmn/llvm-m68k

[Samurai_Crow]

I think mntmn's is ahead but I still only saw the build script was modified and the readme.txt file.

[wawa]

Originally Posted by Samurai_Crow

I think mntmn's is ahead but I still only saw the build script was modified and the readme.txt file.

lukas is trying to build 68k backend? he never mentioned it publicly to my knowledege.

im sorry there isnt some place where all these projects might gather and help each other rather than being spread all over the place to almost no awareness of others.

edit: äh, its for a mac, could be helpful though but only with joined forces, looks like he lost interest in favour of an amiga project.

[Samurai_Crow]

Originally Posted by wawa

lukas is trying to build 68k backend? he never mentioned it publicly to my knowledege.

im sorry there isnt some place where all these projects might gather and help each other rather than being spread all over the place to almost no awareness of others.

edit: äh, its for a mac, could be helpful though but only with joined forces, looks like he lost interest in favour of an amiga project.

The readme.txt file indicates it is for AmigaOS 3.

[nogginthenog]

Resurrecting this thread. Seems to be some activity on LLVM (Sept 2020). Looks like the devs are trying again to merge their changes into the main branch.

https://www.phoronix.com/scan.php?pa...00-Series-2020
(6800 is a typo)

https://lists.llvm.org/pipermail/llv...er/145320.html

https://github.com/M680x0/M680x0-mono-repo

[Samurai_Crow]

Even if it is for Debian 68k, having LLVM would be a good stepping-stone for Rust and WebAssembly on our machines.

[nogginthenog]

Update: LLVM 68k has been merged upstream. This is pretty good news. Congrats to the authors.
It's marked as experimental right now.

[nogginthenog]

Got it to compile (VM needed more swap lol) but can't build anything with clang due to missing functionality. Maybe I jumped the gun. I'll report back if I get further.

[nogginthenog]

There was a recent push that fixed the error I was getting. Now I can compile code (020+ seems broken).

int func1();
int func2(int x);

void main()
{
    func2(func1() * 16);
    func2(func1() * 3);
}

clang -O3 -fomit-frame-pointer -target m68k -S test.c

main:                                   ; @main
        sub.l   #4, %sp
        jsr     func1
        lsl.l   #4, %d0
        move.l  %d0, (%sp)
        jsr     func2

        jsr     func1
        move.l  %d0, %d1
        lsl.l   #1, %d1
        add.l   %d0, %d1
        move.l  %d1, (%sp)
        jsr     func2
        add.l   #4, %sp
        rts

Could be worse! The GCC protos don't work so I can't compile any OS calls. Didn't spend much time on this.

[cbmuser]

Hello!

I'm a long time Amiga user and new to this forum. I have been hanging around mostly at amiga.org and a1k.org.

Having said that, I am Debian's primary maintainer of the m68k port and the person who initiated both the conversion of the m68k GCC backend from CC0 to MODE_CC [1] and therefore saving it from being dropped as well as the effort to add an m68k backend to LLVM [2]. I am myself not a compiler expert, so the work on the backend is done by two young LLVM developers in the US [3] and Taiwan [4].

The initial work for adding the backend was funded by a Bountysource campaign while the current maintenance and continued development is being funded with the help of Patreon [5] and OpenCollective [6]. Funding helps keep both developers motivated to work on the backend since the m68k is not their primary area of interest while the necessary time and effort cannot be underestimated.

Since there seems to be some interested at EAB in the M68k LLVM backend, I was wondering whether anyone here would be willing to support the project with a regular or one-time donation via the Patreon or Open Collective pages.

I am not personally getting any of the money from these two funding pages, all of the money goes to the two LLVM developers. I am actually myself supporting the project with a monthly donation of 10 Euros via Patreon.

For more information, please have a look at the project homepage [7] as well as the LLVM issue [8] and pull request [9] pages which show what is currently being worked on in the backend.

Also, if you prefer talking over IRC, there is a project-related chat at #llvm-m68k on OFTC IRC.

Let me know if you have any questions or comments.

Adrian

> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91851
> [2] https://github.com/M680x0/M680x0-llvm/issues/59
> [3] https://github.com/mshockwave
> [4] https://github.com/0x59616e
> [5] https://www.patreon.com/m68k_llvm
> [6] https://opencollective.com/m68k-llvm-dev
> [7] https://m680x0.github.io
> [8] https://github.com/llvm/llvm-project...is%3Aopen+m68k
> [9] https://github.com/llvm/llvm-project...is%3Aopen+m68k

[E-Penguin]

I do like a fully referenced article in IEEE citation style.