[bug] ld doesn't support ppc

[ChrisNonyminus]

When cctools is configured to target powerpc-apple-darwin, as assembles powerpc assembly without issue, but ld will not link it.

chrisnonyminus@DESKTOP-JLMSNKL:~/test/helloworld$ powerpc-apple-darwin-ld a.out -o test -arch ppc
ld: unknown/unsupported architecture name for: -arch ppc
chrisnonyminus@DESKTOP-JLMSNKL:~/test/helloworld$ powerpc-apple-darwin-ld -v
@(#)PROGRAM:ld  PROJECT:ld64-609
BUILD 17:46:02 Mar  3 2022
configured to support archs: armv6 armv7 armv7s arm64 arm64e arm64_32 i386 x86_64 x86_64h armv6m armv7k armv7m armv7em (tvOS)
LTO support using: LLVM version 10.0.0 (static support for 26, runtime is 26)```

[Gcenx]

ld64-127 was the last version from Apple to support ppc

[Wohlstand]

Recently I attempted to review the difference between the current ld64 and the ld64-127.2, and I attempted to backport PowerPC-related parts but hadn't completed that yet. On another side, I want to get a shot to make the ld64-127 work here. How much the modded ld64 is different from the official ld64?

[Wohlstand]

P.S. I attempting to make a cross-compiler toolchain to run my experimental stuff on my Mac Mini G4, however, it's a pain to build the code on it because of a very weak processor, and therefore a cross-compiler for modern Linux should simplify this. I built everything except GCC: it fails to build the thing because of the ld64...

[Wohlstand]

I made the first attempt here: master...Wohlstand:cctools-port:ppc-revive
I had to compare two versions: 127.2 and 128.2 to see what parts of the code were changed in order to remove PowerPC support. So, I attempted to restore that code to the latest version (which appears in this repository here). Trying some to verify its work and make the thing work...

[VariantXYZ]

@Wohlstand , it's two years later, but I did try your branch and was able to successfully build cctools for powerpc-apple-darwin8, but if I try to use this to build something like 'gcc' seems to fail to actually link against anything in the 10.4 SDK (ld can t map file errno 22 dylib1.o for architecture ppc). Were you ever able to get a cross-compiler targeting powerpc-apple-darwin8 working?

Edit: Ah, actually, it seems like by simply using lipo to extract the mach-o ppc binaries out, ld can use them, this can work!

[Wohlstand]

Hello!
I didn't yet finished my experiment, and the other problem that GCC is tricky to build. I did all the work on Ubuntu 20.04, and it was close to work, but on Ubuntu 22.04 it now fails (gcc itself) to build even worse for unknown reasons, even I trying to build the same version (Maybe I should build older GCC version for native Linux then? 🤔).

Edit: Ah, actually, it seems like by simply using lipo to extract the mach-o ppc binaries out, ld can use them, this can work!

I built the custom Apple's LD thing, and backported support for PPC from old version of ld to new. And even more tricky thing, that when I building GCC cross-compiler to macOS on Linux host, it gets confused how LD should behave (I mean build scripts of GCC itself).

[Un1q32]

For testing I would recommend using Clang 7.1.0 (the last version to support ppc darwin). The build process is much simpler than GCC's. They also have prebuilt binaries on llvm.org for it, they require ncurses 5 though so you'll need to build that.

[VariantXYZ]

I built the custom Apple's LD thing, and backported support for PPC from old version of ld to new. And even more tricky thing, that when I building GCC cross-compiler to macOS on Linux host, it gets confused how LD should behave (I mean build scripts of GCC itself).

I actually have your cctools ppc-review branch and GCC 13.3 building and compiling/linking target binaries, as mentioned I needed to forcibly extract the ppc binaries out of the SDK with the cctools lipo (you do have to rebuild crt1.o if you're using the 10.4 SDK but the 10.5 SDK should work fine if you target 10.4 specifically):

    cd $TARGET_PREFIX/$MAC_SDK_VERSION.sdk/usr/lib
    for i in **/*.dylib **/*.a **/*.o; do
      if file "$i" | grep -q "for architecture ppc)"; then
        mv $i $i.universal
        $TARGET_PREFIX/bin/$TARGET-lipo -thin ppc $i.universal -output $i
      elif file "$i" | grep -q "for architecture ppc7400)"; then
        mv $i $i.universal
        $TARGET_PREFIX/bin/$TARGET-lipo -thin ppc7400 $i.universal -output $i
      fi
    done

Now I'm trying to figure out why ld crashes when debug information is requested (-g passed by the driver). My current theory is that it's this issue with ld not having sufficient stack size by default, but I'll need to dig in when I have time, maybe later tonight.

% powerpc-apple-darwin8-gcc -O2 -g -o conftest.o conftest.c         
powerpc-apple-darwin8-gcc: internal compiler error: Segmentation fault: 11 signal terminated program collect2
Please submit a full bug report, with preprocessed source (by using -freport-bug).
See <https://gcc.gnu.org/bugs/> for instructions.

vs

% powerpc-apple-darwin8-gcc -O2 -o conftest.o conftest.c 
% file conftest.o
conftest.o: Mach-O executable ppc

conftest.c (from libpng) is just:

/* confdefs.h */
#define PACKAGE_NAME "libpng"
#define PACKAGE_TARNAME "libpng"
#define PACKAGE_VERSION "1.6.46.git"
#define PACKAGE_STRING "libpng 1.6.46.git"
#define PACKAGE_BUGREPORT "png-mng-implement@lists.sourceforge.net"
#define PACKAGE_URL ""
#define PACKAGE "libpng"
#define VERSION "1.6.46.git"
/* end confdefs.h.  */
int
main (void)
{
  ;
  return 0;
}

It seems to work if I use the old XCode PPC ld from 3.2, but that locks me into using a Mac instead of like... a Linux docker image running on GHA :P.

If we can figure this last bit out, then it means we'll have a modern-enough cross-compiler for powerpc-apple-darwin8 (Mac OS X Tiger on PPC).

[Un1q32]

Once LLVM 20 rc1 is out I'll try to add back support for PPC darwin into LLVM 20, no promises though, I'm not that experienced with the LLVM codebase yet or compilers in general, but being able to use Clang instead of GCC would be great for me personally.

I probably shouldn't announce projects that I have no idea if I can complete but oh well.

[Un1q32]

This guy seems to have made powerpc darwin work in LLVM 8, that's the newest I can find online https://github.com/iains/LLVM-8-Branch

[VariantXYZ]

Yeah, I'm using his gcc 13.3 branch as well.

Anyway, built dsymutil with llvm-7 and it seemed to do the trick.

Anyway, for anyone who might benefit, I created a Dockerfile to create the cross compiler: https://github.com/VariantXYZ/gcc-powerpc-apple-darwin8/blob/main/.github/workflows/Dockerfile

Unfortunately, I'm struggling to build it in GHA 🥲

[Un1q32]

Is there a reason to not use GCC 14? Also why do you rebuild the Csu when you can just use the one in the SDK?

[VariantXYZ]

The build finally succeeded: https://github.com/VariantXYZ/gcc-powerpc-apple-darwin8/pkgs/container/gcc-powerpc-apple-darwin8

Is there a reason to not use GCC 14

I just started with 13.3 and forgot 14 existed, actually.

Also why do you rebuild the Csu when you can just use the one in the SDK?

We're using the 10.4u SDK which has this issue tpoechtrager/osxcross#50 (comment)

10.5 works too, but I was stubbornly trying to see if I could get just 10.4u working.

For reference, I documented the general reasoning for everything here: https://vxyz.me/00003.html

[VariantXYZ]

Once LLVM 20 rc1 is out I'll try to add back support for PPC darwin into LLVM 20

I think this is definitely the 'cleanest' solution that will probably allow for keeping everything as modern as can be, but for the interim, I think GCC 14.2 + 10.5 SDK targeting 10.4 works (I'm trying to fix my iBook so I can actually test things...). Images for both are here: https://github.com/VariantXYZ/gcc-powerpc-apple-darwin8/pkgs/container/gcc-powerpc-apple-darwin8/versions

I have some workflows here that show how to use them: https://github.com/VariantXYZ/gbdev-powerpc-apple-darwin8/tree/main/.github/workflows

I'm trying to setup QEMU for now to properly test things...

[Un1q32]

I saw that ppc64 with LLVM 7 doesn't work, there will probably be a lot of issues beyond that since the LLVM code for ppc darwin hasn't been updated in like 6 years and my only experience with the LLVM codebase is a few extremely minor patches. Plus I've had serious issues with compiling for OSX 10.5 x86 and iOS 2 with Clang. I imagine GCC will probably be the best option for a while.