For us (Wireshark) the difficulty wasn't with our own codebase, but with getting our dependencies ported over. Most libraries built just fine, but some strongly assumed that "Windows" meant "x86".

It's not just Windows, either. Many libraries (particularly ones that use Autotools) are absolutely blind to the notion that you might want a universal binary on macOS.

▲

astrange 5 days ago | parent [-]

When we ported OpenJDK to macOS, I ended up producing a universal binary by having the Makefile run itself to produce HotSpot twice, and then gluing them together with `lipo` afterwards. There isn't really a better way when the actual project configurations are different.

IIRC it was eventually removed because nobody else needed to do such a thing so it was hard to maintain.

▲

kccqzy 5 days ago | parent [-]

Sure. How else would you build a universal binary then? Given the low-level nature of the language not many tasks can be usefully shared between different architectures.

▲

geraldcombs 5 days ago | parent [-]

For plain C/C++ you can just pass `-arch x86_64 -arch arm64` to clang. CMake takes care of this for you if you specify `CMAKE_OSX_ARCHITECTURES=x86_64;arm64` and IIRC Meson has similar functionality.

▲

kccqzy 5 days ago | parent | next [-]

TIL. I didn't know clang supports this natively.

▲

delta_p_delta_x 4 days ago | parent | next [-]

Clang is natively a cross-compiler. Pass in --sysroot and a corresponding valid sysroot tree for any micro architecture/platform (arm-eabi, macOS, Windows MSVC, PowerPC, Alpine Linux with musl, you name it) and Clang will happily retarget the binary to the correct target platform.

▲

mrpippy 4 days ago | parent | prev [-]

Apple has supported that ‘-arch’ option in their GCC/Clang since at least the PPC->Intel transition, maybe even earlier (PPC64? NeXT/OPENSTEP?)

▲

lukeh 4 days ago | parent [-]

Yes, since NEXTSTEP.

	▲	pjmlp 4 days ago \| parent [-]
		That would be NeXT, and their GCC fork.

▲

zdw 5 days ago | parent | prev | next [-]

I assume this is faster than doing two separate builds, because it can skip certain steps of the complier pipeline, and only the items that are arch specific (codegen, probably others) are done twice?

	▲	astrange 4 days ago \| parent [-]
		They can't really share anything since the preprocessor stage can be different.

▲

bee_rider 4 days ago | parent | prev [-]

How much work does clang have to do for this sort of thing (as opposed to llvm). Hypothetically could we start distributing programs in llvm ir, and compile that locally to ARM, x86, risc-v, or whatever else?

I mean, no, that’s silly, right? But it would be kind of neat…

	▲	dundarious 4 days ago \| parent [-]
		llvm-ir is not architecture independent, even excluding issues like the C pre-processor.