Rethink musl-clang

[geofft]

In both build-cpython.sh and build-zstd.sh we have some workarounds for the fact that musl-clang suppresses the vectorized math intrinsics headers that are shipped with clang. We've also had other issues with it, notably that the command name leaks out in sysconfig and end users may not have a command with that name - there are a few issues on this issue tracker and uv's.(See also discussion in an internal channel starting at https://discord.com/channels/1039017663004942429/1344040998724046920/1346959405081694350 .) Also there's no musl-clang++ and we have a # TODO: Explore a musl-clang++ shim?.

The thing that musl-clang does is relatively uncomplicated and there's probably a way to do it better. It might be as simple as adding to CPPFLAG -nostdinc --sysroot /prefix/to/musl -isystem /prefix/to/musl/include, and to LDFLAGS -static-libgcc --sysroot /prefix/to/musl -nostdlib -lc -Wl,-dynamic-linker,/lib/ld-musl-whatever.so. Or we create a custom sysroot, or something.

This would also save us some subprocesses by avoiding a shell wrapper, marginally speeding up the build.

[indygreg]

Agree that musl-clang isn't rocket science and could likely be reimplemented via common compiler/linker flags in PBS.

FWIW musl-clang and the musl PBS builds in general have always been troublesome. I'm not sure how much of it is musl in general or the use musl+clang. (I always suspected that most of the world uses musl+gcc and problems would go away with GCC.)

When I encounter problems like this, I usually go to the package build definitions from distros to see what they do, trusting Debian+Ubuntu and Fedora the most. In the case of musl, I don't believe there's much diversity in the ecosystem and it is only Alpine Linux using musl. So I suspect that if we're doing anything that Alpine is not, we're likely to run into novel problems that need to be solved.

In the case of musl, I can make a compelling argument that it is just easier to use GCC. Since we build with an ancient Debian, we'd need to build / install a modern GCC toolchain to get modern PGO/LTO/etc and claw back performance loss from ditching modern LLVM.