In certain occasions, both gcc and LLVM emit libcalls to a set of equivalent __atomic_* routines.

• gcc implements these routines in its libatomic module, which programs are expected to link explicitly (-latomic)
• LLVM implements these routines in compiler-rt, but the routines are not built by default

This means specifying useLLVM = true; for a platform yields an stdenv that does not provide these routines anywhere, making it impossible to build certain programs. Trivial example:

let
pkgs = import <nixpkgs> {
  crossSystem = {
    config = "x86_64-unknown-linux-gnu";
    useLLVM = true;
  };
};
source = pkgs.writeText "test.c" ''
  int main() { __int128 src, dst; __atomic_load(&src, &dst, __ATOMIC_SEQ_CST); }
'';
in pkgs.runCommandCC "test" {} ''
  source $stdenv/setup
  $CC ${source} -o $out
''

I'm not sure how we should fix that. Since the set of routines should coincide, perhaps after #132340 we could try using gcc's libatomic in LLVM stdenvs. Assuming we want to use compiler-rt's implementation (which btw does not depend on pthreads), their recommended setup seems to be shipping those symbols in a shared library (which I presume would be named libatomic.so to match gcc's requirements on user builds). But from feedback I got, I think we do not plan on shipping these symbols in a separate library and would instead like to ship them in builtins.a? I'm not sure and this is the first thing I'd like to understand before attempting to fix anything.