iOS fat binaries: should every arch include bitcode?

Question

When building a library that is to be used in bitcode-enabled apps, does each arch in the fat binary need to be built with -fembed-bitcode or only one of them?

One of them because the bitcode is architecture-independent and will just be duplicated?

All of them because pre-processor commands may alter the code based on architecture (e.g. NSInteger width)?

Excluding x86 slices otherwise duplicate symbol _llvm.cmdline happens? - rdar://21884601

Leoleod

Bitcode is not architecture-independent, hence each slice in a fat binary should include bitcode. – Outdo

@Outdo Have you got a reference for that? – Leoleod

Posted an answer with links. – Outdo

Answer 1

Bitcode is just another form of LLVM IR, which is architecture-dependent.

Which means that each slice in your fat binary should contain its own bitcode section.

Update: I wrote a blog-post about Bitcode, you may find there some useful details: Bitcode Demystified

Outdo

Where does it say bitcode is architecture-dependent? – Leoleod

Fair enough. Can't find it in documentation as well. But I can say for sure that's different for each slice. You can check it out manually by calling clang -S -emit-llvm -arch arch_name main.c using different 'arch_name' and discover LLVM IR. – Outdo

I though the whole point of including bitcode was so Apple can compile it for different architectures in the future. – Leoleod

In fact, if you do not include bitcode for each arch, you'll get failures at archive time. – Thrift