I wouldn't say that whether the shim libraries are built into the emulator or statically linked into the transformed output is very important. There are really

I wouldn't say that whether the shim libraries are built into the emulator or statically linked into the transformed output is very important. There are really two mostly orthogonal axes here: - Instruction set: JIT (emulators, Rosetta) vs. static recompilation (only niche projects when the source is a machine ISA, but arguably Java AOT compilers, OdinMonkey, etc. count) vs. doing nothing because the source and target have the same ISA (Wine, VM software for the most part). - APIs: HLE (emulate semantics) vs. LLE (emulate hardware). The more abstract the API, the better the former works, and vice versa. If you are not emulating things like interrupts and register pokes, that is HLE by definition. In some cases you can /directly/ translate source API usage into target API usage without any shims, but only for relatively simple APIs. For anything complicated, the semantics are likely different enough that transforming calls to it would require complicated global analysis, which would be quite pointless from a performance perspective unless the API is called ridiculously often.