This is where things start to get complicated. Different errors have a different number of words for their fault. In the worst case this can be 5 words, so all 4 message registers on arm need to be stored. Without some detailed information to the syscall stub generator of exactly which invocations can cause which errors, you're looking at just always stashing all 4 MRs back into the IPC buffer. Could probably optimize this by checking the label before return and storing the MRs only if an error happened.

I actually cannot think of a compelling reason for why that wouldn't work.... Either there is something I've forgotten, or the solution really is that simple.