This addresses point 3 of #640 (comment). Unfortunately, this saves only 48 bytes of original 480 (10%).

Additional commit to start addressing point 1.

I'll need to review this carefully.

Personally, I'd go for much lower hanging fruit to begin with:

1. Allow mp_call_function_n_kw to do a proper tail call by not having the check on res. There is only 1 case where res might be NULL (MP_OBJ_NULL actually), so check that explicitly in that case (and enforce the rule that ->call must never return MP_OBJ_NULL).
2. mp_execute_bytecode is only called in one place, so copy it into that function to eliminate the call.

I tested it with testsuite and asyncio_micro server under load, so I'm pretty sure the patch is correct, but it's indeed would be nice if you gave it independent review.

mp_execute_bytecode is only called in one place, so copy it into that function to eliminate the call.

Well, that's something what I have in mind, but I probably approach it from somewhat different side. So, what prompted me to work on this is not really a stack usage, but the fact that generators still don't support _args, *_kwargs, etc. And indeed, what should be gone is mp_execute_bytecode(), and flow should be: client code which wants to call a function allocates a state array (either on stack or heap), then fills in state array with args (via a function call), then calls what's currently mp_execute_bytecode2().

But that's not low-hanging fruit at all, but introduction of mp_code_state structure is a way towards that ;-).

Okay, this is very nice.

It even reduced code size! (On stmhal by about 80 bytes, and also unix port, but strangely bare-arm went up by 500 bytes...).