Making the C stack invisible to bytecode by inserting shim frames

[markshannon]

Motivation

The RETURN_VALUE instruction needs check to whether the current frame is an "entry" frame (it was the frame that was passed to PyEval_EvalFrameDefault) and either do an "internal" return (pop the frame, and continue in the caller frame) or an "external" return (do a return in the C code).

This isn't a big deal deal for RETURN_VALUE, but could get cumbersome if we want to do the same for YIELD_VALUE, to allow
effective specialization of FOR_ITER for generators, and specialization of SEND for coroutines.

By inserting a shim frame on entry to PyEval_EvalFrameDefault and in gen.send we can shield RETURN_VALUE and YIELD_VALUE from knowing anything about the C stack.

Shim frame

A shim frame is a _PyInterpreterFrame inserted into the stack on entry to PyEval_EvalFrameDefault. The code object for a shim frame would consist of the single instruction EXIT, which would perform a C return.

Cost/benefit

Every call to PyEval_EvalFrameDefault would suffer additional cost of pushing a _PyInterpreterFrame, whereas every RETURN_VALUE would only save the cost of a single branch. Unless the number of RETURN_VALUEs is a lot more than the number of calls to PyEval_EvalFrameDefault then it clearly isn't worth doing at the moment.

However, once we want to specialize YIELD_VALUE, then this starts to make a lot more sense, as we can implement gen.send in bytecode that converts a yield into a return and wraps a return value in a StopIteration

def gen_send(coro, value):
      LOAD_FAST coro
      LOAD_FAST value
      SEND return_label
      RETURN_VALUE # yield becomes return
return_label:
     # Might need to explicitly clear the coroutine here.
     LOAD_EXCEPTION StopIteration
     PUSH_NULL
     SWAP 3
     CALL 1
     RAISE_VARARGS 1 # return becomes raise StopIteration

Specializations of FOR_ITER and SEND

If we add another "return address" to _PyInterpreterFrame (or just an offset from the current prev_instr) then YIELD_VALUE can jump to that, much as RETURN_VALUE jumps to prev_instr.

The specialization of FOR_ITER for generators would set this additional "return address", push the generator's frame and jump into that frame.
The specialization of SEND for coroutines would do much the same, as FOR_ITER is basically LOAD_CONST None; SEND.

[brandtbucher]

Are you planning on working on this?

If not, I’d like to take it. I’m going to be picking the yield from/await stuff back up today, and I also want to use these shim frames for inlining pure-Python binary operators, etc.

[markshannon]

Go ahead.

I'm working on related stuff, making the frame popping and pushing a bit slicker: #111 (comment), so there might be some merge conflicts.

python/cpython#92960 (done)
python/cpython#92957 (not urgent)