Make all operations leave a value on the stack for easier specialization.

[markshannon]

Most operation produce a value, and thus leave a value on the stack.
This has the nice property that the operation can be substituted with a call which has the same effect.
For example, we already specialize BINARY_SUBSCR with BINARY_SUBSCR_GETITEM for instances of classes that implement __getitem__ in Python. BINARY_SUBSCR_GETITEM pushes a frame and then enters the __getitem__ function without having to make any C calls.
We can do the same for LOAD_ATTR, and any calls (it makes sense to substitute one callable for another in some circumstances).

However, we cannot do easily do this for __setitem__ or __setattr__.
By changing STORE_SUBSCR and STORE_ATTR to leave a value on the stack, we can freely substitute them with a call.
Obviously changing STORE_SUBSCR to push None and then following it with POP_TOP is not very efficient if done naively.

However, it we guarantee that STORE_SUBSCR and STORE_ATTR are always followed by POP_TOP, we can simply change them to bump next_instr by one and skip the following instruction at no cost.

The question is: does the performance gain and simplification of the specialized forms outweigh the small cost of the extra 2 bytes per store?

I expect that it will, but we will need to experiment.

NOTE:
We can do the specialization without changing STORE_SUBSCR, but it requires the specialized form to push an extra frame just to execute an instruction to discard the top of the stack and pop the current frame: a RETURN_VOID instruction.

[gvanrossum]

Are you sure it's a perf gain?

[markshannon]

Am I sure? No:

I expect that it will, but we will need to experiment.