new `TestUopsOptimization` test case failing an assert about cpython

[gvanrossum]

Thanks for finding this. It's this assert:

assert((this_instr + 2)->opcode == _PUSH_FRAME);

in _INIT_CALL_PY_EXACT_ARGS in optimizer_bytecodes.c. I added that a few days ago. I'll come up with a fix.

[Fidget-Spinner]

@gvanrossum I dug into this further and it's actually two bugs:

1. The assert in the optimizer is wrong, it could end with _EXIT_TRACE, but the second bug in the tracer itself is more nefarious and actually crashes uops:
2. With how we create traces for stuff that goes too deep, it seems it's not side exiting to the correct place.

[gvanrossum]

Oh, thanks! If you have time for a fix I can review it.

[gvanrossum]

Maybe it's better to back out the entire sequence of uops generated by the CALL_* instruction containing the push?

[Fidget-Spinner]

Maybe it's better to back out the entire sequence of uops generated by the CALL_* instruction containing the push?

Yeah that seems better. Right now it's deopting after _INIT_CALL_PY_EXACT_ARGS which creates a new frame, this means the stack is inconsistent with CALL_PY_EXACT_ARGS and the whole thing crashes.

[gvanrossum]

In the end I think it's better to emit the _PUSH_TRACE (it was being emitted until I refactored that code a few days ago).

[lazorchakp]

Thanks for getting to this so quickly! This fix should also simplify an edge case for #116168 too, which is helpful for me.