GH-111485: Remove some special cases from the code generator and bytecodes.c

[markshannon]

• Removes a bunch of special cases for tier-1-only instruction detection, making them explicit in bytecodes.c with TIER_ONE_ONLY.
• Don't emit or check formats for psuedo-instructions. They don't exist in the bytecode, so don't have a format.
• Make the counter in JUMP_BACKWARD explicit
• Fix error handling in MAKE_CELL so it can be executed in tier 2.

• Issue: The code generated for the interpreter(s) is too fragile #111485

[brandtbucher]

@gvanrossum, do we know why we were special-casing BEFORE_WITH like this?

I ask because it looks like this PR made BEFORE_ASYNC_WITH (which also uses _PyObject_CallNoArgs) a viable uop. It seems fine, but maybe I'm missing something.

[gvanrossum]

@gvanrossum, do we know why we were special-casing BEFORE_WITH like this?

I ask because it looks like this PR made BEFORE_ASYNC_WITH (which also uses _PyObject_CallNoArgs) a viable uop. It seems fine, but maybe I'm missing something.

Urgh, I have no idea. :-( Agreed that both seem fine. Either both should have the TIER_ONE_ONLY marker, or neither should. Let me dig through some history:

It's been like this since the cases_generator code first landed (#105924). And even then, commenting out that line makes BEFORE_*WITH appear in the output and it compiles just fine. In the PR I found this commit though:
e3471da

The commit message gives a clue:

Fix test_memoryview by excluding BEFORE_WITH

Unclear what's wrong with that opcode."

Commenting out the line, regenerating, and then running ./python.exe -Xuops -m test test_memoryview gives this error:

test test_memoryview failed -- Traceback (most recent call last):
  File "/Users/guido/cpython/Lib/test/test_memoryview.py", line 288, in test_contextmanager
    self._check_released(m, tp)
  File "/Users/guido/cpython/Lib/test/test_memoryview.py", line 273, in _check_released
    with m:
         ^
ValueError: operation forbidden on released memoryview object

Doing the same on main, however, causes no crashes. Quite possibly whatever was defective in the original code generator and Tier 2 interpreter has since been fixed, and the TIER_ONE_ONLY marker can safely be removed from BEFORE_WITH. I don't feel like doing a bisection to figure out the exact commit that made the test_memoryview failure go away.