Who, doesn't this imply that all the OPCODE_HAS_<FLAG> macros are wrong? Just fixing those two occurrences doesn't seem the right fix.

Oh wait, it's worse. The macros apply to Tier 1 opcodes, but were incorrectly applied to Tier 2 here. I'll approve. But we need to think more about how to avoid such mistakes.

But we need to think more about how to avoid such mistakes.

Agreed. Probably the best thing to do is define equivalent macros for uops, and assert the valid ranges for each (although this is tricky, since tier one is in range(0, 256) and tier two is in range(0, MAX_UOP_ID)... so they overlap).

If there's no major convenience added by allowing tier one and tier two to overlap like that, maybe they should each get their own distinct range? Then it would be easier to catch mistakes like this.

Agreed. Probably the best thing to do is define equivalent macros for uops, and assert the valid ranges for each (although this is tricky, since tier one is in range(0, 256) and tier two is in range(0, MAX_UOP_ID)... so they overlap).

If there's no major convenience added by allowing tier one and tier two to overlap like that, maybe they should each get their own distinct range? Then it would be easier to catch mistakes like this.

Actually, currently uops are in the range [300, MAX_UOP_ID] while bytecode ops are in [0, 255]. So they are distinct and we can add range checks to the macros.

However, I believe @markshannon is planning to make the ranges overlapping, so we can store uops in a single byte. The uop_id_generator.py script already has a -n (--namespace) flag to do so.

(The lowest uop is 300 instead of 256 because we need some space for pseudo ops etc.)

I'm seeing numbers that are below 256 in tier two for certain instructions. For example, _IS_OP (tier two) is defined as IS_OP (tier one) which is defined as 76.

I think it's just for instructions that translate 1-1 between tiers? Check out Include/internal/pycore_uop_ids.h.

I think it's just for instructions that translate 1-1 between tiers? Check out Include/internal/pycore_uop_ids.h.

Oh, the work isn't as far along as I thought. Shame on me for not remembering what I saw in that PR. It just introduces separate names, but the numeric IDs aren't separately allocated yet.