Specialize BINARY_OP by refcount and by type of operands

[markshannon]

Instead of the fairly limited ad-hoc specialization we currently use, we should specialize BINARY_OP by refcount and use table lookup to specialize by type.

Doing so would have the following advantages:

• Increased specialization from ~70% to over 95%
• Potential for large performance improvement in tier 2
• Makes it easy to incorporate BINARY_SUBSCR into BINARY_OP freeing up several tier 1 opcodes
• Can use similar techniques for CONTAINS_OP and COMPARE_OP to free up more tier 1 opcodes

I expect the overall performance impact on tier 1 will be negligible. Maybe a little bit faster, maybe a little bit slower.
The improved specialized will (I expect) largely cancel out the degraded performance for the most common operations.

See for faster-cpython/ideas#662 detailed analysis and faster-cpython/ideas#660 for the related tier 2 optimization

Linked PRs

• GH-117581: Specialize binary operators by refcount as well as type. #117627

[encukou]

The PR was abandoned in favour of #128722. Should this be closed?