Code generator: support variable stack effects in macros

[gvanrossum]

Currently you can't split e.g.

inst(CALL_PY_EXACT_ARGS, (unused/1, func_version/2, method, callable, args[oparg] -- unused))
{ ... }

into two micro-ops

op(_CHECK_PY_FUNCTION, (func_version/2, method, callable, args[oparg] -- method, callable, args[oparg]))
{ ... }
op(_CALL_PY_EXACT_ARGS, (method, callable, args[oparg] -- unused))
{ ... }
macro(CALL_PY_EXACT_ARGS) = unused/1 + _CHECK_PY_FUNCTION + _CALL_PY_EXACT_ARGS;

We need at least to handle micro-ops that have the same variable-length input effect as output effect, and a final (action) op that has a variable-length stack input effect and a fixed output effect. This should handle all the CALL specializations. (For UNPACK_SEQUENCE we also need to support a fixed input effect and a variable output effect, in he action op.)

Linked PRs

• gh-106812: Refactor to allow uops with array stack effects #107564
• gh-106812: Fix two tiny bugs in analysis.py #107649
• gh-106812: Small stack effect fixes #107759

[gvanrossum]

This is turning into a bit of a research project. I would like to refactor all the handling of stack effects in the generator.

I'm thinking of having an algorithm that takes a sequence of operations such as pop/push, peek/poke, shrink/grow, and copy, and which applies certain transformations (e.g. push+pop can be replaced by copy under certain conditions), trying to reach an optimal end state, which can then be turned into optimal C code.

But the right set of operations, transformations and conditions is still elusive.