Streamline the `MAKE_WARM` micro-op

[markshannon]

The MAKE_WARM micro-op is inserted into every piece of jitted code and within the loop if there is one.
So it needs to be cheap.

Currently, it does this:

    current_executor->vm_data.warm = true;
    if (--tstate->interp->trace_run_counter == 0) {
        _Py_set_eval_breaker_bit(tstate, _PY_EVAL_JIT_INVALIDATE_COLD_BIT);
    }

we should move the logic for testing for "coldness" onto a less frequently taken path, perhaps in the jit itself, so _MAKE_WARM just needs to do the "warming":

    current_executor->vm_data.warm = true;

Linked PRs

• gh-138050: JIT - Streamline MAKE_WARM - move coldness check to executor creation #138240
• gh-138050: move cold invalidation to Tools/jit/trampoline.c #139904
• gh-138050: Global cooling: use cold flag instead of warm flag in MAKE_WARM #143827

[albertedwardson]

Hi! I’m not claiming this issue- if someone else wants to take it, I’ll step aside, since I don’t have much practice with JIT. But if no one does, I’d like to dig into it and give it a try. I find this topic very interesting, and it would be a good opportunity to get some practice and understand it better

[markshannon]

From a question on the PR:

This is worth doing because:

• MAKE_WARM is 2.6% of all uops executed. https://github.com/faster-cpython/benchmarking-public/blob/main/results/bm-20250719-3.15.0a0-800d37f-PYTHON_UOPS/bm-20250719-azure-x86_64-python-800d37feca2e0ea33439-3.15.0a0-800d37f-pystats.md#uop-execution-stats. Any reduction in its size and number of instructions executed is significant.
• Moving the check from a hot path to a cold path allows us to make the check more complex. This would allow us, for example, to dynamically change the rate at which we check for coldness as we approach some memory consumption limit.
• In general, we always want to perform work on cold paths, not hot paths.

[alonme]

Can this issue be closed now that #138240 has been merged?

[markshannon]

There is one further, small, improvement that can be made.

Operations involving 0 are generally faster than those involving 1, because of how the hardware works.
So we could flip the meaning of the byte from "warm" to "cold", setting it to 0/false on the hot path (MAKE_WARM) and setting to 1/true on the cold path (during executor creation).