JIT/Tier 2 tracing: plans for function-entry hotness triggers (hot callbacks from C loops)?

[113xiaoji]

Title: JIT/Tier 2 tracing: plans for function-entry hotness triggers (hot callbacks from C loops)?

Question

When experimenting with the experimental JIT (PEP 744), I got the impression that trace construction / compilation is primarily triggered by loop/backedge warmup (e.g. JUMP_BACKWARD backoff counters) plus side-exit temperature mechanics, rather than by “hot function entry” call counts.

If that’s correct:

• Is this an intentional design choice for now?
• Are there plans (or prior discussions) to support function-entry hotness triggers too?

Why I’m asking (workloads impacted)

Some real workloads execute a small Python callback many times from C-level loops, so there may be no Python-level backedge to ever warm up. Examples include:

• sorted(..., key=callback) / list.sort(key=callback)
• regex callbacks (re.sub(..., repl=callable))
• some itertools / C extension APIs that repeatedly invoke a Python callable

In these cases, the callback function can be extremely hot, but a loop/backedge-only trigger would not start tracing/compilation for it.

(For reference, other runtimes/JITs (e.g. Cinder) can also trigger compilation from hot functions, not only hot loops.)

Minimal example (no Python loop)

def key(x):
    return (x * 3) ^ 0x1234

data = list(range(500_000))
sorted(data, key=key)   # key() is called many times from C

[Fidget-Spinner]

Please see #142982. We do plan to eventually trace from function entry.