Can you tell me more about the theory of this fix? I was surprised to see UNWIND events handled, since this problem didn't involve exceptions. Indeed, deleting that addition didn't change the outcome. We don't need the UNWIND handling.

What does last_lines do for us here and how does it fix the measurement of with statements?

Can you tell me more about the theory of this fix? I was surprised to see UNWIND events handled, since this problem didn't involve exceptions. Indeed, deleting that addition didn't change the outcome. We don't need the UNWIND handling.

I can back that out, however my understanding is if there's an observed exception during the test suite run, memory will bloat since last_lines won't be "emptied" as a result.

What does last_lines do for us here and how does it fix the measurement of with statements?

It maps each active function call (by frame id) to the last line we saw in it.
The same function can be called multiple times, and each call needs its own tracking.

Why do we need to track stuff? When we see a LINE event, remember it. When we see the next LINE event, record the transition.

An example:

LINE event: line 5 → remember "last line was 5"
LINE event: line 6 → record arc (5→6), remember "last line was 6"
LINE event: line 7 → record arc (6→7), remember "last line was 7"

Previously, sysmon would say "I've seen line 5, don't tell me about it again." Now we need to hear about every line execution to track transitions.

Oh, I hadn't noticed this yet:

        # Don't return DISABLE: keep getting LINE events for arc tracking.
        return None

This undoes the entire benefit of sys.monitoring. If we're going to change that we might as well use the C tracer.

Oh, I hadn't noticed this yet:

        # Don't return DISABLE: keep getting LINE events for arc tracking.
        return None

This undoes the entire benefit of sys.monitoring. If we're going to change that we might as well use the C tracer.

Very good point - shows how much I know about sysmon! 🤦

I took another pass at this, and added a helper method that checks the bytecode and makes a decision based on that, so the sysmon performance would be preserved for most code except those using these deeply-nested context managers.

Let me know what you think!