gh-143183: Link trace to side exits, rather than stop

[Fidget-Spinner]

#143187 caused a pretty serious regression in JIT performance, mainly because we stop the trace prior to the ENTER_EXECUTOR. This PR restores the old behavior, stopping the trace at the ENTER_EXECUTOR, allowing it to link up to the executor.

The perf regression can be clearly seen here (going up means things got slower):

• Issue: JIT: Assertion failure in stopiteration_error #143183

[Fidget-Spinner]

Before:

After (notice how more things are linked):

[diegorusso]

I see that #143187 add a test. Can we have it here as well?

[Fidget-Spinner]

I fixed the test in the test suite so that it catches it. I verified that without the change, the test fails. With the change, the test passes.

[Fidget-Spinner]

Also verified that:

At commit 713684d (before my previous PR), this test passes.
At commit daa9aa4 (the first commit with my PR), this test fails.
At commit a9548b8 (current PR commit), this test passes.

So this does indeed restore the old behavior!

[Fidget-Spinner]

For the following code:

TIER2_THRESHOLD = 4001
def f():
    for x in range(TIER2_THRESHOLD + 3):
        for y in range(TIER2_THRESHOLD + 3):
            z = x + y

import sys
f()
sys._dump_tracelets("hello.gvz")

Current main (buggy):

This PR (notice how the traces now link up):

[diegorusso]

Thanks for fixing the regression and adding the test for it!

[tomasr8]

Damn you're too fast.. was about to hit approve when you merged it 😆

[tomasr8]

In the other PR, I meant to update this comment but GH shows the suggestions in a weird a way for unchanges lines.. The point is the comment mentions _EXIT_TRACE but it should be _DEOPT now :)

// Rewind to previous instruction and replace with _EXIT_TRACE.

[Fidget-Spinner]

Oh yeah. Good point!