pystats: Improve with JIT builds

[mdboom]

Feature or enhancement

Proposal:

At the moment, not all pystats work for a JIT build. This is mostly because certain stat collections exist in the Tier 2 interpreter loop, not in the bytecode instructions themselves. The reason for having parity in the stats is not so much that we expect the numbers to be different vs. a Tier 2 run, but because it would be nice to add some JIT-specific stats (e.g. a code size histogram) and the only way to do that now would be to collect stats twice and combine them.

A list of what's missing:

• traces executed
• uops executed
• trace run length histogram
• uop execution stats

I'm volunteering to do the work, but maybe @brandtbucher could confirm this makes sense?

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-117121: Add pystats to JIT #117346

[mdboom]

Ok -- obviously I didn't understand how magical the JIT is, and I should have tried it first :) It seems that in fact these stats all work. Closing as "already done".

[mdboom]

Ok -- obviously I didn't understand how magical the JIT is, and I should have tried it first :) It seems that in fact these stats all work. Closing as "already done".

This, in fact, was not the case. In my local testing, I forgot to clear the /tmp/py_stats directory.

[brandtbucher]

I think the only tricky part will be run length, since that adds a new local variable... it'll probably need a workaround, unless we want jit_func to have a different signature on stats builds.

Everything else shouldn't be too difficult (but still seems like an interesting project for anyone taking it on).

[brandtbucher]

Leaving open in case someone wants to tackle trace run length histograms.