GeezerVison

@pablogsal and @nedbat

This PR improves coverage performance on 3.11 by 5-7%. Using Ned's benchmarks

# 3.11 branch at 1497d7fdefff8207b8ccde82e96b6f471416c284
Median for bm_sudoku.py, python3.11, cov=none: 10.476s
Median for bm_sudoku.py, python3.11, cov=6.4.1: 69.124s
Median for bm_spectral_norm.py, python3.11, cov=none: 9.072s
Median for bm_spectral_norm.py, python3.11, cov=6.4.1: 72.143s

# 3.11 co_code_cached
Median for bm_sudoku.py, python3.11, cov=none: 10.363s
Median for bm_sudoku.py, python3.11, cov=6.4.1: 64.726s
Median for bm_spectral_norm.py, python3.11, cov=none: 9.325s
Median for bm_spectral_norm.py, python3.11, cov=6.4.1: 69.713s

An observation: bm_spectral_norm improved less than bm_sudoku because the size of its co_code is smaller. So the cost of getting a new one every time isn't as high. I think real-world code sizes are more likely to be like bm_sudoku or larger. So I'd guestimate we will see ~10% improvement in real-world code running coverage in 3.11.

Unfortunately this means that whatever is making Python 3.11 slower with coverage, unfortunately is not only this :(

Great investigation @Fidget-Spinner and thanks for working on this ♥️

I will try to review ASAP but it would be great if @markshannon @iritkatriel @brandtbucher or @ericsnowcurrently can take a look.

I also ran the benchmarks using #93493:

It's a slight improvement, but isn't solving the problem.

It's a slight improvement, but isn't solving the problem.

Yeah I'm aware. I sent this PR in because 10% improvement on macrobenchmarks is still something! Even cProfile slowed down by 60% when profiling code in 3.11. My hunch is that accessing the full PyFrameObject is signifcantly more expensive now in 3.11. However, I can't fix that because it's part of the tracing Py_tracefunc C API.

I also ran the benchmarks using #93493:

It's a slight improvement, but isn't solving the problem.

Hm, it looks like in some cases this actually makes things a bit slower (perhaps due to the extra memory consumption)?

Maybe we should pause this PR until @markshannon has finished reworking the line number calculations, which seem to be the bulk of the issue at this point. Or at least see if it slows down pyperformance at all before merging?

I also ran the benchmarks using #93493:

It's a slight improvement, but isn't solving the problem.

Hm, it looks like in some cases this actually makes things a bit slower (perhaps due to the extra memory consumption)?

Maybe we should pause this PR until @markshannon has finished reworking the line number calculations, which seem to be the bulk of the issue at this point. Or at least see if it slows down pyperformance at all before merging?

When I benchmarked on the main/3.12 branch, it didn't make anything slower #93383 (comment). However, I agree on waiting for a while.

When benchmarking with Ned's benchmarks, I saw a slowdown in bm_sudoku but no slowdown in bm_spectral_norm. I'm not sure what's up with that.

When you're done making the requested changes, leave the comment: I have made the requested changes; please review again.

I have made the requested changes; please review again Mark.

Thanks for making the requested changes!

@markshannon: please review the changes made to this pull request.