Use native bytearray truncation

[lovelydinosaur]

Switching from .compress() to using native buffer truncation.

I figure despite some related work on #115 it's worth looking at this PR in isolation.

We probably want this to be blocked on #116.

Benchmarking on Python 3.7...

Before:

$ PYTHONPATH=. venv/bin/pythonbench/benchmarks/benchmarks.py6901.9requests/sec7055.0requests/sec7084.2requests/sec7070.6requests/sec7107.9requests/sec7075.8requests/sec7079.4requests/sec
$ PYTHONPATH=. venv/bin/pythonbench/benchmarks/benchmarks.py6972.9requests/sec6996.9requests/sec6947.3requests/sec6999.9requests/sec7040.6requests/sec7030.0requests/sec6997.6requests/sec

After:

$ PYTHONPATH=. venv/bin/pythonbench/benchmarks/benchmarks.py7054.6requests/sec7133.3requests/sec7148.3requests/sec7116.4requests/sec7121.0requests/sec7128.2requests/sec7156.5requests/sec
$ PYTHONPATH=. venv/bin/pythonbench/benchmarks/benchmarks.py7018.9requests/sec7074.7requests/sec7075.0requests/sec7070.6requests/sec7028.2requests/sec7118.3requests/sec7131.2requests/sec

[lovelydinosaur]

We also ought to follow up on the action in the comment "make sure PyPy also has the optimization", but I'm not sure where to start with that.

[bluetech]

I think it'd be helpful to keep the paragraph above.

[lovelydinosaur]

Fair point, yes. I've addressed it as an inline comment.

[bluetech]

I think a del would be a little clearer but maybe it's just me. (Also below).

[lovelydinosaur]

Agreed, yup!

[bluetech]

We also ought to follow up on the action in the comment "make sure PyPy also has the optimization", but I'm not sure where to start with that.

With this code:

size=50*2**20b=bytearray(size)
foriinrange(size):
delb[0]

Results:

python3.8: 4.257s
pypy3 version 7.3.2: 0.326s
pypy2 version 7.3.2: 0.326s
python2.7: DNF

So it's either that pypy has this optimization, or it's doing some magic JIT handling of this specific code. I tried some tricks to foil any such optimization and it seems legit.

If we're willing to have a timing-based unit test for this, a size = 3 * 2**20 with a timeout of 5s would do the trick IMO, even with a super fast CPU running the quadratic case or a very slow CPU running the linear case. But it will probably be flaky anyway...

[njsmith]

I think this has been superseded by #115

[pgjones]

Agree with #115 superseding this. (Please reopen if not).