fix: Remove unnecessary string casting of bytes - #276

To call out the explicit: is this a behaviour change or a performance improvement?

If its a performance related change, I would like to see a benchmark showing "time spent" or similar metrics that changed.

Great call! Thanks to your prompt I actually timed the new version and improved it even further (code will be pushed when you are reading this).

So this should keep the same behavior in terms of types. The caveat is, the original version always returns a new bytes object due to the bytes -> str -> bytes conversion whereas the new version simply returns the same object if it is already bytes. This should be okay, even good as the intent was never to get a new copy but worth calling out. Otherwise this is purely a performance optimization with the following numbers where _to_bytes is the old version and _to_bytes_new is the new version proposed in this PR:

>>> s
'someheader'
>>> b
b'someheader'
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.22222996300115483
0.16239697000128217
0.18601619399851188
0.1108450250030728
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.20466226000280585
0.16354602499632165
0.1937305280007422
0.12870584500342375
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.20280992899643024
0.16408630299702054
0.1972411400056444
0.12873706800019136
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.21261542499996722
0.16578178100462537
0.18726225899445126
0.15498205900075845

As you can see it is consistently faster than the old version for both str and bytes inputs and significantly faster for bytes inputs which is what we should be dealing with after python-hyper/h2#1286

These numbers are using Python 3.10.12. When using Python 3.12.4, they are even better:

>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.13931607300037285
0.1050484180013882
0.11498607300018193
0.07148081599734724
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.14760045900038676
0.0974436950054951
0.11569900900212815
0.06948418200045126
>>> timeit(lambda: _to_bytes(s));timeit(lambda: _to_bytes(b));timeit(lambda: _to_bytes_new(s));timeit(lambda: _to_bytes_new(b))
0.13979562999884365
0.09725762299785856
0.11685396800021408
0.075411345998873

@BYK I would also be happy to hear your thoughts about #277 which I just sent - and if it changes anything or points out an previously hidden issue!

@BYK I would also be happy to hear your thoughts about #277 which I just sent - and if it changes anything or points out an previously hidden issue!

Oooh nice! I can rebase my patch on top of that if you wanna go ahead and merge it before. I'll give that a proper look this weekend.