Seemingly slow capturing of int in closures

[stinos]

Originally reported here: https://forum.micropython.org/viewtopic.php?f=2&t=12060; 'seemingly' because I'm not sure what exactly is slow. Reproduce with:

import time

def ret(arg):
    yield None

def test(klass=int, count=10000):
    for n in range(count):
        ret(klass())

def measure(what, name):
    start = time.ticks_ms()
    for _ in range(10):
        what()
    end = time.ticks_ms()
    print(name, time.ticks_diff(end, start))

for typ in [int, float, dict, list]:
    measure(lambda: test(typ), typ)

Which yields numbers like this on windows and unix ports and reportedly does the same on RP2/pico:

<class 'int'> 1090
<class 'float'> 18
<class 'dict'> 14
<class 'list'> 19

Given the extreme difference this seems to indicate an underlying problem.

[yjchun]

Please note that async function is a generator function too. Replace ret() with async_ret() and will see identical results.

async def async_ret(arg):
    pass

This problem is actually more relevant in asyncio.

[dpgeorge]

It looks like this has to do with the GC doing excessive work to find some memory to create the generator (calling ret). If klass() allocates memory then it's fast. If it doesn't allocate memory then it's slow.