Optimize truth value testing for strings

[pranavrajpal]

This causes code that tests for the boolean value of a string to test the length directly instead of going through PyObject_IsTrue, as mentioned in mypyc/mypyc#768.

This should make code that uses this faster, but my benchmarks seem to show that this makes code slower than it would be without this optimization. I was hoping you could take a look at my code and the benchmarks and see if I'm doing anything wrong.

Test Plan

I added a test that compiles and runs code that uses this, as well as an irbuild test, and ran all of the mypyc tests to make sure they were still working.

Benchmarks

This is my benchmark code:

import time

def run(s: str) -> None:
    t0 = time.time()
    for i in range(1_000_000):
        if s:
            pass
    t1 = time.time()
    print(f'{t1 - t0:.3f}')

run('')
run('A' * 100)

My results from running this benchmark several times each:

Interpreted version without mypyc: 0.022 to 0.024 seconds
Compiled version, from master (without these changes): 0.006 to 0.007 seconds
Compiled version, with these changes: 0.008 to 0.009 seconds

[JukkaL]

This is probably the bottleneck. This may be about as fast/slow as PyObject_IsTrue, so you aren't seeing an improvement. You'd need to specialize this operation for str values, and access the size of the string directly using C struct access. There are examples of how to do this for list and tuple values. Strings should be fairly similar (details may be different).

[pranavrajpal]

Ok, I have moved the check for whether a string's length is 0 to a primitive, and this appears to have made it faster.

Using the previously mentioned benchmark, I got these times:
Interpreted: 0.022 to 0.024 seconds
Compiled version, without these changes: 0.006 to 0.007 seconds
Compiled version, with these changes: 0.003 to 0.004 seconds

That appears to show that these changes make that benchmark approximately twice as fast.

[TH3CHARLie]

LGTM, good job!

[JukkaL]

Nice performance win! Thanks!