Add basic binary operators (and, or, xor, not) to IO::Buffer.

[ioquatix]

This introduces both the operators (&, |, ^ and ~) as well as in-place methods: and!, or!, xor!, not!.

string = 'Hello World'
buffer = IO::Buffer.for(string)
mask = IO::Buffer.for('abcd')

pp buffer.xor!(mask).xor!(mask)

[eregon]

% mask_size is likely quite slower than a bit mask (if the size is power of two), or condition with a separate mask_index resetting to 0 if >= mask_size.

[ioquatix]

I checked the assembly and it looks to generate fairly efficient SSE/AVX instructions, so I trust the compiler is doing the right things here. In any case, simple and correct is a good first implementation, and later on we can unroll it by hand if performance is an issue.

[eregon]

It's not about unrolling. It's about a modulo operation is as expensive as division (many cycles), and a lot more than + - * & ! ^.

[ioquatix]

I played around with this: https://godbolt.org/z/7MEfsrz9z - change -Os to -O3 and you see the compiler will unroll the loop and load the data in chunks. My understanding is that by unrolling the loop carefully, some division operations can be skipped.

If you know how we can count the number of divs being performed, that would be awesome, otherwise, I'm fine with slow but correct to start with, and figuring out a faster implementation later as an improvement. If you think adding a branch to the inner loop will ultimately be faster, let's do that? But at that point I think we need to benchmark it. It's entirely possible the compiler would do a worse job at unrolling a loop with an if statement.

[eregon]

I'm fairly confident it does not remove the division operations, it could only do that if it did the branch I suggest, and there are plenty of divl and divq left in that assembly.
Maybe it does a few less, but it could only do that without branches if it can make sure the mask_size is e.g. > 1, and it cannot know that.

I agree, we should benchmark it, then it should be clear. With the explicit condition there would be 0 division left, but the extra branch. It should still unroll fine.
The best would be if we could have the mask_size as a constant in the source, this should be possible by doing an early check and call with a constant for the mask_size we care about + force inline this function.