Consider making the window size non-configurable.

[zenhack]

Right now the window size W is part of the configuration. I think this should actually just be a constant, probably 64. Rationale:

• This is what some of the Go packages already do.
• IIRC, the design of go4.org/rollsum requires W to be a power of 2 for correctness, and exploits this for some efficiency gains.
• For languages that don't support const generics, there is no way to abstract this out in a zero-cost way.
• I can't think of strong use cases for different window sizes
• All else being equal, the less variability allowed by the spec the better; the point is to standardize after all.

Does anyone see any reason not do do this?

//cc @bobg, @cole-miller

[cole-miller]

The only thing I can think of is that if you want to use vector instructions to speed up hashing (which is something I've been thinking about for RRS in particular) then it's necessary helpful for the window size to be a multiple of the SIMD register width. I think the widest registers available on x86 or x86_64 come in at 64 bytes (AVX-512), and the other widths are smaller powers of 2, so fixing the window size at 64 bytes would work well here.

[zenhack]

I went ahead and did this, closed.