two: Avoid branch in computation of idx2

[sjakobi]

No description provided.

[sjakobi]

I haven't been able to find any changes in the benchmarks so far. The generated code looks better though.

[treeowl]

I'm a bit surprised GHC doesn't undo this optimization. If the generated code looks better, go for it.

[sjakobi]

I'm a bit surprised GHC doesn't undo this optimization.

Yeah, GHC is pretty eager to generate a branch here. E.g. idx2 = fromEnum (... < ...) would still generate the branch.

[treeowl]

We have

index :: Hash -> Shift -> Int
index w s = fromIntegral $ unsafeShiftR w s .&. subkeyMask

So I think another option would be to shift the subKeyMask left and use the result for both hashes. I dunno if that would be better, worse, or exactly the same.

[sjakobi]

Yes, there are more low-level tweaks that can be applied here. #468 (comment) is related.