Provide instances for bitvec

[Bodigrim]

Closes #20.

Tests can be found in https://github.com/Bodigrim/hw-rankselect-base/tree/tests, but I do not submit it for merge because of orphan instances.

[Bodigrim]

How does it look, @newhoggy?

[newhoggy]

Sorry, I've been pre-occupied with other things. Looking at this now.

[newhoggy]

Do you by any chance know what this comment in the bitvec docs is referring to?

This module exposes an interface with thread-unsafe writes and flips. Consider using Data.Bit.ThreadSafe, which is thread-safe, but slower (up to 20%).

[Bodigrim]

Sure, most of the current incarnation of bitvec is written by me.

https://github.com/Bodigrim/bitvec/blob/master/README.md#thread-safety gives a bit more details:

• Data.Bit is faster, but writes and flips are thread-unsafe. This is because naive updates are not atomic: read the whole word from memory, then modify a bit, then write the whole word back.
• Data.Bit.ThreadSafe is slower (up to 20%), but writes and flips are thread-safe.

If multiple threads concurrently write neighbouring bits in the same word, they may end up with inconsistent states, effectively losing the work of one of the threads:

thread1: x <- arr[0]
thread2: y <- arr[0]
thread1: x |= 1<<16
thread2: y |= 1<<24
thread1: arr[0] <- x 
thread2: arr[0] <- y

In multithreaded enviroment we must ensure that threads access a mutable vector in turns:

thread1: x <- arr[0]
thread1: x |= 1<<16
thread1: arr[0] <- x 
thread2: y <- arr[0]
thread2: y |= 1<<24
thread2: arr[0] <- y

Data.Bit.ThreadSafe ensures this restriction, using atomic locks. Unfortunately, there is an associated performance cost, non-negligible for some applications. That is why bitvec provides both Data.Bit and Data.Bit.ThreadSafe.

[newhoggy]

Thanks for the explanation.

[newhoggy]

Would you be able to add property tests to verify that the indexing and endian-ness is the same as rank/select for say [Bool] or Data.Vector.Storable.Vector Word64?

[Bodigrim]

As mentioned above, I can add the tests (https://github.com/Bodigrim/hw-rankselect-base/commit/b0c6b5c15e0034b4bb9dfa7d83dba921592b7772), but they involve a (rather hastily written) orphan instance BitRead (DVU.Vector Bit.Bit). Is it OK?

Or I can take a long way and add these instances to hw-bits, if you are happy to accept such contributions.

[newhoggy]

Please add the instance to hw-bits. 😁

[Bodigrim]

I have added some tests, please re-review.

[newhoggy]

Nice! Thanks!