Refactor MLton.Parallel

[ramLlama]

MLton.Parallel is currently a "bag of stuff". It's worth organizing this. I'll go section by section

Reorganization

I propose the following:

• MLton.Parallel: contains numberOfProcessors, processorNumber, registerProcessorFunction, and initializeProcessors
• MLton.Parallel.Atomics: Contains atomics such as CAS and fetch-and-add
• MLton.Parallel.Synchronization: Contains locks. Perhaps should have more substructures such as Spinlock, etc. (although we only have a spinlock for now)

Deprecated

We have a few things in MLton.Parallel.Deprecated that have been there a while. We should handle them.

My thoughts:

• yield: If no one is using this, let us remove it entirely.
• lockInit, takeLock, releaseLock: Move this out of Deprecated. Move into MLton.Parallel.Synchronization, proposed previously

Unsafe

• arrayUninit: This aren't parallel-only and shouldn't be in MLton.Parallel.Unsafe. Perhaps in a new MLton.Unsafe?
• arrayCompareAndSwap, arrayFetchAndAdd: Should be in MLton.Parallel.Atomics, proposed previously.

Top-level

• The CAS operations should perhaps go into a MLton.Parallel.Atomics structure, proposed previously. Perhaps, we should even have a higher level MLton.Atomics structure. In addition, we should expand this to support all primitive types.

[MatthewFluet]

There is already an Unsafe module (basis-library/sml-nj/unsafe.{sig,sml}), although within the Basis Library implementation (i.e., before we constrain the Array module to that specified by the Basis Library), you have Array.arrayUninit available.

[ramLlama]

Isn't that Unsafe a compatibility structure with SML/NJ Unsafe? What has been mainline MLton's thoughts on where to put MLton-specific unsafe things?

Also, yes, we would use the same Array.arrayUninit that is provided in the basis, not a reimplementation. We just a need a place to put it.

[MatthewFluet]

Yes, the exported structure Unsafe is for compatibility with SML/NJ. We haven't needed to export MLton-specific unsafe things. Within the Basis Library implementation, we simply use *Unsafe functions within their natural structures. So, it is really a question of whether these operations are meant for export to user programs or just for internal implementation.