Improve performance

[lehins]

This PR fixes:

• Performance issue that I discovered when benchmarking generation of integral values in a range
• Order of arguments in uniformR function was inconsistent with other range functions.

[curiousleo]

* Performance issue that I discovered when benchmarking generation of integral values in a range

Nice! How big is the performance boost?

[lehins]

@curiousleo Pretty significant ~x20. Problem is when a function is recursive, then it cannot be inlined, which can hurt the performance really bad when that function is being called many many times.
In case of [un]signedBitmaskWithRejectionRM there was no need for them to be recursive.

Haven't tried the internal benchmarks, but running with criterion difference in performance was obvious.

[lehins]

@curiousleo Applied the same fix of removing recursion to Integer generation, got quite a bit of performance boost as well ;)

Benchmarks for ranges are available in the same place: https://github.com/lehins/haskell-benchmarks/tree/new-random/new-random-benchmarks

[curiousleo]

Fixed a semantic merge conflict: 068aa36#diff-5b60458897b8b814de6d596bdc71e75cL210-R210

[curiousleo]

According to the new benchmarks, this is an improvement for some types (e.g. Word64) and a performance regression for others (e.g. Word16).

                                         mean BEFORE    AFTER     error BEFORE     AFTER
pure/uniformR/full/Word16                mean 28.35 μs  17.96 ms  ( +- 1.383 μs  ) ( +- 1.369 ms  )
pure/uniformR/full/Word64                mean 1.552 ms  131.3 μs  ( +- 103.9 μs  ) ( +- 3.496 μs  )

Full comparison in compare.txt.

improve-performance-bb7b99a.txt
interface-to-performance-d03e325.txt

[curiousleo]

Amongst performance improvements, this also contains performance regressions: #103 (comment)

[curiousleo]

Ping @lehins - I've cherry-picked most performance improvements from here onto #116, which is now merged. I've played around with the optimisation you made to uniformIntegerM (making it non-recursive). This has strange side-effects: in my benchmarks, it makes other generators run slower. I won't pursue this particular optimisation further for now.

[lehins]

I am closing this, since most of improvements were cherry picked