Depend on splitmix package or copy generator code?

[curiousleo]

Context

Currently there is a dependency on the splitmix package:

random/random.cabal

Line 38 in a5c70bd

splitmix -any

By default, splitmix depends on random. This dependency can be severed by setting a flag. That is currently necessary, otherwise cabal and stack fail to build random due to a dependency circle:

random/cabal.project

Line 2 in a5c70bd

constraints: splitmix -random

random/stack.yaml

Lines 42 to 44 in a5c70bd

	flags:
	splitmix:
	random: false

Alternative

The alternative to depending on the splitmix package is to inline the SplitMix implementation.

Discussion

Pull in splitmix dependency pros:

• Whenever improvements are made to splitmix, random can easily benefit from them.

Pull in splitmix dependency cons:

• Setting the flag called random to False is required; without it there will be a circular dependency. Flags are supported by cabal and stack, but it looks like e.g. cabal2nix does not support them.
  Q: would users need to set this flag too in order to use random?

Copy algorithm pros:

• No dependency on splitmix and thus no need to set any flags.
• Can reduce the dependency tree overall: some of splitmix' dependencies are not necessary for our purposes and can be removed from the dependency tree.

Copy algorithm cons:

• Code duplication.
• random does not automatically benefit from updates to splitmix. However, note that the algorithm is now stable - the last change to the PRNG itself occurred in May 2018: Bug in splitmix haskellari/splitmix#9.

[lehins]

I opened an issue on splitmix about it: haskellari/splitmix#34 I don't see a good reason to inline the splitmix code into random, except only to break the cycle, so in the end it will be @phadej decision on which way the dependency arrow points.

[lehins]

Great, Oleg is ok with dropping dependency on random, so there is no need to duplicate the full implementation. See comment haskellari/splitmix#34 (comment)