Semantics of randSetSeed and randIntU

[dlunde]

Currently, the intrinsics randSetSeed and randIntU have a somewhat peculiar semantics. Specifically, the first occurrence of randIntU randomly initializes the seed of the pseudo-random number generator (Random.self_init in boot), unless you call randSetSeed before this first occurence. This caused some unexpected behavior in combination with externalSetSeed that, in addition to setting the standard OCaml Random library seed, also initializes random number generators for Owl. For example, the call to randIntU below randomly initializes the seed even though we call externalSetSeed.

externalSetSeed 0 -- Calls Random.init 0 internally
let x = randIntU 0 10 in -- Calls Random.self_init internally
...

I fixed this for my use cases by constructing a function setSeed that internally calls both externalSetSeed and randSetSeed:

setSeed 0 -- Calls externalSetSeed and randSetSeed internally
let x = randIntU 0 10 in -- Does _not_ call Random.self_init internally
...

However, I think that we should change the current randSetSeed and randIntU semantics to one of the following:

1. Simply do not call Random.self_init in randIntU. The result is that the seed is set to a standard value in each execution. If we want random seed initialization, we can add another intrinsic randSeedSelfInit or similar.
2. Call Random.self_init once, in the very beginning of the program, and not in conjunction with randIntU.

[lingmar]

Maybe we could remove randSetSeed and randIntU as intrinsics since the same functionality is implemented in externals?

[dlunde]

Yes, that would probably be the best long-term solution. I think there are some limitations with externals currently? For example, they don't work in boot (but maybe that's fine).

[lingmar]

It might be worth checking if the uses of randIntU in the stdlib can be replaced by externals, then. If they're not part of the first pass of bootstrapping it should work, right?