Add Base.allunique

[tpapp]

Allows testing if elements in a collection are distinct (when compared
with isequal). Terminates early in case of repeated elements, has
methods for collections which are by construction distinct. Also added
tests.

See discussion at issue #15803, which this PR fixes.

[hayd]

This should probably be:

(step(r) != 0) || (length(r) <= 1)

or potentially:

alldistinct{T}(r::Range{T}) = (step(r) !== zero(T)) || (length(r) <= one(T))

and add tests for negative step and zero length ranges.

[tpapp]

Thanks. Made the change, with tests.

[tkelman]

needs to be added to rst docs too

[tpapp]

Done, thanks for the reminder.

[tpapp]

Is there anything preventing a merge? I understand the core devs are busy so it might take time, but in case I missed something that prevents a merge that I can fix please let me know.

[tkelman]

I think this is good. If no one has any more feedback, I'll merge soon.

Note for next time, it can help in keeping things organized if you use a new branch (other than master of your fork) for making each pull request.

[tkelman]

hm, spoke too soon maybe - technically this isn't testing using isequal until #9381 gets resolved, right?

[nalimilan]

?in says that for sets, isequal is used, so that looks OK.

[stevengj]

s/Test if/Return true if/

No comma before "when".

[tpapp]

@stevengj: thanks, pushed minor change correcting that.
@tkelman: Sorry that I did not branch, I am still learning the ropes, will do it next time.

I hope that this can be merged now, but let me know if I should do anything else.

[nalimilan]

Thanks. Can you squash the commits?

[KristofferC]

Can't we start using githubs own squasher?

[stevengj]

Yes, PRs no longer have to squash commits manually, thank goodness.

[StefanKarpinski]

AppVeyor seems to be stuck, but it got to SUCCESS, so I'm calling it a pass.

[tkelman]

when did this get renamed to allunique? I thought alldistinct was clearer

[StefanKarpinski]

There was a whole conversation about it.

[DNF2]

There's a problem here which also popped up in #15009 , namely that you can have 'oversampled' ranges of floats:

>> r = linspace(1.0, nextfloat(1.0), 10)
>> allunique(r)
true
>> allunique(collect(r))
false

As suggested in #15009 (comment) you can prevent this with the following check:

last(r) >= nextfloat(first(r), length(r) - 1)

One more thing: This comparison

length(r) <= one(T)

seems strange to me. length always returns an Int, right? So it should be length(r) <= 1.

(I tried all this on 0.4.5, so apologies if linspace or length have somehow changed in 0.5.)

Oh, also, isn't this a bit pointless: step(r) != zero(T) instead of step(r) != 0?
The compiler figures it out and there is no performance difference. Aren't the zero(T) and one(T) constructs mainly for avoiding type instabilities? No harm, though, I guess.