I don't agree that nonzero is clearer on boolean inputs. where is also used in e.g. IDL.
where_nonzero would be better but I don't think another alias is a good idea either.

My argument would be that where is ambiguous, as you have to count the number of arguments to know what it does.

On the whole, it seems pretty unpythonic to have two basically unrelated functions share the same name - the only thing in common between where(c, x, y) and nonzero(c) is that the cast their first argument to bool. The only exception to this rule in the standard library that I can think of is type(var) vs type(name, bases, dict)

Obviously it's far too late to change our public API there, but I think we should at least discourage the use of np.where(c) and avoid it internally

Yeah, nonzero isn't the best name and where_nonzero would probably be better, but it is entrenched. The dual use of where always seemed weird to me.

Thanks Eric.

Alternatively, perhaps where(c, x, y) has the wrong name. If we want to implement #8994 without dropping support for where(c), then we need a new name for that anyway, so it can be used by __array_ufunc__ (although I guess np.where.ufunc could be that name)