fast methods for any and all for Bool tuples

[matthias314]

The new methods allow for vectorization and are therefore much faster: With nightly, I get

julia> t = ntuple(Returns(true), 32); @b any($t)
2.440 ns

julia> t = ntuple(>(4), 32); @b any($t)
5.014 ns

julia> t = ntuple(Returns(false), 32); @b any($t)
22.961 ns

With this PR,

julia> t = ntuple(Returns(false), 32); @b any($t)
2.638 ns

As for other Tuple methods, this is only done up to length 32. Beyond that, the generic method is called. At present, the cut-off is 3 elements.

I find this quite useful. What you think?

[matthias314]

It seems that some people like this proposal. What about the maintainers?

[nsajko]

Given that the tuple is homogeneous, there's no need for recursion, we can handle inputs of all sizes with a loop. This results in the same LLVM code for a length-32 tuple as with your variant with reduce:

Suggested change

	all(x::NTuple{N,Bool}) where N = N <= 32 ? reduce(&, x; init = true) : _all(x, :)
	any(x::NTuple{N,Bool}) where N = N <= 32 ? reduce(|, x; init = false) : _any(x, :)
	function _all_bool(x::Tuple{Vararg{Bool}})
	@_terminates_locally_meta
	r = true
	for i ∈ eachindex(x)
	r &= getfield(x, i, false) # avoid `getindex` bounds checking to help vectorization
	end
	r
	end
	function _any_bool(x::Tuple{Vararg{Bool}})
	@_terminates_locally_meta
	r = false
	for i ∈ eachindex(x)
	r |= getfield(x, i, false) # avoid `getindex` bounds checking to help vectorization
	end
	r
	end
	any(x::Tuple{Vararg{Bool}}) = _any_bool(x)
	all(x::Tuple{Vararg{Bool}}) = _all_bool(x)
	all(x::Tuple{Bool}) = _all_bool(x) # disambiguate

[nsajko]

There's also a merge conflict now.

[nsajko]

Could also be a bit more general by dispatching on typeof(identity). Then the single-argument methods wouldn't be necessary.

[nsajko]

Here's a more comprehensive and general PR: #57091, @matthias314 hope you don't mind.