Speed up map on tuples. Fixes #15695

[shashi]

No description provided.

[timholy]

While it will be nice for big tuples, this seems likely to hammer performance on small tuples.

[shashi]

It would probably be good to figure out approximately what "small" is and then write generated functions (or just @eval in a for loop) up to that number. Then use this method for all other cases. From my experiments performance takes a hit after some 8 elements. I'm also fine with holding up on this PR till there is a better consolidated clever implementation that Jeff mentioned in #15695

[KristofferC]

I get the following run times comparing the "Vector version" vs a @generated version up to N = 100:

(Forgot y-axis, it is time[s])

The compilation time looks like this:

Pretty interesting that the @generated compilation time is lower than the vector one for < 25 elements.

[JeffBezanson]

Nice data!

We also need to worry about the ability to infer the result type. We'll need some kind of cutoff, perhaps like this one:

julia/base/operators.jl

Line 85 in 7b51368

function afoldl(op,a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,qs...)