Use reverse iteration protocol in mapfoldr_impl

[YingboMa]

This PR addresses the issue #24823. I replace the getindex and endof by the reverse iteration protocol in the function mapfoldr_impl.

[vchuravy]

Thanks for tackling this! I was wondering if there is any difference between mapreduce and mapfoldr_impl?

mapfoldr(f, op, v0, itr) = mapreduce(f, op, v0 , Iterators.reverse(itr))

should be equivalent.

[stevengj]

Two problems with this.

1. mapreduce doesn't define its associativity — it is allowed to associate however it wants.
2. even if you use mapfoldl, you are assuming that op is commutative as well as associative. foldr changes associativity, but does not commute operands.

[stevengj]

Maybe add a test case for something that fails in 0.6, i.e. for an iterator that is reversible but not indexable.

For example,

@test foldr(Pair, 0, Iterators.filter(iseven, 1:6)) === foldr(Pair, 0, 2:2:6) === (2=>(4=>(6=>0)))

[KristofferC]

@nanosoldier runbenchmarks(ALL, vs = ":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[stevengj]

The performance regressions reported by Nanosoldier can't possibly be real; they are in lots of functions (e.g. transpose!) that have nothing to do with this PR. It seems that Nanosoldier is still not working properly?

[KristofferC]

This noise level is nothing out of the ordinary. Benchmarks looks good to me.

[stevengj]

±45% noise, with ±25% on dozens of benchmarks, is far worse than it used to be. It makes nanosoldier almost useless except to catch truly enormous regressions.

[stevengj]

Why was this closed?

[StefanKarpinski]

Unclear. @YingboMa?

[YingboMa]

Oops sorry, I thought that this PR wasn't planned to be merged.