For what it's worth, just sticking to the Fib example, since that's a fairly standard benchmark: struct FibonacciIterable{I} s0::I s1::I end impo

For what it's worth, just sticking to the Fib example, since that's a fairly standard benchmark: struct FibonacciIterable{I} s0::I s1::I end import Base: iterate iterate(iter::FibonacciIterable) = iter.s0, (iter.s0, iter.s1) iterate(iter::FibonacciIterable, state) = state[2], (state[2], sum(state)) function iteratorfib(N) i = 0 x = 0 for F in FibonacciIterable(0,1) i += 1 x = F i > N && break end x end vs function loopfib(N) x0 = 0; x1 = 1 N == 0 && return 0 for i in 2:N x0, x1 = x1, x0+x1 end x1 end Then I get: julia> N = 30; julia> using BenchmarkTools julia> @benchmark iteratorfib($N) BenchmarkTools.Trial: memory estimate: 0 bytes allocs estimate: 0 -------------- minimum time: 31.351 ns (0.00% GC) median time: 32.723 ns (0.00% GC) mean time: 32.891 ns (0.00% GC) maximum time: 54.525 ns (0.00% GC) -------------- samples: 10000 evals/sample: 994 julia> @benchmark loopfib($N) BenchmarkTools.Trial: memory estimate: 0 bytes allocs estimate: 0 -------------- minimum time: 37.831 ns (0.00% GC) median time: 37.871 ns (0.00% GC) mean time: 38.616 ns (0.00% GC) maximum time: 105.626 ns (0.00% GC) -------------- samples: 10000 evals/sample: 992 If instead of writing "N=30" and interpolating it ($N) I just used N, I'd be benchmarking the dynamic dispatch. If I declared N constant, or just wrote 30 directly, the compiler turns into a cheater: julia> @benchmark iteratorfib(30) BenchmarkTools.Trial: memory estimate: 0 bytes allocs estimate: 0 -------------- minimum time: 2.990 ns (0.00% GC) median time: 2.997 ns (0.00% GC) mean time: 3.030 ns (0.00% GC) maximum time: 56.348 ns (0.00% GC) -------------- samples: 10000 evals/sample: 1000 julia> @benchmark loopfib(30) BenchmarkTools.Trial: memory estimate: 0 bytes allocs estimate: 0 -------------- minimum time: 2.990 ns (0.00% GC) median time: 2.997 ns (0.00% GC) mean time: 3.006 ns (0.00% GC) maximum time: 8.908 ns (0.00% GC) -------------- samples: 10000 evals/sample: 1000 Anyway, my points here are: 1) In this benchmark that should be focused on iteration, the custom iteration protocol actually seems faster than the for loop for some reason. 2) While for iterative algorithms like conjugate gradient, it does seem like a nice fit, it's probably not the clearest approach everywhere.