minimum() (foldl(min, ary)) seems to be much slower than it can be

[Moelf]

julia> function normal_findmin(dij)
           vmin = Inf
           for x in dij
               vmin = ifelse(x < vmin, x, vmin)
           end
           return vmin
       end
julia> a = rand(120)

julia> @be minimum($a)
Benchmark: 2911 samples with 71 evaluations
 min    376.620 ns
 median 412.183 ns
 mean   439.258 ns
 max    1.319 μs

julia> @be normal_findmin($a)
Benchmark: 2923 samples with 761 evaluations
 min    37.783 ns
 median 38.403 ns
 mean   40.277 ns
 max    130.361 ns

(Handling isempty() and init isn't enough to account for the difference)

[giordano]

Duplicate of #38558

[Moelf]

I think this is not a 100% duplicate. Yes, solving #38558 will solve this, but we don't need to solve #38558 to solve this -- we can imagine specialize minimum/maximum/maxima to at least make them faster?

[Moelf]

at least maybe don't implement them as mapreduce? for identity, and Colon(), we just need foldl(), but foldl() is still slower, because fold() is also implemented via mapfoldreduce

[jakobnissen]

Your normal_findmin does not take NaN into account, and therefore compiles to more efficient machine code. Handling NaN (by using min instead of the hand-rolled ifelse(x < vmin, x, vmin)) completely closes the speed gap on my computer.

[Moelf]

I will note that right now on aarch64 using min is actually faster, not slower, so maybe still something to be done for x86