Reduce has dreadful performance?

[John Colvin via Digitalmars-d]

On Thursday, 18 June 2015 at 10:27:58 UTC, Russel Winder wrote:
> On a given machine, the code:
>
> double sequential_loop(const int n, const double delta) {
>   auto sum = 0.0;
>   foreach (immutable i; 1 .. n + 1) {
>     immutable x = (i - 0.5) * delta;
>     sum += 1.0 / (1.0 + x * x);
>   }
>   return 4.0 * delta * sum;
> }
>
> runs in about 6.70s. However the code:
>
> double sequential_reduce(const int n, const double delta) {
>   return 4.0 * delta * reduce!((double t, int i){immutable x = 
> (i -
> 0.5) * delta; return t + 1.0 / (1.0 + x * x);})(0.0, iota(1, n 
> + 1));
> }
>
> runs in about 17.03s, whilst:
>
> double sequential_reduce_alt(const int n, const double delta) {
>   return 4.0 * delta * reduce!"a + b"(
>          map!((int i){immutable x = (i - 0.5) * delta; return 
> 1.0 /
> (1.0 + x * x);})(iota(1, n + 1)));
> }
>
> takes about 28.02s. Unless I am missing something (very 
> possible), this
> is not going to create a good advert for D as an imperative 
> language
> with declarative (internal iteration) expression.

What you've forgotten is to turn on your optimisation flags. Even 
DMD gets the first 2 vaguely right with -O -inline -release

gdc is able to vectorize all 3 examples if you give it -Ofast

The 3rd one does appear to be a bit slower in all cases, not sure 
why right now.