Performance penalty for using ranges

[Paul Jurczak]

On Sunday, 25 August 2013 at 18:07:33 UTC, Paul Jurczak wrote:
[..]
> I'm also running it with LDC, but it reports timings too good 
> to be true - something meaningful is getting optimized away and 
> I'm trying to find out why.
[..]

It seems that LDC optimizer is smarter than I expected and it was 
able to optimize away comparison on constant elements of array a 
in my test:

int test(alias F)(int N) {
    int a[10];
    int n = 0;

    foreach (immutable _; 0..N) {
       a[4] = !a[4];
       n += F(a);
    }

    return n;
}

I modified it, so there is no constant elements and all cases of 
pairwise comparison are exercised:

int test(alias F)(int N) {
    enum L = 10;
    int  a[L];
    int  n = 0;

    foreach (int i; 0..N) {
       auto j = (i%L + L/2)%L;
       a[j] = !a[j];
       n += F(a);
    }

    return n;
}

Here are the results:

Xubuntu 13.04 64-bit Core i5 3450S 2.8GHz (3.5GHz turbo):
LDC 0.11.0: ldmd2 -m64 -O -noboundscheck -inline -release
100000  5284
100000  4265
100000  4268
100000  5717
100000  4265

GDC 4.8.1: gdc -m64 -march=native -fno-bounds-check 
-frename-registers -frelease -O3
100000  4612
100000  5717
100000  5718
100000  4984
100000  11546

DMD64 2.063.2: dmd -O -noboundscheck -inline -release
100000  8350
100000  14615
100000  14597
100000  14270
100000  18509

Windows 7 64-bit Core i5 2500 3.4GHz turbo:
DMD 2.063.2 (32-bit) dmd -O -noboundscheck -inline -release
100000  9803
100000  17050
100000  17031
100000  20851
100000  20283

I'm stunned by LDC ability to optimize isPalindrome4 to the best 
performance level of the whole group. Other compilers placed it 
at the bottom of performance rank.