Inlining Code Test

[Craig Black]

"Iain Buclaw" <ibuclaw at ubuntu.com> wrote in message 
news:ie6ahr$1coc$1 at digitalmars.com...
> == Quote from Craig Black (craigblack2 at cox.net)'s article
>> > Testing your C++ program (altered getCycle() for GCC)
>> >
>> > Times I get:
>> > -------
>> > Sorting with Array: 46869.159840
>> > Sorting with pointers: 38688.937320
>> > 17.453316 percent faster
>> >
>> > Sorting with Array: 46631.903760
>> > Sorting with pointers: 38520.609360
>> > 17.394302 percent faster
>> >
>> > Sorting with Array: 46674.330720
>> > Sorting with pointers: 38545.202520
>> > 17.416700 percent faster
>> > -------
>> >
>> >
>> > On a , I thought I might try an older version of GCC for the D program.
>> > Really surprisingly, I got:
>> >
>> > -------
>> > Sorting with Array.opIndex: 43075.059840
>> > Sorting with pointers: 40019.701920
>> > 7.093102 percent faster
>> >
>> > Sorting with Array.opIndex: 42940.085640
>> > Sorting with pointers: 39594.089040
>> > 7.792245 percent faster
>> >
>> > Sorting with Array.opIndex: 44389.127280
>> > Sorting with pointers: 41159.016960
>> > 7.276805 percent faster
>> > -------
>> >
>> > This will need some thinking through as to just what happened between
>> > GDC-4.3 -> GDC-4.4 :~)
>> >
>> > Regards
>> Curious benches there.  Seems GCC isn't inlining (or other optimization) 
>> as
>> well as Visual C++ here.  I'm getting a neglible ~2% difference.
>> -Craig
>
> OK, found and fixed the problem with GDC-4.4. My results are now:
>
> g++-4.4 -O3 -march=native -ftree-vectorize
> -------
> Sorting with Array: 45861.599640
> Sorting with pointers: 38417.632680
> 16.231372 percent faster
>
> Sorting with Array: 46421.979120
> Sorting with pointers: 38474.709840
> 17.119626 percent faster
>
> Sorting with Array: 48622.885200
> Sorting with pointers: 39593.385600
> 18.570473 percent faster
> -------
>
> gdc-4.4 -frelease -O3 -march=native -ftree-vectorize
> -------
> Sorting with Array.opIndex: 41454.052200
> Sorting with pointers: 38331.517560
> 7.532520 percent faster
>
> Sorting with Array.opIndex: 41109.852720
> Sorting with pointers: 37836.883320
> 7.961521 percent faster
>
> Sorting with Array.opIndex: 40587.349320
> Sorting with pointers: 37390.488120
> 7.876497 percent faster
> -------
>
> dmd-2.050 -release -O -inline
> -------
> Sorting with Array.opIndex: 53561.598720
> Sorting with pointers: 48348.626760
> 9.732667 percent faster
>
> Sorting with Array.opIndex: 55861.658280
> Sorting with pointers: 49909.880760
> 10.654495 percent faster
>
> Sorting with Array.opIndex: 56887.660800
> Sorting with pointers: 51470.453400
> 9.522640 percent faster
> -------
>
> I noticed that with dmd -inline needs to be explicitly set. Otherwise 
> there appears to be nothing to suggest a wild difference
> between using opIndex and pointers. :~)
>
> Regards

Thanks for your help and interest.  Basically you are saying I should use 
GDC instead?  I've heard rumors that GDC is not as stable.  Can you comment 
on this?  I am still getting different results than you with dmd, but maybe 
that's because we have different machines?

My results on a core i7 1.86 GHz notebook:

dmd-2.050 -O -release -inline

Sorting with Array.opIndex: 19160.2
Sorting with pointers: 15006.4
21.6791 percent faster

Sorting with Array.opIndex: 20200.8
Sorting with pointers: 15699.1
22.2887 percent faster

Sorting with Array.opIndex: 19563.3
Sorting with pointers: 15418.4
21.1872 percent faster

-Craig