I've done a few tests on how DMD and GDC handle tight loops and was surprised by the slow code DMD generates for "foreach".

results and source code: http://dstress.kuehne.cn/benchmark/tight_loops/sum.html

Thomas

Not Found

The requested URL /benchmark/tight_loops/sum.html was not found on this server.

-Craig

That link is dead.

Thomas Kuehne wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> I've done a few tests on how DMD and GDC handle tight loops and was surprised by the slow code DMD generates for "foreach".
> 
> results and source code: http://dstress.kuehne.cn/benchmark/tight_loops/sum.html
> 
> Thomas
> 
> 
> -----BEGIN PGP SIGNATURE-----
> 
> iD8DBQFGJkFyLK5blCcjpWoRAgM4AJ9JnbxtHTlBFwXdWfwfcuShxi/2iACfVaxI
> ULxRdOI+28mzZXCbv8RiPO8=
> =Z0K0
> -----END PGP SIGNATURE-----

Invalid URL. I tried looking at http://dstress.kuehne.cn/benchmark/ and there is not even a tight_loops directory!

Craig Black schrieb am 2007-04-18:
> Not Found
>
> The requested URL /benchmark/tight_loops/sum.html was not found on this server.

Seems to be some server issue, backup can be found here: http://svn.berlios.de/svnroot/repos/dstress/trunk/benchmark/tight_loops/sum.html

Thomas

Good work on this.  It is certainly a problem for Walter if he sticks with his "trust the compiler" philiosphy.  Since the DMD documentation claims that foreach should emit optimized code, and it clearly doesn't in these cases, I think you could submit this as a bug.

-Craig

Indeed, this is clearly an optimization problem.  I agree that we should put our faith in the compiler because it's bad to undermine it to try to write optimal code when the compiler may still be improved to solve the problem *properly*

That said, DMDScript undermines associative arrays by completely re-implementing them with a tweak - instead of using opIndex and opIndexAssign.  : p

But yes, let's work on getting the compiler to Do The Right Thing(tm) so we don't have to kludge our code, mm?

Craig Black wrote:
> Good work on this.  It is certainly a problem for Walter if he sticks with his "trust the compiler" philiosphy.  Since the DMD documentation claims 

Walter consistently mentioned that a) most of the DM compiler back-end is 20 years old and b) is not optimized for D yet. I think the "trust the compiler" philosophy is for stability (20 years old), not optimizations. Plus for integer stuff it seems to do pretty well (GCC is no slouch).

That said, both foreach and FP should perform better!

> that foreach should emit optimized code, and it clearly doesn't in these cases, I think you could submit this as a bug.
> 
> -Craig 
> 

Thomas Kuehne wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Craig Black schrieb am 2007-04-18:
>> Not Found
>>
>> The requested URL /benchmark/tight_loops/sum.html was not found on this server.
> 
> Seems to be some server issue, backup can be found here:
> http://svn.berlios.de/svnroot/repos/dstress/trunk/benchmark/tight_loops/sum.html
> 
> Thomas
> 

Nice, but didn't the 'result' variable overflow several times during the loop for the int tests?

> 
> -----BEGIN PGP SIGNATURE-----
> 
> iD8DBQFGJkfJLK5blCcjpWoRAnMnAJ0dabQGW2edXBXlXVbXqqUSlgUmjACgqJUh
> kh2pP8ZmgultrR2U+Cusgyc=
> =tUSf
> -----END PGP SIGNATURE-----

Thomas Kuehne wrote

surprised by the slow code DMD generates for "foreach".

Not confirmed.

On my machine foreach is about 15% faster than pointer_2a.

-manfred