Re: [Widelands-public] Loops / Efficiency

[Florian W. <Flo...@pf...>]

The following mail thread (from 2003) is quite interesting in this regard,=
=20
especially the consideration of data and instruction caches:
http://compilers.iecc.com/comparch/article/03-02-073

Am Donnerstag, 17. August 2006 00:00 schrieb Christoph Mallon:

> Disclaimer: IAACC (I Am A Compiler Constructer)

OK, I can't and won't compete with that background, I'm a mere user :-) It=
=20
also seems my information was out of date (or maybe not, cf. the above mail=
=20
thread)

> IIRC gcc only includes -fomit-frame-pointer automatically on
> architectures where doing so doesn't interfere with debugging, i.e. on
> x86 it doesn't do it automatically.

At -O1, yes. At -O2/3 it's always on.

> I don't consider calculation the loop bounds directly before the loop
> "unreadable",

Neither do I.=20

> sometimes quite contrary: it can break one monster line=20
> into two or three short lines.

Which it does decidedly not do in this case. Tell me about readability of t=
he=20
following blocks (that's rhethorical; I hereby rest my case).

>>>       const Descr_Maintainer<Building_Descr>::Index nr_buildings =3D
>>>               get_nr_buildings();
>>>       for (Descr_Maintainer<Building_Descr>::Index i =3D 0;
>>>               i < nr_buildings; ++i) {
>>>-----------------------------------------------
>>>       for (Descr_Maintainer<Building_Descr>::Index i =3D 0;
>>>               i < get_nr_buildings(); ++i) {



> When you've got the best algorithm, all that's left are the constants.
> And for the reasons I stated above the statement "most compilers will do
> anyway" is bold.

OK, I was sufficiently interested to blow some time for creating numbers. I=
=20
stopped very quickly though.

Tribe_Descr::parse_buildings (which is the context we're talking about) is=
=20
called once when loading a tribe, and nowhere else. This is only done once=
=20
when starting a new game. The actual for loop is being called 63 times for=
=20
the barbarians (which might itself be a sign of a problem because there's=20
only 45 barbarian buildings).

So unless this method does a lot of work (which it doesn't - gprof shows <1=
=20
microsecond total accumulated time) I totally ignore those 63 calls. It's n=
ot=20
worth the trouble.

<pause>

Yes, I really mean that. If the profiler doesn't tell me otherwise, I don't=
=20
even _consider_ performance on a sub-"which algorithm/datastructure"-level.=
=20

http://www.cookcomputing.com/blog/archives/000084.html summarizes my point=
=20
better than I could.

I think highly of Erik for considering performance. My *personal* view is t=
hat=20
*in this case* readability takes the upper hand. I sincerely hope I didn't=
=20
express this more strongly than it it was meant.

=2D-=20
Mit freundlichen Gr=FC=DFen,
=46lorian Weber

PGP key id: A34C32F9

=2D-=20
With best regards,
=46lorian Weber

PGP key id: A34C32F9