[OT] - Interesting optimization papers

[mandrav]

Here's a site I stumbled upon the other day with a few interesting papers regarding C++ code optimization.

http://www.agner.org/optimize/

[Game_Ender]

That is very good site.  One of the parts I don't like is where basically says to use C++ like its C with classes, ie: No run time type information or exception handling.

Good to see you back mandrav, I hope whatever caused your absence went very well (never did learn why you were away).

[mandrav]

Good to see you back mandrav, I hope whatever caused your absence went very well (never did learn why you were away).

Hum, I was on vacations  :oops: . Thanks anyway

[sethjackson]

Good to see you back mandrav, I hope whatever caused your absence went very well (never did learn why you were away).

Hum, I was on vacations  :oops: . Thanks anyway

Now everyone has to get back to work. :lol:

Nice to see you back.

[thomas]

One of the parts I don't like is where basically says to use C++ like its C with classes, ie: No run time type information or exception handling.

From the strict viewpoint of optimisation, he is right, though. It is no secret that both of these consume considerable CPU time. I am not denying the useful applications and advantages of exceptions, but they are not free.

I don't have figures at hand for RTTI, but I have timed the overhead caused by enabling exceptions with gcc 3.4 some months ago. Simply turning on exceptions (without actually throwing anything) makes your code about 6-8% slower.
The time for throwing and catching an exception (which luckily does not happen in "normal" operation) is unpredictable, it could take milliseconds in the worst case. In the "normal" case, you can assume it is something on the order of 100 to 1000 times more expensive than a conditional branch (depends on how much stack has to be unwound and on what happens in destructors).

Thus, from the strict viewpoint of optimisation, it is not entirely unreasonable discourage exceptions where they are not absolutely necessary.

[Game_Ender]

I agree, I just think those are two very useful features of the language.  And its worth even a 10% performance loss for the better code it produces, from the standpoint of read ability and maintainability.  Properly designed exception handling code can be cleaner and easy to understand.

[Vampyre_Dark]

The author is right, users don't like long load times...
Here is a demo: http://www.youtube.com/watch?v=kBVmfIUR1DA  :lol:

-edit-

And here he is optimized

http://www.youtube.com/watch?v=nTDxahJj2Qw