Re: c++ optimisation fails badly

[Teddy Todorov <Teddy dot Todorov at cern dot ch>]

Andrew Dorrell wrote:
> Notice in this case test 2 (user 0.564s) takes 7 times as long as test 3 
> (user    0m0.074s) which performs exactly the same amount of work!
> 
> Please Please Please! Am I missing something or is this a real optimiser 
> problem.  Is this a known issue?  Am I doing something significantly wrong 
> in my code?  This is driving me to distraction as I am coding for image 
> processing and just can't accept that kind of performance hit.

It is a known issue, and I had a similar problem. However, I _did_ manage to 
get optimal performance from gcc (version 3.2.3 and later). I tried to apply 
the tricks that helped me to your test case, namely
 - declare the global scope operators "inline". I am not sure if the need for 
this is a compiler bug.
 - Add a copy constructor that does exactly what the compiler-generated one 
does. This is definitely a gcc bug, and I filed a bug report (number 12685), 
which turned out to duplicate a few other (see bug entry 6880), and which was 
indeed fixed in gcc tree-ssa, but not in 3.3.x or 3.4.x

Adding the "inline" did speed up your example a little, but not by much.
Adding a copy constructor did not change the timing.
The only remaining difference with my code I could think of is that you use 
integer arithmetics while I use floating point.
Indeed, when I converted your test to float the difference between cases 2 and 
3 became small:

real    0m0.146s
user    0m0.057s
sys     0m0.083s

real    0m0.468s
user    0m0.173s
sys     0m0.284s

real    0m0.424s
user    0m0.148s
sys     0m0.264s

I guess I was lucky (and so was the Stepanov test of cost of abstraction etc.)
But this probably does not solve your problem (unless you can use a unreleased 
version of gcc).

         Teddy Todorov