Re: [boost] Google SoC 2007 Big Integer project

From: Jeff F (TriumphSprint2000_at_[hidden])
Date: 2007-03-26 09:35:07

Hugh Wimberly wrote:

...

> I don't understand how not using exceptions allow better efficiency.
>> If you want to detect the error, you have to make a test in both
>> cases.
>
> I think I was confusing myself. Arithmetic errors aren't handled using
> exceptions, so code like
>    try {  ... }
>    catch (Exception e) {throw DivideByZero}
> won't work if the try block contains just the division algorithm. I
> was thinking of comparing the above code to
>    if (d == 0) throw DivdeByZero;
>    else { ... }
> Which would be less silly and more efficient, but I have seen the
> former, especially in Java. In this case, I think I'm supposed to be
> comparing the second example to something that doesn't check for or
> throw errors at all, which seems like a bad practice to me, but is
> how ordinary integer
> operations are handled. Right? Am I still crazy?

In my experience, at least with earlier versions of MSVC under Windows, 
using a floating point math error handler to generate C++ exceptions 
improved performance of our RK5 integrator by >40% versus doing the divide 
by zero check in the inner loop. Also you wouldn't put the try catch block 
inside that inner loop, it is at the executive level, determining whether 
the overal operation completed, or failed. This is where the decision is 
made to decrease step size, or carry out some other operation. Actually that 
40% was the overall frame time improvment, the time spent in the integrator 
was even less.

Jeff Flinn