[Blitz-support] strict-aliasing problems

[Torquil M. S. <to...@gm...>]

Hi!

I'm suddenly getting incorrect numerical results when using -O2 with g++ 4.6.1. 
The culprit seems to be -fstrict-alisasing, since disabling this fixes everything.

I posted about my problem on the GCC mailing list:

http://gcc.gnu.org/ml/gcc-help/2011-08/msg00168.html

Basically, the following program prints the correct result (prints 1.666) when 
compiling with "-O2 -no-strict-aliasing", but incorrect (prints 0) when using 
"-O2", with g++ 4.6.1. Using g++ 4.5.3 there is no problem.

------------------------------
#include "iostream"
#include "blitz/array.h"

int main()
{
     blitz::Array<double,1> X(3); X = 0.0;

     blitz::Array<double,1> var(1); var = 0.0;

     for(int i = 0; i != 1; ++i) {
         for(int c = 0; c != 3; ++c) X(c) = double(c);

         var(i) = 0.0;
         for(int c = 0; c != 3; ++c) var(i) += X(c)*X(c);
         var(i) /= 3.0;
     }

     std::cout << var(0) << std::endl;

     return(0);
}
---------------------------

The program is a little bit strange, but that only because it is the result of 
simplifying something else in order to get a minimal example that triggers the 
error.

Initially, my impression was that it is a compiler bug since it worked with 
g++4.5.3, however I'm posting here as well due to this response from one of the 
guys at the GCC list:

> In every case I've looked at, when code gives the expected answer only
> when using -fno-strict-aliasing, it is because the code is using a type
> cast in an invalid way.
>
> In this case I would assume that the invalid type cast occurs in the
> blitz library code.  Look for invalid type casts there.
>
> Ian Lance Taylor

What do you Blitz+-guys think? I'm using Blitz++ v0.9 on Debian Sid on amd64 
architecture. Debian version number is 1:0.9-12 for the  libblitz0-dev package. 
My g++ has Debian version 4.6.1-6.

Best regards
Torquil Sørensen