[Mingw-users] Re: GCC 3.4 Optimization Problem?

[<rr...@cs...>]

I'm a bit confused here, Brian pointed out an aliasing problem in your
code that I can't find in your message on Sourceforge's archive on
this list.  But I can find another aliasing problem in the code that
appears in the archive:

>#if 1
>    uSum += *((u_short*)&uOldValue) + (~(*((u_short*)&uNewValue)) & 
>0xFFFF) + *((u_short*)&uOldValue + 1) + (~(*((u_short*)&uNewValue + 1)) & 
>0xFFFF);
>#else
>    uNewValue = ~uNewValue;
>    uSum += (uOldValue & 0xFFFF) + (uOldValue >> 16);
>    uSum += (uNewValue & 0xFFFF) + (uNewValue >> 16);
>#endif

In the first alternative, you're breaking strict aliasing rules and in
the seconfd alternative you're not.

In Standard C the effect of accessing an object, whether reading or
writing, using a pointer of different type than the object's type is
undefined unless you use a pointer to unsigned char.

There's really no reason to use the first alternative.  The second
alternative is easier to read and understand, and likely to be faster
since it doesn't force the values into memory.

						Ross Ridge