Re: Aliasing bug in GCC 4.3.1? Resolved in 4.3.3?

[Andrew Haley <aph at redhat dot com>]

Andreas Lubensky wrote:
> Hello,
> i worked with g++ 4.3.1 (linux/586) and experienced some aliasing issues.
> Basically, i have a base class and an derived class and i am doing downcasts 
> in a number of places using static_cast<derived*>(base_pointer) (no other 
> casting is involved, except implicit casting from derived* to base*).
> To my understanding, this should not be a problem, since both types are 
> related.
> Nevertheless, using -O2 or -O3 resulted in broken code. Apparently the order 
> of execution was mixed up in some places. Using -fno-strict-aliasing or 
> applying the may_alias attribute to these classes resolved the problem.
> The whole issue is very hard to reproduce, so i could not boil it down to a 
> simple example (as the optimizer propably acts very unpredictable on 
> different code).
> I did not have this problem using some GCC 4.1 and 4.2. I just upgraded to 
> 4.3.3 and the problem vanished again. 
> Nevertheless, i am worried the bug just *happens* to not show up, because the 
> optimizer may work a little different.
> So: are there any aliasing issues that are known to be explicitly resolved by 
> 4.3.3?

It's hard to know exactly what caused your problem, but yes, there are a
number of aliasing changes on the 4.3 branch: I counted twelve patches.

Andrew.