EduardoS (no.delete@this.spam.com) on November 28, 2012 7:33 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on November 28, 2012 6:15 am wrote:
> > Huh?
> > Are both of you talking about out of bound array accesses or about some other C++ undefined behavior?
>
> No, I am not talking about out of bound array accesses (there isn't much the C++ compiler can do about this,
> anyway) but about all those undefined behaviours in C++ specs that, in theory, could allow for many optimizations
> that nobody (sane) does, you know, pointer aliasing, overlapping memcpy, (int*)(void*)(short*), etc.
>

Overlapping memcpy? Do you mean, Linus's battle # n+1 with gcc developers from year or so ago? I think, what they did in this particular case was pessimization which is bad. But I agree with gcc developers that reliance on particular order of execution of memcpy is a dangerous practice, at least as long as organization does not have full control on compiler. Since both memcpy and memmove present in the standard, so there is probably a god reason for that.

In the last part, do you refer to reliance of gcc optimizer on C99 strict aliasing riles?
I don't like them too. But C99, and to less extent C++, has other, perfectly sane although not 100% sufficient, ways by which programmer can carry information about absence of aliasing to compiler. Besides, in many cases C/C++ compiler can figure it out by itself during LTCG/GPO, which, I think, is perfectly sane and very desirable.