Re: Linux 2.0.3x & gcc 2.8 [better use egcs for now]

Andi Kleen (ak@muc.de)
07 Feb 1998 09:26:44 +0100


wmglo@dent.med.uni-muenchen.de (Wolfram Gloger) writes:

> I just checked this code:
>
> int f(long e)
> {
> *(&e) = e & 0x1234;
> return 0;
> }
>
> and found that egcs-1.0.1 _never_ optimizes the assignment away (no
> volatile is needed). So, people should maybe use that rather than
> gcc-2.8.0 for compiling 2.0.x.

That is because egcs 1.0.1 doesn't contain the addressof optimization.
Post 1.0.1-release versions of egcs have merged that code from 2.8.0.

>
> > Kenner admited that it was a bug in the new addressof optimization (which
> > finally supports *(&x) = y) and promised a fix for 2.8.1.
>
> IMHO, a single assignment to *(&x) can _never_ be optimized away
> (multiple such assignments could be folded into one possibly). I
> can't see why a volatile specification or cast should make a
> difference either, since the variable isn't written to by another
> context of execution.

It can be optimized away when x is dead and there are no aliases - and it is
certainly dead when it is the last operation on a local variable or an argument.

For threads you need to use volatile variables.

-Andi

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu