type aliasing (was: gcc 3.3 -03 produces bad code)

[Daniel Jacobowitz]

On Fri, Apr 18, 2003 at 12:59:45AM +0200, Tolga Dalman wrote:
> On Thu, 17 Apr 2003 18:41:49 -0400 Daniel Jacobowitz <drow@mvista.com> wrote:
> 
> > On Fri, Apr 18, 2003 at 12:32:24AM +0200, Tolga Dalman wrote:
> > > hello, 
> > > 
> > > i missed to reply to the group, only Daniel got the message. sorry about
> > > that.
> > > 
> > > On Thu, 17 Apr 2003 18:25:24 -0400 Daniel Jacobowitz <drow@mvista.com>
> > > wrote:
> > > 
> > > > On Fri, Apr 18, 2003 at 12:21:54AM +0200, Tolga Dalman wrote:
> > > > > hello, 
> > > > > 
> > > > > On Thu, 17 Apr 2003 16:04:24 -0400 Daniel Jacobowitz <drow@mvista.com>
> > > > > wrote:
> > > > > 
> > > > > > On Thu, Apr 17, 2003 at 09:32:23PM +0200, Ronald Kukuck wrote:
> > > > > > > Hello,
> > > > > > > 
> > > > > > > I'm using:
> > > > > > > 
> > > > > > > gcc version 3.3 20030226 (prerelease) (SuSE Linux)
> > > > > > > 
> > > > > > > A large application (~25MB), which worked fine with the gcc included
> > > > > > > in Suse 8.1, is now crashing in some modules if I compile it with
> > > > > > > "-O3".
> > > > > > > 
> > > > > > > I have tried to make the bug reproducible in a small file (test1.C).
> > > > > > > To see it working OK use:
> > > > > > > 
> > > > > > > g++ -O -o test test1.C
> > > > > > > 
> > > > > > > to see it fail use:
> > > > > > > 
> > > > > > > g++ -O3 -funroll-loops -mcpu=pentiumpro -o test test1.C
> > > > > > > 
> > > > > > > or use simply:
> > > > > > > 
> > > > > > > g++ -O3 -o test test1.C
> > > > > > > 
> > > > > > > which is producing different results then above one.
> > > > > > > 
> > > > > > > 
> > > > > > > Is there a workaround for this problem or is there a place where I
> > > > > > > can download a new binary?
> > > > > > 
> > > > > > > inline int ReadDbRecHead(int *pred,int *succ,int *next, short
> > > > > > > *infopa, const int rec_no, const int dbunit) {
> > > > > > >       int *adress, *iptr;
> > > > > > >       
> > > > > > >      // DbsFindReadAdr(dbunit, rec_no);
> > > > > > >       adress = (int*)&s_in;
> > > > > > >       iptr=(int*)infopa;
> > > > > > >       *pred=*adress++;
> > > > > > >       *succ=*adress++;
> > > > > > >       *next=*adress++;
> > > > > > >       *iptr++=*adress++;
> > > > > > >       *iptr=*adress;
> > > > > > >       return(*adress);
> > > > > > >    }
> > > > > > 
> > > > > > You can't cast a short * to an int * that way.  I recommend searching
> > > > > > for information on type aliasing, or seeing the effect of
> > > > > > -fno-strict-aliasing on your code.
> > > > > 
> > > > > well, i was wondering about type aliasing and the effect of
> > > > > -fno-strict-aliasing for a long time. could you give a brief
> > > > > explaination about aliasing? the example in the man page was not really
> > > > > enlightening to me :( when "should" i use -fno-strict-aliasing and when
> > > > > not?
> > > > 
> > > > You should never use it - fix the code.  I really can't explain it any
> > > > better than that example.
> > > 
> > > can't there be a case, where it can't be avoided to violate aliasing rules?
> > > i haven't got an example handy, but as far as i understood, i can't access
> > > to the low 8-bit data of a 32-bit variable directly, right?
> > 
> > If you use "char *", that's not an aliasing violation.  It's often not
> > the best way to do it, but it's not an aliasing violation.
> 
> sorry for insisting, but what is it then? my example was not meant to be a
> "char*", but how about:
> 
> uint32_t i = /* a value */;
> uint8_t  l; 
> 
> l = ((uint8_t*) &i)[1];
> 
> would that be a violation? or this one:
> 
> char* c = "teststring";
> int* p;
> 
> for (p = (int*) c; p; p++) {
>      /* do something */
> }

If it's a char *, then it's OK.  If it's a uint8_t, then it's a
violation.  I'm not sure if the second one would be a problem but I
think it might be, since you're accessing char data through an int*. 
You have to access an object through a pointer of the appropriate type,
OR a pointer to char.

I seriously recommend picking up a copy of the language standard, which
explains this quite well.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer