Re: [PATCH][RFC] Adjust the middle-end memory model

[Richard Guenther <rguenther@suse.de>]

On Wed, 20 May 2009, Mark Mitchell wrote:

> Richard Guenther wrote:
> 
> >>   void f(float *f, int *n) {
> >>     for (int i = 0; i < *n; ++i) {
> >>       f[i] *= 2;
> >>     }
> >>   }
> 
> > The difference is if you want to sink a load from *n beyond the
> > store to f[i] - in which case you ask if there is an anti-dependence
> > which we cannot exclude in this case (no TBAA is allowed here).
> 
> By "not allowed", you don't mean "would be an invalid optimization", but
> rather "will no longer be done by GCC", right?

Right, not invalid in the above case but nevertheless no longer being
done by GCC.  This is to properly support

int i;
float f;
void foo()
{
  int *p = (int *)malloc(sizeof(int));
  *p = 1;
  i = *p;
  float *q = (float *)p;
  *q = 2.0;
  f = *q;
}

where we need to avoid scheduling the store via *q before the load
from *p.  The above is valid as I read C99 6.5/6, it is an object
with no declared type (obtained via malloc) and has type int
due to the store via *p "for that access and for subsequent accesses
_that do not modify the stored value_." (emphasis mine).  So for
objects without a declared type C can do "placement new" by simply
storing with a different type.  In C++ we of course have the
usual placement new situations.

Richard.