Re: [PATCH]: Fix PR tree-optimization/21407

[Mike Stump <mrs at apple dot com>]

On May 10, 2005, at 11:07 AM, Mark Mitchell wrote:
> I'm sure Geoff and Mike will share their reasoning.

Easier to do in person.

> So, I think that the argument for invalidity has to hinge on the  
> fact that "(int*)((char*) &s + offsetof (struct S, j)))" is  
> different from "&s.j", even though the standard says those two  
> expressions have the same type and the same value.

No, you got it wrong.

&s.j == (int *) ((char*) &s + offsetof (struct S, j))   in all ways  
we can ever detect.

this is not in dispute.

Geoff's contention is that you cannot decrement it to point to before  
&s.j on the grounds that object, means, the s.j object, and it  
alone.  The notion of object, is one the optimizer comes up with  
though analysis; it's a runtime thing.  Also, one cannot increment a  
pointer to point far past (the second next field if there where two  
after j) &s.j, because this is beyond the end of the object, the  
object j.

I disagree.  I think that when the value representation of two  
pointers are the same, then, there cannot be a difference, unless the  
standard is exceptionally clear they meant to make a difference; that  
the object, is the complete object.  To me, object means, &s for  
sizeof (s) bytes.  Then trivially, one can wonder around in it all  
day long, and also, as long as the language mandates that the value  
representation is the same as &s.m, for any given m, then one can  
dereference that value representation when converted to a pointer to  
the type of m and that this is in all ways indistinguishable from * 
(&s.m).

Apologies to Geoff if I am mis-remembering, I cannot seem to find the  
thread in my email archive.