I think that the point of the article (example 1 vs example 3) is that you are not allowed to cast or unionize pointers.  You must store and fetch from a location that’s declared as a union.

 

Now what if you cast, not from T1* to T2*, but from T1* to U*?  Does the union itself have to be defined as a union, or can you cast the original typed pointer to the union pointer, perhaps as the parameter to a function call so that only the U* pointer is in scope and being dereferenced while T1* in the caller is dormant and assumed to be “changed” by virtue of the call.

 

Perhaps a close reading of the spec will indicate that a “barrier” is possible where two pointers are not used at the same time.  I’m certain that works in practice though:  upon calling foo(p);, the compiler assumes that *p has been changed.  Within the body of foo, p is not used at all but q is used, which has the same address and different type.  If p is not used, you don’t have to worry about it being assumed not to change just because q is used.  However, inlining complicates things.

 

I suppose to be ratified, a portable solution is needed.  But individual compiler-specific solutions would implement the alias_cast by saying “yes this will alias” in a compiler-specific manner.  The general solution is only for formality and bootstrapping a port to a different compiler/version, so it doesn’t need to be that efficient.