This is probably difficult and unnecessary to implement, but it seems correct to me, based on the definition of restrict in N1256:
I believe that a dereferenced T const *restrict qualified pointer actually is guaranteed to have the same UB as a reference, within the domain the access can be reordered.
Specifically, const guarantees[^1] that this pointer does not modify the value, and restrict guarantees that, if accessed and modified it is only accessed via this pointer.
In contrast, a reference is dereferenceable and read only, except in UnsafeCell.
Thus if a T const *restrict pointer is accessed, within the C-block it is declared in, it is dereferenceable and read only[^1].
Similarly, if a T *restrict qualified pointer is written to, then within the C-block it is declared in it is at least as strong as a &Cell<T> as it is immutable or only accessed via this pointer, and might be as strong as &mut T depending on the precise rules for sending &mut T around[^2].
[^1]: except via locally casting away const?
[^2]: specifically I am worried about floating pointer provenance, restrict should be enough to require disjoint accesses.
This is probably difficult and unnecessary to implement, but it seems correct to me, based on the definition of restrict in N1256:
I believe that a dereferenced
T const *restrictqualified pointer actually is guaranteed to have the same UB as a reference, within the domain the access can be reordered. Specifically,constguarantees[^1] that this pointer does not modify the value, andrestrictguarantees that, if accessed and modified it is only accessed via this pointer. In contrast, a reference is dereferenceable and read only, except inUnsafeCell. Thus if aT const *restrictpointer is accessed, within the C-block it is declared in, it is dereferenceable and read only[^1].Similarly, if a
T *restrictqualified pointer is written to, then within the C-block it is declared in it is at least as strong as a&Cell<T>as it is immutable or only accessed via this pointer, and might be as strong as&mut Tdepending on the precise rules for sending&mut Taround[^2].[^1]: except via locally casting away
const? [^2]: specifically I am worried about floating pointer provenance,restrictshould be enough to require disjoint accesses.