Rollup merge of #124921 - RalfJung:offset-from-same-addr, r=oli-obk

offset_from: always allow pointers to point to the same address

This PR implements the last remaining part of the t-opsem consensus in rust-lang/unsafe-code-guidelines#472: always permits offset_from when both pointers have the same address, no matter how they are computed. This is required to achieve *provenance monotonicity*.

Tracking issue: rust-lang/rust#117945

### What is provenance monotonicity and why does it matter?

Provenance monotonicity is the property that adding arbitrary provenance to any no-provenance pointer must never make the program UB. More specifically, in the program state, data in memory is stored as a sequence of [abstract bytes](https://rust-lang.github.io/unsafe-code-guidelines/glossary.html#abstract-byte), where each byte can optionally carry provenance. When a pointer is stored in memory, all of the bytes it is stored in carry that provenance. Provenance monotonicity means: if we take some byte that does not have provenance, and give it some arbitrary provenance, then that cannot change program behavior or introduce UB into a UB-free program.

We care about provenance monotonicity because we want to allow the optimizer to remove provenance-stripping operations. Removing a provenance-stripping operation effectively means the program after the optimization has provenance where the program before the optimization did not -- since the provenance removal does not happen in the optimized program. IOW, the compiler transformation added provenance to previously provenance-free bytes. This is exactly what provenance monotonicity lets us do.

We care about removing provenance-stripping operations because `*ptr = *ptr` is, in general, (likely) a provenance-stripping operation. Specifically, consider `ptr: *mut usize` (or any integer type), and imagine the data at `*ptr` is actually a pointer (i.e., we are type-punning between pointers and integers). Then `*ptr` on the right-hand side evaluates to the data in memory *without* any provenance (because [integers do not have provenance](https://rust-lang.github.io/rfcs/3559-rust-has-provenance.html#integers-do-not-have-provenance)). Storing that back to `*ptr` means that the abstract bytes `ptr` points to are the same as before, except their provenance is now gone. This makes  `*ptr = *ptr`  a provenance-stripping operation  (Here we assume `*ptr` is fully initialized. If it is not initialized, evaluating `*ptr` to a value is UB, so removing `*ptr = *ptr` is trivially correct.)

### What does `offset_from` have to do with provenance monotonicity?

With `ptr = without_provenance(N)`, `ptr.offset_from(ptr)` is always well-defined and returns 0. By provenance monotonicity, I can now add provenance to the two arguments of `offset_from` and it must still be well-defined. Crucially, I can add *different* provenance to the two arguments, and it must still be well-defined. In other words, this must always be allowed: `ptr1.with_addr(N).offset_from(ptr2.with_addr(N))` (and it returns 0). But the current spec for `offset_from` says that the two pointers must either both be derived from an integer or both be derived from the same allocation, which is not in general true for arbitrary `ptr1`, `ptr2`.

To obtain provenance monotonicity, this PR hence changes the spec for offset_from to say that if both pointers have the same address, the function is always well-defined.

### What further consequences does this have?

It means the compiler can no longer transform `end2 = begin.offset(end.offset_from(begin))` into `end2 = end`. However, it can still be transformed into `end2 = begin.with_addr(end.addr())`, which later parts of the backend (when provenance has been erased) can trivially turn into `end2 = end`.

The only alternative I am aware of is a fundamentally different handling of zero-sized accesses, where a "no provenance" pointer is not allowed to do zero-sized accesses and instead we have a special provenance that indicates "may be used for zero-sized accesses (and nothing else)". `offset` and `offset_from` would then always be UB on a "no provenance" pointer, and permit zero-sized offsets on a "zero-sized provenance" pointer. This achieves provenance monotonicity. That is, however, a breaking change as it contradicts what we landed in rust-lang/rust#117329. It's also a whole bunch of extra UB, which doesn't seem worth it just to achieve that transformation.

### What about the backend?

LLVM currently doesn't have an intrinsic for pointer difference, so we anyway cast to integer and subtract there. That's never UB so it is compatible with any relaxation we may want to apply.

If LLVM gets a `ptrsub` in the future, then plausibly it will be consistent with `ptradd` and [consider two equal pointers to be inbounds](rust-lang/rust#124921 (comment)).

tests/pass/zero-sized-accesses-and-offsets.rs

@@ -39,8 +39,6 @@ fn test_ptr(ptr: *mut ()) {
         // Distance.
         let ptr = ptr.cast::<i32>();
         ptr.offset_from(ptr);
-        /*
-        FIXME: this is disabled for now as these cases are not yet allowed.
         // Distance from other "bad" pointers that have the same address, but different provenance. Some
         // of this is library UB, but we don't want it to be language UB since that would violate
         // provenance monotonicity: if we allow computing the distance between two ptrs with no
@@ -54,6 +52,5 @@ fn test_ptr(ptr: *mut ()) {
         // - Distance from use-after-free pointer
         drop(b);
         ptr.offset_from(other_ptr.with_addr(ptr.addr()));
-        */
     }
 }