Member

oli-obk commented 17 days ago

Is this testable within this repo or do we need miri for it?

r=me with a test if possible

RalfJung commented 17 days ago

Good point, I added a test.

@bors r=oli-obk

bors commented 17 days ago

Commit f76bae9 has been approved by oli-obk

Is that a guarantee that we will want to make going forward?

I tried to create a *const dyn Debug from a *const DynamicallySized<[usize]> and failed (see playground) as apparently it's not possible to create a pointer to trait from a pointer to DST right now.

It's unclear whether this is a restriction that will stay, or not. If it doesn't, however, then size_of_val_raw may, on certain types, require accessing the data-pointer of the trait.

As such, it seems that if we wish to keep our options open regarding the conversion from DST to trait, then we should enforce that size_of_val_raw require a valid data-pointer.

RalfJung commented 8 days ago

Is that a guarantee that we will want to make going forward?

The guarantee is what is documented for align_of_val_raw_/size_of_val_raw. In particular, for sized types, the pointers may dangle. So what Miri did previously was definitely wrong.

I think Miri also implements the rest of those docs correctly, but if there's a bug, please let me know. Note the last clause in the documentation:

otherwise, it is conservatively not allowed to call this function.

Currently, this is impossible to reach (there are no other kinds of DST), but that is not a stable assumptions -- so if one has no clue what the unsized tail is, one has to use size_of_val still.