The `const_raw_ptr_to_usize_cast` feature has been removed

[timothee-haudebourg]

This feature has been recently removed so the crate does not compile anymore.

   |
51 |   const_raw_ptr_to_usize_cast,
   |   ^^^^^^^^^^^^^^^^^^^^^^^^^^^ feature has been removed
   |
   = note: at compile-time, pointers do not have an integer value, so these casts cannot be properly supported

[timothee-haudebourg]

@slightlyoutofphase can we expect a fix sometime soon along with #49?

[ohmyarch]

Same.

[RalfJung]

I am curious how this crate even used const_raw_ptr_to_usize_cast. Almost all ways of using it would lead to an error when actually evaluating the constant.

[slightlyoutofphase]

Hey guys, I've been dealing with an extended family emergency recently. Looking into fixing everything ASAP, now.

@RalfJung:

I am curious how this crate even used const_raw_ptr_to_usize_cast. Almost all ways of using it would lead to an error when actually evaluating the constant.

I frankly couldn't tell you off the top of my head right at this very moment. It was definitely doing something practical that I would have made sure to directly use both in the test suite and example programs, though. In any case, it appears I have a fair chunk of updating to do as far as feature flags added and removed since the last commit.

[slightlyoutofphase]

After refreshing my memory by re-reading the code, the use of const_raw_ptr_to_usize_cast was largely related to directly accounting for the possibility of ZSTs in a non-panicking way. It did work as intended in the places it was used, even under Miri.

[RalfJung]

Ah so those must have been pointers that started as integers cast to raw pointers, and later cast back. That is the only way you would not have gotten errors. (Miri and rustc execute compile-time code the same way.)

Do you still need that feature or can it be replaced by other alternatives? If there is something that you used to be able to do that is now no longer possible, please let us know!

[slightlyoutofphase]

I'm still looking into "what's new" in the core and standard libraries in terms of functionality since I last worked on the crate, but it seems like I should be able to use functionality from there for a few things I previously had to hand-roll, as there's stuff that's now const which definitely wasn't const last time I looked at it.

I'm still not 100% sure how I'll keep the compile-time-quicksort functionality going ultimately, but I think it'll be possible. At worst it might just have to panic outright if it encounters a ZST, or something along those lines. May require a larger version bump to 0.11 for the crate, when all is said and done, but that's not a huge deal I guess.

[slightlyoutofphase]

Ok, everything is mostly working. Going to leave all these issues open until I'm 100% satisfied with the state of things, though.

[slightlyoutofphase]

Ah so those must have been pointers that started as integers cast to raw pointers, and later cast back. That is the only way you would not have gotten errors. (Miri and rustc execute compile-time code the same way.)

Do you still need that feature or can it be replaced by other alternatives? If there is something that you used to be able to do that is now no longer possible, please let us know!

Getting back into working on this again a bit more, I do seem to basically be facing a problem where it is currently no longer possible for the following function from src/utils.rs to be const in general:

/// An internal function for calculating pointer offsets as usizes, while accounting
/// directly for possible ZSTs. This is used specifically in the iterator implementations.
#[inline(always)]
pub(crate) /*const*/ fn distance_between<T>(dest: *const T, origin: *const T) -> usize {
  match size_of::<T>() {
    0 => (dest as usize).wrapping_sub(origin as usize),
    // Safety: this function is used strictly with linear inputs
    // where dest is known to come after origin.
    _ => unsafe { ptr_offset_from(dest, origin) as usize },
  }
}

I'm aware that this would have (in most cases at least) produced one of those any use of this value will cause an error compilation failures if it was actually used in a specifically const context where size_of::<T>() really was 0, however, that would have been fine with me as it would not have amounted to anything useful in that scenario anyways.

Basically I only actually cared about the const-context usability of the function as far as it usefully related to non-ZSTs, but still needed it to correctly handle ZSTs without panicking the way ptr_offset_from does if / when encountering them in regular non-const contexts, if that makes any sense.

[RalfJung]

Would casting the pointer to *const u8 and then calling ptr_offset_from work? Or is the problem that the pointers are not actually in-bounds of the same allocation in the ZST case?

Cc @oli-obk @fee1-dead in case they have some more ideas

[slightlyoutofphase]

To clarify, I guess the intrinsic form of ptr_offset_from doesn't actually "panic" on typed ZST pointers the way the normal offset_from method does, but rather just either crashes or returns a random number depending if you're in release mode or debug mode (not that any of that is the issue here specifically).

In any case, your *const u8 suggestion seems likely to amount to the same thing that the usize-cast subtraction did (which I believe was always just returning 0, as you'd expect when operating on a sizeless array of nothing at all). I'll give it a shot. Thanks!

[slightlyoutofphase]

Yep, that works perfectly. Wish I'd thought to do it that way in the first place, honestly! Thanks again. I'd say this particular issue can be consider fully solved, now.

[RalfJung]

Note the side-conditions of offset_from though: https://doc.rust-lang.org/nightly/std/primitive.pointer.html#method.offset_from.

In particular, (20usize as *const u8).offset_from(10usize as *const u8) is UB since these pointers are not both in-bounds of the same allocated object.