Array-walking is aligned

[workingjubilee]

The alignment-handling code has been correct for some time.

Remove the questioning assert, extract the align-up pattern into a function on layout::Align, and precompute the fixed-size offset to use for fixed-size elements in arrays.

[workingjubilee]

This addresses one assert failure in #1185

[eeeebbbbrrrr]

This looks fine. I guess the problem here is that we were continuing to assert something that's actually not true, or otherwise properly handled by the follow-up changes after zero-copy-arrays first landed?

[workingjubilee]

Correct. I am not even sure if it was still unhandled by the end of the first PR, but somewhere after the Box<dyn ChaChaSlide> refactor, it was definitely not a concern.

[workingjubilee]

I need to add an assert elsewhere, I think?

[workingjubilee]

Added appropriate debug asserts for "YES this pointer is inbounds of the array!" These will be lighter weight than certain other notorious debug asserts we had.

[workingjubilee]

Oh lol, this debug assert is wrong.

[thomcc]

Note that the comparison is wrong if ptr wrapped around the address space and other such edge cases. Might not be worth giving a shit about, since handling that correctly is annoying.

Actually, I have some code in a non-work project that cares about that, so I tweaked it for pgrx (hopefully not breaking it in the process) which gives https://gist.github.com/thomcc/b2730e3b1a56c4d8cdebf29814ea2351. It can probably be simplified for our use though (we perhaps don't need the size param), and has not been tested after I tweaked it for pgrx. Actually, Something like this on RawArray is simpler but still sufficient.

#[inline]
pub(crate) fn is_in_bounds(&self, p: *const u8) -> bool {
    let base = self.ptr.as_ptr();
    let mem_size = unsafe { crate::varlena::varsize_any(base.cast()) };
    // Note: This intentionally underflows if `p < base`.
    // If that happens, the wraparound will cause the result to
    // be greater than `mem_size` (we assume that the size postgres
    // gave us is reasonable, e.g. less than `isize::MAX`).
    (p as usize).wrapping_sub(base as usize) <= mem_size
}

[thomcc]

If you go that route, we have a couple other places where we have similar debug assertions comparing against end_ptr, might be worth fixing them too.

[workingjubilee]

Can we land that in a separate PR? I probably want it but we should discuss it as its own beast because that raises questions about more of our codebase and this patch has already scope-creeped a bit.

[thomcc]

Sure. We talked out of band (I don't think it raises other questions about most of the code base, just places where we might have broken address arithmetic)

[workingjubilee]

Carrying a full usize is unnecessary here and in any case makes the result that much slightly-less typesafe.