Dynamically select alignment of a `dyn Trait` local value

[pnkfelix]

Spawned off of #71416

Specifically this comment: #71416 (comment)

quoted here:

Unlike the other unsound features, this is one where we don't even have a plan for how to soundly implement it. Not sure how much I like keeping that around.

Can't we do something dynamic where we allocate size + (align - 1) and align it ourselves?
Also, this has been discussed before, hasn't it?

[pnkfelix]

@rustbot prioritize

[RalfJung]

@pnkfelix (in #71694)

which actually I guess would have to be done via alloca if we cannot put an upper-bound on align

size is also statically unknown here, so yes this needs alloca, that is indeed the entire point. ;)

[spastorino]

Assigning P-medium as discussed as part of the Prioritization Working Group process and removing I-prioritize.

Assigning to myself as a followup of #68304.

[ldr709]

Can't we do something dynamic where we allocate size + (align - 1) and align it ourselves?

Here's a small optimization: pick some default alignment, maybe 16, then allocate size + max(align - default_align, 0) bytes. If align >= default_align then align % default_align == 0 because they are both powers of two, so the offset required to align to align bytes is i * default_align for i < align / default_align, which is at most (align / default_align - 1) * default_align = align - default_align. If align < default_align then the allocation will already be properly aligned, no offset required.