codegen: Fix struct padding alignment regression

[emilio]

See #3406.

[ojeda]

Thanks for looking at this so quickly.

What about aligned typedefs:

typedef int AlignedInt __attribute__((aligned(16)));

struct Outer {
    long long before;
    AlignedInt inner;
};

? 0.72.1 and 0.73.1 would give a different result (inner's offset (16 vs. 8) and Outer's size (32 vs. 16)) than 445dee5.

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[emilio]

@ojeda Hmm, fun one... This patch doesn't fix it, but that never worked quite right, tho, did it? 0.72.1 generates broken code for this:

typedef int AlignedInt __attribute__((aligned(16)));

struct Outer {
    int before;
    AlignedInt one;
    AlignedInt two;
};

Or even something simpler like:

typedef int AlignedInt __attribute__((aligned(8)));

struct Outer {
    int before;
    AlignedInt inner;
};

I guess the crux of the issue here is that Rust doesn't really have a way of provide the alignment on a typedef. For the struct member case, manually injecting the padding works, I suppose, but it won't work for things like arguments and such.

So to make this work properly I think we'd need to generate something like a wrapped type (but that is also not quite right, because you do want #[repr(transparent)].

So I think I'd consider that a technically pre-existing, separate bug. I can fix the struct layout by manually padding, though it's a bit ugly.

[ojeda]

Yeah, if those were broken, agreed, of course.

Just to clarify: the one I mentioned was "working" before, no? i.e. it is still a change (which perhaps doesn't break anybody, but it may be surprising).

Thanks!

[emilio]

Just to clarify: the one I mentioned was "working" before, no? i.e. it is still a change (which perhaps doesn't break anybody, but it may be surprising).

Yeah, it was working, but just by chance, because we were deciding to manually pad...

I'll look at fixing those in a follow-up, I think it might not be too annoying, at least from the struct layout perspective...