Alignment not respected on the heap

[cgaebel]

We have a cool intrinsic called pref_align_of. We don't respect this when allocating memory. This means things like heap-allocated SIMD will be impossible to construct without some pretty crazy hackery.

Ideally, all calls to malloc should be replaced with calls to posix_memalign. This would exploit the undocumented "feature" of glibc that memory returned by posix_memalign can be realloc'd with the same alignment (but an assert and/or unit test for this should definitely be added).

[huonw]

This is something we need to consider in our Allocator trait design.

[Aatch]

It's worth noting on the SIMD support that there are instructions for doing unaligned loads (on x86 at least) into SIMD registers. So the issue is more that they are going to be slow, rather than be impossible.

[cgaebel]

Sometimes, I like to use the bottom few bits of pointers for various tags. This only works when alignment is respected in the allocator. It's also "morally" wrong that every type has an alignment property, but for all intents and purposes, heap allocated memory is just 8-byte aligned.

[alexcrichton]

Nominating.

[thestinger]

There's no equivalent to realloc with the posix_memalign API, so that presents an issue. If we go back to using jemalloc, then we can make use of the jemalloc-specific API for this.

@cgaebel: It's 16-byte aligned on x86_64 Linux. I don't know about what it is elsewhere.

[cgaebel]

that's only an implementation detail though, like being able to realloc memaligned memory. Technically it only needs to be "suitably aligned for any kind of variable", which I interpret as 8 bytes.

[thestinger]

It's an official part of the Linux ABI on x86_64. Technically, __m256 exists in the System V x86_64 ABI definition but it's marked optional so the minimum required alignment for malloc is 16 bytes.

[cgaebel]

What section is this in?

[thestinger]

3.1 (scalar types)

[thestinger]

(stack frames and arrays of at least 16 bytes also have a 16-byte alignment requirement - I think it was done for SSE)

[cgaebel]

I see. Interestingly enough, the only non-optional type there with more than 8-byte alignment is _Decimal128. Cool.

EDIT: oops forgot about long double.

[brson]

Required for SIMD

[pnkfelix]

We can fix this backwards-compatibly, I think. We need it for SIMD, and it should be kept in mind when developing the (still to-be-written) Allocator RFC.

But it is nonetheless a corner case that can be fixed later, so marking as P-low (and not putting on the 1.0 milestone).