Figure out rules for minimal alignment of system allocator

[RalfJung]

libstd assumes that the system allocator (which does not have an explicitly alignment parameter) still provides some minimal alignment, given by MIN_ALIGN:

// The minimum alignment guaranteed by the architecture. This value is used to
// add fast paths for low alignment values.
#[cfg(all(any(target_arch = "x86",
              target_arch = "arm",
              target_arch = "mips",
              target_arch = "powerpc",
              target_arch = "powerpc64",
              target_arch = "asmjs",
              target_arch = "wasm32")))]
pub const MIN_ALIGN: usize = 8;
#[cfg(all(any(target_arch = "x86_64",
              target_arch = "aarch64",
              target_arch = "mips64",
              target_arch = "s390x",
              target_arch = "sparc64")))]
pub const MIN_ALIGN: usize = 16;

However, at least on Unix, MIN_ALIGN is only exploited for allocations where the size is at least as big as the alignment. On Windows, OTOH, MIN_ALIGN is exploited regardless of the size of the allocation.

Does that mean that we can reduce the alignment we provide for small allocations?

[RalfJung]

This behavior on Unix platforms was introduced by rust-lang/rust@21d8992 to fix rust-lang/rust#45955 (I totally forgot I reported this...). Curiously, the Windows allocator was left unchanged.

Cc @SimonSapin

[RalfJung]

https://support.microsoft.com/en-us/help/286470/how-to-use-pageheap-exe-in-windows-xp-windows-2000-and-windows-server seems to indicate that Windows indeed guarantees alignment regardless of size:

The Windows heap managers (all versions) have always guaranteed that the heap allocations have a start address that is 8-byte aligned (on 64-bit platforms the alignment is 16-bytes).

it doesn't say anything about size.
So maybe we should make Miri give out more liberally aligned small allocations everywhere except on Windows.

[SimonSapin]

It’s been a while, but not changing Windows may have been an oversight. I did not look at all into the guarantees provided by HeapAlloc like I did for malloc.

[RalfJung]

See jemalloc/jemalloc#1533.