neon vst1_* intrinsics copy many more bytes than they should.

[HEnquist]

This was introduces in this commit:
907cfb2

For example:

stdarch/crates/core_arch/src/aarch64/neon/mod.rs

Line 1129 in 89b0e35

pub unsafe fn vst1q_f64(ptr: *mut f64, a: float64x2_t) {

pub unsafe fn vst1q_f64(ptr: *mut f64, a: float64x2_t) {
    copy_nonoverlapping(
        &a as *const float64x2_t as *const f64,
        ptr as *mut f64,
        size_of::<float64x2_t>(),
    )
}

The docs for copy_nonoverlapping says "Copies count * size_of::() bytes from src to dst"
The example above will get called with type T = f64 (with size 8 bytes), and count = size_of::<float64x2_t> = 16. So it will end up copying 8*16=128 bytes. That's quite a bit more than the 16 bytes it should have been.
The result is that it writes garbage to memory locations it should not touch.

https://doc.rust-lang.org/core/ptr/fn.copy_nonoverlapping.html

[hkratz]

There is no instance of copy_nonoverlapping() remaining in the armv7/aarch64 intrinsics, so that all should be fixed in master, see e.g.

stdarch/crates/core_arch/src/aarch64/neon/mod.rs

Lines 937 to 939 in 5fdbc47

	pub unsafe fn vst1q_f64(ptr: *mut f64, a: float64x2_t) {
	write_unaligned(ptr.cast(), a);
	}

But the stdarch submodule should be updated in the main Rust repo asap.

[HEnquist]

Ah thanks, the submodule thing threw me off. Let's keep this open then until the changes have made it into the main Rust repo.

[pthariensflame]

Seems like rust-lang/rust#89145 is now merged!

[HEnquist]

This works fine now in the latest nightly. Closing!