move to array simd #1422
move to array simd #1422
Conversation
|
Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @Amanieu (or someone else) soon. |
|
Gah, different architectures cause pain here. |
| pub(crate) const fn new($($elem_name: $elem_ty),*) -> Self { | ||
| $id($($elem_name),*) | ||
| pub(crate) const fn new(val: [$ety; $ecount]) -> Self { | ||
| $id(val) |
There was a problem hiding this comment.
From what I've heard, the plan is to make any direct field accesses (including calling the constructor like this) illegal on #[repr(simd)] types and only allow creation of the type and access to the inner array through transmute or pointer operations, so I think it might make sense to make the struct field private and to do something similar to rust-lang/portable-simd#339 or rust-lang/portable-simd#342 in new and splat.
There was a problem hiding this comment.
I think it'd also be fine for that to be a separate PR, if you prefer.
There was a problem hiding this comment.
direct field accesses (including calling the constructor like this) illegal
Emphasis on plan. This isn't happening with the current MCP, we are just removing the field-style #[repr(simd)]. In the longer term, I think scottmcm and other people have plans for that but not immediately.
|
|
||
| #[allow(clippy::use_self)] | ||
| impl $id { | ||
| #[inline(always)] | ||
| pub(crate) const fn new($($elem_name: $elem_ty),*) -> Self { |
There was a problem hiding this comment.
That is less ergonomic, shall we consider adding From implementation to support both tuple and array and keep new as it is now?
There was a problem hiding this comment.
Note that this was always pub(crate), so users couldn't call it anyway.
They can only be created using intrinsics (or memory stuff like transmutes) in stable today: https://doc.rust-lang.org/core/arch/x86/struct.__m128i.html.
(The portable-simd types do have nice safe public constructors. But these arch ones intentionally don't.)
Ezrashaw
force-pushed
the
move-to-array-simd
branch
from
461a87b to
f2d8f7c
Compare
|
@Ezrashaw With rust-lang/rust#110672 merged a while ago, are we good to go on this again? |
Yeah, the technical side of this pull request should be good to go. Personally, I haven't been contributing to Rust recently, so if someone wants to take this up, then free totally free. However, I'll try to find some time to complete this PR in maybe a week or so. |
As part of compiler-team/MCP#621, this repo must be updated to use the new-style array fields instead.
Note that a very recent nightly is needed which fixed a bug where inline assembly could not be used with the array style.
Note also that this cannot be pulled to
rust-lang/rustuntil beta updates with the bugfix.cc @scottmcm