NEON instructions on `aarch64`?

[torokati44]

I know the README says this:

... and on other architectures this is done by carefully writing functions so that LLVM hopefully does the right thing

But what magic incantation do I have to yell at it to make it actually vectorize my code written using wide? :/ Is NEON never emitted by rustc unless directly using the arch intrinsics?

[Lokathor]

https://rust-lang.github.io/packed_simd/perf-guide/target-feature/rustflags.html#target-feature

"+neon" should do it

[torokati44]

Yeah, thanks, I knew about that! And sorry, I should have mentioned that right at the beginning... The thing is, the neon target feature is actually enabled by default on aarch64-linux-android:

$ rustc --target aarch64-linux-android --print cfg
[...]
target_feature="neon"
target_feature="pmuv3"
[...]

So adding that flag won't do anything.

But even if I do add it, I get autovectorization on x86_64, but not on aarch64: https://godbolt.org/z/91TrexsvP Same thing with C++, with either Clang or GCC: https://godbolt.org/z/8qqfjWd7f

[Lokathor]

I'm not sure much else can be done right now unfortunately.

Other than like, join the portable simd working group and help them get that api to stable faster

[torokati44]

Yeah, that's one way, I guess... :/

But really, I shared your hope that LLVM would take care of it on its own, and I find it really weird that it does not, given how popular and important aarch64 is becoming...

[mcroomp]

aarch64 neon support is now in version 0.7.7

[Lokathor]

Actually, the main branch's Cargo.toml version doesn't get bumped until the end of development, so it's 0.7.8

But I released wide-0.7.8 just now, so it should be available by the time anyone reads this.