add vneg neon instructions #1087
Conversation
|
r? @Amanieu (rust-highfive has picked a reviewer for you, use r? to override) |
| let b: i64x1 = i64x1::new(0); | ||
| simd_sub(transmute(b), a) |
There was a problem hiding this comment.
Nice, I was actually looking into this one as well earlier and "gave up on it", because there didn't seem to be a way other than 0 - the_vec. Good to see I didn't miss anything xD
Though that still begs the question if there should just be a simd_neg instead, cause that's how you are supposed to get to this instruction via LLVM.
There was a problem hiding this comment.
It seems that core_arch does not have an instruction such as simd_neg, I don’t know how to add it, since I don’t know the compiler very well. :(
There was a problem hiding this comment.
You will need to add to the function at https://github.com/rust-lang/rust/blob/f24ce9b0140d9be5a336954e878d0c1522966bb8/compiler/rustc_codegen_llvm/src/intrinsic.rs#L730 It is probably similar to https://github.com/rust-lang/rust/blob/f24ce9b0140d9be5a336954e878d0c1522966bb8/compiler/rustc_codegen_llvm/src/intrinsic.rs#L1631-L1661 except a single argument instead of two arguments.
There was a problem hiding this comment.
Thanks for guiding. I will try later. I think the implementation of sub(0, a) is sufficient for now. Since the implementation in Clang is the same
There was a problem hiding this comment.
They are different if you look at the code for floating-point numbers. 0.0 - x and -x have different behavior for floats with regards to negative zeros, NaNs, etc.
There was a problem hiding this comment.
Thanks for pointing. It looks like we do need to add simd_neg.
|
It looks like we do need to add simd_neg. I suggest to close this PR, and then raise an issue to track its implementation. |
|
I'll close the PR for now.
|
|
@Amanieu I submitted the PR #rust-lang/rust/83271 for adding simd_neg to the compiler. Could I ask for your reviewing? |
Add simd_neg platform intrinsic Stdarch needs to add simd_neg to support the implementation of vneg neon instructions. Look [here](rust-lang/stdarch#1087)
Fixed-point and Floating-point negate instructions