Remove Simd: Index

[scottmcm]

(Inspired by this conversation on Discord https://discord.com/channels/273534239310479360/592856094527848449/1452407229843247197)

This PR removes the Simd<_, _>: Index and Simd<_, _>: IndexMut trait implementations, and instead adds get and set methods.

IMHO this is a good idea as v[1] = 2; looks innocuous but is actually surprisingly bad for codegen, as it forces the vector to be in-memory in order to return the &T -- and there's no way to return a T from Index. One can always v.as_mut_array()[1] = 2; if you want (that still works) but that extra speed bump is, I posit, a good thing to help push people to better methods.

What are those better methods? Well, this also adds Simd::set and Simd::get using the simd_{insert,extract}_dyn intrinsics internally. Thus instead of let x = v[i];, you'd write let x = v.get(i);. Doing that solves the optimization problem that the person was asking about on discord, and seeing that it did fix it (and that there's no use of simd_extract_dyn in the current library) is why I went and made this PR 🙂

Discussion topics:

• Should Index(Mut) be left there for a bit, to help smooth the transition? Can we even mark impls as deprecated?
• What should the names be? I like set/get for shortness, but maybe something domain-specific would be better. I considered using insert/extract, but https://doc.rust-lang.org/std/simd/struct.Simd.html#method.extract already exists. (It could be extract_vector or something, though.)
• Should set be &mut self or self? I noticed that https://doc.rust-lang.org/std/simd/struct.Mask.html#method.set is &mut self, but using &mut self also forces it through memory in codegen, albeit in a way much easier for the optimizer to remove.

[workingjubilee]

Oh, I thought we already added get and set? I guess I forgot.

[scottmcm]

@workingjubilee I was surprised too, TBH :P

I was expecting to say "oh, don't use as_mut_array, use set" and then it didn't exist...

[calebzulawski]

Looks good to me. I think the _dyn variations of the intrinsics are "new" which is why we didn't originally implement it (we did implement it for masks, since they can't implement Index). Implementing Index was always a bit controversial--I expected the optimizer to handle it but if not, removing it is probably best.

[scottmcm]

I think the _dyn variations of the intrinsics are "new"

They are, but helpfully they have fallback mir so they'll get working implementation for free, and that way so long as they've been in nightly long enough we don't need to worry about things like going and implementing them in cg_clif.

I expected the optimizer to handle it but if not

I'm surprised it doesn't, for what seems like a simple case, but overall I think discouraging address versions of things is good anyway -- we'll still have AsRef and AsMut and as_array and as_mut_array, so we're not blocking anything, just nudging the "wait, this is really not what simd is particularly good at" more strongly.

[scottmcm]

No idea what the problem was, but LLVM was very unhappy for some reason

rustc-LLVM ERROR: Cannot select: 0x7ff0c871e2a0: v16i8 = setcc 0x7ff0c871ee00, 0x7ff0ca14b000, setne:ch
  0x7ff0c871ee00: v8i16,ch = CopyFromReg 0x7ff0c871e8c0:1, Register:v8i16 %21
  0x7ff0ca14b000: v8i16 = xor 0x7ff0c8720a80, 0x7ff0c871db60
    0x7ff0c8720a80: v8i16,ch = CopyFromReg 0x7ff0c871e070:1, Register:v8i16 %11
    0x7ff0c871db60: v8i16 = splat_vector Constant:i32<-1>