Portable SIMD support

[RalfJung]

It'd be great if Miri would support at least a reasonable subset of Rust's portable SIMD API -- and from what I heard in the past, that is much more feasible than supporting things from core::arch. Now that rust-lang/rust#89167 has landed I'd like to add corresponding tests in Miri to ensure we cover the required intrinsics.

@workingjubilee what would be a good small-ish set of tests to start with here, that cover as much of the underlying support infrastructure as possible?

[RalfJung]

I guess we can start with the hello world example. :D

#![feature(portable_simd)]
use std::simd::*;

fn main() {
    let a = f32x4::splat(10.0);
    let b = f32x4::from_array([1.0, 2.0, 3.0, 4.0]);
    println!("{:?}", a + b);
}

error: unsupported operation: unimplemented intrinsic: simd_add
   --> /home/r/.rustup/toolchains/miri/lib/rustlib/src/rust/library/core/src/../../portable-simd/crates/core_simd/src/ops.rs:644:1
    |
644 | impl_float_ops! { f32, f64 }
    | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ unimplemented intrinsic: simd_add
    |
    = help: this is likely not a bug in the program; it indicates that the program performed an operation that the interpreter does not support
            
    = note: inside `core::core_simd::ops::<impl std::ops::Add for std::simd::Simd<f32, 4_usize>>::add` at /home/r/.rustup/toolchains/miri/lib/rustlib/src/rust/library/core/src/../../portable-simd/crates/core_simd/src/ops.rs:230:25
note: inside `main` at tests/run-pass/portable-simd.rs:7:22
   --> tests/run-pass/portable-simd.rs:7:22
    |
7   |     println!("{:?}", a + b);
    |                      ^^^^^

I also found this list of intrinsics so I guess once the basics work we'll want to ensure that we have a test that calls all these intrinsics.

[workingjubilee]

Yes, if you support that suite of LLVM-esque intrinsics, you will support essentially all of the higher-level constructs we derive from that, at least for the moment.

In general, we don't get very far with integers without the SIMD arithmetic and bitops.

Almost all higher-level constructs after there then gate on the availability of the masks, which require simd_lt and simd_{eq,ne} to construct, and then are used with simd_select.

[RalfJung]

Sounds good. Supporting most of that list should not be too hard -- we can basically apply the primitive MIR operators in a loop, I think (unless there are extra restrictions for the SIMD intrinsics that we have to check for).

at least for the moment.

Now I am worried. ;)

Is there a longer-term plan to rely more directly on the arch intrinsics, or significantly grow the set of these simd_* intrinsics?

[workingjubilee]

Is there a longer-term plan to rely more directly on the arch intrinsics,

No.

or significantly grow the set of these simd_* intrinsics?

Significantly? Not really. We will only add instructions to this "Rust SIMD Extension ISA" with great reluctance since we know the burden of supporting them is high. When we do, it will be aimed mostly at "we know this is fairly simple and there is a clear, direct lowering for it", e.g. we have discussed simd_min and simd_max before.

[workingjubilee]

Also by the time we are seriously looking at adding to this list we likely will have finally advanced on our plan of adding a """library""" in Rust that can handle these instructions to the compiler (by converting them to scalarized loops on arrays, yes), so that there is always a valid lowering a backend can use.

[RalfJung]

I see, sounds great! Miri being an interpreter it has slightly different needs than regular codegen backends, but that infrastructure can probably be made to handle both. :) (And I know a lot of formal verification people are also interested in having these things lowered away before their MIR/LLVM-based tools run.)