diff --git a/library/portable-simd/.github/workflows/ci.yml b/library/portable-simd/.github/workflows/ci.yml index d50dfa1be4c..acd47a3da72 100644 --- a/library/portable-simd/.github/workflows/ci.yml +++ b/library/portable-simd/.github/workflows/ci.yml @@ -241,6 +241,10 @@ jobs: - "--features std" - "--features generic_const_exprs" - "--features std --features generic_const_exprs" + - "--features all_lane_counts" + - "--features all_lane_counts --features std" + - "--features all_lane_counts --features generic_const_exprs" + - "--features all_lane_counts --features std --features generic_const_exprs" steps: - uses: actions/checkout@v2 diff --git a/library/portable-simd/README.md b/library/portable-simd/README.md index db0af2da606..e8ac600debe 100644 --- a/library/portable-simd/README.md +++ b/library/portable-simd/README.md @@ -24,19 +24,10 @@ or by setting up `rustup default nightly` or else with `cargo +nightly {build,te ```bash cargo new hellosimd ``` -to create a new crate. Edit `hellosimd/Cargo.toml` to be -```toml -[package] -name = "hellosimd" -version = "0.1.0" -edition = "2018" -[dependencies] -core_simd = { git = "https://github.com/rust-lang/portable-simd" } -``` - -and finally write this in `src/main.rs`: +to create a new crate. Finally write this in `src/main.rs`: ```rust -use core_simd::*; +#![feature(portable_simd)] +use std::simd::f32x4; fn main() { let a = f32x4::splat(10.0); let b = f32x4::from_array([1.0, 2.0, 3.0, 4.0]); @@ -44,24 +35,23 @@ fn main() { } ``` -Explanation: We import all the bindings from the crate with the first line. Then, we construct our SIMD vectors with methods like `splat` or `from_array`. Finally, we can use operators on them like `+` and the appropriate SIMD instructions will be carried out. When we run `cargo run` you should get `[11.0, 12.0, 13.0, 14.0]`. +Explanation: We construct our SIMD vectors with methods like `splat` or `from_array`. Next, we can use operators like `+` on them, and the appropriate SIMD instructions will be carried out. When we run `cargo run` you should get `[11.0, 12.0, 13.0, 14.0]`. -## Code Organization +## Supported vectors -Currently the crate is organized so that each element type is a file, and then the 64-bit, 128-bit, 256-bit, and 512-bit vectors using those types are contained in said file. - -All types are then exported as a single, flat module. +Currently, vectors may have up to 64 elements, but aliases are provided only up to 512-bit vectors. Depending on the size of the primitive type, the number of lanes the vector will have varies. For example, 128-bit vectors have four `f32` lanes and two `f64` lanes. The supported element types are as follows: * **Floating Point:** `f32`, `f64` -* **Signed Integers:** `i8`, `i16`, `i32`, `i64`, `i128`, `isize` -* **Unsigned Integers:** `u8`, `u16`, `u32`, `u64`, `u128`, `usize` -* **Masks:** `mask8`, `mask16`, `mask32`, `mask64`, `mask128`, `masksize` +* **Signed Integers:** `i8`, `i16`, `i32`, `i64`, `isize` (`i128` excluded) +* **Unsigned Integers:** `u8`, `u16`, `u32`, `u64`, `usize` (`u128` excluded) +* **Pointers:** `*const T` and `*mut T` (zero-sized metadata only) +* **Masks:** 8-bit, 16-bit, 32-bit, 64-bit, and `usize`-sized masks -Floating point, signed integers, and unsigned integers are the [primitive types](https://doc.rust-lang.org/core/primitive/index.html) you're already used to. -The `mask` types are "truthy" values, but they use the number of bits in their name instead of just 1 bit like a normal `bool` uses. +Floating point, signed integers, unsigned integers, and pointers are the [primitive types](https://doc.rust-lang.org/core/primitive/index.html) you're already used to. +The mask types have elements that are "truthy" values, like `bool`, but have an unspecified layout because different architectures prefer different layouts for mask types. [simd-guide]: ./beginners-guide.md [zulip-project-portable-simd]: https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd diff --git a/library/portable-simd/crates/core_simd/Cargo.toml b/library/portable-simd/crates/core_simd/Cargo.toml index 8a29cf15696..d1a3a515a7e 100644 --- a/library/portable-simd/crates/core_simd/Cargo.toml +++ b/library/portable-simd/crates/core_simd/Cargo.toml @@ -13,12 +13,11 @@ default = ["as_crate"] as_crate = [] std = [] generic_const_exprs = [] +all_lane_counts = [] -[target.'cfg(target_arch = "wasm32")'.dev-dependencies.wasm-bindgen] -version = "0.2" - -[dev-dependencies.wasm-bindgen-test] -version = "0.3" +[target.'cfg(target_arch = "wasm32")'.dev-dependencies] +wasm-bindgen = "0.2" +wasm-bindgen-test = "0.3" [dev-dependencies.proptest] version = "0.10" diff --git a/library/portable-simd/crates/core_simd/examples/README.md b/library/portable-simd/crates/core_simd/examples/README.md new file mode 100644 index 00000000000..82747f1b5a6 --- /dev/null +++ b/library/portable-simd/crates/core_simd/examples/README.md @@ -0,0 +1,13 @@ +### `stdsimd` examples + +This crate is a port of example uses of `stdsimd`, mostly taken from the `packed_simd` crate. + +The examples contain, as in the case of `dot_product.rs`, multiple ways of solving the problem, in order to show idiomatic uses of SIMD and iteration of performance designs. + +Run the tests with the command + +``` +cargo run --example dot_product +``` + +and verify the code for `dot_product.rs` on your machine. diff --git a/library/portable-simd/crates/core_simd/examples/dot_product.rs b/library/portable-simd/crates/core_simd/examples/dot_product.rs new file mode 100644 index 00000000000..391f08f55a0 --- /dev/null +++ b/library/portable-simd/crates/core_simd/examples/dot_product.rs @@ -0,0 +1,169 @@ +// Code taken from the `packed_simd` crate +// Run this code with `cargo test --example dot_product` +//use std::iter::zip; + +#![feature(array_chunks)] +#![feature(slice_as_chunks)] +// Add these imports to use the stdsimd library +#![feature(portable_simd)] +use core_simd::simd::*; + +// This is your barebones dot product implementation: +// Take 2 vectors, multiply them element wise and *then* +// go along the resulting array and add up the result. +// In the next example we will see if there +// is any difference to adding and multiplying in tandem. +pub fn dot_prod_scalar_0(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + + a.iter().zip(b.iter()).map(|(a, b)| a * b).sum() +} + +// When dealing with SIMD, it is very important to think about the amount +// of data movement and when it happens. We're going over simple computation examples here, and yet +// it is not trivial to understand what may or may not contribute to performance +// changes. Eventually, you will need tools to inspect the generated assembly and confirm your +// hypothesis and benchmarks - we will mention them later on. +// With the use of `fold`, we're doing a multiplication, +// and then adding it to the sum, one element from both vectors at a time. +pub fn dot_prod_scalar_1(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + a.iter() + .zip(b.iter()) + .fold(0.0, |a, zipped| a + zipped.0 * zipped.1) +} + +// We now move on to the SIMD implementations: notice the following constructs: +// `array_chunks::<4>`: mapping this over the vector will let use construct SIMD vectors +// `f32x4::from_array`: construct the SIMD vector from a slice +// `(a * b).reduce_sum()`: Multiply both f32x4 vectors together, and then reduce them. +// This approach essentially uses SIMD to produce a vector of length N/4 of all the products, +// and then add those with `sum()`. This is suboptimal. +// TODO: ASCII diagrams +pub fn dot_prod_simd_0(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + // TODO handle remainder when a.len() % 4 != 0 + a.array_chunks::<4>() + .map(|&a| f32x4::from_array(a)) + .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b))) + .map(|(a, b)| (a * b).reduce_sum()) + .sum() +} + +// There's some simple ways to improve the previous code: +// 1. Make a `zero` `f32x4` SIMD vector that we will be accumulating into +// So that there is only one `sum()` reduction when the last `f32x4` has been processed +// 2. Exploit Fused Multiply Add so that the multiplication, addition and sinking into the reduciton +// happen in the same step. +// If the arrays are large, minimizing the data shuffling will lead to great perf. +// If the arrays are small, handling the remainder elements when the length isn't a multiple of 4 +// Can become a problem. +pub fn dot_prod_simd_1(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + // TODO handle remainder when a.len() % 4 != 0 + a.array_chunks::<4>() + .map(|&a| f32x4::from_array(a)) + .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b))) + .fold(f32x4::splat(0.0), |acc, zipped| acc + zipped.0 * zipped.1) + .reduce_sum() +} + +// A lot of knowledgeable use of SIMD comes from knowing specific instructions that are +// available - let's try to use the `mul_add` instruction, which is the fused-multiply-add we were looking for. +use std_float::StdFloat; +pub fn dot_prod_simd_2(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + // TODO handle remainder when a.len() % 4 != 0 + let mut res = f32x4::splat(0.0); + a.array_chunks::<4>() + .map(|&a| f32x4::from_array(a)) + .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b))) + .for_each(|(a, b)| { + res = a.mul_add(b, res); + }); + res.reduce_sum() +} + +// Finally, we will write the same operation but handling the loop remainder. +const LANES: usize = 4; +pub fn dot_prod_simd_3(a: &[f32], b: &[f32]) -> f32 { + assert_eq!(a.len(), b.len()); + + let (a_extra, a_chunks) = a.as_rchunks(); + let (b_extra, b_chunks) = b.as_rchunks(); + + // These are always true, but for emphasis: + assert_eq!(a_chunks.len(), b_chunks.len()); + assert_eq!(a_extra.len(), b_extra.len()); + + let mut sums = [0.0; LANES]; + for ((x, y), d) in std::iter::zip(a_extra, b_extra).zip(&mut sums) { + *d = x * y; + } + + let mut sums = f32x4::from_array(sums); + std::iter::zip(a_chunks, b_chunks).for_each(|(x, y)| { + sums += f32x4::from_array(*x) * f32x4::from_array(*y); + }); + + sums.reduce_sum() +} + +// Finally, we present an iterator version for handling remainders in a scalar fashion at the end of the loop. +// Unfortunately, this is allocating 1 `XMM` register on the order of `~len(a)` - we'll see how we can get around it in the +// next example. +pub fn dot_prod_simd_4(a: &[f32], b: &[f32]) -> f32 { + let mut sum = a + .array_chunks::<4>() + .map(|&a| f32x4::from_array(a)) + .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b))) + .map(|(a, b)| a * b) + .fold(f32x4::splat(0.0), std::ops::Add::add) + .reduce_sum(); + let remain = a.len() - (a.len() % 4); + sum += a[remain..] + .iter() + .zip(&b[remain..]) + .map(|(a, b)| a * b) + .sum::(); + sum +} + +// This version allocates a single `XMM` register for accumulation, and the folds don't allocate on top of that. +// Notice the the use of `mul_add`, which can do a multiply and an add operation ber iteration. +pub fn dot_prod_simd_5(a: &[f32], b: &[f32]) -> f32 { + a.array_chunks::<4>() + .map(|&a| f32x4::from_array(a)) + .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b))) + .fold(f32x4::splat(0.), |acc, (a, b)| a.mul_add(b, acc)) + .reduce_sum() +} + +fn main() { + // Empty main to make cargo happy +} + +#[cfg(test)] +mod tests { + #[test] + fn smoke_test() { + use super::*; + let a: Vec = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]; + let b: Vec = vec![-8.0, -7.0, -6.0, -5.0, 4.0, 3.0, 2.0, 1.0]; + let x: Vec = [0.5; 1003].to_vec(); + let y: Vec = [2.0; 1003].to_vec(); + + // Basic check + assert_eq!(0.0, dot_prod_scalar_0(&a, &b)); + assert_eq!(0.0, dot_prod_scalar_1(&a, &b)); + assert_eq!(0.0, dot_prod_simd_0(&a, &b)); + assert_eq!(0.0, dot_prod_simd_1(&a, &b)); + assert_eq!(0.0, dot_prod_simd_2(&a, &b)); + assert_eq!(0.0, dot_prod_simd_3(&a, &b)); + assert_eq!(0.0, dot_prod_simd_4(&a, &b)); + assert_eq!(0.0, dot_prod_simd_5(&a, &b)); + + // We can handle vectors that are non-multiples of 4 + assert_eq!(1003.0, dot_prod_simd_3(&x, &y)); + } +} diff --git a/library/portable-simd/crates/core_simd/src/alias.rs b/library/portable-simd/crates/core_simd/src/alias.rs new file mode 100644 index 00000000000..23f121c4619 --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/alias.rs @@ -0,0 +1,227 @@ +macro_rules! number { + { 1 } => { "one" }; + { 2 } => { "two" }; + { 4 } => { "four" }; + { 8 } => { "eight" }; + { $x:literal } => { stringify!($x) }; +} + +macro_rules! plural { + { 1 } => { "" }; + { $x:literal } => { "s" }; +} + +macro_rules! alias { + { + $( + $element_ty:ty = { + $($alias:ident $num_elements:tt)* + } + )* + } => { + $( + $( + #[doc = concat!("A SIMD vector with ", number!($num_elements), " element", plural!($num_elements), " of type [`", stringify!($element_ty), "`].")] + #[allow(non_camel_case_types)] + pub type $alias = $crate::simd::Simd<$element_ty, $num_elements>; + )* + )* + } +} + +macro_rules! mask_alias { + { + $( + $element_ty:ty : $size:literal = { + $($alias:ident $num_elements:tt)* + } + )* + } => { + $( + $( + #[doc = concat!("A SIMD mask with ", number!($num_elements), " element", plural!($num_elements), " for vectors with ", $size, " element types.")] + /// + #[doc = concat!( + "The layout of this type is unspecified, and may change between platforms and/or Rust versions, and code should not assume that it is equivalent to `[", + stringify!($element_ty), "; ", $num_elements, "]`." + )] + #[allow(non_camel_case_types)] + pub type $alias = $crate::simd::Mask<$element_ty, $num_elements>; + )* + )* + } +} + +alias! { + i8 = { + i8x1 1 + i8x2 2 + i8x4 4 + i8x8 8 + i8x16 16 + i8x32 32 + i8x64 64 + } + + i16 = { + i16x1 1 + i16x2 2 + i16x4 4 + i16x8 8 + i16x16 16 + i16x32 32 + i16x64 64 + } + + i32 = { + i32x1 1 + i32x2 2 + i32x4 4 + i32x8 8 + i32x16 16 + i32x32 32 + i32x64 64 + } + + i64 = { + i64x1 1 + i64x2 2 + i64x4 4 + i64x8 8 + i64x16 16 + i64x32 32 + i64x64 64 + } + + isize = { + isizex1 1 + isizex2 2 + isizex4 4 + isizex8 8 + isizex16 16 + isizex32 32 + isizex64 64 + } + + u8 = { + u8x1 1 + u8x2 2 + u8x4 4 + u8x8 8 + u8x16 16 + u8x32 32 + u8x64 64 + } + + u16 = { + u16x1 1 + u16x2 2 + u16x4 4 + u16x8 8 + u16x16 16 + u16x32 32 + u16x64 64 + } + + u32 = { + u32x1 1 + u32x2 2 + u32x4 4 + u32x8 8 + u32x16 16 + u32x32 32 + u32x64 64 + } + + u64 = { + u64x1 1 + u64x2 2 + u64x4 4 + u64x8 8 + u64x16 16 + u64x32 32 + u64x64 64 + } + + usize = { + usizex1 1 + usizex2 2 + usizex4 4 + usizex8 8 + usizex16 16 + usizex32 32 + usizex64 64 + } + + f32 = { + f32x1 1 + f32x2 2 + f32x4 4 + f32x8 8 + f32x16 16 + f32x32 32 + f32x64 64 + } + + f64 = { + f64x1 1 + f64x2 2 + f64x4 4 + f64x8 8 + f64x16 16 + f64x32 32 + f64x64 64 + } +} + +mask_alias! { + i8 : "8-bit" = { + mask8x1 1 + mask8x2 2 + mask8x4 4 + mask8x8 8 + mask8x16 16 + mask8x32 32 + mask8x64 64 + } + + i16 : "16-bit" = { + mask16x1 1 + mask16x2 2 + mask16x4 4 + mask16x8 8 + mask16x16 16 + mask16x32 32 + mask16x64 64 + } + + i32 : "32-bit" = { + mask32x1 1 + mask32x2 2 + mask32x4 4 + mask32x8 8 + mask32x16 16 + mask32x32 32 + mask32x64 64 + } + + i64 : "64-bit" = { + mask64x1 1 + mask64x2 2 + mask64x4 4 + mask64x8 8 + mask64x16 16 + mask64x32 32 + mask64x64 64 + } + + isize : "pointer-sized" = { + masksizex1 1 + masksizex2 2 + masksizex4 4 + masksizex8 8 + masksizex16 16 + masksizex32 32 + masksizex64 64 + } +} diff --git a/library/portable-simd/crates/core_simd/src/cast.rs b/library/portable-simd/crates/core_simd/src/cast.rs new file mode 100644 index 00000000000..65a3f845ffc --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/cast.rs @@ -0,0 +1,55 @@ +use crate::simd::SimdElement; + +/// Supporting trait for `Simd::cast`. Typically doesn't need to be used directly. +/// +/// # Safety +/// Implementing this trait asserts that the type is a valid vector element for the `simd_cast` or +/// `simd_as` intrinsics. +pub unsafe trait SimdCast: SimdElement {} + +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for i8 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for i16 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for i32 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for i64 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for isize {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for u8 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for u16 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for u32 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for u64 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for usize {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for f32 {} +// Safety: primitive number types can be cast to other primitive number types +unsafe impl SimdCast for f64 {} + +/// Supporting trait for `Simd::cast_ptr`. Typically doesn't need to be used directly. +/// +/// # Safety +/// Implementing this trait asserts that the type is a valid vector element for the `simd_cast_ptr` +/// intrinsic. +pub unsafe trait SimdCastPtr {} + +// Safety: pointers can be cast to other pointer types +unsafe impl SimdCastPtr for *const U +where + U: core::ptr::Pointee, + T: core::ptr::Pointee, +{ +} +// Safety: pointers can be cast to other pointer types +unsafe impl SimdCastPtr for *mut U +where + U: core::ptr::Pointee, + T: core::ptr::Pointee, +{ +} diff --git a/library/portable-simd/crates/core_simd/src/elements.rs b/library/portable-simd/crates/core_simd/src/elements.rs index 701eb66b248..dc7f52a4d57 100644 --- a/library/portable-simd/crates/core_simd/src/elements.rs +++ b/library/portable-simd/crates/core_simd/src/elements.rs @@ -1,11 +1,15 @@ +mod const_ptr; mod float; mod int; +mod mut_ptr; mod uint; mod sealed { pub trait Sealed {} } +pub use const_ptr::*; pub use float::*; pub use int::*; +pub use mut_ptr::*; pub use uint::*; diff --git a/library/portable-simd/crates/core_simd/src/elements/const_ptr.rs b/library/portable-simd/crates/core_simd/src/elements/const_ptr.rs new file mode 100644 index 00000000000..0ef9802b5e2 --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/elements/const_ptr.rs @@ -0,0 +1,141 @@ +use super::sealed::Sealed; +use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount}; + +/// Operations on SIMD vectors of constant pointers. +pub trait SimdConstPtr: Copy + Sealed { + /// Vector of `usize` with the same number of lanes. + type Usize; + + /// Vector of `isize` with the same number of lanes. + type Isize; + + /// Vector of mutable pointers to the same type. + type MutPtr; + + /// Mask type used for manipulating this SIMD vector type. + type Mask; + + /// Returns `true` for each lane that is null. + fn is_null(self) -> Self::Mask; + + /// Changes constness without changing the type. + /// + /// Equivalent to calling [`pointer::cast_mut`] on each lane. + fn cast_mut(self) -> Self::MutPtr; + + /// Gets the "address" portion of the pointer. + /// + /// This method discards pointer semantic metadata, so the result cannot be + /// directly cast into a valid pointer. + /// + /// This method semantically discards *provenance* and + /// *address-space* information. To properly restore that information, use [`Self::with_addr`]. + /// + /// Equivalent to calling [`pointer::addr`] on each lane. + fn addr(self) -> Self::Usize; + + /// Creates a new pointer with the given address. + /// + /// This performs the same operation as a cast, but copies the *address-space* and + /// *provenance* of `self` to the new pointer. + /// + /// Equivalent to calling [`pointer::with_addr`] on each lane. + fn with_addr(self, addr: Self::Usize) -> Self; + + /// Gets the "address" portion of the pointer, and "exposes" the provenance part for future use + /// in [`Self::from_exposed_addr`]. + fn expose_addr(self) -> Self::Usize; + + /// Convert an address back to a pointer, picking up a previously "exposed" provenance. + /// + /// Equivalent to calling [`core::ptr::from_exposed_addr`] on each lane. + fn from_exposed_addr(addr: Self::Usize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_offset`] on each lane. + fn wrapping_offset(self, offset: Self::Isize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_add`] on each lane. + fn wrapping_add(self, count: Self::Usize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_sub`] on each lane. + fn wrapping_sub(self, count: Self::Usize) -> Self; +} + +impl Sealed for Simd<*const T, LANES> where + LaneCount: SupportedLaneCount +{ +} + +impl SimdConstPtr for Simd<*const T, LANES> +where + LaneCount: SupportedLaneCount, +{ + type Usize = Simd; + type Isize = Simd; + type MutPtr = Simd<*mut T, LANES>; + type Mask = Mask; + + #[inline] + fn is_null(self) -> Self::Mask { + Simd::splat(core::ptr::null()).simd_eq(self) + } + + #[inline] + fn cast_mut(self) -> Self::MutPtr { + self.cast_ptr() + } + + #[inline] + fn addr(self) -> Self::Usize { + // FIXME(strict_provenance_magic): I am magic and should be a compiler intrinsic. + // SAFETY: Pointer-to-integer transmutes are valid (if you are okay with losing the + // provenance). + unsafe { core::mem::transmute_copy(&self) } + } + + #[inline] + fn with_addr(self, addr: Self::Usize) -> Self { + // FIXME(strict_provenance_magic): I am magic and should be a compiler intrinsic. + // + // In the mean-time, this operation is defined to be "as if" it was + // a wrapping_offset, so we can emulate it as such. This should properly + // restore pointer provenance even under today's compiler. + self.cast_ptr::<*const u8>() + .wrapping_offset(addr.cast::() - self.addr().cast::()) + .cast_ptr() + } + + #[inline] + fn expose_addr(self) -> Self::Usize { + // Safety: `self` is a pointer vector + unsafe { intrinsics::simd_expose_addr(self) } + } + + #[inline] + fn from_exposed_addr(addr: Self::Usize) -> Self { + // Safety: `self` is a pointer vector + unsafe { intrinsics::simd_from_exposed_addr(addr) } + } + + #[inline] + fn wrapping_offset(self, count: Self::Isize) -> Self { + // Safety: simd_arith_offset takes a vector of pointers and a vector of offsets + unsafe { intrinsics::simd_arith_offset(self, count) } + } + + #[inline] + fn wrapping_add(self, count: Self::Usize) -> Self { + self.wrapping_offset(count.cast()) + } + + #[inline] + fn wrapping_sub(self, count: Self::Usize) -> Self { + self.wrapping_offset(-count.cast::()) + } +} diff --git a/library/portable-simd/crates/core_simd/src/elements/mut_ptr.rs b/library/portable-simd/crates/core_simd/src/elements/mut_ptr.rs new file mode 100644 index 00000000000..d87986b4a09 --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/elements/mut_ptr.rs @@ -0,0 +1,136 @@ +use super::sealed::Sealed; +use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount}; + +/// Operations on SIMD vectors of mutable pointers. +pub trait SimdMutPtr: Copy + Sealed { + /// Vector of `usize` with the same number of lanes. + type Usize; + + /// Vector of `isize` with the same number of lanes. + type Isize; + + /// Vector of constant pointers to the same type. + type ConstPtr; + + /// Mask type used for manipulating this SIMD vector type. + type Mask; + + /// Returns `true` for each lane that is null. + fn is_null(self) -> Self::Mask; + + /// Changes constness without changing the type. + /// + /// Equivalent to calling [`pointer::cast_const`] on each lane. + fn cast_const(self) -> Self::ConstPtr; + + /// Gets the "address" portion of the pointer. + /// + /// This method discards pointer semantic metadata, so the result cannot be + /// directly cast into a valid pointer. + /// + /// Equivalent to calling [`pointer::addr`] on each lane. + fn addr(self) -> Self::Usize; + + /// Creates a new pointer with the given address. + /// + /// This performs the same operation as a cast, but copies the *address-space* and + /// *provenance* of `self` to the new pointer. + /// + /// Equivalent to calling [`pointer::with_addr`] on each lane. + fn with_addr(self, addr: Self::Usize) -> Self; + + /// Gets the "address" portion of the pointer, and "exposes" the provenance part for future use + /// in [`Self::from_exposed_addr`]. + fn expose_addr(self) -> Self::Usize; + + /// Convert an address back to a pointer, picking up a previously "exposed" provenance. + /// + /// Equivalent to calling [`core::ptr::from_exposed_addr_mut`] on each lane. + fn from_exposed_addr(addr: Self::Usize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_offset`] on each lane. + fn wrapping_offset(self, offset: Self::Isize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_add`] on each lane. + fn wrapping_add(self, count: Self::Usize) -> Self; + + /// Calculates the offset from a pointer using wrapping arithmetic. + /// + /// Equivalent to calling [`pointer::wrapping_sub`] on each lane. + fn wrapping_sub(self, count: Self::Usize) -> Self; +} + +impl Sealed for Simd<*mut T, LANES> where LaneCount: SupportedLaneCount +{} + +impl SimdMutPtr for Simd<*mut T, LANES> +where + LaneCount: SupportedLaneCount, +{ + type Usize = Simd; + type Isize = Simd; + type ConstPtr = Simd<*const T, LANES>; + type Mask = Mask; + + #[inline] + fn is_null(self) -> Self::Mask { + Simd::splat(core::ptr::null_mut()).simd_eq(self) + } + + #[inline] + fn cast_const(self) -> Self::ConstPtr { + self.cast_ptr() + } + + #[inline] + fn addr(self) -> Self::Usize { + // FIXME(strict_provenance_magic): I am magic and should be a compiler intrinsic. + // SAFETY: Pointer-to-integer transmutes are valid (if you are okay with losing the + // provenance). + unsafe { core::mem::transmute_copy(&self) } + } + + #[inline] + fn with_addr(self, addr: Self::Usize) -> Self { + // FIXME(strict_provenance_magic): I am magic and should be a compiler intrinsic. + // + // In the mean-time, this operation is defined to be "as if" it was + // a wrapping_offset, so we can emulate it as such. This should properly + // restore pointer provenance even under today's compiler. + self.cast_ptr::<*mut u8>() + .wrapping_offset(addr.cast::() - self.addr().cast::()) + .cast_ptr() + } + + #[inline] + fn expose_addr(self) -> Self::Usize { + // Safety: `self` is a pointer vector + unsafe { intrinsics::simd_expose_addr(self) } + } + + #[inline] + fn from_exposed_addr(addr: Self::Usize) -> Self { + // Safety: `self` is a pointer vector + unsafe { intrinsics::simd_from_exposed_addr(addr) } + } + + #[inline] + fn wrapping_offset(self, count: Self::Isize) -> Self { + // Safety: simd_arith_offset takes a vector of pointers and a vector of offsets + unsafe { intrinsics::simd_arith_offset(self, count) } + } + + #[inline] + fn wrapping_add(self, count: Self::Usize) -> Self { + self.wrapping_offset(count.cast()) + } + + #[inline] + fn wrapping_sub(self, count: Self::Usize) -> Self { + self.wrapping_offset(-count.cast::()) + } +} diff --git a/library/portable-simd/crates/core_simd/src/eq.rs b/library/portable-simd/crates/core_simd/src/eq.rs index c7111f720a8..80763c07272 100644 --- a/library/portable-simd/crates/core_simd/src/eq.rs +++ b/library/portable-simd/crates/core_simd/src/eq.rs @@ -1,4 +1,6 @@ -use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdElement, SupportedLaneCount}; +use crate::simd::{ + intrinsics, LaneCount, Mask, Simd, SimdConstPtr, SimdElement, SimdMutPtr, SupportedLaneCount, +}; /// Parallel `PartialEq`. pub trait SimdPartialEq { @@ -71,3 +73,37 @@ macro_rules! impl_mask { } impl_mask! { i8, i16, i32, i64, isize } + +impl SimdPartialEq for Simd<*const T, LANES> +where + LaneCount: SupportedLaneCount, +{ + type Mask = Mask; + + #[inline] + fn simd_eq(self, other: Self) -> Self::Mask { + self.addr().simd_eq(other.addr()) + } + + #[inline] + fn simd_ne(self, other: Self) -> Self::Mask { + self.addr().simd_ne(other.addr()) + } +} + +impl SimdPartialEq for Simd<*mut T, LANES> +where + LaneCount: SupportedLaneCount, +{ + type Mask = Mask; + + #[inline] + fn simd_eq(self, other: Self) -> Self::Mask { + self.addr().simd_eq(other.addr()) + } + + #[inline] + fn simd_ne(self, other: Self) -> Self::Mask { + self.addr().simd_ne(other.addr()) + } +} diff --git a/library/portable-simd/crates/core_simd/src/fmt.rs b/library/portable-simd/crates/core_simd/src/fmt.rs index dbd9839c4bf..b7317969cbb 100644 --- a/library/portable-simd/crates/core_simd/src/fmt.rs +++ b/library/portable-simd/crates/core_simd/src/fmt.rs @@ -1,39 +1,21 @@ use crate::simd::{LaneCount, Simd, SimdElement, SupportedLaneCount}; use core::fmt; -macro_rules! impl_fmt_trait { - { $($trait:ident,)* } => { - $( - impl fmt::$trait for Simd - where - LaneCount: SupportedLaneCount, - T: SimdElement + fmt::$trait, - { - fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { - #[repr(transparent)] - struct Wrapper<'a, T: fmt::$trait>(&'a T); - - impl fmt::Debug for Wrapper<'_, T> { - fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { - self.0.fmt(f) - } - } - - f.debug_list() - .entries(self.as_array().iter().map(|x| Wrapper(x))) - .finish() - } - } - )* +impl fmt::Debug for Simd +where + LaneCount: SupportedLaneCount, + T: SimdElement + fmt::Debug, +{ + /// A `Simd` has a debug format like the one for `[T]`: + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd::Simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd::Simd; + /// let floats = Simd::::splat(-1.0); + /// assert_eq!(format!("{:?}", [-1.0; 4]), format!("{:?}", floats)); + /// ``` + #[inline] + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + <[T] as fmt::Debug>::fmt(self.as_array(), f) } } - -impl_fmt_trait! { - Debug, - Binary, - LowerExp, - UpperExp, - Octal, - LowerHex, - UpperHex, -} diff --git a/library/portable-simd/crates/core_simd/src/intrinsics.rs b/library/portable-simd/crates/core_simd/src/intrinsics.rs index 704e6ed0159..dd6698e2ba5 100644 --- a/library/portable-simd/crates/core_simd/src/intrinsics.rs +++ b/library/portable-simd/crates/core_simd/src/intrinsics.rs @@ -61,9 +61,6 @@ extern "platform-intrinsic" { /// xor pub(crate) fn simd_xor(x: T, y: T) -> T; - /// getelementptr (without inbounds) - pub(crate) fn simd_arith_offset(ptrs: T, offsets: U) -> T; - /// fptoui/fptosi/uitofp/sitofp /// casting floats to integers is truncating, so it is safe to convert values like e.g. 1.5 /// but the truncated value must fit in the target type or the result is poison. @@ -150,4 +147,17 @@ extern "platform-intrinsic" { pub(crate) fn simd_select(m: M, yes: T, no: T) -> T; #[allow(unused)] pub(crate) fn simd_select_bitmask(m: M, yes: T, no: T) -> T; + + /// getelementptr (without inbounds) + /// equivalent to wrapping_offset + pub(crate) fn simd_arith_offset(ptr: T, offset: U) -> T; + + /// equivalent to `T as U` semantics, specifically for pointers + pub(crate) fn simd_cast_ptr(ptr: T) -> U; + + /// expose a pointer as an address + pub(crate) fn simd_expose_addr(ptr: T) -> U; + + /// convert an exposed address back to a pointer + pub(crate) fn simd_from_exposed_addr(addr: T) -> U; } diff --git a/library/portable-simd/crates/core_simd/src/lane_count.rs b/library/portable-simd/crates/core_simd/src/lane_count.rs index 63723e2ec13..2b91eb9e800 100644 --- a/library/portable-simd/crates/core_simd/src/lane_count.rs +++ b/library/portable-simd/crates/core_simd/src/lane_count.rs @@ -23,24 +23,20 @@ pub trait SupportedLaneCount: Sealed { impl Sealed for LaneCount {} -impl SupportedLaneCount for LaneCount<1> { - type BitMask = [u8; 1]; -} -impl SupportedLaneCount for LaneCount<2> { - type BitMask = [u8; 1]; -} -impl SupportedLaneCount for LaneCount<4> { - type BitMask = [u8; 1]; -} -impl SupportedLaneCount for LaneCount<8> { - type BitMask = [u8; 1]; -} -impl SupportedLaneCount for LaneCount<16> { - type BitMask = [u8; 2]; -} -impl SupportedLaneCount for LaneCount<32> { - type BitMask = [u8; 4]; -} -impl SupportedLaneCount for LaneCount<64> { - type BitMask = [u8; 8]; +macro_rules! supported_lane_count { + ($($lanes:literal),+) => { + $( + impl SupportedLaneCount for LaneCount<$lanes> { + type BitMask = [u8; ($lanes + 7) / 8]; + } + )+ + }; } + +supported_lane_count!(1, 2, 4, 8, 16, 32, 64); +#[cfg(feature = "all_lane_counts")] +supported_lane_count!( + 3, 5, 6, 7, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, + 31, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, + 56, 57, 58, 59, 60, 61, 62, 63 +); diff --git a/library/portable-simd/crates/core_simd/src/lib.rs b/library/portable-simd/crates/core_simd/src/lib.rs index 715f258f617..e5307de2155 100644 --- a/library/portable-simd/crates/core_simd/src/lib.rs +++ b/library/portable-simd/crates/core_simd/src/lib.rs @@ -1,5 +1,8 @@ #![no_std] #![feature( + const_refs_to_cell, + const_maybe_uninit_as_mut_ptr, + const_mut_refs, convert_float_to_int, decl_macro, intra_doc_pointers, @@ -7,7 +10,9 @@ repr_simd, simd_ffi, staged_api, - stdsimd + stdsimd, + strict_provenance, + ptr_metadata )] #![cfg_attr(feature = "generic_const_exprs", feature(generic_const_exprs))] #![cfg_attr(feature = "generic_const_exprs", allow(incomplete_features))] @@ -19,4 +24,3 @@ #[path = "mod.rs"] mod core_simd; pub use self::core_simd::simd; -pub use simd::*; diff --git a/library/portable-simd/crates/core_simd/src/masks.rs b/library/portable-simd/crates/core_simd/src/masks.rs index c36c336d8a2..e58df80fca8 100644 --- a/library/portable-simd/crates/core_simd/src/masks.rs +++ b/library/portable-simd/crates/core_simd/src/masks.rs @@ -55,6 +55,7 @@ pub unsafe trait MaskElement: SimdElement + Sealed {} macro_rules! impl_element { { $ty:ty } => { impl Sealed for $ty { + #[inline] fn valid(value: Simd) -> bool where LaneCount: SupportedLaneCount, @@ -62,6 +63,7 @@ macro_rules! impl_element { (value.simd_eq(Simd::splat(0 as _)) | value.simd_eq(Simd::splat(-1 as _))).all() } + #[inline] fn eq(self, other: Self) -> bool { self == other } const TRUE: Self = -1; @@ -83,7 +85,9 @@ impl_element! { isize } /// /// Masks represent boolean inclusion/exclusion on a per-lane basis. /// -/// The layout of this type is unspecified. +/// The layout of this type is unspecified, and may change between platforms +/// and/or Rust versions, and code should not assume that it is equivalent to +/// `[T; LANES]`. #[repr(transparent)] pub struct Mask(mask_impl::Mask) where @@ -102,6 +106,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn clone(&self) -> Self { *self } @@ -113,11 +118,13 @@ where LaneCount: SupportedLaneCount, { /// Construct a mask by setting all lanes to the given value. + #[inline] pub fn splat(value: bool) -> Self { Self(mask_impl::Mask::splat(value)) } /// Converts an array of bools to a SIMD mask. + #[inline] pub fn from_array(array: [bool; LANES]) -> Self { // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of // true: 0b_0000_0001 @@ -134,6 +141,7 @@ where } /// Converts a SIMD mask to an array of bools. + #[inline] pub fn to_array(self) -> [bool; LANES] { // This follows mostly the same logic as from_array. // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of @@ -261,6 +269,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn from(array: [bool; LANES]) -> Self { Self::from_array(array) } @@ -271,6 +280,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn from(vector: Mask) -> Self { vector.to_array() } @@ -520,60 +530,6 @@ where } } -/// A mask for SIMD vectors with eight elements of 8 bits. -pub type mask8x8 = Mask; - -/// A mask for SIMD vectors with 16 elements of 8 bits. -pub type mask8x16 = Mask; - -/// A mask for SIMD vectors with 32 elements of 8 bits. -pub type mask8x32 = Mask; - -/// A mask for SIMD vectors with 64 elements of 8 bits. -pub type mask8x64 = Mask; - -/// A mask for SIMD vectors with four elements of 16 bits. -pub type mask16x4 = Mask; - -/// A mask for SIMD vectors with eight elements of 16 bits. -pub type mask16x8 = Mask; - -/// A mask for SIMD vectors with 16 elements of 16 bits. -pub type mask16x16 = Mask; - -/// A mask for SIMD vectors with 32 elements of 16 bits. -pub type mask16x32 = Mask; - -/// A mask for SIMD vectors with two elements of 32 bits. -pub type mask32x2 = Mask; - -/// A mask for SIMD vectors with four elements of 32 bits. -pub type mask32x4 = Mask; - -/// A mask for SIMD vectors with eight elements of 32 bits. -pub type mask32x8 = Mask; - -/// A mask for SIMD vectors with 16 elements of 32 bits. -pub type mask32x16 = Mask; - -/// A mask for SIMD vectors with two elements of 64 bits. -pub type mask64x2 = Mask; - -/// A mask for SIMD vectors with four elements of 64 bits. -pub type mask64x4 = Mask; - -/// A mask for SIMD vectors with eight elements of 64 bits. -pub type mask64x8 = Mask; - -/// A mask for SIMD vectors with two elements of pointer width. -pub type masksizex2 = Mask; - -/// A mask for SIMD vectors with four elements of pointer width. -pub type masksizex4 = Mask; - -/// A mask for SIMD vectors with eight elements of pointer width. -pub type masksizex8 = Mask; - macro_rules! impl_from { { $from:ty => $($to:ty),* } => { $( @@ -581,6 +537,7 @@ macro_rules! impl_from { where LaneCount: SupportedLaneCount, { + #[inline] fn from(value: Mask<$from, LANES>) -> Self { value.cast() } diff --git a/library/portable-simd/crates/core_simd/src/masks/bitmask.rs b/library/portable-simd/crates/core_simd/src/masks/bitmask.rs index 365ecc0a325..20465ba9b07 100644 --- a/library/portable-simd/crates/core_simd/src/masks/bitmask.rs +++ b/library/portable-simd/crates/core_simd/src/masks/bitmask.rs @@ -26,6 +26,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn clone(&self) -> Self { *self } @@ -36,6 +37,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn eq(&self, other: &Self) -> bool { self.0.as_ref() == other.0.as_ref() } @@ -46,6 +48,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn partial_cmp(&self, other: &Self) -> Option { self.0.as_ref().partial_cmp(other.0.as_ref()) } @@ -63,6 +66,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn cmp(&self, other: &Self) -> core::cmp::Ordering { self.0.as_ref().cmp(other.0.as_ref()) } diff --git a/library/portable-simd/crates/core_simd/src/masks/full_masks.rs b/library/portable-simd/crates/core_simd/src/masks/full_masks.rs index b5ba198e504..1d13c45b8e7 100644 --- a/library/portable-simd/crates/core_simd/src/masks/full_masks.rs +++ b/library/portable-simd/crates/core_simd/src/masks/full_masks.rs @@ -37,6 +37,7 @@ where T: MaskElement + PartialEq, LaneCount: SupportedLaneCount, { + #[inline] fn eq(&self, other: &Self) -> bool { self.0.eq(&other.0) } @@ -47,6 +48,7 @@ where T: MaskElement + PartialOrd, LaneCount: SupportedLaneCount, { + #[inline] fn partial_cmp(&self, other: &Self) -> Option { self.0.partial_cmp(&other.0) } @@ -64,6 +66,7 @@ where T: MaskElement + Ord, LaneCount: SupportedLaneCount, { + #[inline] fn cmp(&self, other: &Self) -> core::cmp::Ordering { self.0.cmp(&other.0) } @@ -262,6 +265,7 @@ where T: MaskElement, LaneCount: SupportedLaneCount, { + #[inline] fn from(value: Mask) -> Self { value.0 } diff --git a/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs b/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs index 2235f016c71..fc7d6b781f2 100644 --- a/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs +++ b/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs @@ -48,10 +48,12 @@ macro_rules! impl_integer_intrinsic { impl ToBitMask for Mask { type BitMask = $int; + #[inline] fn to_bitmask(self) -> $int { self.0.to_bitmask_integer() } + #[inline] fn from_bitmask(bitmask: $int) -> Self { Self(mask_impl::Mask::from_bitmask_integer(bitmask)) } @@ -83,10 +85,12 @@ where { const BYTES: usize = bitmask_len(LANES); + #[inline] fn to_bitmask_array(self) -> [u8; Self::BYTES] { self.0.to_bitmask_array() } + #[inline] fn from_bitmask_array(bitmask: [u8; Self::BYTES]) -> Self { Mask(mask_impl::Mask::from_bitmask_array(bitmask)) } diff --git a/library/portable-simd/crates/core_simd/src/mod.rs b/library/portable-simd/crates/core_simd/src/mod.rs index b472aa3abe2..35c659b7a42 100644 --- a/library/portable-simd/crates/core_simd/src/mod.rs +++ b/library/portable-simd/crates/core_simd/src/mod.rs @@ -6,6 +6,8 @@ pub(crate) mod intrinsics; #[cfg(feature = "generic_const_exprs")] mod to_bytes; +mod alias; +mod cast; mod elements; mod eq; mod fmt; @@ -15,6 +17,7 @@ mod masks; mod ops; mod ord; mod select; +mod swizzle_dyn; mod vector; mod vendor; @@ -22,11 +25,14 @@ mod vendor; pub mod simd { pub(crate) use crate::core_simd::intrinsics; + pub use crate::core_simd::alias::*; + pub use crate::core_simd::cast::*; pub use crate::core_simd::elements::*; pub use crate::core_simd::eq::*; pub use crate::core_simd::lane_count::{LaneCount, SupportedLaneCount}; pub use crate::core_simd::masks::*; pub use crate::core_simd::ord::*; pub use crate::core_simd::swizzle::*; + pub use crate::core_simd::swizzle_dyn::*; pub use crate::core_simd::vector::*; } diff --git a/library/portable-simd/crates/core_simd/src/ord.rs b/library/portable-simd/crates/core_simd/src/ord.rs index 9a87bc2e344..1ae9cd061fb 100644 --- a/library/portable-simd/crates/core_simd/src/ord.rs +++ b/library/portable-simd/crates/core_simd/src/ord.rs @@ -1,4 +1,6 @@ -use crate::simd::{intrinsics, LaneCount, Mask, Simd, SimdPartialEq, SupportedLaneCount}; +use crate::simd::{ + intrinsics, LaneCount, Mask, Simd, SimdConstPtr, SimdMutPtr, SimdPartialEq, SupportedLaneCount, +}; /// Parallel `PartialOrd`. pub trait SimdPartialOrd: SimdPartialEq { @@ -211,3 +213,101 @@ macro_rules! impl_mask { } impl_mask! { i8, i16, i32, i64, isize } + +impl SimdPartialOrd for Simd<*const T, LANES> +where + LaneCount: SupportedLaneCount, +{ + #[inline] + fn simd_lt(self, other: Self) -> Self::Mask { + self.addr().simd_lt(other.addr()) + } + + #[inline] + fn simd_le(self, other: Self) -> Self::Mask { + self.addr().simd_le(other.addr()) + } + + #[inline] + fn simd_gt(self, other: Self) -> Self::Mask { + self.addr().simd_gt(other.addr()) + } + + #[inline] + fn simd_ge(self, other: Self) -> Self::Mask { + self.addr().simd_ge(other.addr()) + } +} + +impl SimdOrd for Simd<*const T, LANES> +where + LaneCount: SupportedLaneCount, +{ + #[inline] + fn simd_max(self, other: Self) -> Self { + self.simd_lt(other).select(other, self) + } + + #[inline] + fn simd_min(self, other: Self) -> Self { + self.simd_gt(other).select(other, self) + } + + #[inline] + fn simd_clamp(self, min: Self, max: Self) -> Self { + assert!( + min.simd_le(max).all(), + "each lane in `min` must be less than or equal to the corresponding lane in `max`", + ); + self.simd_max(min).simd_min(max) + } +} + +impl SimdPartialOrd for Simd<*mut T, LANES> +where + LaneCount: SupportedLaneCount, +{ + #[inline] + fn simd_lt(self, other: Self) -> Self::Mask { + self.addr().simd_lt(other.addr()) + } + + #[inline] + fn simd_le(self, other: Self) -> Self::Mask { + self.addr().simd_le(other.addr()) + } + + #[inline] + fn simd_gt(self, other: Self) -> Self::Mask { + self.addr().simd_gt(other.addr()) + } + + #[inline] + fn simd_ge(self, other: Self) -> Self::Mask { + self.addr().simd_ge(other.addr()) + } +} + +impl SimdOrd for Simd<*mut T, LANES> +where + LaneCount: SupportedLaneCount, +{ + #[inline] + fn simd_max(self, other: Self) -> Self { + self.simd_lt(other).select(other, self) + } + + #[inline] + fn simd_min(self, other: Self) -> Self { + self.simd_gt(other).select(other, self) + } + + #[inline] + fn simd_clamp(self, min: Self, max: Self) -> Self { + assert!( + min.simd_le(max).all(), + "each lane in `min` must be less than or equal to the corresponding lane in `max`", + ); + self.simd_max(min).simd_min(max) + } +} diff --git a/library/portable-simd/crates/core_simd/src/swizzle.rs b/library/portable-simd/crates/core_simd/src/swizzle.rs index 22999d24950..68f20516cf5 100644 --- a/library/portable-simd/crates/core_simd/src/swizzle.rs +++ b/library/portable-simd/crates/core_simd/src/swizzle.rs @@ -265,16 +265,13 @@ where /// Interleave two vectors. /// - /// Produces two vectors with lanes taken alternately from `self` and `other`. + /// The resulting vectors contain lanes taken alternatively from `self` and `other`, first + /// filling the first result, and then the second. /// - /// The first result contains the first `LANES / 2` lanes from `self` and `other`, - /// alternating, starting with the first lane of `self`. - /// - /// The second result contains the last `LANES / 2` lanes from `self` and `other`, - /// alternating, starting with the lane `LANES / 2` from the start of `self`. + /// The reverse of this operation is [`Simd::deinterleave`]. /// /// ``` - /// #![feature(portable_simd)] + /// # #![feature(portable_simd)] /// # use core::simd::Simd; /// let a = Simd::from_array([0, 1, 2, 3]); /// let b = Simd::from_array([4, 5, 6, 7]); @@ -285,29 +282,17 @@ where #[inline] #[must_use = "method returns a new vector and does not mutate the original inputs"] pub fn interleave(self, other: Self) -> (Self, Self) { - const fn lo() -> [Which; LANES] { + const fn interleave(high: bool) -> [Which; LANES] { let mut idx = [Which::First(0); LANES]; let mut i = 0; while i < LANES { - let offset = i / 2; - idx[i] = if i % 2 == 0 { - Which::First(offset) + // Treat the source as a concatenated vector + let dst_index = if high { i + LANES } else { i }; + let src_index = dst_index / 2 + (dst_index % 2) * LANES; + idx[i] = if src_index < LANES { + Which::First(src_index) } else { - Which::Second(offset) - }; - i += 1; - } - idx - } - const fn hi() -> [Which; LANES] { - let mut idx = [Which::First(0); LANES]; - let mut i = 0; - while i < LANES { - let offset = (LANES + i) / 2; - idx[i] = if i % 2 == 0 { - Which::First(offset) - } else { - Which::Second(offset) + Which::Second(src_index % LANES) }; i += 1; } @@ -318,11 +303,11 @@ where struct Hi; impl Swizzle2 for Lo { - const INDEX: [Which; LANES] = lo::(); + const INDEX: [Which; LANES] = interleave::(false); } impl Swizzle2 for Hi { - const INDEX: [Which; LANES] = hi::(); + const INDEX: [Which; LANES] = interleave::(true); } (Lo::swizzle2(self, other), Hi::swizzle2(self, other)) @@ -336,8 +321,10 @@ where /// The second result takes every other lane of `self` and then `other`, starting with /// the second lane. /// + /// The reverse of this operation is [`Simd::interleave`]. + /// /// ``` - /// #![feature(portable_simd)] + /// # #![feature(portable_simd)] /// # use core::simd::Simd; /// let a = Simd::from_array([0, 4, 1, 5]); /// let b = Simd::from_array([2, 6, 3, 7]); @@ -348,22 +335,17 @@ where #[inline] #[must_use = "method returns a new vector and does not mutate the original inputs"] pub fn deinterleave(self, other: Self) -> (Self, Self) { - const fn even() -> [Which; LANES] { + const fn deinterleave(second: bool) -> [Which; LANES] { let mut idx = [Which::First(0); LANES]; let mut i = 0; - while i < LANES / 2 { - idx[i] = Which::First(2 * i); - idx[i + LANES / 2] = Which::Second(2 * i); - i += 1; - } - idx - } - const fn odd() -> [Which; LANES] { - let mut idx = [Which::First(0); LANES]; - let mut i = 0; - while i < LANES / 2 { - idx[i] = Which::First(2 * i + 1); - idx[i + LANES / 2] = Which::Second(2 * i + 1); + while i < LANES { + // Treat the source as a concatenated vector + let src_index = i * 2 + second as usize; + idx[i] = if src_index < LANES { + Which::First(src_index) + } else { + Which::Second(src_index % LANES) + }; i += 1; } idx @@ -373,11 +355,11 @@ where struct Odd; impl Swizzle2 for Even { - const INDEX: [Which; LANES] = even::(); + const INDEX: [Which; LANES] = deinterleave::(false); } impl Swizzle2 for Odd { - const INDEX: [Which; LANES] = odd::(); + const INDEX: [Which; LANES] = deinterleave::(true); } (Even::swizzle2(self, other), Odd::swizzle2(self, other)) diff --git a/library/portable-simd/crates/core_simd/src/swizzle_dyn.rs b/library/portable-simd/crates/core_simd/src/swizzle_dyn.rs new file mode 100644 index 00000000000..3eb80d5dca1 --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/swizzle_dyn.rs @@ -0,0 +1,157 @@ +use crate::simd::{LaneCount, Simd, SupportedLaneCount}; +use core::mem; + +impl Simd +where + LaneCount: SupportedLaneCount, +{ + /// Swizzle a vector of bytes according to the index vector. + /// Indices within range select the appropriate byte. + /// Indices "out of bounds" instead select 0. + /// + /// Note that the current implementation is selected during build-time + /// of the standard library, so `cargo build -Zbuild-std` may be necessary + /// to unlock better performance, especially for larger vectors. + /// A planned compiler improvement will enable using `#[target_feature]` instead. + #[inline] + pub fn swizzle_dyn(self, idxs: Simd) -> Self { + #![allow(unused_imports, unused_unsafe)] + #[cfg(target_arch = "aarch64")] + use core::arch::aarch64::{uint8x8_t, vqtbl1q_u8, vtbl1_u8}; + #[cfg(all(target_arch = "arm", target_feature = "v7"))] + use core::arch::arm::{uint8x8_t, vtbl1_u8}; + #[cfg(target_arch = "wasm32")] + use core::arch::wasm32 as wasm; + #[cfg(target_arch = "x86")] + use core::arch::x86; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64 as x86; + // SAFETY: Intrinsics covered by cfg + unsafe { + match N { + #[cfg(target_feature = "neon")] + 8 => transize(vtbl1_u8, self, idxs), + #[cfg(target_feature = "ssse3")] + 16 => transize(x86::_mm_shuffle_epi8, self, idxs), + #[cfg(target_feature = "simd128")] + 16 => transize(wasm::i8x16_swizzle, self, idxs), + #[cfg(all(target_arch = "aarch64", target_feature = "neon"))] + 16 => transize(vqtbl1q_u8, self, idxs), + #[cfg(all(target_feature = "avx2", not(target_feature = "avx512vbmi")))] + 32 => transize_raw(avx2_pshufb, self, idxs), + #[cfg(target_feature = "avx512vl,avx512vbmi")] + 32 => transize(x86::_mm256_permutexvar_epi8, self, idxs), + // Notable absence: avx512bw shuffle + // If avx512bw is available, odds of avx512vbmi are good + // FIXME: initial AVX512VBMI variant didn't actually pass muster + // #[cfg(target_feature = "avx512vbmi")] + // 64 => transize(x86::_mm512_permutexvar_epi8, self, idxs), + _ => { + let mut array = [0; N]; + for (i, k) in idxs.to_array().into_iter().enumerate() { + if (k as usize) < N { + array[i] = self[k as usize]; + }; + } + array.into() + } + } + } + } +} + +/// "vpshufb like it was meant to be" on AVX2 +/// +/// # Safety +/// This requires AVX2 to work +#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] +#[target_feature(enable = "avx2")] +#[allow(unused)] +#[inline] +#[allow(clippy::let_and_return)] +unsafe fn avx2_pshufb(bytes: Simd, idxs: Simd) -> Simd { + use crate::simd::SimdPartialOrd; + #[cfg(target_arch = "x86")] + use core::arch::x86; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64 as x86; + use x86::_mm256_permute2x128_si256 as avx2_cross_shuffle; + use x86::_mm256_shuffle_epi8 as avx2_half_pshufb; + let mid = Simd::splat(16u8); + let high = mid + mid; + // SAFETY: Caller promised AVX2 + unsafe { + // This is ordering sensitive, and LLVM will order these how you put them. + // Most AVX2 impls use ~5 "ports", and only 1 or 2 are capable of permutes. + // But the "compose" step will lower to ops that can also use at least 1 other port. + // So this tries to break up permutes so composition flows through "open" ports. + // Comparative benches should be done on multiple AVX2 CPUs before reordering this + + let hihi = avx2_cross_shuffle::<0x11>(bytes.into(), bytes.into()); + let hi_shuf = Simd::from(avx2_half_pshufb( + hihi, // duplicate the vector's top half + idxs.into(), // so that using only 4 bits of an index still picks bytes 16-31 + )); + // A zero-fill during the compose step gives the "all-Neon-like" OOB-is-0 semantics + let compose = idxs.simd_lt(high).select(hi_shuf, Simd::splat(0)); + let lolo = avx2_cross_shuffle::<0x00>(bytes.into(), bytes.into()); + let lo_shuf = Simd::from(avx2_half_pshufb(lolo, idxs.into())); + // Repeat, then pick indices < 16, overwriting indices 0-15 from previous compose step + let compose = idxs.simd_lt(mid).select(lo_shuf, compose); + compose + } +} + +/// This sets up a call to an architecture-specific function, and in doing so +/// it persuades rustc that everything is the correct size. Which it is. +/// This would not be needed if one could convince Rust that, by matching on N, +/// N is that value, and thus it would be valid to substitute e.g. 16. +/// +/// # Safety +/// The correctness of this function hinges on the sizes agreeing in actuality. +#[allow(dead_code)] +#[inline(always)] +unsafe fn transize( + f: unsafe fn(T, T) -> T, + bytes: Simd, + idxs: Simd, +) -> Simd +where + LaneCount: SupportedLaneCount, +{ + let idxs = zeroing_idxs(idxs); + // SAFETY: Same obligation to use this function as to use mem::transmute_copy. + unsafe { mem::transmute_copy(&f(mem::transmute_copy(&bytes), mem::transmute_copy(&idxs))) } +} + +/// Make indices that yield 0 for this architecture +#[inline(always)] +fn zeroing_idxs(idxs: Simd) -> Simd +where + LaneCount: SupportedLaneCount, +{ + // On x86, make sure the top bit is set. + #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] + let idxs = { + use crate::simd::SimdPartialOrd; + idxs.simd_lt(Simd::splat(N as u8)) + .select(idxs, Simd::splat(u8::MAX)) + }; + // Simply do nothing on most architectures. + idxs +} + +/// As transize but no implicit call to `zeroing_idxs`. +#[allow(dead_code)] +#[inline(always)] +unsafe fn transize_raw( + f: unsafe fn(T, T) -> T, + bytes: Simd, + idxs: Simd, +) -> Simd +where + LaneCount: SupportedLaneCount, +{ + // SAFETY: Same obligation to use this function as to use mem::transmute_copy. + unsafe { mem::transmute_copy(&f(mem::transmute_copy(&bytes), mem::transmute_copy(&idxs))) } +} diff --git a/library/portable-simd/crates/core_simd/src/vector.rs b/library/portable-simd/crates/core_simd/src/vector.rs index d52d1ac4d3a..3809cc96151 100644 --- a/library/portable-simd/crates/core_simd/src/vector.rs +++ b/library/portable-simd/crates/core_simd/src/vector.rs @@ -1,60 +1,63 @@ -mod float; -mod int; -mod uint; - -pub use float::*; -pub use int::*; -pub use uint::*; - -// Vectors of pointers are not for public use at the current time. -pub(crate) mod ptr; - use crate::simd::{ - intrinsics, LaneCount, Mask, MaskElement, SimdPartialOrd, SupportedLaneCount, Swizzle, + intrinsics, LaneCount, Mask, MaskElement, SimdCast, SimdCastPtr, SimdConstPtr, SimdMutPtr, + SimdPartialOrd, SupportedLaneCount, Swizzle, }; +use core::convert::{TryFrom, TryInto}; -/// A SIMD vector of `LANES` elements of type `T`. `Simd` has the same shape as [`[T; N]`](array), but operates like `T`. +/// A SIMD vector with the shape of `[T; N]` but the operations of `T`. /// -/// Two vectors of the same type and length will, by convention, support the operators (+, *, etc.) that `T` does. -/// These take the lanes at each index on the left-hand side and right-hand side, perform the operation, -/// and return the result in the same lane in a vector of equal size. For a given operator, this is equivalent to zipping -/// the two arrays together and mapping the operator over each lane. +/// `Simd` supports the operators (+, *, etc.) that `T` does in "elementwise" fashion. +/// These take the element at each index from the left-hand side and right-hand side, +/// perform the operation, then return the result in the same index in a vector of equal size. +/// However, `Simd` differs from normal iteration and normal arrays: +/// - `Simd` executes `N` operations in a single step with no `break`s +/// - `Simd` can have an alignment greater than `T`, for better mechanical sympathy +/// +/// By always imposing these constraints on `Simd`, it is easier to compile elementwise operations +/// into machine instructions that can themselves be executed in parallel. /// /// ```rust -/// # #![feature(array_zip, portable_simd)] +/// # #![feature(portable_simd)] /// # use core::simd::{Simd}; -/// let a0: [i32; 4] = [-2, 0, 2, 4]; -/// let a1 = [10, 9, 8, 7]; -/// let zm_add = a0.zip(a1).map(|(lhs, rhs)| lhs + rhs); -/// let zm_mul = a0.zip(a1).map(|(lhs, rhs)| lhs * rhs); +/// # use core::array; +/// let a: [i32; 4] = [-2, 0, 2, 4]; +/// let b = [10, 9, 8, 7]; +/// let sum = array::from_fn(|i| a[i] + b[i]); +/// let prod = array::from_fn(|i| a[i] * b[i]); /// /// // `Simd` implements `From<[T; N]>` -/// let (v0, v1) = (Simd::from(a0), Simd::from(a1)); +/// let (v, w) = (Simd::from(a), Simd::from(b)); /// // Which means arrays implement `Into>`. -/// assert_eq!(v0 + v1, zm_add.into()); -/// assert_eq!(v0 * v1, zm_mul.into()); +/// assert_eq!(v + w, sum.into()); +/// assert_eq!(v * w, prod.into()); /// ``` /// -/// `Simd` with integers has the quirk that these operations are also inherently wrapping, as if `T` was [`Wrapping`]. +/// +/// `Simd` with integer elements treats operators as wrapping, as if `T` was [`Wrapping`]. /// Thus, `Simd` does not implement `wrapping_add`, because that is the default behavior. /// This means there is no warning on overflows, even in "debug" builds. /// For most applications where `Simd` is appropriate, it is "not a bug" to wrap, /// and even "debug builds" are unlikely to tolerate the loss of performance. /// You may want to consider using explicitly checked arithmetic if such is required. -/// Division by zero still causes a panic, so you may want to consider using floating point numbers if that is unacceptable. +/// Division by zero on integers still causes a panic, so +/// you may want to consider using `f32` or `f64` if that is unacceptable. /// /// [`Wrapping`]: core::num::Wrapping /// /// # Layout -/// `Simd` has a layout similar to `[T; N]` (identical "shapes"), but with a greater alignment. +/// `Simd` has a layout similar to `[T; N]` (identical "shapes"), with a greater alignment. /// `[T; N]` is aligned to `T`, but `Simd` will have an alignment based on both `T` and `N`. -/// It is thus sound to [`transmute`] `Simd` to `[T; N]`, and will typically optimize to zero cost, -/// but the reverse transmutation is more likely to require a copy the compiler cannot simply elide. +/// Thus it is sound to [`transmute`] `Simd` to `[T; N]` and should optimize to "zero cost", +/// but the reverse transmutation may require a copy the compiler cannot simply elide. /// /// # ABI "Features" -/// Due to Rust's safety guarantees, `Simd` is currently passed to and from functions via memory, not SIMD registers, -/// except as an optimization. `#[inline]` hints are recommended on functions that accept `Simd` or return it. -/// The need for this may be corrected in the future. +/// Due to Rust's safety guarantees, `Simd` is currently passed and returned via memory, +/// not SIMD registers, except as an optimization. Using `#[inline]` on functions that accept +/// `Simd` or return it is recommended, at the cost of code generation time, as +/// inlining SIMD-using functions can omit a large function prolog or epilog and thus +/// improve both speed and code size. The need for this may be corrected in the future. +/// +/// Using `#[inline(always)]` still requires additional care. /// /// # Safe SIMD with Unsafe Rust /// @@ -65,18 +68,22 @@ use crate::simd::{ /// Thus, when using `unsafe` Rust to read and write `Simd` through [raw pointers], it is a good idea to first try with /// [`read_unaligned`] and [`write_unaligned`]. This is because: /// - [`read`] and [`write`] require full alignment (in this case, `Simd`'s alignment) -/// - the likely source for reading or destination for writing `Simd` is [`[T]`](slice) and similar types, aligned to `T` -/// - combining these actions would violate the `unsafe` contract and explode the program into a puff of **undefined behavior** -/// - the compiler can implicitly adjust layouts to make unaligned reads or writes fully aligned if it sees the optimization -/// - most contemporary processors suffer no performance penalty for "unaligned" reads and writes that are aligned at runtime +/// - `Simd` is often read from or written to [`[T]`](slice) and other types aligned to `T` +/// - combining these actions violates the `unsafe` contract and explodes the program into +/// a puff of **undefined behavior** +/// - the compiler can implicitly adjust layouts to make unaligned reads or writes fully aligned +/// if it sees the optimization +/// - most contemporary processors with "aligned" and "unaligned" read and write instructions +/// exhibit no performance difference if the "unaligned" variant is aligned at runtime /// -/// By imposing less obligations, unaligned functions are less likely to make the program unsound, +/// Less obligations mean unaligned reads and writes are less likely to make the program unsound, /// and may be just as fast as stricter alternatives. -/// When trying to guarantee alignment, [`[T]::as_simd`][as_simd] is an option for converting `[T]` to `[Simd]`, -/// and allows soundly operating on an aligned SIMD body, but it may cost more time when handling the scalar head and tail. -/// If these are not sufficient, then it is most ideal to design data structures to be already aligned -/// to the `Simd` you wish to use before using `unsafe` Rust to read or write. -/// More conventional ways to compensate for these facts, like materializing `Simd` to or from an array first, +/// When trying to guarantee alignment, [`[T]::as_simd`][as_simd] is an option for +/// converting `[T]` to `[Simd]`, and allows soundly operating on an aligned SIMD body, +/// but it may cost more time when handling the scalar head and tail. +/// If these are not enough, it is most ideal to design data structures to be already aligned +/// to `mem::align_of::>()` before using `unsafe` Rust to read or write. +/// Other ways to compensate for these facts, like materializing `Simd` to or from an array first, /// are handled by safe methods like [`Simd::from_array`] and [`Simd::from_slice`]. /// /// [`transmute`]: core::mem::transmute @@ -86,21 +93,26 @@ use crate::simd::{ /// [`read`]: pointer::read /// [`write`]: pointer::write /// [as_simd]: slice::as_simd +// +// NOTE: Accessing the inner array directly in any way (e.g. by using the `.0` field syntax) or +// directly constructing an instance of the type (i.e. `let vector = Simd(array)`) should be +// avoided, as it will likely become illegal on `#[repr(simd)]` structs in the future. It also +// causes rustc to emit illegal LLVM IR in some cases. #[repr(simd)] -pub struct Simd([T; LANES]) +pub struct Simd([T; N]) where - T: SimdElement, - LaneCount: SupportedLaneCount; + LaneCount: SupportedLaneCount, + T: SimdElement; -impl Simd +impl Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { - /// Number of lanes in this vector. - pub const LANES: usize = LANES; + /// Number of elements in this vector. + pub const LANES: usize = N; - /// Returns the number of lanes in this SIMD vector. + /// Returns the number of elements in this SIMD vector. /// /// # Examples /// @@ -111,10 +123,10 @@ where /// assert_eq!(v.lanes(), 4); /// ``` pub const fn lanes(&self) -> usize { - LANES + Self::LANES } - /// Constructs a new SIMD vector with all lanes set to the given value. + /// Constructs a new SIMD vector with all elements set to the given value. /// /// # Examples /// @@ -125,11 +137,11 @@ where /// assert_eq!(v.as_array(), &[8, 8, 8, 8]); /// ``` pub fn splat(value: T) -> Self { - // This is preferred over `[value; LANES]`, since it's explicitly a splat: + // This is preferred over `[value; N]`, since it's explicitly a splat: // https://github.com/rust-lang/rust/issues/97804 struct Splat; - impl Swizzle<1, LANES> for Splat { - const INDEX: [usize; LANES] = [0; LANES]; + impl Swizzle<1, N> for Splat { + const INDEX: [usize; N] = [0; N]; } Splat::swizzle(Simd::::from([value])) } @@ -144,32 +156,100 @@ where /// let v: u64x4 = Simd::from_array([0, 1, 2, 3]); /// assert_eq!(v.as_array(), &[0, 1, 2, 3]); /// ``` - pub const fn as_array(&self) -> &[T; LANES] { - &self.0 + pub const fn as_array(&self) -> &[T; N] { + // SAFETY: `Simd` is just an overaligned `[T; N]` with + // potential padding at the end, so pointer casting to a + // `&[T; N]` is safe. + // + // NOTE: This deliberately doesn't just use `&self.0`, see the comment + // on the struct definition for details. + unsafe { &*(self as *const Self as *const [T; N]) } } /// Returns a mutable array reference containing the entire SIMD vector. - pub fn as_mut_array(&mut self) -> &mut [T; LANES] { - &mut self.0 + pub fn as_mut_array(&mut self) -> &mut [T; N] { + // SAFETY: `Simd` is just an overaligned `[T; N]` with + // potential padding at the end, so pointer casting to a + // `&mut [T; N]` is safe. + // + // NOTE: This deliberately doesn't just use `&mut self.0`, see the comment + // on the struct definition for details. + unsafe { &mut *(self as *mut Self as *mut [T; N]) } + } + + /// Load a vector from an array of `T`. + /// + /// This function is necessary since `repr(simd)` has padding for non-power-of-2 vectors (at the time of writing). + /// With padding, `read_unaligned` will read past the end of an array of N elements. + /// + /// # Safety + /// Reading `ptr` must be safe, as if by `<*const [T; N]>::read_unaligned`. + const unsafe fn load(ptr: *const [T; N]) -> Self { + // There are potentially simpler ways to write this function, but this should result in + // LLVM `load ` + + let mut tmp = core::mem::MaybeUninit::::uninit(); + // SAFETY: `Simd` always contains `N` elements of type `T`. It may have padding + // which does not need to be initialized. The safety of reading `ptr` is ensured by the + // caller. + unsafe { + core::ptr::copy_nonoverlapping(ptr, tmp.as_mut_ptr().cast(), 1); + tmp.assume_init() + } + } + + /// Store a vector to an array of `T`. + /// + /// See `load` as to why this function is necessary. + /// + /// # Safety + /// Writing to `ptr` must be safe, as if by `<*mut [T; N]>::write_unaligned`. + const unsafe fn store(self, ptr: *mut [T; N]) { + // There are potentially simpler ways to write this function, but this should result in + // LLVM `store ` + + // Creating a temporary helps LLVM turn the memcpy into a store. + let tmp = self; + // SAFETY: `Simd` always contains `N` elements of type `T`. The safety of writing + // `ptr` is ensured by the caller. + unsafe { core::ptr::copy_nonoverlapping(tmp.as_array(), ptr, 1) } } /// Converts an array to a SIMD vector. - pub const fn from_array(array: [T; LANES]) -> Self { - Self(array) + pub const fn from_array(array: [T; N]) -> Self { + // SAFETY: `&array` is safe to read. + // + // FIXME: We currently use a pointer load instead of `transmute_copy` because `repr(simd)` + // results in padding for non-power-of-2 vectors (so vectors are larger than arrays). + // + // NOTE: This deliberately doesn't just use `Self(array)`, see the comment + // on the struct definition for details. + unsafe { Self::load(&array) } } /// Converts a SIMD vector to an array. - pub const fn to_array(self) -> [T; LANES] { - self.0 + pub const fn to_array(self) -> [T; N] { + let mut tmp = core::mem::MaybeUninit::uninit(); + // SAFETY: writing to `tmp` is safe and initializes it. + // + // FIXME: We currently use a pointer store instead of `transmute_copy` because `repr(simd)` + // results in padding for non-power-of-2 vectors (so vectors are larger than arrays). + // + // NOTE: This deliberately doesn't just use `self.0`, see the comment + // on the struct definition for details. + unsafe { + self.store(tmp.as_mut_ptr()); + tmp.assume_init() + } } - /// Converts a slice to a SIMD vector containing `slice[..LANES]`. + /// Converts a slice to a SIMD vector containing `slice[..N]`. /// /// # Panics /// - /// Panics if the slice's length is less than the vector's `Simd::LANES`. + /// Panics if the slice's length is less than the vector's `Simd::N`. /// - /// # Examples + /// # Example /// /// ``` /// # #![feature(portable_simd)] @@ -180,22 +260,49 @@ where /// ``` #[must_use] pub const fn from_slice(slice: &[T]) -> Self { - assert!(slice.len() >= LANES, "slice length must be at least the number of lanes"); - let mut array = [slice[0]; LANES]; - let mut i = 0; - while i < LANES { - array[i] = slice[i]; - i += 1; - } - Self(array) + assert!( + slice.len() >= Self::LANES, + "slice length must be at least the number of elements" + ); + // SAFETY: We just checked that the slice contains + // at least `N` elements. + unsafe { Self::load(slice.as_ptr().cast()) } } - /// Performs lanewise conversion of a SIMD vector's elements to another SIMD-valid type. + /// Writes a SIMD vector to the first `N` elements of a slice. /// - /// This follows the semantics of Rust's `as` conversion for casting - /// integers to unsigned integers (interpreting as the other type, so `-1` to `MAX`), - /// and from floats to integers (truncating, or saturating at the limits) for each lane, - /// or vice versa. + /// # Panics + /// + /// Panics if the slice's length is less than the vector's `Simd::N`. + /// + /// # Example + /// + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd; + /// # use simd::u32x4; + /// let mut dest = vec![0; 6]; + /// let v = u32x4::from_array([1, 2, 3, 4]); + /// v.copy_to_slice(&mut dest); + /// assert_eq!(&dest, &[1, 2, 3, 4, 0, 0]); + /// ``` + pub fn copy_to_slice(self, slice: &mut [T]) { + assert!( + slice.len() >= Self::LANES, + "slice length must be at least the number of elements" + ); + // SAFETY: We just checked that the slice contains + // at least `N` elements. + unsafe { self.store(slice.as_mut_ptr().cast()) } + } + + /// Performs elementwise conversion of a SIMD vector's elements to another SIMD-valid type. + /// + /// This follows the semantics of Rust's `as` conversion for casting integers between + /// signed and unsigned (interpreting integers as 2s complement, so `-1` to `U::MAX` and + /// `1 << (U::BITS -1)` becoming `I::MIN` ), and from floats to integers (truncating, + /// or saturating at the limits) for each element. /// /// # Examples /// ``` @@ -215,11 +322,26 @@ where /// ``` #[must_use] #[inline] - pub fn cast(self) -> Simd { - // Safety: The input argument is a vector of a valid SIMD element type. + pub fn cast(self) -> Simd + where + T: SimdCast, + { + // Safety: supported types are guaranteed by SimdCast unsafe { intrinsics::simd_as(self) } } + /// Casts a vector of pointers to another pointer type. + #[must_use] + #[inline] + pub fn cast_ptr(self) -> Simd + where + T: SimdCastPtr, + U: SimdElement, + { + // Safety: supported types are guaranteed by SimdCastPtr + unsafe { intrinsics::simd_cast_ptr(self) } + } + /// Rounds toward zero and converts to the same-width integer type, assuming that /// the value is finite and fits in that type. /// @@ -235,90 +357,90 @@ where /// /// [cast]: Simd::cast #[inline] - pub unsafe fn to_int_unchecked(self) -> Simd + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces + pub unsafe fn to_int_unchecked(self) -> Simd where - T: core::convert::FloatToInt, - I: SimdElement, + T: core::convert::FloatToInt + SimdCast, + I: SimdCast, { - // Safety: `self` is a vector, and `FloatToInt` ensures the type can be casted to - // an integer. + // Safety: supported types are guaranteed by SimdCast, the caller is responsible for the extra invariants unsafe { intrinsics::simd_cast(self) } } /// Reads from potentially discontiguous indices in `slice` to construct a SIMD vector. - /// If an index is out-of-bounds, the lane is instead selected from the `or` vector. + /// If an index is out-of-bounds, the element is instead selected from the `or` vector. /// /// # Examples /// ``` /// # #![feature(portable_simd)] /// # use core::simd::Simd; /// let vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 5]); + /// let idxs = Simd::from_array([9, 3, 0, 5]); // Note the index that is out-of-bounds /// let alt = Simd::from_array([-5, -4, -3, -2]); /// - /// let result = Simd::gather_or(&vec, idxs, alt); // Note the lane that is out-of-bounds. + /// let result = Simd::gather_or(&vec, idxs, alt); /// assert_eq!(result, Simd::from_array([-5, 13, 10, 15])); /// ``` #[must_use] #[inline] - pub fn gather_or(slice: &[T], idxs: Simd, or: Self) -> Self { + pub fn gather_or(slice: &[T], idxs: Simd, or: Self) -> Self { Self::gather_select(slice, Mask::splat(true), idxs, or) } - /// Reads from potentially discontiguous indices in `slice` to construct a SIMD vector. - /// If an index is out-of-bounds, the lane is set to the default value for the type. + /// Reads from indices in `slice` to construct a SIMD vector. + /// If an index is out-of-bounds, the element is set to the default given by `T: Default`. /// /// # Examples /// ``` /// # #![feature(portable_simd)] /// # use core::simd::Simd; /// let vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 5]); + /// let idxs = Simd::from_array([9, 3, 0, 5]); // Note the index that is out-of-bounds /// - /// let result = Simd::gather_or_default(&vec, idxs); // Note the lane that is out-of-bounds. + /// let result = Simd::gather_or_default(&vec, idxs); /// assert_eq!(result, Simd::from_array([0, 13, 10, 15])); /// ``` #[must_use] #[inline] - pub fn gather_or_default(slice: &[T], idxs: Simd) -> Self + pub fn gather_or_default(slice: &[T], idxs: Simd) -> Self where T: Default, { Self::gather_or(slice, idxs, Self::splat(T::default())) } - /// Reads from potentially discontiguous indices in `slice` to construct a SIMD vector. - /// The mask `enable`s all `true` lanes and disables all `false` lanes. - /// If an index is disabled or is out-of-bounds, the lane is selected from the `or` vector. + /// Reads from indices in `slice` to construct a SIMD vector. + /// The mask `enable`s all `true` indices and disables all `false` indices. + /// If an index is disabled or is out-of-bounds, the element is selected from the `or` vector. /// /// # Examples /// ``` /// # #![feature(portable_simd)] /// # use core::simd::{Simd, Mask}; /// let vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 5]); + /// let idxs = Simd::from_array([9, 3, 0, 5]); // Includes an out-of-bounds index /// let alt = Simd::from_array([-5, -4, -3, -2]); - /// let enable = Mask::from_array([true, true, true, false]); // Note the mask of the last lane. + /// let enable = Mask::from_array([true, true, true, false]); // Includes a masked element /// - /// let result = Simd::gather_select(&vec, enable, idxs, alt); // Note the lane that is out-of-bounds. + /// let result = Simd::gather_select(&vec, enable, idxs, alt); /// assert_eq!(result, Simd::from_array([-5, 13, 10, -2])); /// ``` #[must_use] #[inline] pub fn gather_select( slice: &[T], - enable: Mask, - idxs: Simd, + enable: Mask, + idxs: Simd, or: Self, ) -> Self { - let enable: Mask = enable & idxs.simd_lt(Simd::splat(slice.len())); - // Safety: We have masked-off out-of-bounds lanes. + let enable: Mask = enable & idxs.simd_lt(Simd::splat(slice.len())); + // Safety: We have masked-off out-of-bounds indices. unsafe { Self::gather_select_unchecked(slice, enable, idxs, or) } } - /// Reads from potentially discontiguous indices in `slice` to construct a SIMD vector. - /// The mask `enable`s all `true` lanes and disables all `false` lanes. - /// If an index is disabled, the lane is selected from the `or` vector. + /// Reads from indices in `slice` to construct a SIMD vector. + /// The mask `enable`s all `true` indices and disables all `false` indices. + /// If an index is disabled, the element is selected from the `or` vector. /// /// # Safety /// @@ -332,57 +454,123 @@ where /// # #[cfg(not(feature = "as_crate"))] use core::simd; /// # use simd::{Simd, SimdPartialOrd, Mask}; /// let vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 5]); + /// let idxs = Simd::from_array([9, 3, 0, 5]); // Includes an out-of-bounds index /// let alt = Simd::from_array([-5, -4, -3, -2]); - /// let enable = Mask::from_array([true, true, true, false]); // Note the final mask lane. + /// let enable = Mask::from_array([true, true, true, false]); // Includes a masked element /// // If this mask was used to gather, it would be unsound. Let's fix that. /// let enable = enable & idxs.simd_lt(Simd::splat(vec.len())); /// - /// // We have masked the OOB lane, so it's safe to gather now. + /// // The out-of-bounds index has been masked, so it's safe to gather now. /// let result = unsafe { Simd::gather_select_unchecked(&vec, enable, idxs, alt) }; /// assert_eq!(result, Simd::from_array([-5, 13, 10, -2])); /// ``` /// [undefined behavior]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html #[must_use] #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces pub unsafe fn gather_select_unchecked( slice: &[T], - enable: Mask, - idxs: Simd, + enable: Mask, + idxs: Simd, or: Self, ) -> Self { - let base_ptr = crate::simd::ptr::SimdConstPtr::splat(slice.as_ptr()); + let base_ptr = Simd::<*const T, N>::splat(slice.as_ptr()); // Ferris forgive me, I have done pointer arithmetic here. let ptrs = base_ptr.wrapping_add(idxs); - // Safety: The ptrs have been bounds-masked to prevent memory-unsafe reads insha'allah - unsafe { intrinsics::simd_gather(or, ptrs, enable.to_int()) } + // Safety: The caller is responsible for determining the indices are okay to read + unsafe { Self::gather_select_ptr(ptrs, enable, or) } + } + + /// Read elementwise from pointers into a SIMD vector. + /// + /// # Safety + /// + /// Each read must satisfy the same conditions as [`core::ptr::read`]. + /// + /// # Example + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd; + /// # use simd::{Simd, SimdConstPtr}; + /// let values = [6, 2, 4, 9]; + /// let offsets = Simd::from_array([1, 0, 0, 3]); + /// let source = Simd::splat(values.as_ptr()).wrapping_add(offsets); + /// let gathered = unsafe { Simd::gather_ptr(source) }; + /// assert_eq!(gathered, Simd::from_array([2, 6, 6, 9])); + /// ``` + #[must_use] + #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces + pub unsafe fn gather_ptr(source: Simd<*const T, N>) -> Self + where + T: Default, + { + // TODO: add an intrinsic that doesn't use a passthru vector, and remove the T: Default bound + // Safety: The caller is responsible for upholding all invariants + unsafe { Self::gather_select_ptr(source, Mask::splat(true), Self::default()) } + } + + /// Conditionally read elementwise from pointers into a SIMD vector. + /// The mask `enable`s all `true` pointers and disables all `false` pointers. + /// If a pointer is disabled, the element is selected from the `or` vector, + /// and no read is performed. + /// + /// # Safety + /// + /// Enabled elements must satisfy the same conditions as [`core::ptr::read`]. + /// + /// # Example + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd; + /// # use simd::{Mask, Simd, SimdConstPtr}; + /// let values = [6, 2, 4, 9]; + /// let enable = Mask::from_array([true, true, false, true]); + /// let offsets = Simd::from_array([1, 0, 0, 3]); + /// let source = Simd::splat(values.as_ptr()).wrapping_add(offsets); + /// let gathered = unsafe { Simd::gather_select_ptr(source, enable, Simd::splat(0)) }; + /// assert_eq!(gathered, Simd::from_array([2, 6, 0, 9])); + /// ``` + #[must_use] + #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces + pub unsafe fn gather_select_ptr( + source: Simd<*const T, N>, + enable: Mask, + or: Self, + ) -> Self { + // Safety: The caller is responsible for upholding all invariants + unsafe { intrinsics::simd_gather(or, source, enable.to_int()) } } /// Writes the values in a SIMD vector to potentially discontiguous indices in `slice`. - /// If two lanes in the scattered vector would write to the same index - /// only the last lane is guaranteed to actually be written. + /// If an index is out-of-bounds, the write is suppressed without panicking. + /// If two elements in the scattered vector would write to the same index + /// only the last element is guaranteed to actually be written. /// /// # Examples /// ``` /// # #![feature(portable_simd)] /// # use core::simd::Simd; /// let mut vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 0]); + /// let idxs = Simd::from_array([9, 3, 0, 0]); // Note the duplicate index. /// let vals = Simd::from_array([-27, 82, -41, 124]); /// - /// vals.scatter(&mut vec, idxs); // index 0 receives two writes. + /// vals.scatter(&mut vec, idxs); // two logical writes means the last wins. /// assert_eq!(vec, vec![124, 11, 12, 82, 14, 15, 16, 17, 18]); /// ``` #[inline] - pub fn scatter(self, slice: &mut [T], idxs: Simd) { + pub fn scatter(self, slice: &mut [T], idxs: Simd) { self.scatter_select(slice, Mask::splat(true), idxs) } - /// Writes the values in a SIMD vector to multiple potentially discontiguous indices in `slice`. - /// The mask `enable`s all `true` lanes and disables all `false` lanes. - /// If an enabled index is out-of-bounds, the lane is not written. - /// If two enabled lanes in the scattered vector would write to the same index, - /// only the last lane is guaranteed to actually be written. + /// Writes values from a SIMD vector to multiple potentially discontiguous indices in `slice`. + /// The mask `enable`s all `true` indices and disables all `false` indices. + /// If an enabled index is out-of-bounds, the write is suppressed without panicking. + /// If two enabled elements in the scattered vector would write to the same index, + /// only the last element is guaranteed to actually be written. /// /// # Examples /// ``` @@ -391,29 +579,24 @@ where /// # #[cfg(not(feature = "as_crate"))] use core::simd; /// # use simd::{Simd, Mask}; /// let mut vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; - /// let idxs = Simd::from_array([9, 3, 0, 0]); + /// let idxs = Simd::from_array([9, 3, 0, 0]); // Includes an out-of-bounds index /// let vals = Simd::from_array([-27, 82, -41, 124]); - /// let enable = Mask::from_array([true, true, true, false]); // Note the mask of the last lane. + /// let enable = Mask::from_array([true, true, true, false]); // Includes a masked element /// - /// vals.scatter_select(&mut vec, enable, idxs); // index 0's second write is masked, thus omitted. + /// vals.scatter_select(&mut vec, enable, idxs); // The last write is masked, thus omitted. /// assert_eq!(vec, vec![-41, 11, 12, 82, 14, 15, 16, 17, 18]); /// ``` #[inline] - pub fn scatter_select( - self, - slice: &mut [T], - enable: Mask, - idxs: Simd, - ) { - let enable: Mask = enable & idxs.simd_lt(Simd::splat(slice.len())); - // Safety: We have masked-off out-of-bounds lanes. + pub fn scatter_select(self, slice: &mut [T], enable: Mask, idxs: Simd) { + let enable: Mask = enable & idxs.simd_lt(Simd::splat(slice.len())); + // Safety: We have masked-off out-of-bounds indices. unsafe { self.scatter_select_unchecked(slice, enable, idxs) } } - /// Writes the values in a SIMD vector to multiple potentially discontiguous indices in `slice`. - /// The mask `enable`s all `true` lanes and disables all `false` lanes. - /// If two enabled lanes in the scattered vector would write to the same index, - /// only the last lane is guaranteed to actually be written. + /// Writes values from a SIMD vector to multiple potentially discontiguous indices in `slice`. + /// The mask `enable`s all `true` indices and disables all `false` indices. + /// If two enabled elements in the scattered vector would write to the same index, + /// only the last element is guaranteed to actually be written. /// /// # Safety /// @@ -429,22 +612,23 @@ where /// let mut vec: Vec = vec![10, 11, 12, 13, 14, 15, 16, 17, 18]; /// let idxs = Simd::from_array([9, 3, 0, 0]); /// let vals = Simd::from_array([-27, 82, -41, 124]); - /// let enable = Mask::from_array([true, true, true, false]); // Note the mask of the last lane. + /// let enable = Mask::from_array([true, true, true, false]); // Masks the final index /// // If this mask was used to scatter, it would be unsound. Let's fix that. /// let enable = enable & idxs.simd_lt(Simd::splat(vec.len())); /// - /// // We have masked the OOB lane, so it's safe to scatter now. + /// // We have masked the OOB index, so it's safe to scatter now. /// unsafe { vals.scatter_select_unchecked(&mut vec, enable, idxs); } - /// // index 0's second write is masked, thus was omitted. + /// // The second write to index 0 was masked, thus omitted. /// assert_eq!(vec, vec![-41, 11, 12, 82, 14, 15, 16, 17, 18]); /// ``` /// [undefined behavior]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces pub unsafe fn scatter_select_unchecked( self, slice: &mut [T], - enable: Mask, - idxs: Simd, + enable: Mask, + idxs: Simd, ) { // Safety: This block works with *mut T derived from &mut 'a [T], // which means it is delicate in Rust's borrowing model, circa 2021: @@ -458,36 +642,89 @@ where // 3. &mut [T] which will become our base ptr. unsafe { // Now Entering ☢️ *mut T Zone - let base_ptr = crate::simd::ptr::SimdMutPtr::splat(slice.as_mut_ptr()); + let base_ptr = Simd::<*mut T, N>::splat(slice.as_mut_ptr()); // Ferris forgive me, I have done pointer arithmetic here. let ptrs = base_ptr.wrapping_add(idxs); // The ptrs have been bounds-masked to prevent memory-unsafe writes insha'allah - intrinsics::simd_scatter(self, ptrs, enable.to_int()) + self.scatter_select_ptr(ptrs, enable); // Cleared ☢️ *mut T Zone } } + + /// Write pointers elementwise into a SIMD vector. + /// + /// # Safety + /// + /// Each write must satisfy the same conditions as [`core::ptr::write`]. + /// + /// # Example + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd; + /// # use simd::{Simd, SimdMutPtr}; + /// let mut values = [0; 4]; + /// let offset = Simd::from_array([3, 2, 1, 0]); + /// let ptrs = Simd::splat(values.as_mut_ptr()).wrapping_add(offset); + /// unsafe { Simd::from_array([6, 3, 5, 7]).scatter_ptr(ptrs); } + /// assert_eq!(values, [7, 5, 3, 6]); + /// ``` + #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces + pub unsafe fn scatter_ptr(self, dest: Simd<*mut T, N>) { + // Safety: The caller is responsible for upholding all invariants + unsafe { self.scatter_select_ptr(dest, Mask::splat(true)) } + } + + /// Conditionally write pointers elementwise into a SIMD vector. + /// The mask `enable`s all `true` pointers and disables all `false` pointers. + /// If a pointer is disabled, the write to its pointee is skipped. + /// + /// # Safety + /// + /// Enabled pointers must satisfy the same conditions as [`core::ptr::write`]. + /// + /// # Example + /// ``` + /// # #![feature(portable_simd)] + /// # #[cfg(feature = "as_crate")] use core_simd::simd; + /// # #[cfg(not(feature = "as_crate"))] use core::simd; + /// # use simd::{Mask, Simd, SimdMutPtr}; + /// let mut values = [0; 4]; + /// let offset = Simd::from_array([3, 2, 1, 0]); + /// let ptrs = Simd::splat(values.as_mut_ptr()).wrapping_add(offset); + /// let enable = Mask::from_array([true, true, false, false]); + /// unsafe { Simd::from_array([6, 3, 5, 7]).scatter_select_ptr(ptrs, enable); } + /// assert_eq!(values, [0, 0, 3, 6]); + /// ``` + #[inline] + #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces + pub unsafe fn scatter_select_ptr(self, dest: Simd<*mut T, N>, enable: Mask) { + // Safety: The caller is responsible for upholding all invariants + unsafe { intrinsics::simd_scatter(self, dest, enable.to_int()) } + } } -impl Copy for Simd +impl Copy for Simd where + LaneCount: SupportedLaneCount, T: SimdElement, - LaneCount: SupportedLaneCount, { } -impl Clone for Simd +impl Clone for Simd where + LaneCount: SupportedLaneCount, T: SimdElement, - LaneCount: SupportedLaneCount, { fn clone(&self) -> Self { *self } } -impl Default for Simd +impl Default for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + Default, { #[inline] @@ -496,20 +733,20 @@ where } } -impl PartialEq for Simd +impl PartialEq for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + PartialEq, { #[inline] fn eq(&self, other: &Self) -> bool { // Safety: All SIMD vectors are SimdPartialEq, and the comparison produces a valid mask. let mask = unsafe { - let tfvec: Simd<::Mask, LANES> = intrinsics::simd_eq(*self, *other); + let tfvec: Simd<::Mask, N> = intrinsics::simd_eq(*self, *other); Mask::from_int_unchecked(tfvec) }; - // Two vectors are equal if all lanes tested true for vertical equality. + // Two vectors are equal if all elements are equal when compared elementwise mask.all() } @@ -518,18 +755,18 @@ where fn ne(&self, other: &Self) -> bool { // Safety: All SIMD vectors are SimdPartialEq, and the comparison produces a valid mask. let mask = unsafe { - let tfvec: Simd<::Mask, LANES> = intrinsics::simd_ne(*self, *other); + let tfvec: Simd<::Mask, N> = intrinsics::simd_ne(*self, *other); Mask::from_int_unchecked(tfvec) }; - // Two vectors are non-equal if any lane tested true for vertical non-equality. + // Two vectors are non-equal if any elements are non-equal when compared elementwise mask.any() } } -impl PartialOrd for Simd +impl PartialOrd for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + PartialOrd, { #[inline] @@ -539,16 +776,16 @@ where } } -impl Eq for Simd +impl Eq for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + Eq, { } -impl Ord for Simd +impl Ord for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + Ord, { #[inline] @@ -558,9 +795,9 @@ where } } -impl core::hash::Hash for Simd +impl core::hash::Hash for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement + core::hash::Hash, { #[inline] @@ -573,72 +810,96 @@ where } // array references -impl AsRef<[T; LANES]> for Simd +impl AsRef<[T; N]> for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { #[inline] - fn as_ref(&self) -> &[T; LANES] { - &self.0 + fn as_ref(&self) -> &[T; N] { + self.as_array() } } -impl AsMut<[T; LANES]> for Simd +impl AsMut<[T; N]> for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { #[inline] - fn as_mut(&mut self) -> &mut [T; LANES] { - &mut self.0 + fn as_mut(&mut self) -> &mut [T; N] { + self.as_mut_array() } } // slice references -impl AsRef<[T]> for Simd +impl AsRef<[T]> for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { #[inline] fn as_ref(&self) -> &[T] { - &self.0 + self.as_array() } } -impl AsMut<[T]> for Simd +impl AsMut<[T]> for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { #[inline] fn as_mut(&mut self) -> &mut [T] { - &mut self.0 + self.as_mut_array() } } // vector/array conversion -impl From<[T; LANES]> for Simd +impl From<[T; N]> for Simd where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { - fn from(array: [T; LANES]) -> Self { - Self(array) + fn from(array: [T; N]) -> Self { + Self::from_array(array) } } -impl From> for [T; LANES] +impl From> for [T; N] where - LaneCount: SupportedLaneCount, + LaneCount: SupportedLaneCount, T: SimdElement, { - fn from(vector: Simd) -> Self { + fn from(vector: Simd) -> Self { vector.to_array() } } +impl TryFrom<&[T]> for Simd +where + LaneCount: SupportedLaneCount, + T: SimdElement, +{ + type Error = core::array::TryFromSliceError; + + fn try_from(slice: &[T]) -> Result { + Ok(Self::from_array(slice.try_into()?)) + } +} + +impl TryFrom<&mut [T]> for Simd +where + LaneCount: SupportedLaneCount, + T: SimdElement, +{ + type Error = core::array::TryFromSliceError; + + fn try_from(slice: &mut [T]) -> Result { + Ok(Self::from_array(slice.try_into()?)) + } +} + mod sealed { pub trait Sealed {} } @@ -740,3 +1001,27 @@ impl Sealed for f64 {} unsafe impl SimdElement for f64 { type Mask = i64; } + +impl Sealed for *const T {} + +// Safety: (thin) const pointers are valid SIMD element types, and are supported by this API +// +// Fat pointers may be supported in the future. +unsafe impl SimdElement for *const T +where + T: core::ptr::Pointee, +{ + type Mask = isize; +} + +impl Sealed for *mut T {} + +// Safety: (thin) mut pointers are valid SIMD element types, and are supported by this API +// +// Fat pointers may be supported in the future. +unsafe impl SimdElement for *mut T +where + T: core::ptr::Pointee, +{ + type Mask = isize; +} diff --git a/library/portable-simd/crates/core_simd/src/vector/float.rs b/library/portable-simd/crates/core_simd/src/vector/float.rs deleted file mode 100644 index f836c99b1e2..00000000000 --- a/library/portable-simd/crates/core_simd/src/vector/float.rs +++ /dev/null @@ -1,24 +0,0 @@ -#![allow(non_camel_case_types)] - -use crate::simd::Simd; - -/// A 64-bit SIMD vector with two elements of type `f32`. -pub type f32x2 = Simd; - -/// A 128-bit SIMD vector with four elements of type `f32`. -pub type f32x4 = Simd; - -/// A 256-bit SIMD vector with eight elements of type `f32`. -pub type f32x8 = Simd; - -/// A 512-bit SIMD vector with 16 elements of type `f32`. -pub type f32x16 = Simd; - -/// A 128-bit SIMD vector with two elements of type `f64`. -pub type f64x2 = Simd; - -/// A 256-bit SIMD vector with four elements of type `f64`. -pub type f64x4 = Simd; - -/// A 512-bit SIMD vector with eight elements of type `f64`. -pub type f64x8 = Simd; diff --git a/library/portable-simd/crates/core_simd/src/vector/int.rs b/library/portable-simd/crates/core_simd/src/vector/int.rs deleted file mode 100644 index 20e56c7dc64..00000000000 --- a/library/portable-simd/crates/core_simd/src/vector/int.rs +++ /dev/null @@ -1,63 +0,0 @@ -#![allow(non_camel_case_types)] - -use crate::simd::Simd; - -/// A SIMD vector with two elements of type `isize`. -pub type isizex2 = Simd; - -/// A SIMD vector with four elements of type `isize`. -pub type isizex4 = Simd; - -/// A SIMD vector with eight elements of type `isize`. -pub type isizex8 = Simd; - -/// A 32-bit SIMD vector with two elements of type `i16`. -pub type i16x2 = Simd; - -/// A 64-bit SIMD vector with four elements of type `i16`. -pub type i16x4 = Simd; - -/// A 128-bit SIMD vector with eight elements of type `i16`. -pub type i16x8 = Simd; - -/// A 256-bit SIMD vector with 16 elements of type `i16`. -pub type i16x16 = Simd; - -/// A 512-bit SIMD vector with 32 elements of type `i16`. -pub type i16x32 = Simd; - -/// A 64-bit SIMD vector with two elements of type `i32`. -pub type i32x2 = Simd; - -/// A 128-bit SIMD vector with four elements of type `i32`. -pub type i32x4 = Simd; - -/// A 256-bit SIMD vector with eight elements of type `i32`. -pub type i32x8 = Simd; - -/// A 512-bit SIMD vector with 16 elements of type `i32`. -pub type i32x16 = Simd; - -/// A 128-bit SIMD vector with two elements of type `i64`. -pub type i64x2 = Simd; - -/// A 256-bit SIMD vector with four elements of type `i64`. -pub type i64x4 = Simd; - -/// A 512-bit SIMD vector with eight elements of type `i64`. -pub type i64x8 = Simd; - -/// A 32-bit SIMD vector with four elements of type `i8`. -pub type i8x4 = Simd; - -/// A 64-bit SIMD vector with eight elements of type `i8`. -pub type i8x8 = Simd; - -/// A 128-bit SIMD vector with 16 elements of type `i8`. -pub type i8x16 = Simd; - -/// A 256-bit SIMD vector with 32 elements of type `i8`. -pub type i8x32 = Simd; - -/// A 512-bit SIMD vector with 64 elements of type `i8`. -pub type i8x64 = Simd; diff --git a/library/portable-simd/crates/core_simd/src/vector/ptr.rs b/library/portable-simd/crates/core_simd/src/vector/ptr.rs deleted file mode 100644 index fa756344db9..00000000000 --- a/library/portable-simd/crates/core_simd/src/vector/ptr.rs +++ /dev/null @@ -1,51 +0,0 @@ -//! Private implementation details of public gather/scatter APIs. -use crate::simd::intrinsics; -use crate::simd::{LaneCount, Simd, SupportedLaneCount}; - -/// A vector of *const T. -#[derive(Debug, Copy, Clone)] -#[repr(simd)] -pub(crate) struct SimdConstPtr([*const T; LANES]); - -impl SimdConstPtr -where - LaneCount: SupportedLaneCount, - T: Sized, -{ - #[inline] - #[must_use] - pub fn splat(ptr: *const T) -> Self { - Self([ptr; LANES]) - } - - #[inline] - #[must_use] - pub fn wrapping_add(self, addend: Simd) -> Self { - // Safety: this intrinsic doesn't have a precondition - unsafe { intrinsics::simd_arith_offset(self, addend) } - } -} - -/// A vector of *mut T. Be very careful around potential aliasing. -#[derive(Debug, Copy, Clone)] -#[repr(simd)] -pub(crate) struct SimdMutPtr([*mut T; LANES]); - -impl SimdMutPtr -where - LaneCount: SupportedLaneCount, - T: Sized, -{ - #[inline] - #[must_use] - pub fn splat(ptr: *mut T) -> Self { - Self([ptr; LANES]) - } - - #[inline] - #[must_use] - pub fn wrapping_add(self, addend: Simd) -> Self { - // Safety: this intrinsic doesn't have a precondition - unsafe { intrinsics::simd_arith_offset(self, addend) } - } -} diff --git a/library/portable-simd/crates/core_simd/src/vector/uint.rs b/library/portable-simd/crates/core_simd/src/vector/uint.rs deleted file mode 100644 index b4a69c44363..00000000000 --- a/library/portable-simd/crates/core_simd/src/vector/uint.rs +++ /dev/null @@ -1,63 +0,0 @@ -#![allow(non_camel_case_types)] - -use crate::simd::Simd; - -/// A SIMD vector with two elements of type `usize`. -pub type usizex2 = Simd; - -/// A SIMD vector with four elements of type `usize`. -pub type usizex4 = Simd; - -/// A SIMD vector with eight elements of type `usize`. -pub type usizex8 = Simd; - -/// A 32-bit SIMD vector with two elements of type `u16`. -pub type u16x2 = Simd; - -/// A 64-bit SIMD vector with four elements of type `u16`. -pub type u16x4 = Simd; - -/// A 128-bit SIMD vector with eight elements of type `u16`. -pub type u16x8 = Simd; - -/// A 256-bit SIMD vector with 16 elements of type `u16`. -pub type u16x16 = Simd; - -/// A 512-bit SIMD vector with 32 elements of type `u16`. -pub type u16x32 = Simd; - -/// A 64-bit SIMD vector with two elements of type `u32`. -pub type u32x2 = Simd; - -/// A 128-bit SIMD vector with four elements of type `u32`. -pub type u32x4 = Simd; - -/// A 256-bit SIMD vector with eight elements of type `u32`. -pub type u32x8 = Simd; - -/// A 512-bit SIMD vector with 16 elements of type `u32`. -pub type u32x16 = Simd; - -/// A 128-bit SIMD vector with two elements of type `u64`. -pub type u64x2 = Simd; - -/// A 256-bit SIMD vector with four elements of type `u64`. -pub type u64x4 = Simd; - -/// A 512-bit SIMD vector with eight elements of type `u64`. -pub type u64x8 = Simd; - -/// A 32-bit SIMD vector with four elements of type `u8`. -pub type u8x4 = Simd; - -/// A 64-bit SIMD vector with eight elements of type `u8`. -pub type u8x8 = Simd; - -/// A 128-bit SIMD vector with 16 elements of type `u8`. -pub type u8x16 = Simd; - -/// A 256-bit SIMD vector with 32 elements of type `u8`. -pub type u8x32 = Simd; - -/// A 512-bit SIMD vector with 64 elements of type `u8`. -pub type u8x64 = Simd; diff --git a/library/portable-simd/crates/core_simd/tests/autoderef.rs b/library/portable-simd/crates/core_simd/tests/autoderef.rs index 9359da16ee5..3181826ef59 100644 --- a/library/portable-simd/crates/core_simd/tests/autoderef.rs +++ b/library/portable-simd/crates/core_simd/tests/autoderef.rs @@ -1,6 +1,6 @@ // Test that we handle all our "auto-deref" cases correctly. #![feature(portable_simd)] -use core_simd::f32x4; +use core_simd::simd::f32x4; #[cfg(target_arch = "wasm32")] use wasm_bindgen_test::*; diff --git a/library/portable-simd/crates/core_simd/tests/mask_ops_impl/mask_macros.rs b/library/portable-simd/crates/core_simd/tests/mask_ops_impl/mask_macros.rs index 795f9e27c44..faafa5fa51f 100644 --- a/library/portable-simd/crates/core_simd/tests/mask_ops_impl/mask_macros.rs +++ b/library/portable-simd/crates/core_simd/tests/mask_ops_impl/mask_macros.rs @@ -2,7 +2,7 @@ macro_rules! mask_tests { { $vector:ident, $lanes:literal } => { #[cfg(test)] mod $vector { - use core_simd::$vector as Vector; + use core_simd::simd::$vector as Vector; const LANES: usize = $lanes; #[cfg(target_arch = "wasm32")] diff --git a/library/portable-simd/crates/core_simd/tests/masks.rs b/library/portable-simd/crates/core_simd/tests/masks.rs index 673d0db93fe..9f8bad1c36c 100644 --- a/library/portable-simd/crates/core_simd/tests/masks.rs +++ b/library/portable-simd/crates/core_simd/tests/masks.rs @@ -13,11 +13,13 @@ macro_rules! test_mask_api { #[cfg(target_arch = "wasm32")] use wasm_bindgen_test::*; + use core_simd::simd::Mask; + #[test] #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)] fn set_and_test() { let values = [true, false, false, true, false, false, true, false]; - let mut mask = core_simd::Mask::<$type, 8>::splat(false); + let mut mask = Mask::<$type, 8>::splat(false); for (lane, value) in values.iter().copied().enumerate() { mask.set(lane, value); } @@ -29,7 +31,7 @@ macro_rules! test_mask_api { #[test] #[should_panic] fn set_invalid_lane() { - let mut mask = core_simd::Mask::<$type, 8>::splat(false); + let mut mask = Mask::<$type, 8>::splat(false); mask.set(8, true); let _ = mask; } @@ -37,24 +39,24 @@ macro_rules! test_mask_api { #[test] #[should_panic] fn test_invalid_lane() { - let mask = core_simd::Mask::<$type, 8>::splat(false); + let mask = Mask::<$type, 8>::splat(false); let _ = mask.test(8); } #[test] fn any() { - assert!(!core_simd::Mask::<$type, 8>::splat(false).any()); - assert!(core_simd::Mask::<$type, 8>::splat(true).any()); - let mut v = core_simd::Mask::<$type, 8>::splat(false); + assert!(!Mask::<$type, 8>::splat(false).any()); + assert!(Mask::<$type, 8>::splat(true).any()); + let mut v = Mask::<$type, 8>::splat(false); v.set(2, true); assert!(v.any()); } #[test] fn all() { - assert!(!core_simd::Mask::<$type, 8>::splat(false).all()); - assert!(core_simd::Mask::<$type, 8>::splat(true).all()); - let mut v = core_simd::Mask::<$type, 8>::splat(false); + assert!(!Mask::<$type, 8>::splat(false).all()); + assert!(Mask::<$type, 8>::splat(true).all()); + let mut v = Mask::<$type, 8>::splat(false); v.set(2, true); assert!(!v.all()); } @@ -62,57 +64,57 @@ macro_rules! test_mask_api { #[test] fn roundtrip_int_conversion() { let values = [true, false, false, true, false, false, true, false]; - let mask = core_simd::Mask::<$type, 8>::from_array(values); + let mask = Mask::<$type, 8>::from_array(values); let int = mask.to_int(); assert_eq!(int.to_array(), [-1, 0, 0, -1, 0, 0, -1, 0]); - assert_eq!(core_simd::Mask::<$type, 8>::from_int(int), mask); + assert_eq!(Mask::<$type, 8>::from_int(int), mask); } #[test] fn roundtrip_bitmask_conversion() { - use core_simd::ToBitMask; + use core_simd::simd::ToBitMask; let values = [ true, false, false, true, false, false, true, false, true, true, false, false, false, false, false, true, ]; - let mask = core_simd::Mask::<$type, 16>::from_array(values); + let mask = Mask::<$type, 16>::from_array(values); let bitmask = mask.to_bitmask(); assert_eq!(bitmask, 0b1000001101001001); - assert_eq!(core_simd::Mask::<$type, 16>::from_bitmask(bitmask), mask); + assert_eq!(Mask::<$type, 16>::from_bitmask(bitmask), mask); } #[test] fn roundtrip_bitmask_conversion_short() { - use core_simd::ToBitMask; + use core_simd::simd::ToBitMask; let values = [ false, false, false, true, ]; - let mask = core_simd::Mask::<$type, 4>::from_array(values); + let mask = Mask::<$type, 4>::from_array(values); let bitmask = mask.to_bitmask(); assert_eq!(bitmask, 0b1000); - assert_eq!(core_simd::Mask::<$type, 4>::from_bitmask(bitmask), mask); + assert_eq!(Mask::<$type, 4>::from_bitmask(bitmask), mask); let values = [true, false]; - let mask = core_simd::Mask::<$type, 2>::from_array(values); + let mask = Mask::<$type, 2>::from_array(values); let bitmask = mask.to_bitmask(); assert_eq!(bitmask, 0b01); - assert_eq!(core_simd::Mask::<$type, 2>::from_bitmask(bitmask), mask); + assert_eq!(Mask::<$type, 2>::from_bitmask(bitmask), mask); } #[test] fn cast() { - fn cast_impl() + fn cast_impl() where - core_simd::Mask<$type, 8>: Into>, + Mask<$type, 8>: Into>, { let values = [true, false, false, true, false, false, true, false]; - let mask = core_simd::Mask::<$type, 8>::from_array(values); + let mask = Mask::<$type, 8>::from_array(values); let cast_mask = mask.cast::(); assert_eq!(values, cast_mask.to_array()); - let into_mask: core_simd::Mask = mask.into(); + let into_mask: Mask = mask.into(); assert_eq!(values, into_mask.to_array()); } @@ -126,15 +128,15 @@ macro_rules! test_mask_api { #[cfg(feature = "generic_const_exprs")] #[test] fn roundtrip_bitmask_array_conversion() { - use core_simd::ToBitMaskArray; + use core_simd::simd::ToBitMaskArray; let values = [ true, false, false, true, false, false, true, false, true, true, false, false, false, false, false, true, ]; - let mask = core_simd::Mask::<$type, 16>::from_array(values); + let mask = Mask::<$type, 16>::from_array(values); let bitmask = mask.to_bitmask_array(); assert_eq!(bitmask, [0b01001001, 0b10000011]); - assert_eq!(core_simd::Mask::<$type, 16>::from_bitmask_array(bitmask), mask); + assert_eq!(Mask::<$type, 16>::from_bitmask_array(bitmask), mask); } } } @@ -150,9 +152,10 @@ mod mask_api { #[test] fn convert() { + use core_simd::simd::Mask; let values = [true, false, false, true, false, false, true, false]; assert_eq!( - core_simd::Mask::::from_array(values), - core_simd::Mask::::from_array(values).into() + Mask::::from_array(values), + Mask::::from_array(values).into() ); } diff --git a/library/portable-simd/crates/core_simd/tests/ops_macros.rs b/library/portable-simd/crates/core_simd/tests/ops_macros.rs index f759394d075..3a02f3f01e1 100644 --- a/library/portable-simd/crates/core_simd/tests/ops_macros.rs +++ b/library/portable-simd/crates/core_simd/tests/ops_macros.rs @@ -7,7 +7,7 @@ macro_rules! impl_unary_op_test { test_helpers::test_lanes! { fn $fn() { test_helpers::test_unary_elementwise( - & as core::ops::$trait>::$fn, + & as core::ops::$trait>::$fn, &$scalar_fn, &|_| true, ); @@ -27,7 +27,7 @@ macro_rules! impl_binary_op_test { { $scalar:ty, $trait:ident :: $fn:ident, $trait_assign:ident :: $fn_assign:ident, $scalar_fn:expr } => { mod $fn { use super::*; - use core_simd::Simd; + use core_simd::simd::Simd; test_helpers::test_lanes! { fn normal() { @@ -64,7 +64,7 @@ macro_rules! impl_binary_checked_op_test { { $scalar:ty, $trait:ident :: $fn:ident, $trait_assign:ident :: $fn_assign:ident, $scalar_fn:expr, $check_fn:expr } => { mod $fn { use super::*; - use core_simd::Simd; + use core_simd::simd::Simd; test_helpers::test_lanes! { fn normal() { @@ -173,7 +173,7 @@ macro_rules! impl_signed_tests { { $scalar:tt } => { mod $scalar { use core_simd::simd::SimdInt; - type Vector = core_simd::Simd; + type Vector = core_simd::simd::Simd; type Scalar = $scalar; impl_common_integer_tests! { Vector, Scalar } @@ -314,7 +314,7 @@ macro_rules! impl_unsigned_tests { { $scalar:tt } => { mod $scalar { use core_simd::simd::SimdUint; - type Vector = core_simd::Simd; + type Vector = core_simd::simd::Simd; type Scalar = $scalar; impl_common_integer_tests! { Vector, Scalar } @@ -348,8 +348,8 @@ macro_rules! impl_unsigned_tests { macro_rules! impl_float_tests { { $scalar:tt, $int_scalar:tt } => { mod $scalar { - use core_simd::SimdFloat; - type Vector = core_simd::Simd; + use core_simd::simd::SimdFloat; + type Vector = core_simd::simd::Simd; type Scalar = $scalar; impl_unary_op_test!(Scalar, Neg::neg); diff --git a/library/portable-simd/crates/core_simd/tests/pointers.rs b/library/portable-simd/crates/core_simd/tests/pointers.rs new file mode 100644 index 00000000000..0ae8f83b8b9 --- /dev/null +++ b/library/portable-simd/crates/core_simd/tests/pointers.rs @@ -0,0 +1,111 @@ +#![feature(portable_simd, strict_provenance)] + +use core_simd::simd::{Simd, SimdConstPtr, SimdMutPtr}; + +macro_rules! common_tests { + { $constness:ident } => { + test_helpers::test_lanes! { + fn is_null() { + test_helpers::test_unary_mask_elementwise( + &Simd::<*$constness u32, LANES>::is_null, + &<*$constness u32>::is_null, + &|_| true, + ); + } + + fn addr() { + test_helpers::test_unary_elementwise( + &Simd::<*$constness u32, LANES>::addr, + &<*$constness u32>::addr, + &|_| true, + ); + } + + fn with_addr() { + test_helpers::test_binary_elementwise( + &Simd::<*$constness u32, LANES>::with_addr, + &<*$constness u32>::with_addr, + &|_, _| true, + ); + } + + fn expose_addr() { + test_helpers::test_unary_elementwise( + &Simd::<*$constness u32, LANES>::expose_addr, + &<*$constness u32>::expose_addr, + &|_| true, + ); + } + + fn wrapping_offset() { + test_helpers::test_binary_elementwise( + &Simd::<*$constness u32, LANES>::wrapping_offset, + &<*$constness u32>::wrapping_offset, + &|_, _| true, + ); + } + + fn wrapping_add() { + test_helpers::test_binary_elementwise( + &Simd::<*$constness u32, LANES>::wrapping_add, + &<*$constness u32>::wrapping_add, + &|_, _| true, + ); + } + + fn wrapping_sub() { + test_helpers::test_binary_elementwise( + &Simd::<*$constness u32, LANES>::wrapping_sub, + &<*$constness u32>::wrapping_sub, + &|_, _| true, + ); + } + } + } +} + +mod const_ptr { + use super::*; + common_tests! { const } + + test_helpers::test_lanes! { + fn cast_mut() { + test_helpers::test_unary_elementwise( + &Simd::<*const u32, LANES>::cast_mut, + &<*const u32>::cast_mut, + &|_| true, + ); + } + + fn from_exposed_addr() { + test_helpers::test_unary_elementwise( + &Simd::<*const u32, LANES>::from_exposed_addr, + &core::ptr::from_exposed_addr::, + &|_| true, + ); + } + } +} + +mod mut_ptr { + use super::*; + common_tests! { mut } + + test_helpers::test_lanes! { + fn cast_const() { + test_helpers::test_unary_elementwise( + &Simd::<*mut u32, LANES>::cast_const, + &<*mut u32>::cast_const, + &|_| true, + ); + } + + fn from_exposed_addr() { + test_helpers::test_unary_elementwise( + &Simd::<*mut u32, LANES>::from_exposed_addr, + &core::ptr::from_exposed_addr_mut::, + &|_| true, + ); + } + } +} diff --git a/library/portable-simd/crates/core_simd/tests/round.rs b/library/portable-simd/crates/core_simd/tests/round.rs index 484fd5bf47d..8b9638ad466 100644 --- a/library/portable-simd/crates/core_simd/tests/round.rs +++ b/library/portable-simd/crates/core_simd/tests/round.rs @@ -5,7 +5,7 @@ macro_rules! float_rounding_test { mod $scalar { use std_float::StdFloat; - type Vector = core_simd::Simd<$scalar, LANES>; + type Vector = core_simd::simd::Simd<$scalar, LANES>; type Scalar = $scalar; type IntScalar = $int_scalar; diff --git a/library/portable-simd/crates/core_simd/tests/swizzle.rs b/library/portable-simd/crates/core_simd/tests/swizzle.rs index 51c63611aba..8cd7c33e823 100644 --- a/library/portable-simd/crates/core_simd/tests/swizzle.rs +++ b/library/portable-simd/crates/core_simd/tests/swizzle.rs @@ -1,5 +1,5 @@ #![feature(portable_simd)] -use core_simd::{Simd, Swizzle}; +use core_simd::simd::{Simd, Swizzle}; #[cfg(target_arch = "wasm32")] use wasm_bindgen_test::*; @@ -60,3 +60,17 @@ fn interleave() { assert_eq!(even, a); assert_eq!(odd, b); } + +// portable-simd#298 +#[test] +#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)] +fn interleave_one() { + let a = Simd::from_array([0]); + let b = Simd::from_array([1]); + let (lo, hi) = a.interleave(b); + assert_eq!(lo.to_array(), [0]); + assert_eq!(hi.to_array(), [1]); + let (even, odd) = lo.deinterleave(hi); + assert_eq!(even, a); + assert_eq!(odd, b); +} diff --git a/library/portable-simd/crates/core_simd/tests/swizzle_dyn.rs b/library/portable-simd/crates/core_simd/tests/swizzle_dyn.rs new file mode 100644 index 00000000000..646cd5f3383 --- /dev/null +++ b/library/portable-simd/crates/core_simd/tests/swizzle_dyn.rs @@ -0,0 +1,74 @@ +#![feature(portable_simd)] +use core::{fmt, ops::RangeInclusive}; +use proptest; +use test_helpers::{self, biteq, make_runner, prop_assert_biteq}; + +fn swizzle_dyn_scalar_ver(values: [u8; N], idxs: [u8; N]) -> [u8; N] { + let mut array = [0; N]; + for (i, k) in idxs.into_iter().enumerate() { + if (k as usize) < N { + array[i] = values[k as usize]; + }; + } + array +} + +test_helpers::test_lanes! { + fn swizzle_dyn() { + match_simd_with_fallback( + &core_simd::simd::Simd::::swizzle_dyn, + &swizzle_dyn_scalar_ver, + &|_, _| true, + ); + } +} + +fn match_simd_with_fallback( + fv: &dyn Fn(Vector, Vector) -> VectorResult, + fs: &dyn Fn([Scalar; N], [Scalar; N]) -> [ScalarResult; N], + check: &dyn Fn([Scalar; N], [Scalar; N]) -> bool, +) where + Scalar: Copy + fmt::Debug + SwizzleStrategy, + ScalarResult: Copy + biteq::BitEq + fmt::Debug + SwizzleStrategy, + Vector: Into<[Scalar; N]> + From<[Scalar; N]> + Copy, + VectorResult: Into<[ScalarResult; N]> + From<[ScalarResult; N]> + Copy, +{ + test_swizzles_2(&|x: [Scalar; N], y: [Scalar; N]| { + proptest::prop_assume!(check(x, y)); + let result_v: [ScalarResult; N] = fv(x.into(), y.into()).into(); + let result_s: [ScalarResult; N] = fs(x, y); + crate::prop_assert_biteq!(result_v, result_s); + Ok(()) + }); +} + +fn test_swizzles_2( + f: &dyn Fn(A, B) -> proptest::test_runner::TestCaseResult, +) { + let mut runner = make_runner(); + runner + .run( + &(A::swizzled_strategy(), B::swizzled_strategy()), + |(a, b)| f(a, b), + ) + .unwrap(); +} + +pub trait SwizzleStrategy { + type Strategy: proptest::strategy::Strategy; + fn swizzled_strategy() -> Self::Strategy; +} + +impl SwizzleStrategy for u8 { + type Strategy = RangeInclusive; + fn swizzled_strategy() -> Self::Strategy { + 0..=64 + } +} + +impl SwizzleStrategy for [T; N] { + type Strategy = test_helpers::array::UniformArrayStrategy; + fn swizzled_strategy() -> Self::Strategy { + Self::Strategy::new(T::swizzled_strategy()) + } +} diff --git a/library/portable-simd/crates/core_simd/tests/to_bytes.rs b/library/portable-simd/crates/core_simd/tests/to_bytes.rs index debb4335e2c..be0ee4349c5 100644 --- a/library/portable-simd/crates/core_simd/tests/to_bytes.rs +++ b/library/portable-simd/crates/core_simd/tests/to_bytes.rs @@ -2,7 +2,7 @@ #![allow(incomplete_features)] #![cfg(feature = "generic_const_exprs")] -use core_simd::Simd; +use core_simd::simd::Simd; #[test] fn byte_convert() { diff --git a/library/portable-simd/crates/core_simd/tests/try_from_slice.rs b/library/portable-simd/crates/core_simd/tests/try_from_slice.rs new file mode 100644 index 00000000000..859e3b94f2c --- /dev/null +++ b/library/portable-simd/crates/core_simd/tests/try_from_slice.rs @@ -0,0 +1,25 @@ +#![feature(portable_simd)] + +#[cfg(target_arch = "wasm32")] +use wasm_bindgen_test::*; + +#[cfg(target_arch = "wasm32")] +wasm_bindgen_test_configure!(run_in_browser); + +use core_simd::simd::i32x4; + +#[test] +#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)] +fn try_from_slice() { + // Equal length + assert_eq!( + i32x4::try_from([1, 2, 3, 4].as_slice()).unwrap(), + i32x4::from_array([1, 2, 3, 4]) + ); + + // Slice length > vector length + assert!(i32x4::try_from([1, 2, 3, 4, 5].as_slice()).is_err()); + + // Slice length < vector length + assert!(i32x4::try_from([1, 2, 3].as_slice()).is_err()); +} diff --git a/library/portable-simd/crates/test_helpers/Cargo.toml b/library/portable-simd/crates/test_helpers/Cargo.toml index a04b0961d7f..1d2bc8b519a 100644 --- a/library/portable-simd/crates/test_helpers/Cargo.toml +++ b/library/portable-simd/crates/test_helpers/Cargo.toml @@ -8,3 +8,6 @@ publish = false version = "0.10" default-features = false features = ["alloc"] + +[features] +all_lane_counts = [] diff --git a/library/portable-simd/crates/test_helpers/src/array.rs b/library/portable-simd/crates/test_helpers/src/array.rs index 5ffc9226976..984a427320d 100644 --- a/library/portable-simd/crates/test_helpers/src/array.rs +++ b/library/portable-simd/crates/test_helpers/src/array.rs @@ -41,6 +41,7 @@ where fn new_tree(&self, runner: &mut TestRunner) -> NewTree { let tree: [S::Tree; LANES] = unsafe { + #[allow(clippy::uninit_assumed_init)] let mut tree: [MaybeUninit; LANES] = MaybeUninit::uninit().assume_init(); for t in tree.iter_mut() { *t = MaybeUninit::new(self.strategy.new_tree(runner)?) @@ -60,6 +61,7 @@ impl ValueTree for ArrayValueTree<[T; LANES]> fn current(&self) -> Self::Value { unsafe { + #[allow(clippy::uninit_assumed_init)] let mut value: [MaybeUninit; LANES] = MaybeUninit::uninit().assume_init(); for (tree_elem, value_elem) in self.tree.iter().zip(value.iter_mut()) { *value_elem = MaybeUninit::new(tree_elem.current()); diff --git a/library/portable-simd/crates/test_helpers/src/biteq.rs b/library/portable-simd/crates/test_helpers/src/biteq.rs index 00350e22418..7d91260d838 100644 --- a/library/portable-simd/crates/test_helpers/src/biteq.rs +++ b/library/portable-simd/crates/test_helpers/src/biteq.rs @@ -55,6 +55,26 @@ macro_rules! impl_float_biteq { impl_float_biteq! { f32, f64 } +impl BitEq for *const T { + fn biteq(&self, other: &Self) -> bool { + self == other + } + + fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result { + write!(f, "{:?}", self) + } +} + +impl BitEq for *mut T { + fn biteq(&self, other: &Self) -> bool { + self == other + } + + fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result { + write!(f, "{:?}", self) + } +} + impl BitEq for [T; N] { fn biteq(&self, other: &Self) -> bool { self.iter() diff --git a/library/portable-simd/crates/test_helpers/src/lib.rs b/library/portable-simd/crates/test_helpers/src/lib.rs index 141bee18a9a..b26cdc311a2 100644 --- a/library/portable-simd/crates/test_helpers/src/lib.rs +++ b/library/portable-simd/crates/test_helpers/src/lib.rs @@ -38,6 +38,28 @@ impl_num! { usize } impl_num! { f32 } impl_num! { f64 } +impl DefaultStrategy for *const T { + type Strategy = proptest::strategy::Map *const T>; + fn default_strategy() -> Self::Strategy { + fn map(x: isize) -> *const T { + x as _ + } + use proptest::strategy::Strategy; + proptest::num::isize::ANY.prop_map(map) + } +} + +impl DefaultStrategy for *mut T { + type Strategy = proptest::strategy::Map *mut T>; + fn default_strategy() -> Self::Strategy { + fn map(x: isize) -> *mut T { + x as _ + } + use proptest::strategy::Strategy; + proptest::num::isize::ANY.prop_map(map) + } +} + #[cfg(not(target_arch = "wasm32"))] impl DefaultStrategy for u128 { type Strategy = proptest::num::u128::Any; @@ -135,21 +157,21 @@ pub fn test_unary_elementwise ScalarResult, check: &dyn Fn([Scalar; LANES]) -> bool, ) where - Scalar: Copy + Default + core::fmt::Debug + DefaultStrategy, - ScalarResult: Copy + Default + biteq::BitEq + core::fmt::Debug + DefaultStrategy, + Scalar: Copy + core::fmt::Debug + DefaultStrategy, + ScalarResult: Copy + biteq::BitEq + core::fmt::Debug + DefaultStrategy, Vector: Into<[Scalar; LANES]> + From<[Scalar; LANES]> + Copy, VectorResult: Into<[ScalarResult; LANES]> + From<[ScalarResult; LANES]> + Copy, { test_1(&|x: [Scalar; LANES]| { proptest::prop_assume!(check(x)); let result_1: [ScalarResult; LANES] = fv(x.into()).into(); - let result_2: [ScalarResult; LANES] = { - let mut result = [ScalarResult::default(); LANES]; - for (i, o) in x.iter().zip(result.iter_mut()) { - *o = fs(*i); - } - result - }; + let result_2: [ScalarResult; LANES] = x + .iter() + .copied() + .map(fs) + .collect::>() + .try_into() + .unwrap(); crate::prop_assert_biteq!(result_1, result_2); Ok(()) }); @@ -162,7 +184,7 @@ pub fn test_unary_mask_elementwise( fs: &dyn Fn(Scalar) -> bool, check: &dyn Fn([Scalar; LANES]) -> bool, ) where - Scalar: Copy + Default + core::fmt::Debug + DefaultStrategy, + Scalar: Copy + core::fmt::Debug + DefaultStrategy, Vector: Into<[Scalar; LANES]> + From<[Scalar; LANES]> + Copy, Mask: Into<[bool; LANES]> + From<[bool; LANES]> + Copy, { @@ -196,9 +218,9 @@ pub fn test_binary_elementwise< fs: &dyn Fn(Scalar1, Scalar2) -> ScalarResult, check: &dyn Fn([Scalar1; LANES], [Scalar2; LANES]) -> bool, ) where - Scalar1: Copy + Default + core::fmt::Debug + DefaultStrategy, - Scalar2: Copy + Default + core::fmt::Debug + DefaultStrategy, - ScalarResult: Copy + Default + biteq::BitEq + core::fmt::Debug + DefaultStrategy, + Scalar1: Copy + core::fmt::Debug + DefaultStrategy, + Scalar2: Copy + core::fmt::Debug + DefaultStrategy, + ScalarResult: Copy + biteq::BitEq + core::fmt::Debug + DefaultStrategy, Vector1: Into<[Scalar1; LANES]> + From<[Scalar1; LANES]> + Copy, Vector2: Into<[Scalar2; LANES]> + From<[Scalar2; LANES]> + Copy, VectorResult: Into<[ScalarResult; LANES]> + From<[ScalarResult; LANES]> + Copy, @@ -206,13 +228,14 @@ pub fn test_binary_elementwise< test_2(&|x: [Scalar1; LANES], y: [Scalar2; LANES]| { proptest::prop_assume!(check(x, y)); let result_1: [ScalarResult; LANES] = fv(x.into(), y.into()).into(); - let result_2: [ScalarResult; LANES] = { - let mut result = [ScalarResult::default(); LANES]; - for ((i1, i2), o) in x.iter().zip(y.iter()).zip(result.iter_mut()) { - *o = fs(*i1, *i2); - } - result - }; + let result_2: [ScalarResult; LANES] = x + .iter() + .copied() + .zip(y.iter().copied()) + .map(|(x, y)| fs(x, y)) + .collect::>() + .try_into() + .unwrap(); crate::prop_assert_biteq!(result_1, result_2); Ok(()) }); @@ -333,6 +356,39 @@ pub fn test_ternary_elementwise< ); } +#[doc(hidden)] +#[macro_export] +macro_rules! test_lanes_helper { + ($($(#[$meta:meta])* $fn_name:ident $lanes:literal;)+) => { + $( + #[test] + $(#[$meta])* + fn $fn_name() { + implementation::<$lanes>(); + } + )+ + }; + ( + $(#[$meta:meta])+; + $($(#[$meta_before:meta])+ $fn_name_before:ident $lanes_before:literal;)* + $fn_name:ident $lanes:literal; + $($fn_name_rest:ident $lanes_rest:literal;)* + ) => { + $crate::test_lanes_helper!( + $(#[$meta])+; + $($(#[$meta_before])+ $fn_name_before $lanes_before;)* + $(#[$meta])+ $fn_name $lanes; + $($fn_name_rest $lanes_rest;)* + ); + }; + ( + $(#[$meta_ignored:meta])+; + $($(#[$meta:meta])+ $fn_name:ident $lanes:literal;)+ + ) => { + $crate::test_lanes_helper!($($(#[$meta])+ $fn_name $lanes;)+); + }; +} + /// Expand a const-generic test into separate tests for each possible lane count. #[macro_export] macro_rules! test_lanes { @@ -345,57 +401,96 @@ macro_rules! test_lanes { fn implementation() where - core_simd::LaneCount<$lanes>: core_simd::SupportedLaneCount, + core_simd::simd::LaneCount<$lanes>: core_simd::simd::SupportedLaneCount, $body #[cfg(target_arch = "wasm32")] wasm_bindgen_test::wasm_bindgen_test_configure!(run_in_browser); - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] - fn lanes_1() { - implementation::<1>(); - } + $crate::test_lanes_helper!( + #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]; + lanes_1 1; + lanes_2 2; + lanes_4 4; + ); - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] - fn lanes_2() { - implementation::<2>(); - } - - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] - fn lanes_4() { - implementation::<4>(); - } - - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow - fn lanes_8() { - implementation::<8>(); - } + $crate::test_lanes_helper!( + #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]; + lanes_8 8; + lanes_16 16; + lanes_32 32; + lanes_64 64; + ); - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] - #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow - fn lanes_16() { - implementation::<16>(); - } + #[cfg(feature = "all_lane_counts")] + $crate::test_lanes_helper!( + // test some odd and even non-power-of-2 lengths on miri + #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]; + lanes_3 3; + lanes_5 5; + lanes_6 6; + ); - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] + #[cfg(feature = "all_lane_counts")] #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow - fn lanes_32() { - implementation::<32>(); - } - - #[test] - #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)] - #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow - fn lanes_64() { - implementation::<64>(); - } + $crate::test_lanes_helper!( + #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]; + lanes_7 7; + lanes_9 9; + lanes_10 10; + lanes_11 11; + lanes_12 12; + lanes_13 13; + lanes_14 14; + lanes_15 15; + lanes_17 17; + lanes_18 18; + lanes_19 19; + lanes_20 20; + lanes_21 21; + lanes_22 22; + lanes_23 23; + lanes_24 24; + lanes_25 25; + lanes_26 26; + lanes_27 27; + lanes_28 28; + lanes_29 29; + lanes_30 30; + lanes_31 31; + lanes_33 33; + lanes_34 34; + lanes_35 35; + lanes_36 36; + lanes_37 37; + lanes_38 38; + lanes_39 39; + lanes_40 40; + lanes_41 41; + lanes_42 42; + lanes_43 43; + lanes_44 44; + lanes_45 45; + lanes_46 46; + lanes_47 47; + lanes_48 48; + lanes_49 49; + lanes_50 50; + lanes_51 51; + lanes_52 52; + lanes_53 53; + lanes_54 54; + lanes_55 55; + lanes_56 56; + lanes_57 57; + lanes_58 58; + lanes_59 59; + lanes_60 60; + lanes_61 61; + lanes_62 62; + lanes_63 63; + ); } )* } @@ -413,50 +508,93 @@ macro_rules! test_lanes_panic { fn implementation() where - core_simd::LaneCount<$lanes>: core_simd::SupportedLaneCount, + core_simd::simd::LaneCount<$lanes>: core_simd::simd::SupportedLaneCount, $body - #[test] - #[should_panic] - fn lanes_1() { - implementation::<1>(); - } + $crate::test_lanes_helper!( + #[should_panic]; + lanes_1 1; + lanes_2 2; + lanes_4 4; + ); - #[test] - #[should_panic] - fn lanes_2() { - implementation::<2>(); - } + #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow + $crate::test_lanes_helper!( + #[should_panic]; + lanes_8 8; + lanes_16 16; + lanes_32 32; + lanes_64 64; + ); - #[test] - #[should_panic] - fn lanes_4() { - implementation::<4>(); - } + #[cfg(feature = "all_lane_counts")] + $crate::test_lanes_helper!( + // test some odd and even non-power-of-2 lengths on miri + #[should_panic]; + lanes_3 3; + lanes_5 5; + lanes_6 6; + ); - #[test] - #[should_panic] - fn lanes_8() { - implementation::<8>(); - } - - #[test] - #[should_panic] - fn lanes_16() { - implementation::<16>(); - } - - #[test] - #[should_panic] - fn lanes_32() { - implementation::<32>(); - } - - #[test] - #[should_panic] - fn lanes_64() { - implementation::<64>(); - } + #[cfg(feature = "all_lane_counts")] + #[cfg(not(miri))] // Miri intrinsic implementations are uniform and larger tests are sloooow + $crate::test_lanes_helper!( + #[should_panic]; + lanes_7 7; + lanes_9 9; + lanes_10 10; + lanes_11 11; + lanes_12 12; + lanes_13 13; + lanes_14 14; + lanes_15 15; + lanes_17 17; + lanes_18 18; + lanes_19 19; + lanes_20 20; + lanes_21 21; + lanes_22 22; + lanes_23 23; + lanes_24 24; + lanes_25 25; + lanes_26 26; + lanes_27 27; + lanes_28 28; + lanes_29 29; + lanes_30 30; + lanes_31 31; + lanes_33 33; + lanes_34 34; + lanes_35 35; + lanes_36 36; + lanes_37 37; + lanes_38 38; + lanes_39 39; + lanes_40 40; + lanes_41 41; + lanes_42 42; + lanes_43 43; + lanes_44 44; + lanes_45 45; + lanes_46 46; + lanes_47 47; + lanes_48 48; + lanes_49 49; + lanes_50 50; + lanes_51 51; + lanes_52 52; + lanes_53 53; + lanes_54 54; + lanes_55 55; + lanes_56 56; + lanes_57 57; + lanes_58 58; + lanes_59 59; + lanes_60 60; + lanes_61 61; + lanes_62 62; + lanes_63 63; + ); } )* }