Auto merge of #93146 - workingjubilee:use-std-simd, r=Mark-Simulacrum

pub use std::simd::StdFloat;

Syncs portable-simd up to commit rust-lang/portable-simd@03f6fbb21e,
Diff: 533f0fc81a...03f6fbb21e

This sync requires a little bit more legwork because it also introduces a trait into `std::simd`, so that it is no longer simply a reexport of `core::simd`. Out of simple-minded consistency and to allow more options, I replicated the pattern for the way `core::simd` is integrated in the first place, however this is not necessary if it doesn't acquire any interdependencies inside `std`: it could be a simple crate reexport. I just don't know yet if that will happen or not.

To summarize other misc changes:
- Shifts no longer panic, now wrap on too-large shifts (like `Simd` integers usually do!)
- mask16x32 will now be many i16s, not many i32s... 🙃
- `#[must_use]` is spread around generously
- Adjusts division, float min/max, and `Mask::{from,to}_array` internally to be faster
- Adds the much-requested `Simd::cast::<U>` function (equivalent to `simd.to_array().map(|lane| lane as U)`)

library/portable-simd/Cargo.toml

@@ -2,5 +2,6 @@
 
 members = [
     "crates/core_simd",
+    "crates/std_float",
     "crates/test_helpers",
 ]

library/portable-simd/crates/core_simd/Cargo.toml

@@ -26,3 +26,6 @@ features = ["alloc"]
 
 [dev-dependencies.test_helpers]
 path = "../test_helpers"
+
+[dev-dependencies]
+std_float = { path = "../std_float/", features = ["as_crate"] }

library/portable-simd/crates/core_simd/examples/nbody.rs

@@ -1,11 +1,13 @@
-#![cfg_attr(feature = "std", feature(portable_simd))]
+#![feature(portable_simd)]
+extern crate std_float;
 
 /// Benchmarks game nbody code
 /// Taken from the `packed_simd` crate
 /// Run this benchmark with `cargo test --example nbody`
-#[cfg(feature = "std")]
 mod nbody {
-    use core_simd::*;
+    use core_simd::simd::*;
+    #[allow(unused)] // False positive?
+    use std_float::StdFloat;
 
     use std::f64::consts::PI;
     const SOLAR_MASS: f64 = 4.0 * PI * PI;
@@ -167,7 +169,6 @@ mod nbody {
     }
 }
 
-#[cfg(feature = "std")]
 #[cfg(test)]
 mod tests {
     // Good enough for demonstration purposes, not going for strictness here.
@@ -184,7 +185,6 @@ mod tests {
 }
 
 fn main() {
-    #[cfg(feature = "std")]
     {
         let (energy_before, energy_after) = nbody::run(1000);
         println!("Energy before: {}", energy_before);

library/portable-simd/crates/core_simd/src/intrinsics.rs

@@ -39,6 +39,10 @@ extern "platform-intrinsic" {
 
     /// fptoui/fptosi/uitofp/sitofp
     pub(crate) fn simd_cast<T, U>(x: T) -> U;
+    /// follows Rust's `T as U` semantics, including saturating float casts
+    /// which amounts to the same as `simd_cast` for many cases
+    #[cfg(not(bootstrap))]
+    pub(crate) fn simd_as<T, U>(x: T) -> U;
 
     /// neg/fneg
     pub(crate) fn simd_neg<T>(x: T) -> T;
@@ -46,6 +50,10 @@ extern "platform-intrinsic" {
     /// fabs
     pub(crate) fn simd_fabs<T>(x: T) -> T;
 
+    // minnum/maxnum
+    pub(crate) fn simd_fmin<T>(x: T, y: T) -> T;
+    pub(crate) fn simd_fmax<T>(x: T, y: T) -> T;
+
     pub(crate) fn simd_eq<T, U>(x: T, y: T) -> U;
     pub(crate) fn simd_ne<T, U>(x: T, y: T) -> U;
     pub(crate) fn simd_lt<T, U>(x: T, y: T) -> U;
@@ -87,29 +95,3 @@ extern "platform-intrinsic" {
     #[allow(unused)]
     pub(crate) fn simd_select_bitmask<M, T>(m: M, a: T, b: T) -> T;
 }
-
-#[cfg(feature = "std")]
-mod std {
-    extern "platform-intrinsic" {
-        // ceil
-        pub(crate) fn simd_ceil<T>(x: T) -> T;
-
-        // floor
-        pub(crate) fn simd_floor<T>(x: T) -> T;
-
-        // round
-        pub(crate) fn simd_round<T>(x: T) -> T;
-
-        // trunc
-        pub(crate) fn simd_trunc<T>(x: T) -> T;
-
-        // fsqrt
-        pub(crate) fn simd_fsqrt<T>(x: T) -> T;
-
-        // fma
-        pub(crate) fn simd_fma<T>(x: T, y: T, z: T) -> T;
-    }
-}
-
-#[cfg(feature = "std")]
-pub(crate) use crate::simd::intrinsics::std::*;

library/portable-simd/crates/core_simd/src/masks.rs

@@ -12,9 +12,10 @@
 )]
 mod mask_impl;
 
+use crate::simd::intrinsics;
 use crate::simd::{LaneCount, Simd, SimdElement, SupportedLaneCount};
 use core::cmp::Ordering;
-use core::fmt;
+use core::{fmt, mem};
 
 mod sealed {
     use super::*;
@@ -105,22 +106,39 @@ where
         Self(mask_impl::Mask::splat(value))
     }
 
-    /// Converts an array to a SIMD vector.
+    /// Converts an array of bools to a SIMD mask.
     pub fn from_array(array: [bool; LANES]) -> Self {
-        let mut vector = Self::splat(false);
-        for (i, v) in array.iter().enumerate() {
-            vector.set(i, *v);
+        // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of
+        //     true:    0b_0000_0001
+        //     false:   0b_0000_0000
+        // Thus, an array of bools is also a valid array of bytes: [u8; N]
+        // This would be hypothetically valid as an "in-place" transmute,
+        // but these are "dependently-sized" types, so copy elision it is!
+        unsafe {
+            let bytes: [u8; LANES] = mem::transmute_copy(&array);
+            let bools: Simd<i8, LANES> =
+                intrinsics::simd_ne(Simd::from_array(bytes), Simd::splat(0u8));
+            Mask::from_int_unchecked(intrinsics::simd_cast(bools))
         }
-        vector
     }
 
-    /// Converts a SIMD vector to an array.
+    /// Converts a SIMD mask to an array of bools.
     pub fn to_array(self) -> [bool; LANES] {
-        let mut array = [false; LANES];
-        for (i, v) in array.iter_mut().enumerate() {
-            *v = self.test(i);
+        // This follows mostly the same logic as from_array.
+        // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of
+        //     true:    0b_0000_0001
+        //     false:   0b_0000_0000
+        // Thus, an array of bools is also a valid array of bytes: [u8; N]
+        // Since our masks are equal to integers where all bits are set,
+        // we can simply convert them to i8s, and then bitand them by the
+        // bitpattern for Rust's "true" bool.
+        // This would be hypothetically valid as an "in-place" transmute,
+        // but these are "dependently-sized" types, so copy elision it is!
+        unsafe {
+            let mut bytes: Simd<i8, LANES> = intrinsics::simd_cast(self.to_int());
+            bytes &= Simd::splat(1i8);
+            mem::transmute_copy(&bytes)
         }
-        array
     }
 
     /// Converts a vector of integers to a mask, where 0 represents `false` and -1
@@ -516,7 +534,7 @@ pub type mask16x8 = Mask<i16, 8>;
 pub type mask16x16 = Mask<i16, 16>;
 
 /// Vector of 32 16-bit masks
-pub type mask16x32 = Mask<i32, 32>;
+pub type mask16x32 = Mask<i16, 32>;
 
 /// Vector of two 32-bit masks
 pub type mask32x2 = Mask<i32, 2>;

library/portable-simd/crates/core_simd/src/ops.rs

@@ -1,4 +1,3 @@
-use crate::simd::intrinsics;
 use crate::simd::{LaneCount, Simd, SimdElement, SupportedLaneCount};
 use core::ops::{Add, Mul};
 use core::ops::{BitAnd, BitOr, BitXor};
@@ -32,232 +31,211 @@ where
     }
 }
 
-/// Checks if the right-hand side argument of a left- or right-shift would cause overflow.
-fn invalid_shift_rhs<T>(rhs: T) -> bool
-where
-    T: Default + PartialOrd + core::convert::TryFrom<usize>,
-    <T as core::convert::TryFrom<usize>>::Error: core::fmt::Debug,
-{
-    let bits_in_type = T::try_from(8 * core::mem::size_of::<T>()).unwrap();
-    rhs < T::default() || rhs >= bits_in_type
+macro_rules! unsafe_base {
+    ($lhs:ident, $rhs:ident, {$simd_call:ident}, $($_:tt)*) => {
+        unsafe { $crate::simd::intrinsics::$simd_call($lhs, $rhs) }
+    };
 }
 
-/// Automatically implements operators over references in addition to the provided operator.
-macro_rules! impl_ref_ops {
-    // binary op
-    {
-        impl<const $lanes:ident: usize> core::ops::$trait:ident<$rhs:ty> for $type:ty
-        where
-            LaneCount<$lanes2:ident>: SupportedLaneCount,
+/// SAFETY: This macro should not be used for anything except Shl or Shr, and passed the appropriate shift intrinsic.
+/// It handles performing a bitand in addition to calling the shift operator, so that the result
+/// is well-defined: LLVM can return a poison value if you shl, lshr, or ashr if rhs >= <Int>::BITS
+/// At worst, this will maybe add another instruction and cycle,
+/// at best, it may open up more optimization opportunities,
+/// or simply be elided entirely, especially for SIMD ISAs which default to this.
+///
+// FIXME: Consider implementing this in cg_llvm instead?
+// cg_clif defaults to this, and scalar MIR shifts also default to wrapping
+macro_rules! wrap_bitshift {
+    ($lhs:ident, $rhs:ident, {$simd_call:ident}, $int:ident) => {
+        unsafe {
+            $crate::simd::intrinsics::$simd_call(
+                $lhs,
+                $rhs.bitand(Simd::splat(<$int>::BITS as $int - 1)),
+            )
+        }
+    };
+}
+
+// Division by zero is poison, according to LLVM.
+// So is dividing the MIN value of a signed integer by -1,
+// since that would return MAX + 1.
+// FIXME: Rust allows <SInt>::MIN / -1,
+// so we should probably figure out how to make that safe.
+macro_rules! int_divrem_guard {
+    (   $lhs:ident,
+        $rhs:ident,
+        {   const PANIC_ZERO: &'static str = $zero:literal;
+            const PANIC_OVERFLOW: &'static str = $overflow:literal;
+            $simd_call:ident
+        },
+        $int:ident ) => {
+        if $rhs.lanes_eq(Simd::splat(0)).any() {
+            panic!($zero);
+        } else if <$int>::MIN != 0
+            && ($lhs.lanes_eq(Simd::splat(<$int>::MIN))
+                // type inference can break here, so cut an SInt to size
+                & $rhs.lanes_eq(Simd::splat(-1i64 as _))).any()
         {
-            type Output = $output:ty;
-
-            $(#[$attrs:meta])*
-            fn $fn:ident($self_tok:ident, $rhs_arg:ident: $rhs_arg_ty:ty) -> Self::Output $body:tt
-        }
-    } => {
-        impl<const $lanes: usize> core::ops::$trait<$rhs> for $type
-        where
-            LaneCount<$lanes2>: SupportedLaneCount,
-        {
-            type Output = $output;
-
-            $(#[$attrs])*
-            fn $fn($self_tok, $rhs_arg: $rhs_arg_ty) -> Self::Output $body
+            panic!($overflow);
+        } else {
+            unsafe { $crate::simd::intrinsics::$simd_call($lhs, $rhs) }
         }
     };
 }
 
-/// Automatically implements operators over vectors and scalars for a particular vector.
-macro_rules! impl_op {
-    { impl Add for $scalar:ty } => {
-        impl_op! { @binary $scalar, Add::add, simd_add }
-    };
-    { impl Sub for $scalar:ty } => {
-        impl_op! { @binary $scalar, Sub::sub, simd_sub }
-    };
-    { impl Mul for $scalar:ty } => {
-        impl_op! { @binary $scalar, Mul::mul, simd_mul }
-    };
-    { impl Div for $scalar:ty } => {
-        impl_op! { @binary $scalar, Div::div, simd_div }
-    };
-    { impl Rem for $scalar:ty } => {
-        impl_op! { @binary $scalar, Rem::rem, simd_rem }
-    };
-    { impl Shl for $scalar:ty } => {
-        impl_op! { @binary $scalar, Shl::shl, simd_shl }
-    };
-    { impl Shr for $scalar:ty } => {
-        impl_op! { @binary $scalar, Shr::shr, simd_shr }
-    };
-    { impl BitAnd for $scalar:ty } => {
-        impl_op! { @binary $scalar, BitAnd::bitand, simd_and }
-    };
-    { impl BitOr for $scalar:ty } => {
-        impl_op! { @binary $scalar, BitOr::bitor, simd_or }
-    };
-    { impl BitXor for $scalar:ty } => {
-        impl_op! { @binary $scalar, BitXor::bitxor, simd_xor }
-    };
+macro_rules! for_base_types {
+    (   T = ($($scalar:ident),*);
+        type Lhs = Simd<T, N>;
+        type Rhs = Simd<T, N>;
+        type Output = $out:ty;
 
-    // generic binary op with assignment when output is `Self`
-    { @binary $scalar:ty, $trait:ident :: $trait_fn:ident, $intrinsic:ident } => {
-        impl_ref_ops! {
-            impl<const LANES: usize> core::ops::$trait<Self> for Simd<$scalar, LANES>
-            where
-                LaneCount<LANES>: SupportedLaneCount,
-            {
-                type Output = Self;
+        impl $op:ident::$call:ident {
+            $macro_impl:ident $inner:tt
+        }) => {
+            $(
+                impl<const N: usize> $op<Self> for Simd<$scalar, N>
+                where
+                    $scalar: SimdElement,
+                    LaneCount<N>: SupportedLaneCount,
+                {
+                    type Output = $out;
 
-                #[inline]
-                fn $trait_fn(self, rhs: Self) -> Self::Output {
-                    unsafe {
-                        intrinsics::$intrinsic(self, rhs)
+                    #[inline]
+                    #[must_use = "operator returns a new vector without mutating the inputs"]
+                    fn $call(self, rhs: Self) -> Self::Output {
+                        $macro_impl!(self, rhs, $inner, $scalar)
                     }
-                }
-            }
+                })*
+    }
+}
+
+// A "TokenTree muncher": takes a set of scalar types `T = {};`
+// type parameters for the ops it implements, `Op::fn` names,
+// and a macro that expands into an expr, substituting in an intrinsic.
+// It passes that to for_base_types, which expands an impl for the types,
+// using the expanded expr in the function, and recurses with itself.
+//
+// tl;dr impls a set of ops::{Traits} for a set of types
+macro_rules! for_base_ops {
+    (
+        T = $types:tt;
+        type Lhs = Simd<T, N>;
+        type Rhs = Simd<T, N>;
+        type Output = $out:ident;
+        impl $op:ident::$call:ident
+            $inner:tt
+        $($rest:tt)*
+    ) => {
+        for_base_types! {
+            T = $types;
+            type Lhs = Simd<T, N>;
+            type Rhs = Simd<T, N>;
+            type Output = $out;
+            impl $op::$call
+                $inner
+        }
+        for_base_ops! {
+            T = $types;
+            type Lhs = Simd<T, N>;
+            type Rhs = Simd<T, N>;
+            type Output = $out;
+            $($rest)*
         }
     };
+    ($($done:tt)*) => {
+        // Done.
+    }
 }
 
-/// Implements floating-point operators for the provided types.
-macro_rules! impl_float_ops {
-    { $($scalar:ty),* } => {
-        $(
-            impl_op! { impl Add for $scalar }
-            impl_op! { impl Sub for $scalar }
-            impl_op! { impl Mul for $scalar }
-            impl_op! { impl Div for $scalar }
-            impl_op! { impl Rem for $scalar }
-        )*
-    };
+// Integers can always accept add, mul, sub, bitand, bitor, and bitxor.
+// For all of these operations, simd_* intrinsics apply wrapping logic.
+for_base_ops! {
+    T = (i8, i16, i32, i64, isize, u8, u16, u32, u64, usize);
+    type Lhs = Simd<T, N>;
+    type Rhs = Simd<T, N>;
+    type Output = Self;
+
+    impl Add::add {
+        unsafe_base { simd_add }
+    }
+
+    impl Mul::mul {
+        unsafe_base { simd_mul }
+    }
+
+    impl Sub::sub {
+        unsafe_base { simd_sub }
+    }
+
+    impl BitAnd::bitand {
+        unsafe_base { simd_and }
+    }
+
+    impl BitOr::bitor {
+        unsafe_base { simd_or }
+    }
+
+    impl BitXor::bitxor {
+        unsafe_base { simd_xor }
+    }
+
+    impl Div::div {
+        int_divrem_guard {
+            const PANIC_ZERO: &'static str = "attempt to divide by zero";
+            const PANIC_OVERFLOW: &'static str = "attempt to divide with overflow";
+            simd_div
+        }
+    }
+
+    impl Rem::rem {
+        int_divrem_guard {
+            const PANIC_ZERO: &'static str = "attempt to calculate the remainder with a divisor of zero";
+            const PANIC_OVERFLOW: &'static str = "attempt to calculate the remainder with overflow";
+            simd_rem
+        }
+    }
+
+    // The only question is how to handle shifts >= <Int>::BITS?
+    // Our current solution uses wrapping logic.
+    impl Shl::shl {
+        wrap_bitshift { simd_shl }
+    }
+
+    impl Shr::shr {
+        wrap_bitshift {
+            // This automatically monomorphizes to lshr or ashr, depending,
+            // so it's fine to use it for both UInts and SInts.
+            simd_shr
+        }
+    }
 }
 
-/// Implements unsigned integer operators for the provided types.
-macro_rules! impl_unsigned_int_ops {
-    { $($scalar:ty),* } => {
-        $(
-            impl_op! { impl Add for $scalar }
-            impl_op! { impl Sub for $scalar }
-            impl_op! { impl Mul for $scalar }
-            impl_op! { impl BitAnd for $scalar }
-            impl_op! { impl BitOr  for $scalar }
-            impl_op! { impl BitXor for $scalar }
+// We don't need any special precautions here:
+// Floats always accept arithmetic ops, but may become NaN.
+for_base_ops! {
+    T = (f32, f64);
+    type Lhs = Simd<T, N>;
+    type Rhs = Simd<T, N>;
+    type Output = Self;
 
-            // Integers panic on divide by 0
-            impl_ref_ops! {
-                impl<const LANES: usize> core::ops::Div<Self> for Simd<$scalar, LANES>
-                where
-                    LaneCount<LANES>: SupportedLaneCount,
-                {
-                    type Output = Self;
+    impl Add::add {
+        unsafe_base { simd_add }
+    }
 
-                    #[inline]
-                    fn div(self, rhs: Self) -> Self::Output {
-                        if rhs.as_array()
-                            .iter()
-                            .any(|x| *x == 0)
-                        {
-                            panic!("attempt to divide by zero");
-                        }
+    impl Mul::mul {
+        unsafe_base { simd_mul }
+    }
 
-                        // Guards for div(MIN, -1),
-                        // this check only applies to signed ints
-                        if <$scalar>::MIN != 0 && self.as_array().iter()
-                                .zip(rhs.as_array().iter())
-                                .any(|(x,y)| *x == <$scalar>::MIN && *y == -1 as _) {
-                            panic!("attempt to divide with overflow");
-                        }
-                        unsafe { intrinsics::simd_div(self, rhs) }
-                    }
-                }
-            }
+    impl Sub::sub {
+        unsafe_base { simd_sub }
+    }
 
-            // remainder panics on zero divisor
-            impl_ref_ops! {
-                impl<const LANES: usize> core::ops::Rem<Self> for Simd<$scalar, LANES>
-                where
-                    LaneCount<LANES>: SupportedLaneCount,
-                {
-                    type Output = Self;
+    impl Div::div {
+        unsafe_base { simd_div }
+    }
 
-                    #[inline]
-                    fn rem(self, rhs: Self) -> Self::Output {
-                        if rhs.as_array()
-                            .iter()
-                            .any(|x| *x == 0)
-                        {
-                            panic!("attempt to calculate the remainder with a divisor of zero");
-                        }
-
-                        // Guards for rem(MIN, -1)
-                        // this branch applies the check only to signed ints
-                        if <$scalar>::MIN != 0 && self.as_array().iter()
-                                .zip(rhs.as_array().iter())
-                                .any(|(x,y)| *x == <$scalar>::MIN && *y == -1 as _) {
-                            panic!("attempt to calculate the remainder with overflow");
-                        }
-                        unsafe { intrinsics::simd_rem(self, rhs) }
-                    }
-                }
-            }
-
-            // shifts panic on overflow
-            impl_ref_ops! {
-                impl<const LANES: usize> core::ops::Shl<Self> for Simd<$scalar, LANES>
-                where
-                    LaneCount<LANES>: SupportedLaneCount,
-                {
-                    type Output = Self;
-
-                    #[inline]
-                    fn shl(self, rhs: Self) -> Self::Output {
-                        // TODO there is probably a better way of doing this
-                        if rhs.as_array()
-                            .iter()
-                            .copied()
-                            .any(invalid_shift_rhs)
-                        {
-                            panic!("attempt to shift left with overflow");
-                        }
-                        unsafe { intrinsics::simd_shl(self, rhs) }
-                    }
-                }
-            }
-
-            impl_ref_ops! {
-                impl<const LANES: usize> core::ops::Shr<Self> for Simd<$scalar, LANES>
-                where
-                    LaneCount<LANES>: SupportedLaneCount,
-                {
-                    type Output = Self;
-
-                    #[inline]
-                    fn shr(self, rhs: Self) -> Self::Output {
-                        // TODO there is probably a better way of doing this
-                        if rhs.as_array()
-                            .iter()
-                            .copied()
-                            .any(invalid_shift_rhs)
-                        {
-                            panic!("attempt to shift with overflow");
-                        }
-                        unsafe { intrinsics::simd_shr(self, rhs) }
-                    }
-                }
-            }
-        )*
-    };
+    impl Rem::rem {
+        unsafe_base { simd_rem }
+    }
 }
-
-/// Implements unsigned integer operators for the provided types.
-macro_rules! impl_signed_int_ops {
-    { $($scalar:ty),* } => {
-        impl_unsigned_int_ops! { $($scalar),* }
-    };
-}
-
-impl_unsigned_int_ops! { u8, u16, u32, u64, usize }
-impl_signed_int_ops! { i8, i16, i32, i64, isize }
-impl_float_ops! { f32, f64 }

library/portable-simd/crates/core_simd/src/round.rs

@@ -5,47 +5,6 @@ macro_rules! implement {
     {
         $type:ty, $int_type:ty
     } => {
-        #[cfg(feature = "std")]
-        impl<const LANES: usize> Simd<$type, LANES>
-        where
-            LaneCount<LANES>: SupportedLaneCount,
-        {
-            /// Returns the smallest integer greater than or equal to each lane.
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[inline]
-            pub fn ceil(self) -> Self {
-                unsafe { intrinsics::simd_ceil(self) }
-            }
-
-            /// Returns the largest integer value less than or equal to each lane.
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[inline]
-            pub fn floor(self) -> Self {
-                unsafe { intrinsics::simd_floor(self) }
-            }
-
-            /// Rounds to the nearest integer value. Ties round toward zero.
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[inline]
-            pub fn round(self) -> Self {
-                unsafe { intrinsics::simd_round(self) }
-            }
-
-            /// Returns the floating point's integer value, with its fractional part removed.
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[inline]
-            pub fn trunc(self) -> Self {
-                unsafe { intrinsics::simd_trunc(self) }
-            }
-
-            /// Returns the floating point's fractional value, with its integer part removed.
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[inline]
-            pub fn fract(self) -> Self {
-                self - self.trunc()
-            }
-        }
-
         impl<const LANES: usize> Simd<$type, LANES>
         where
             LaneCount<LANES>: SupportedLaneCount,

library/portable-simd/crates/core_simd/src/vector.rs

@@ -75,6 +75,36 @@ where
         Self(array)
     }
 
+    /// Performs lanewise conversion of a SIMD vector's elements to another SIMD-valid type.
+    /// 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.
+    ///
+    /// # Examples
+    /// ```
+    /// # #![feature(portable_simd)]
+    /// # #[cfg(feature = "std")] use core_simd::Simd;
+    /// # #[cfg(not(feature = "std"))] use core::simd::Simd;
+    /// let floats: Simd<f32, 4> = Simd::from_array([1.9, -4.5, f32::INFINITY, f32::NAN]);
+    /// let ints = floats.cast::<i32>();
+    /// assert_eq!(ints, Simd::from_array([1, -4, i32::MAX, 0]));
+    ///
+    /// // Formally equivalent, but `Simd::cast` can optimize better.
+    /// assert_eq!(ints, Simd::from_array(floats.to_array().map(|x| x as i32)));
+    ///
+    /// // The float conversion does not round-trip.
+    /// let floats_again = ints.cast();
+    /// assert_ne!(floats, floats_again);
+    /// assert_eq!(floats_again, Simd::from_array([1.0, -4.0, 2147483647.0, 0.0]));
+    /// ```
+    #[must_use]
+    #[inline]
+    #[cfg(not(bootstrap))]
+    pub fn cast<U: SimdElement>(self) -> Simd<U, LANES> {
+        unsafe { intrinsics::simd_as(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.
     ///

library/portable-simd/crates/core_simd/src/vector/float.rs

@@ -38,29 +38,6 @@ macro_rules! impl_float_vector {
                 unsafe { intrinsics::simd_fabs(self) }
             }
 
-            /// Fused multiply-add.  Computes `(self * a) + b` with only one rounding error,
-            /// yielding a more accurate result than an unfused multiply-add.
-            ///
-            /// Using `mul_add` *may* be more performant than an unfused multiply-add if the target
-            /// architecture has a dedicated `fma` CPU instruction.  However, this is not always
-            /// true, and will be heavily dependent on designing algorithms with specific target
-            /// hardware in mind.
-            #[cfg(feature = "std")]
-            #[inline]
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            pub fn mul_add(self, a: Self, b: Self) -> Self {
-                unsafe { intrinsics::simd_fma(self, a, b) }
-            }
-
-            /// Produces a vector where every lane has the square root value
-            /// of the equivalently-indexed lane in `self`
-            #[inline]
-            #[must_use = "method returns a new vector and does not mutate the original value"]
-            #[cfg(feature = "std")]
-            pub fn sqrt(self) -> Self {
-                unsafe { intrinsics::simd_fsqrt(self) }
-            }
-
             /// Takes the reciprocal (inverse) of each lane, `1/x`.
             #[inline]
             #[must_use = "method returns a new vector and does not mutate the original value"]
@@ -128,8 +105,8 @@ macro_rules! impl_float_vector {
                 self.abs().lanes_ne(Self::splat(0.0)) & (self.to_bits() & Self::splat(<$type>::INFINITY).to_bits()).lanes_eq(Simd::splat(0))
             }
 
-            /// Returns true for each lane if its value is neither neither zero, infinite,
-            /// subnormal, or `NaN`.
+            /// Returns true for each lane if its value is neither zero, infinite,
+            /// subnormal, nor `NaN`.
             #[inline]
             #[must_use = "method returns a new mask and does not mutate the original value"]
             pub fn is_normal(self) -> Mask<$mask_ty, LANES> {
@@ -164,11 +141,7 @@ macro_rules! impl_float_vector {
             #[inline]
             #[must_use = "method returns a new vector and does not mutate the original value"]
             pub fn min(self, other: Self) -> Self {
-                // TODO consider using an intrinsic
-                self.is_nan().select(
-                    other,
-                    self.lanes_ge(other).select(other, self)
-                )
+                unsafe { intrinsics::simd_fmin(self, other) }
             }
 
             /// Returns the maximum of each lane.
@@ -177,11 +150,7 @@ macro_rules! impl_float_vector {
             #[inline]
             #[must_use = "method returns a new vector and does not mutate the original value"]
             pub fn max(self, other: Self) -> Self {
-                // TODO consider using an intrinsic
-                self.is_nan().select(
-                    other,
-                    self.lanes_le(other).select(other, self)
-                )
+                unsafe { intrinsics::simd_fmax(self, other) }
             }
 
             /// Restrict each lane to a certain interval unless it is NaN.

library/portable-simd/crates/core_simd/tests/cast.rs

@@ -0,0 +1,37 @@
+#![feature(portable_simd)]
+macro_rules! cast_types {
+    ($start:ident, $($target:ident),*) => {
+        mod $start {
+            use core_simd::simd::Simd;
+            type Vector<const N: usize> = Simd<$start, N>;
+            $(
+                mod $target {
+                    use super::*;
+                    test_helpers::test_lanes! {
+                        fn cast_as<const N: usize>() {
+                            test_helpers::test_unary_elementwise(
+                                &Vector::<N>::cast::<$target>,
+                                &|x| x as $target,
+                                &|_| true,
+                            )
+                        }
+                    }
+                }
+            )*
+        }
+    };
+}
+
+// The hypothesis is that widening conversions aren't terribly interesting.
+cast_types!(f32, f64, i8, u8, usize, isize);
+cast_types!(f64, f32, i8, u8, usize, isize);
+cast_types!(i8, u8, f32);
+cast_types!(u8, i8, f32);
+cast_types!(i16, u16, i8, u8, f32);
+cast_types!(u16, i16, i8, u8, f32);
+cast_types!(i32, u32, i8, u8, f32, f64);
+cast_types!(u32, i32, i8, u8, f32, f64);
+cast_types!(i64, u64, i8, u8, isize, usize, f32, f64);
+cast_types!(u64, i64, i8, u8, isize, usize, f32, f64);
+cast_types!(isize, usize, i8, u8, f32, f64);
+cast_types!(usize, isize, i8, u8, f32, f64);

library/portable-simd/crates/core_simd/tests/ops_macros.rs

@@ -546,6 +546,8 @@ macro_rules! impl_float_tests {
 
             #[cfg(feature = "std")]
             mod std {
+                use std_float::StdFloat;
+
                 use super::*;
                 test_helpers::test_lanes! {
                     fn sqrt<const LANES: usize>() {

library/portable-simd/crates/core_simd/tests/round.rs

@@ -3,6 +3,8 @@
 macro_rules! float_rounding_test {
     { $scalar:tt, $int_scalar:tt } => {
         mod $scalar {
+            use std_float::StdFloat;
+
             type Vector<const LANES: usize> = core_simd::Simd<$scalar, LANES>;
             type Scalar = $scalar;
             type IntScalar = $int_scalar;

library/portable-simd/crates/std_float/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "std_float"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+core_simd = { path = "../core_simd" }
+
+[features]
+default = ["as_crate"]
+as_crate = []

library/portable-simd/crates/std_float/src/lib.rs

@@ -0,0 +1,165 @@
+#![cfg_attr(feature = "as_crate", no_std)] // We are std!
+#![cfg_attr(
+    feature = "as_crate",
+    feature(platform_intrinsics),
+    feature(portable_simd)
+)]
+#[cfg(not(feature = "as_crate"))]
+use core::simd;
+#[cfg(feature = "as_crate")]
+use core_simd::simd;
+
+use simd::{LaneCount, Simd, SupportedLaneCount};
+
+#[cfg(feature = "as_crate")]
+mod experimental {
+    pub trait Sealed {}
+}
+
+#[cfg(feature = "as_crate")]
+use experimental as sealed;
+
+use crate::sealed::Sealed;
+
+// "platform intrinsics" are essentially "codegen intrinsics"
+// each of these may be scalarized and lowered to a libm call
+extern "platform-intrinsic" {
+    // ceil
+    fn simd_ceil<T>(x: T) -> T;
+
+    // floor
+    fn simd_floor<T>(x: T) -> T;
+
+    // round
+    fn simd_round<T>(x: T) -> T;
+
+    // trunc
+    fn simd_trunc<T>(x: T) -> T;
+
+    // fsqrt
+    fn simd_fsqrt<T>(x: T) -> T;
+
+    // fma
+    fn simd_fma<T>(x: T, y: T, z: T) -> T;
+}
+
+/// This trait provides a possibly-temporary implementation of float functions
+/// that may, in the absence of hardware support, canonicalize to calling an
+/// operating system's `math.h` dynamically-loaded library (also known as a
+/// shared object). As these conditionally require runtime support, they
+/// should only appear in binaries built assuming OS support: `std`.
+///
+/// However, there is no reason SIMD types, in general, need OS support,
+/// as for many architectures an embedded binary may simply configure that
+/// support itself. This means these types must be visible in `core`
+/// but have these functions available in `std`.
+///
+/// [`f32`] and [`f64`] achieve a similar trick by using "lang items", but
+/// due to compiler limitations, it is harder to implement this approach for
+/// abstract data types like [`Simd`]. From that need, this trait is born.
+///
+/// It is possible this trait will be replaced in some manner in the future,
+/// when either the compiler or its supporting runtime functions are improved.
+/// For now this trait is available to permit experimentation with SIMD float
+/// operations that may lack hardware support, such as `mul_add`.
+pub trait StdFloat: Sealed + Sized {
+    /// Fused multiply-add.  Computes `(self * a) + b` with only one rounding error,
+    /// yielding a more accurate result than an unfused multiply-add.
+    ///
+    /// Using `mul_add` *may* be more performant than an unfused multiply-add if the target
+    /// architecture has a dedicated `fma` CPU instruction.  However, this is not always
+    /// true, and will be heavily dependent on designing algorithms with specific target
+    /// hardware in mind.
+    #[inline]
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    fn mul_add(self, a: Self, b: Self) -> Self {
+        unsafe { simd_fma(self, a, b) }
+    }
+
+    /// Produces a vector where every lane has the square root value
+    /// of the equivalently-indexed lane in `self`
+    #[inline]
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    fn sqrt(self) -> Self {
+        unsafe { simd_fsqrt(self) }
+    }
+
+    /// Returns the smallest integer greater than or equal to each lane.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn ceil(self) -> Self {
+        unsafe { simd_ceil(self) }
+    }
+
+    /// Returns the largest integer value less than or equal to each lane.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn floor(self) -> Self {
+        unsafe { simd_floor(self) }
+    }
+
+    /// Rounds to the nearest integer value. Ties round toward zero.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn round(self) -> Self {
+        unsafe { simd_round(self) }
+    }
+
+    /// Returns the floating point's integer value, with its fractional part removed.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn trunc(self) -> Self {
+        unsafe { simd_trunc(self) }
+    }
+
+    /// Returns the floating point's fractional value, with its integer part removed.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    fn fract(self) -> Self;
+}
+
+impl<const N: usize> Sealed for Simd<f32, N> where LaneCount<N>: SupportedLaneCount {}
+impl<const N: usize> Sealed for Simd<f64, N> where LaneCount<N>: SupportedLaneCount {}
+
+// We can safely just use all the defaults.
+impl<const N: usize> StdFloat for Simd<f32, N>
+where
+    LaneCount<N>: SupportedLaneCount,
+{
+    /// Returns the floating point's fractional value, with its integer part removed.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn fract(self) -> Self {
+        self - self.trunc()
+    }
+}
+
+impl<const N: usize> StdFloat for Simd<f64, N>
+where
+    LaneCount<N>: SupportedLaneCount,
+{
+    /// Returns the floating point's fractional value, with its integer part removed.
+    #[must_use = "method returns a new vector and does not mutate the original value"]
+    #[inline]
+    fn fract(self) -> Self {
+        self - self.trunc()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use simd::*;
+
+    #[test]
+    fn everything_works() {
+        let x = f32x4::from_array([0.1, 0.5, 0.6, -1.5]);
+        let x2 = x + x;
+        let _xc = x.ceil();
+        let _xf = x.floor();
+        let _xr = x.round();
+        let _xt = x.trunc();
+        let _xfma = x.mul_add(x, x);
+        let _xsqrt = x.sqrt();
+        let _ = x2.abs() * x2;
+    }
+}

library/std/src/lib.rs

@@ -311,6 +311,7 @@
 #![feature(panic_can_unwind)]
 #![feature(panic_unwind)]
 #![feature(pin_static_ref)]
+#![feature(platform_intrinsics)]
 #![feature(portable_simd)]
 #![feature(prelude_import)]
 #![feature(ptr_as_uninit)]
@@ -456,8 +457,6 @@ pub use core::pin;
 pub use core::ptr;
 #[stable(feature = "rust1", since = "1.0.0")]
 pub use core::result;
-#[unstable(feature = "portable_simd", issue = "86656")]
-pub use core::simd;
 #[unstable(feature = "async_stream", issue = "79024")]
 pub use core::stream;
 #[stable(feature = "i128", since = "1.26.0")]
@@ -504,6 +503,25 @@ pub mod time;
 #[unstable(feature = "once_cell", issue = "74465")]
 pub mod lazy;
 
+// Pull in `std_float` crate  into libstd. The contents of
+// `std_float` are in a different repository: rust-lang/portable-simd.
+#[path = "../../portable-simd/crates/std_float/src/lib.rs"]
+#[allow(missing_debug_implementations, dead_code, unsafe_op_in_unsafe_fn, unused_unsafe)]
+#[allow(rustdoc::bare_urls)]
+#[unstable(feature = "portable_simd", issue = "86656")]
+#[cfg(not(all(miri, doctest)))] // Miri does not support all SIMD intrinsics
+mod std_float;
+
+#[cfg(not(all(miri, doctest)))] // Miri does not support all SIMD intrinsics
+#[doc = include_str!("../../portable-simd/crates/core_simd/src/core_simd_docs.md")]
+#[unstable(feature = "portable_simd", issue = "86656")]
+pub mod simd {
+    #[doc(inline)]
+    pub use crate::std_float::StdFloat;
+    #[doc(inline)]
+    pub use core::simd::*;
+}
+
 #[stable(feature = "futures_api", since = "1.36.0")]
 pub mod task {
     //! Types and Traits for working with asynchronous tasks.

src/test/ui/simd/libm_std_can_float.rs

@@ -0,0 +1,23 @@
+// run-pass
+
+// This is the converse of the other libm test.
+#![feature(portable_simd)]
+use std::simd::f32x4;
+use std::simd::StdFloat;
+
+// For SIMD float ops, the LLIR version which is used to implement the portable
+// forms of them may become calls to math.h AKA libm. So, we can't guarantee
+// we can compile them for #![no_std] crates.
+//
+// However, we can expose some of these ops via an extension trait.
+fn main() {
+    let x = f32x4::from_array([0.1, 0.5, 0.6, -1.5]);
+    let x2 = x + x;
+    let _xc = x.ceil();
+    let _xf = x.floor();
+    let _xr = x.round();
+    let _xt = x.trunc();
+    let _xfma = x.mul_add(x, x);
+    let _xsqrt = x.sqrt();
+    let _ = x2.abs() * x2;
+}

src/test/ui/simd/portable-intrinsics-arent-exposed.rs

@@ -1,7 +1,8 @@
 // May not matter, since people can use them with a nightly feature.
 // However this tests to guarantee they don't leak out via portable_simd,
 // and thus don't accidentally get stabilized.
-use std::simd::intrinsics; //~ERROR E0603
+use core::simd::intrinsics; //~ERROR E0433
+use std::simd::intrinsics; //~ERROR E0432
 
 fn main() {
     ()

src/test/ui/simd/portable-intrinsics-arent-exposed.stderr

@@ -1,15 +1,16 @@
-error[E0603]: module `intrinsics` is private
-  --> $DIR/portable-intrinsics-arent-exposed.rs:4:16
+error[E0433]: failed to resolve: maybe a missing crate `core`?
+  --> $DIR/portable-intrinsics-arent-exposed.rs:4:5
+   |
+LL | use core::simd::intrinsics;
+   |     ^^^^ maybe a missing crate `core`?
+
+error[E0432]: unresolved import `std::simd::intrinsics`
+  --> $DIR/portable-intrinsics-arent-exposed.rs:5:5
    |
 LL | use std::simd::intrinsics;
-   |                ^^^^^^^^^^ private module
-   |
-note: the module `intrinsics` is defined here
-  --> $SRC_DIR/core/src/lib.rs:LL:COL
-   |
-LL |     pub use crate::core_simd::simd::*;
-   |             ^^^^^^^^^^^^^^^^^^^^^^^^^
+   |     ^^^^^^^^^^^^^^^^^^^^^ no `intrinsics` in `simd`
 
-error: aborting due to previous error
+error: aborting due to 2 previous errors
 
-For more information about this error, try `rustc --explain E0603`.
+Some errors have detailed explanations: E0432, E0433.
+For more information about an error, try `rustc --explain E0432`.