diff --git a/library/core/src/intrinsics.rs b/library/core/src/intrinsics.rs index 3406112ddb111..5107ba1a9e1be 100644 --- a/library/core/src/intrinsics.rs +++ b/library/core/src/intrinsics.rs @@ -59,6 +59,7 @@ use crate::marker::Tuple; use crate::mem; pub mod mir; +pub mod simd; // These imports are used for simplifying intra-doc links #[allow(unused_imports)] diff --git a/library/core/src/intrinsics/simd.rs b/library/core/src/intrinsics/simd.rs new file mode 100644 index 0000000000000..68c8a335b405c --- /dev/null +++ b/library/core/src/intrinsics/simd.rs @@ -0,0 +1,473 @@ +//! SIMD compiler intrinsics. +//! +//! In this module, a "vector" is any `repr(simd)` type. + +extern "platform-intrinsic" { + /// Add two simd vectors elementwise. + /// + /// `T` must be a vector of integer or floating point primitive types. + pub fn simd_add(x: T, y: T) -> T; + + /// Subtract `rhs` from `lhs` elementwise. + /// + /// `T` must be a vector of integer or floating point primitive types. + pub fn simd_sub(lhs: T, rhs: T) -> T; + + /// Multiply two simd vectors elementwise. + /// + /// `T` must be a vector of integer or floating point primitive types. + pub fn simd_mul(x: T, y: T) -> T; + + /// Divide `lhs` by `rhs` elementwise. + /// + /// `T` must be a vector of integer or floating point primitive types. + /// + /// # Safety + /// For integers, `rhs` must not contain any zero elements. + /// Additionally for signed integers, `::MIN / -1` is undefined behavior. + pub fn simd_div(lhs: T, rhs: T) -> T; + + /// Remainder of two vectors elementwise + /// + /// `T` must be a vector of integer or floating point primitive types. + /// + /// # Safety + /// For integers, `rhs` must not contain any zero elements. + /// Additionally for signed integers, `::MIN / -1` is undefined behavior. + pub fn simd_rem(lhs: T, rhs: T) -> T; + + /// Elementwise vector left shift, with UB on overflow. + /// + /// Shift `lhs` left by `rhs`, shifting in sign bits for signed types. + /// + /// `T` must be a vector of integer primitive types. + /// + /// # Safety + /// + /// Each element of `rhs` must be less than `::BITS`. + pub fn simd_shl(lhs: T, rhs: T) -> T; + + /// Elementwise vector right shift, with UB on overflow. + /// + /// `T` must be a vector of integer primitive types. + /// + /// Shift `lhs` right by `rhs`, shifting in sign bits for signed types. + /// + /// # Safety + /// + /// Each element of `rhs` must be less than `::BITS`. + pub fn simd_shr(lhs: T, rhs: T) -> T; + + /// Elementwise vector "and". + /// + /// `T` must be a vector of integer primitive types. + pub fn simd_and(x: T, y: T) -> T; + + /// Elementwise vector "or". + /// + /// `T` must be a vector of integer primitive types. + pub fn simd_or(x: T, y: T) -> T; + + /// Elementwise vector "exclusive or". + /// + /// `T` must be a vector of integer primitive types. + pub fn simd_xor(x: T, y: T) -> T; + + /// Numerically cast a vector, elementwise. + /// + /// `T` and `U` must be vectors of integer or floating point primitive types, and must have the + /// same length. + /// + /// When casting floats to integers, the result is truncated. Out-of-bounds result lead to UB. + /// When casting integers to floats, the result is rounded. + /// Otherwise, truncates or extends the value, maintaining the sign for signed integers. + /// + /// # Safety + /// Casting from integer types is always safe. + /// Casting between two float types is also always safe. + /// + /// Casting floats to integers truncates, following the same rules as `to_int_unchecked`. + /// Specifically, each element must: + /// * Not be `NaN` + /// * Not be infinite + /// * Be representable in the return type, after truncating off its fractional part + pub fn simd_cast(x: T) -> U; + + /// Numerically cast a vector, elementwise. + /// + /// `T` and `U` be a vectors of integer or floating point primitive types, and must have the + /// same length. + /// + /// Like `simd_cast`, but saturates float-to-integer conversions (NaN becomes 0). + /// This matches regular `as` and is always safe. + /// + /// When casting floats to integers, the result is truncated. + /// When casting integers to floats, the result is rounded. + /// Otherwise, truncates or extends the value, maintaining the sign for signed integers. + pub fn simd_as(x: T) -> U; + + /// Elementwise negation of a vector. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// Rust panics for `-::Min` due to overflow, but it is not UB with this intrinsic. + pub fn simd_neg(x: T) -> T; + + /// Elementwise absolute value of a vector. + /// + /// `T` must be a vector of floating-point primitive types. + pub fn simd_fabs(x: T) -> T; + + /// Elementwise minimum of a vector. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// Follows IEEE-754 `minNum` semantics. + pub fn simd_fmin(x: T, y: T) -> T; + + /// Elementwise maximum of a vector. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// Follows IEEE-754 `maxNum` semantics. + pub fn simd_fmax(x: T, y: T) -> T; + + /// Tests elementwise equality of two vectors. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_eq(x: T, y: T) -> U; + + /// Tests elementwise inequality equality of two vectors. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_ne(x: T, y: T) -> U; + + /// Tests if `x` is less than `y`, elementwise. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_lt(x: T, y: T) -> U; + + /// Tests if `x` is less than or equal to `y`, elementwise. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_le(x: T, y: T) -> U; + + /// Tests if `x` is greater than `y`, elementwise. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_gt(x: T, y: T) -> U; + + /// Tests if `x` is greater than or equal to `y`, elementwise. + /// + /// `T` must be a vector of floating-point primitive types. + /// + /// `U` must be a vector of integers with the same number of elements and element size as `T`. + /// + /// Returns `0` for false and `!0` for true. + pub fn simd_ge(x: T, y: T) -> U; + + /// Shuffle two vectors by const indices. + /// + /// `T` must be a vector. + /// + /// `U` must be a const array of `i32`s. + /// + /// `V` must be a vector with the same element type as `T` and the same length as `U`. + /// + /// Concatenates `x` and `y`, then returns a new vector such that each element is selected from + /// the concatenation by the matching index in `idx`. + pub fn simd_shuffle(x: T, y: T, idx: U) -> V; + + /// Read a vector of pointers. + /// + /// `T` must be a vector. + /// + /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. + /// + /// `V` must be a vector of integers with the same length as `T` (but any element size). + /// + /// `idx` must be a constant: either naming a constant item, or an inline + /// `const {}` expression. + /// + /// For each pointer in `ptr`, if the corresponding value in `mask` is `!0`, read the pointer. + /// Otherwise if the corresponding value in `mask` is `0`, return the corresponding value from + /// `val`. + /// + /// # Safety + /// Unmasked values in `T` must be readable as if by `::read` (e.g. aligned to the element + /// type). + /// + /// `mask` must only contain `0` or `!0` values. + pub fn simd_gather(val: T, ptr: U, mask: V) -> T; + + /// Write to a vector of pointers. + /// + /// `T` must be a vector. + /// + /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. + /// + /// `V` must be a vector of integers with the same length as `T` (but any element size). + /// + /// For each pointer in `ptr`, if the corresponding value in `mask` is `!0`, write the + /// corresponding value in `val` to the pointer. + /// Otherwise if the corresponding value in `mask` is `0`, do nothing. + /// + /// # Safety + /// Unmasked values in `T` must be writeable as if by `::write` (e.g. aligned to the element + /// type). + /// + /// `mask` must only contain `0` or `!0` values. + pub fn simd_scatter(val: T, ptr: U, mask: V); + + /// Read a vector of pointers. + /// + /// `T` must be a vector. + /// + /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. + /// + /// `V` must be a vector of integers with the same length as `T` (but any element size). + /// + /// For each element, if the corresponding value in `mask` is `!0`, read the corresponding + /// pointer from `ptr`. + /// Otherwise if the corresponding value in `mask` is `0`, return the corresponding value from + /// `val`. + /// + /// # Safety + /// Unmasked values in `T` must be readable as if by `::read` (e.g. aligned to the element + /// type). + /// + /// `mask` must only contain `0` or `!0` values. + #[cfg(not(bootstrap))] + pub fn simd_masked_load(mask: V, ptr: U, val: T) -> T; + + /// Write to a vector of pointers. + /// + /// `T` must be a vector. + /// + /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. + /// + /// `V` must be a vector of integers with the same length as `T` (but any element size). + /// + /// For each element, if the corresponding value in `mask` is `!0`, write the corresponding + /// value in `val` to the pointer. + /// Otherwise if the corresponding value in `mask` is `0`, do nothing. + /// + /// # Safety + /// Unmasked values in `T` must be writeable as if by `::write` (e.g. aligned to the element + /// type). + /// + /// `mask` must only contain `0` or `!0` values. + #[cfg(not(bootstrap))] + pub fn simd_masked_store(mask: V, ptr: U, val: T); + + /// Add two simd vectors elementwise, with saturation. + /// + /// `T` must be a vector of integer primitive types. + pub fn simd_saturating_add(x: T, y: T) -> T; + + /// Subtract two simd vectors elementwise, with saturation. + /// + /// `T` must be a vector of integer primitive types. + /// + /// Subtract `rhs` from `lhs`. + pub fn simd_saturating_sub(lhs: T, rhs: T) -> T; + + /// Add elements within a vector from left to right. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + /// + /// Starting with the value `y`, add the elements of `x` and accumulate. + pub fn simd_reduce_add_ordered(x: T, y: U) -> U; + + /// Multiply elements within a vector from left to right. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + /// + /// Starting with the value `y`, multiply the elements of `x` and accumulate. + pub fn simd_reduce_mul_ordered(x: T, y: U) -> U; + + /// Check if all mask values are true. + /// + /// `T` must be a vector of integer primitive types. + /// + /// # Safety + /// `x` must contain only `0` or `!0`. + pub fn simd_reduce_all(x: T) -> bool; + + /// Check if all mask values are true. + /// + /// `T` must be a vector of integer primitive types. + /// + /// # Safety + /// `x` must contain only `0` or `!0`. + pub fn simd_reduce_any(x: T) -> bool; + + /// Return the maximum element of a vector. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + /// + /// For floating-point values, uses IEEE-754 `maxNum`. + pub fn simd_reduce_max(x: T) -> U; + + /// Return the minimum element of a vector. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + /// + /// For floating-point values, uses IEEE-754 `minNum`. + pub fn simd_reduce_min(x: T) -> U; + + /// Logical "and" all elements together. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + pub fn simd_reduce_and(x: T) -> U; + + /// Logical "or" all elements together. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + pub fn simd_reduce_or(x: T) -> U; + + /// Logical "exclusive or" all elements together. + /// + /// `T` must be a vector of integer or floating-point primitive types. + /// + /// `U` must be the element type of `T`. + pub fn simd_reduce_xor(x: T) -> U; + + /// Truncate an integer vector to a bitmask. + /// + /// `T` must be an integer vector. + /// + /// `U` must be either the smallest unsigned integer with at least as many bits as the length + /// of `T`, or the smallest array of `u8` with as many bits as the length of `T`. + /// + /// Each element is truncated to a single bit and packed into the result. + /// + /// No matter whether the output is an array or an unsigned integer, it is treated as a single + /// contiguous list of bits. The bitmask is always packed on the least-significant side of the + /// output, and padded with 0s in the most-significant bits. The order of the bits depends on + /// endianess: + /// + /// * On little endian, the least significant bit corresponds to the first vector element. + /// * On big endian, the least significant bit corresponds to the last vector element. + /// + /// For example, `[!0, 0, !0, !0]` packs to `0b1101` on little endian and `0b1011` on big + /// endian. + /// + /// To consider a larger example, `[!0, 0, 0, 0, 0, 0, 0, 0, !0, !0, 0, 0, 0, 0, !0, 0]` packs + /// to `[0b00000001, 0b01000011]` or `0b0100001100000001` on little endian, and `[0b10000000, + /// 0b11000010]` or `0b1000000011000010` on big endian. + /// + /// # Safety + /// `x` must contain only `0` and `!0`. + pub fn simd_bitmask(x: T) -> U; + + /// Select elements from a mask. + /// + /// `M` must be an integer vector. + /// + /// `T` must be a vector with the same number of elements as `M`. + /// + /// For each element, if the corresponding value in `mask` is `!0`, select the element from + /// `if_true`. If the corresponding value in `mask` is `0`, select the element from + /// `if_false`. + /// + /// # Safety + /// `mask` must only contain `0` and `!0`. + pub fn simd_select(mask: M, if_true: T, if_false: T) -> T; + + /// Select elements from a bitmask. + /// + /// `M` must be an unsigned integer or array of `u8`, matching `simd_bitmask`. + /// + /// `T` must be a vector. + /// + /// For each element, if the bit in `mask` is `1`, select the element from + /// `if_true`. If the corresponding bit in `mask` is `0`, select the element from + /// `if_false`. + /// + /// The bitmask bit order matches `simd_bitmask`. + /// + /// # Safety + /// Padding bits must be all zero. + pub fn simd_select_bitmask(m: M, yes: T, no: T) -> T; + + /// Elementwise calculates the offset from a pointer vector, potentially wrapping. + /// + /// `T` must be a vector of pointers. + /// + /// `U` must be a vector of `isize` or `usize` with the same number of elements as `T`. + /// + /// Operates as if by `::wrapping_offset`. + pub fn simd_arith_offset(ptr: T, offset: U) -> T; + + /// Cast a vector of pointers. + /// + /// `T` and `U` must be vectors of pointers with the same number of elements. + pub fn simd_cast_ptr(ptr: T) -> U; + + /// Expose a vector of pointers as a vector of addresses. + /// + /// `T` must be a vector of pointers. + /// + /// `U` must be a vector of `usize` with the same length as `T`. + pub fn simd_expose_addr(ptr: T) -> U; + + /// Create a vector of pointers from a vector of addresses. + /// + /// `T` must be a vector of `usize`. + /// + /// `U` must be a vector of pointers, with the same length as `T`. + pub fn simd_from_exposed_addr(addr: T) -> U; + + /// Swap bytes of each element. + /// + /// `T` must be a vector of integers. + pub fn simd_bswap(x: T) -> T; + + /// Reverse bits of each element. + /// + /// `T` must be a vector of integers. + pub fn simd_bitreverse(x: T) -> T; + + /// Count the leading zeros of each element. + /// + /// `T` must be a vector of integers. + pub fn simd_ctlz(x: T) -> T; + + /// Count the trailing zeros of each element. + /// + /// `T` must be a vector of integers. + pub fn simd_cttz(x: T) -> T; +}