nordic-dev.net/rust
nordic-dev.net/rust
2
0
Fork 0
mirror of https://github.com/rust-lang/rust.git synced 2024-10-30 22:12:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Auto merge of #126171 - RalfJung:simd_bitmask_multibyte, r=workingjubilee

Browse Source 
simd_bitmask intrinsic: add a non-power-of-2 multi-byte example

r? `@calebzulawski` `@workingjubilee`
This commit is contained in:
bors 2024-07-05 01:58:22 +00:00
commit 51917ba8f2
4 changed files with 144 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
compiler/rustc_codegen_llvm/src/intrinsic.rs
View File

@ -1121,8 +1121,8 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
if name == sym::simd_select_bitmask {
let (len, _) = require_simd!(arg_tys[1], SimdArgument);
let expected_int_bits = (len.max(8) - 1).next_power_of_two();
let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
let expected_int_bits = len.max(8).next_power_of_two();
let expected_bytes = len.div_ceil(8);
let mask_ty = arg_tys[0];
let mask = match mask_ty.kind() {
@ -1379,17 +1379,16 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}
if name == sym::simd_bitmask {
// The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
// vector mask and returns the most significant bit (MSB) of each lane in the form
// of either:
// The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a vector mask and
// returns one bit for each lane (which must all be `0` or `!0`) in the form of either:
// * an unsigned integer
// * an array of `u8`
// If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
//
// The bit order of the result depends on the byte endianness, LSB-first for little
// endian and MSB-first for big endian.
let expected_int_bits = in_len.max(8);
let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
let expected_int_bits = in_len.max(8).next_power_of_two();
let expected_bytes = in_len.div_ceil(8);
// Integer vector <i{in_bitwidth} x in_len>:
let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
@ -1409,7 +1408,8 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}),
};
// Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
// LLVM doesn't always know the inputs are `0` or `!0`, so we shift here so it optimizes to
// `pmovmskb` and similar on x86.
let shift_indices =
vec![
bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);

19
library/core/src/intrinsics/simd.rs
View File

@ -460,7 +460,7 @@ extern "rust-intrinsic" {
/// `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`.
/// of `T`, or the smallest array of `u8` with at least as many bits as the length of `T`.
///
/// Each element is truncated to a single bit and packed into the result.
///
@ -472,12 +472,19 @@ extern "rust-intrinsic" {
/// * 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.
/// For example, `[!0, 0, !0, !0]` packs to
/// - `0b1101u8` or `[0b1101]` on little endian, and
/// - `0b1011u8` or `[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.
/// To consider a larger example,
/// `[!0, 0, 0, 0, 0, 0, 0, 0, !0, !0, 0, 0, 0, 0, !0, 0]` packs to
/// - `0b0100001100000001u16` or `[0b00000001, 0b01000011]` on little endian, and
/// - `0b1000000011000010u16` or `[0b10000000, 0b11000010]` on big endian.
///
/// And finally, a non-power-of-2 example with multiple bytes:
/// `[!0, !0, 0, !0, 0, 0, !0, 0, !0, 0]` packs to
/// - `0b0101001011u16` or `[0b01001011, 0b01]` on little endian, and
/// - `0b1101001010u16` or `[0b11, 0b01001010]` on big endian.
///
/// # Safety
/// `x` must contain only `0` and `!0`.

90
tests/ui/simd/simd-bitmask-notpow2.rs Normal file
View File

@ -0,0 +1,90 @@
//@run-pass
// SEGFAULTS on LLVM 17. This should be merged into `simd-bitmask` once we require LLVM 18.
//@ min-llvm-version: 18
// FIXME: broken codegen on big-endian (https://github.com/rust-lang/rust/issues/127205)
//@ ignore-endian-big
#![feature(repr_simd, intrinsics)]
extern "rust-intrinsic" {
fn simd_bitmask<T, U>(v: T) -> U;
fn simd_select_bitmask<T, U>(m: T, a: U, b: U) -> U;
}
fn main() {
// Non-power-of-2 multi-byte mask.
#[repr(simd, packed)]
#[allow(non_camel_case_types)]
#[derive(Copy, Clone, Debug, PartialEq)]
struct i32x10([i32; 10]);
impl i32x10 {
fn splat(x: i32) -> Self {
Self([x; 10])
}
}
unsafe {
let mask = i32x10([!0, !0, 0, !0, 0, 0, !0, 0, !0, 0]);
let mask_bits = if cfg!(target_endian = "little") { 0b0101001011 } else { 0b1101001010 };
let mask_bytes =
if cfg!(target_endian = "little") { [0b01001011, 0b01] } else { [0b11, 0b01001010] };
let bitmask1: u16 = simd_bitmask(mask);
let bitmask2: [u8; 2] = simd_bitmask(mask);
assert_eq!(bitmask1, mask_bits);
assert_eq!(bitmask2, mask_bytes);
let selected1 = simd_select_bitmask::<u16, _>(
mask_bits,
i32x10::splat(!0), // yes
i32x10::splat(0), // no
);
let selected2 = simd_select_bitmask::<[u8; 2], _>(
mask_bytes,
i32x10::splat(!0), // yes
i32x10::splat(0), // no
);
assert_eq!(selected1, mask);
assert_eq!(selected2, mask);
}
// Test for a mask where the next multiple of 8 is not a power of two.
#[repr(simd, packed)]
#[allow(non_camel_case_types)]
#[derive(Copy, Clone, Debug, PartialEq)]
struct i32x20([i32; 20]);
impl i32x20 {
fn splat(x: i32) -> Self {
Self([x; 20])
}
}
unsafe {
let mask = i32x20([!0, !0, 0, !0, 0, 0, !0, 0, !0, 0, 0, 0, 0, !0, !0, !0, !0, !0, !0, !0]);
let mask_bits = if cfg!(target_endian = "little") {
0b11111110000101001011
} else {
0b11010010100001111111
};
let mask_bytes = if cfg!(target_endian = "little") {
[0b01001011, 0b11100001, 0b1111]
} else {
[0b1101, 0b00101000, 0b01111111]
};
let bitmask1: u32 = simd_bitmask(mask);
let bitmask2: [u8; 3] = simd_bitmask(mask);
assert_eq!(bitmask1, mask_bits);
assert_eq!(bitmask2, mask_bytes);
let selected1 = simd_select_bitmask::<u32, _>(
mask_bits,
i32x20::splat(!0), // yes
i32x20::splat(0), // no
);
let selected2 = simd_select_bitmask::<[u8; 3], _>(
mask_bytes,
i32x20::splat(!0), // yes
i32x20::splat(0), // no
);
assert_eq!(selected1, mask);
assert_eq!(selected2, mask);
}
}

39
tests/ui/simd/simd-bitmask.rs
View File

@ -1,5 +1,4 @@
//@run-pass
//@ignore-endian-big behavior of simd_select_bitmask is endian-specific
#![feature(repr_simd, intrinsics)]
extern "rust-intrinsic" {
@ -17,36 +16,58 @@ fn main() {
let i: u8 = simd_bitmask(v);
let a: [u8; 1] = simd_bitmask(v);
if cfg!(target_endian = "little") {
assert_eq!(i, 0b0101);
assert_eq!(a, [0b0101]);
} else {
assert_eq!(i, 0b1010);
assert_eq!(a, [0b1010]);
}
let v = Simd::<i8, 16>([0, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0]);
let i: u16 = simd_bitmask(v);
let a: [u8; 2] = simd_bitmask(v);
if cfg!(target_endian = "little") {
assert_eq!(i, 0b0101000000001100);
assert_eq!(a, [0b1100, 0b01010000]);
assert_eq!(a, [0b00001100, 0b01010000]);
} else {
assert_eq!(i, 0b0011000000001010);
assert_eq!(a, [0b00110000, 0b00001010]);
}
}
unsafe {
let a = Simd::<i32, 8>([0, 1, 2, 3, 4, 5, 6, 7]);
let b = Simd::<i32, 8>([8, 9, 10, 11, 12, 13, 14, 15]);
let e = [0, 9, 2, 11, 12, 13, 14, 15];
let a = Simd::<i32, 4>([0, 1, 2, 3]);
let b = Simd::<i32, 4>([8, 9, 10, 11]);
let e = [0, 9, 2, 11];
let r = simd_select_bitmask(0b0101u8, a, b);
let mask = if cfg!(target_endian = "little") { 0b0101u8 } else { 0b1010u8 };
let r = simd_select_bitmask(mask, a, b);
assert_eq!(r.0, e);
let r = simd_select_bitmask([0b0101u8], a, b);
let mask = if cfg!(target_endian = "little") { [0b0101u8] } else { [0b1010u8] };
let r = simd_select_bitmask(mask, a, b);
assert_eq!(r.0, e);
let a = Simd::<i32, 16>([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]);
let b = Simd::<i32, 16>([16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]);
let e = [16, 17, 2, 3, 20, 21, 22, 23, 24, 25, 26, 27, 12, 29, 14, 31];
let r = simd_select_bitmask(0b0101000000001100u16, a, b);
let mask = if cfg!(target_endian = "little") {
0b0101000000001100u16
} else {
0b0011000000001010u16
};
let r = simd_select_bitmask(mask, a, b);
assert_eq!(r.0, e);
let r = simd_select_bitmask([0b1100u8, 0b01010000u8], a, b);
let mask = if cfg!(target_endian = "little") {
[0b00001100u8, 0b01010000u8]
} else {
[0b00110000u8, 0b00001010u8]
};
let r = simd_select_bitmask(mask, a, b);
assert_eq!(r.0, e);
}
}