Compiler Explorer

Source code

#![feature(rustc_attrs, repr_simd, aarch64_target_feature, platform_intrinsics)]
use std::hint::unreachable_unchecked;

#[repr(simd)] pub struct uint8x8_t(pub u8, pub u8, pub u8, pub u8, pub u8, pub u8, pub u8, pub u8);
#[repr(simd)] pub struct uint64x2_t(pub u64, pub u64);
#[repr(simd)] pub struct uint64x1_t(pub u64);
#[repr(simd)] pub struct uint16x8_t(pub u16, pub u16, pub u16, pub u16, pub u16, pub u16, pub u16, pub u16);
#[repr(simd)] pub struct uint32x4_t(pub u32, pub u32, pub u32, pub u32);

// vlane_get should return a umov as assemlby.

extern "platform-intrinsic" {
    fn simd_extract<T, E>(x: T, i: u32) -> E;
}

//immediate value: 0:7
#[allow(unused)]
macro_rules! constify_imm3 {
    ($imm8:expr, $expand:ident) => {
        #[allow(overflowing_literals)]
        match ($imm8) & 0b111 {
            0 => $expand!(0),
            1 => $expand!(1),
            2 => $expand!(2),
            3 => $expand!(3),
            4 => $expand!(4),
            5 => $expand!(5),
            6 => $expand!(6),
            _ => $expand!(7),
        }
    };
}

//immediate value: 0:31
#[allow(unused)]
macro_rules! constify_imm5 {
    ($imm8:expr, $expand:ident) => {
        #[allow(overflowing_literals)]
        match ($imm8) & 0b1_1111 {
            0 => $expand!(0),
            1 => $expand!(1),
            2 => $expand!(2),
            3 => $expand!(3),
            4 => $expand!(4),
            5 => $expand!(5),
            6 => $expand!(6),
            7 => $expand!(7),
            8 => $expand!(8),
            9 => $expand!(9),
            10 => $expand!(10),
            11 => $expand!(11),
            12 => $expand!(12),
            13 => $expand!(13),
            14 => $expand!(14),
            15 => $expand!(15),
            16 => $expand!(16),
            17 => $expand!(17),
            18 => $expand!(18),
            19 => $expand!(19),
            20 => $expand!(20),
            21 => $expand!(21),
            22 => $expand!(22),
            23 => $expand!(23),
            24 => $expand!(24),
            25 => $expand!(25),
            26 => $expand!(26),
            27 => $expand!(27),
            28 => $expand!(28),
            29 => $expand!(29),
            30 => $expand!(30),
            _ => $expand!(31),
        }
    };
}

#[allow(unused)]
macro_rules! arm_vget_lane {
    ($name:ident, $from:ty, $to:ty, $lanes:literal) => {
        #[cfg(target_feature = "neon")]
        #[cfg(target_arch = "aarch64")]
        pub unsafe fn $name(v: $from, lane: i32) -> $to {
                macro_rules! call {
                    ($imm3:expr) => {
                        if ($imm3) < 0 || ($imm3) > ($lanes - 1) {
                            unreachable_unchecked()
                        } else {
                            simd_extract(v, $imm3)
                        }
                    }
                }
                constify_imm5!(lane, call)
        }
    };
}

// uint8_t vget_lane_u8 (uint8x8_t v, const int lane)
arm_vget_lane!(vget_lane_u8_1, uint8x8_t, u8, 8);

#[cfg(target_arch = "aarch64")]
#[cfg_attr(target_arch = "arm", target_feature(enable = "v7"))]
#[rustc_args_required_const(1)]
pub unsafe fn vget_lane_u8_2(v: uint8x8_t, imm5: i32) -> u8 {
    if (imm5) < 0 || (imm5) > 7 {
        unreachable_unchecked()
    } 
    let imm5 = (imm5 & 7) as u32;
    simd_extract(v, imm5)
}

// uint64_t vget_lane_u64 (uint64x1_t v, const int lane)
arm_vget_lane!(vget_lane_u64_1, uint64x1_t, u64, 1);
#[cfg(target_arch = "aarch64")]
#[cfg_attr(target_arch = "arm", target_feature(enable = "v7"))]
#[rustc_args_required_const(1)]
pub unsafe fn vget_lane_u64_2(v: uint64x1_t, imm5: i32) -> u64 {
    if (imm5) < 0 || (imm5) > 0 {
        unreachable_unchecked()
    }
    let imm5 = (imm5 & 0) as u32;
    simd_extract(v, imm5)
}

// uint16_t vgetq_lane_u16 (uint16x8_t v, const int lane)
arm_vget_lane!(vgetq_lane_u16_1, uint16x8_t, u16, 8);

#[cfg(target_feature = "neon")]
#[cfg(target_arch = "aarch64")]
#[rustc_args_required_const(1)]
pub unsafe fn vgetq_lane_u16_2(v: uint16x8_t, imm5: i32) -> u16 {
    if (imm5) < 0 || (imm5) > 8 {
        unreachable_unchecked()
    }
    let imm5 = (imm5 & 7) as u32;
    simd_extract(v, imm5)
}

// uint32_t vgetq_lane_u32 (uint32x4_t v, const int lane)/
arm_vget_lane!(vgetq_lane_u32_1, uint32x4_t, u32, 4);

#[cfg(target_feature = "neon")]
#[cfg(target_arch = "aarch64")]
#[rustc_args_required_const(1)]
pub unsafe fn vgetq_lane_u32_2(v: uint32x4_t, imm5: i32) -> u32 {
    if (imm5) < 0 || (imm5) > 3 {
        unreachable_unchecked()
    }
    let imm5 = (imm5 & 3) as u32;
    simd_extract(v, imm5)
}

#[cfg(target_feature = "neon")]
#[cfg(target_arch = "aarch64")]
pub unsafe fn vgetq_lane_u32_3(v: uint32x4_t, imm5: i32) -> u32 {
    if (imm5) < 0 || (imm5) > 3 {
        unreachable_unchecked()
    }
    match imm5 & 3 {
        0 => simd_extract(v, 0),
        1 => simd_extract(v, 1),
        2 => simd_extract(v, 2),
        _ => simd_extract(v, 3),
    }
}
//  uint64_t vgetq_lane_u64 (uint64x2_t v, const int lane)
arm_vget_lane!(vgetq_lane_u64_1, uint64x2_t, u64, 2);

#[cfg(target_feature = "neon")]
#[cfg(target_arch = "aarch64")]
#[rustc_args_required_const(1)]
pub unsafe fn vgetq_lane_u64_2(v: uint64x2_t, imm5: i32) -> u64 {
    let imm5 = (imm5 & 1) as u32;
    simd_extract(v, imm5)
}